The #805 server-side fixes (emit-time batching, _ListenerQueue poison,
out-of-band closing) already cover every SSE stream, but the client-side
companions lived only in the interactive pane. Port them to coordinator.js
and extract the drift-prone pure core into a shared module (closes#806).
- shared_static/sse_overflow.js (new): storm-guard constants +
overflowWindowTripped + degradedCooldownStep, imported by both panes so the
trip threshold and cooldown ladder have one source of truth. interactive.js
imports these instead of holding local copies; the two node runtime probes
move to tests/test_sse_overflow_js.py.
- coordinator.js: handle the stream_overflow frame (storm guard -> degraded
catch-up with a doubling cooldown; the reconnect replays from the ring, or
falls to the replay_truncated -> /history floor); add the close-on-hide /
replay-on-show visibilitychange handler plus a document.hidden guard at the
connectSSE chokepoint; add drop-vs-render-wedge counters (onmessage now wraps
the dispatch in try/catch -- the coordinator previously had no wedge guard,
so a handler throw silently poisoned every later turn).
- After a stream gap the children/tasks sidebar re-syncs only when the ring
replay cannot cover it: no resume cursor, a replay_truncated envelope, a gap
beyond the cursor-trust window, or a live event id below the saved cursor (a
process restart reset the counter, which the replay path reports as a false
replay_ok). child_ws_*/task events are ordinary ring entries, so an ordinary
short reconnect heals the sidebar through the live handlers with no REST
rebuild -- a momentary blur/focus under close-on-hide rebuilds nothing.
- Close-session teardown detaches the visibility handler before the close POST
so a hide/show mid-close can't resurrect a dying stream. A replay_truncated
seen mid-stream is deferred (not dropped) and re-synced from /history on the
next idle -- repairing both a ring-evicted gap and a turn stranded by
close-on-hide (stream_end evicted while hidden), matching interactive.js's
_pendingTruncatedResync.
The extraction stops at the pure core: interactive.js's stateful glue is
hard-pinned by its source-assertion suite, so its class-method shape stays put
and the coordinator reimplements the equivalent glue as closure functions.
Tests: new test_sse_overflow_js.py (module exports + the two runtime probes);
coordinator parity + lifecycle pins in test_app_js.py (replay-aware sidebar
refresh, restart detection, truncated-resync deferral, close-session
visibility detach); interactive's moved probes replaced by an extraction pin.
All JS-source suites green.
(cherry picked from commit 2a32211e4a)
HYPOTHESIS.md:
- New appendix entry "Daemons (the recurrent harness)": a daemon as the
regenerative process of concatenated runs — ready-set recurrence,
renewal-reward lifting exactly at regeneration points, accumulation as
what breaks regeneration (cross-cycle provenance meet, renewal events
that reset accumulated risk), and authority under intermittence
(owner contact as a renewal point for authority; TOCTOU at cycle
scale).
- New body section "The loop": the task-dispatching outer loop as the
harness construction applied one level out — the composition
correspondence read at the top level, the daemon as its single-agent
special case, the bare while-loop as the trivial-group harness one
level up. Flagged as a sketch; outer fail-closed/reach-avoid
treatment deferred to later rounds.
- Veto caveat threaded to match: judge-as-veto safety scoped to the
authority lattice, and the nonblocking escape degrades to an
always-enabled safe halt when the principal is unreachable.
- Consistency: Grounding's Asserted tier now covers "The loop";
"always-enabled escalation" -> "escape" (the appendix's own term, now
that the escape has an unattended form); brace the one unbraced \bot
subscript (linter section-B HIT).
PRIMER.md: new plain-language companion — same object, no symbols, the
formal doc wins every disagreement. README's entry link now points at
the primer, which links onward to HYPOTHESIS.md.
(cherry picked from commit d115111756)
highlight.js, KaTeX and Mermaid all run on every surface via the shared
renderer (renderer.js), but their theme/wrapper CSS lived only in
ui/static/style.css. The console and coordinator load /static/style.css from
console/static/ — a different file on a different server — so hljs token spans
fell back to --fg (flat monospace for several releases), and the KaTeX/Mermaid
wrappers lacked their overflow containers, letting wide equations/diagrams
overflow the pane.
Move the hljs theme, .katex-display/.katex-error and the .mermaid-* wrappers
into shared_static/chat.css, which every surface loads via /shared/chat.css.
Restate the mermaid width-clamp for the preview pane (.preview-markdown) too,
since its content isn't a .msg.assistant message.
Drop the redundant background on .msg.assistant pre code.hljs so the <pre>
carries the code surface on every surface — otherwise the console/coordinator
(where the pre is --panel, not --code-bg) showed a darker box inside a lighter
padding band.
(cherry picked from commit 5dcf66c284)
A scheduled task could pin the model and skill of the workstream each
firing creates; it can now also pin its persona and project, so a
schedule can run under, e.g., the researcher persona attached to a
specific project's memory bucket.
The two values live on scheduled_tasks (migration 066, Text NOT NULL
default '') and are passed verbatim to create_workstream at dispatch,
where the node resolves the persona for the workstream kind and gates
the project attach. Empty means "kind-default persona / no project",
resolved late at each firing (mirrors how empty model/skill already
behave) -- existing schedules keep byte-identical dispatch behaviour,
so there is no backfill.
Also fixes a latent bug this feature depends on: admin_create_schedule
read created_by from request.state.user_id, which AuthMiddleware never
sets, so every scheduled task stored created_by=''. It now reads
auth_result.user_id like every other console endpoint. This is now
load-bearing -- the scheduler dispatches under created_by and the node
gates the project attach against it. admin_update_schedule adopts the
editing admin as owner when a project is assigned to a pre-fix orphaned
('') schedule, and re-validates persona/project only when they change
so a since-disabled persona or lost membership does not block unrelated
edits (the node re-checks at dispatch either way).
Wired through: schema + migration (up/down + parity tested), both
storage backends, API schemas, SDK create_workstream and console
create_schedule/update_schedule, scheduler dispatch, and the admin
schedule shelf (persona + project pickers, current value preserved so
an edit cannot silently clear a filtered-out selection).
(cherry picked from commit c328bebecd)
The URL-extraction call hard-coded max_tokens=8192 and rode the "low"
reasoning default, which broke local-inference models whose registry entry
advertises a tighter output limit or a different reasoning config. Inherit
the session/registry max_tokens and reasoning_effort instead (temperature
already was) — the same knobs the main turn uses.
max_tokens is capped to context_window // 4, the ~25% output slice Phase 2
already reserves, matching the main turn's response reserve
(_remaining_token_budget), so a large operator budget can't push
prompt + output past a small context window on strict runtimes.
(cherry picked from commit d5ddc95e9f)
github-code-quality flagged 8 spots where tests imported
turnstone.core.session_ui_base both as `from ... import` and `import ... as
suib` (the alias was only there to monkeypatch the module-level batch
constants). Drop the alias and patch via string target
(`monkeypatch.setattr("turnstone.core.session_ui_base._TOKEN_BATCH_WINDOW_SECS", ...)`),
which resolves to the same module global — behavior-identical. The one test
that READS the constant imports the symbol directly. Test-only, no
production change.
(cherry picked from commit 026c646116)
PR #805 review (Copilot + the round-3 finding it corroborates):
- connectSSE opened a new EventSource even when the tab was already hidden
(e.g. a first load in a background tab), where the close-on-hide handler
never fires because there is no open stream to close — so a throttled
hidden tab could still become the slow consumer this PR prevents. Add the
document.hidden guard at the single connect chokepoint, after the wsId
assignment + visibilitychange-handler install (so the show edge reconnects)
and before new EventSource (so nothing opens). The timer callbacks keep
their own pre-checks (the recover beat's also gates failCount); this closes
the fresh-connect path they never covered.
- Fix the stale _ListenerQueue.closing docstring: it claimed the drain loop
checks closing BEFORE poisoned, but the round-2 fix moved that check INSIDE
the poison branch (poisoned+closing -> clean close; a healthy closing queue
drains its tail to the ws_closed sentinel). Wording now matches the code.
(cherry picked from commit dfe09d029b)
A long live session driven by a fast local model (500-2000 tok/s) showed
corrupted / missing spans of assistant text while the backend stayed
healthy. Root cause: on_content_token/on_reasoning_token enqueued one SSE
event per model delta, so the per-listener queue (cap 500) overflowed
against any slow consumer; put_nowait on a full queue silently dropped the
newest event. Once saturated, drops scatter (the consumer keeps freeing
single slots), so the client's lastEventId sails past the holes and
reconnect-replay (eid > last_event_id) can never heal them. A dropped
fence-closer reshapes all downstream markdown -> reads as heavy corruption.
Fix B (primary) - emit-time micro-batching:
Coalesce content/reasoning fragments over a ~25 ms window (or 4 KB) into
one _enqueue, cutting the wire event rate ~10-20x at local-inference
speeds. A batch is assembled before it gets an _event_id, so it is one
ordinary ring entry no cursor can fall inside (unlike the forbidden
in-ring coalesce). Two conditions are load-bearing and pinned:
1. The inflight-buffer append and the enqueue are one _ws_lock section,
so a snapshot's snap_seq stays a true high-water mark for its text.
Splitting them lets a straddling snapshot double-render (the client
content path is a blind +=, no dedup).
2. Every non-token emit flushes the pending batch first, enforced at the
single _enqueue choke point, so stream_end/tool_*/state_change can't
overtake trailing content and repaint it into a new bubble.
Fix A (recovery net) - poison-at-first-overflow:
_ListenerQueue latches `poisoned` atomically at the FIRST rejected put and
refuses every later put, freezing its contents as a contiguous prefix; the
drain loop closes the stream after an id-less stream_overflow frame and the
native EventSource reconnect replays the whole gap from the ring buffer.
Poisoning at the first full (not after N) is required: any deliver-while-
dropping window advances lastEventId past interior holes that reconnect
can't replay. A ws teardown that races the overflow sets an out-of-band
`closing` flag (mark_closing), checked inside the drain loop's poison
branch: a poisoned+closing queue returns clean (no false overflow frame),
while a healthy closing queue still drains its full tail FIFO to the in-band
ws_closed sentinel -- so a slow-but-unpoisoned client never loses the turn's
final content batch + stream_end at teardown.
Client (interactive.js):
- Reconnect storm guard: after 3 overflow closes in 60 s the pane drops to
a degraded catch-up (stop live streaming, "connection is slow" state,
reconnect after a doubling 15->120 s cooldown that resyncs from the ring
or the uncapped /history floor). The cooldown ladder is keyed off a
last-trip timestamp, not the overflow-window array (which the trip
clears), so the escalation survives its own backoff.
- Close-on-hide / replay-on-show: a visibilitychange handler closes the
EventSource on tab-hide (a throttled hidden tab is the likeliest slow
consumer) and reconnects with the saved Last-Event-ID on show. The
factory recovery beat defers when hidden, and giveUp() detaches the
handler, so a dead or backgrounded controller can't reopen a stream.
- Drop-vs-render-wedge counters distinguish this bug (server overflow
closes) from the handler-wedge class (render/finalize throws) in the
field. No global gap-detector: live ids are not strictly monotonic
across concurrent tool+content emit, so a naive id!=last+1 check would
false-positive; recovery is server-signalled instead.
Corrects the stale _resolve_event_buffer_max comment that justified the
50k ring on a "PR-G closes connections on hide" mitigation that never
existed (the close-on-hide handler above is the real one).
Negative-tested (revert the guarantee, confirm the pin fails, restore):
per-token inflight append -> snapshot straddle double-render; removed
choke-point flush -> stream_end split; no poison latch -> silent drops;
top-of-loop closing check -> healthy-close tail loss; missing mark_closing
wiring / drain closing check -> clean close mis-reported as overflow;
_noteStreamOverflow cooldown reset -> ladder never escalates; removed
hidden-tab recovery guard / giveUp handler removal -> hidden-tab reconnect.
(cherry picked from commit 5083f67e96)
Copilot review on PR #804:
- An indented closing fence line (" ```") left its leading spaces as a
trailing whitespace-only line inside the rendered code block: the content
capture runs up to the backtick run and the close-line indent precedes it, so
it was captured as content. Strip a trailing newline PLUS any trailing indent
(/\n[ \t]*$/ instead of /\n$/); a column-0 close is unaffected. Red-green
pinned (content is exactly " x = 1", no trailing whitespace line).
- Correct a stale test docstring claiming the fence open anchor allows "up to 3
spaces" of indent — it allows arbitrary indent (the 4-space case is pinned
separately).
(cherry picked from commit e5e48a788a)
The markdown renderer protects structural blocks with in-band NUL-framed
sentinels (chr(0)+tag+index+chr(0)). escapeHtml preserves U+0000, so
model/tool text could forge sentinels, and recursively-rendered <details>
bodies re-rendered against fresh block arrays and lost their content. This
lands the ordered containment fixes from the render-containment brief.
Fixes (each pinned in tests/test_renderer_js.py; all NUL-sensitive cases also
confirmed in real headless Chrome, which drops a U+0000 token the node harness
preserves):
- B1/B2/B3 — forged sentinels: strip U+0000 at the TOP-LEVEL render entry only
(_fnDepth === 0). renderer.js is the sole NUL producer and every restore
regex is NUL-framed, so removing NUL closes every forgery path (block
duplication/relocation, out-of-range "undefined", cross-container injection)
while generated sentinels in recursive frames survive. Only NUL is stripped,
so a code fence still shows pasted control bytes (ESC/FF/VT/DEL) verbatim.
- B4 — blockquote-in-fence (the common one): the code-fence pass now runs
before the line-based blockquote pass. Its open matches at line start after
optional indent and an optional list marker (`- `, `1. `), and re-emits that
indent+marker before the sentinel so the fence keeps its document position
(a nested-list item stays nested; a fence continuing a footnote definition
keeps the indent its continuation scan needs). A blockquoted fence (`> ```)
is not matched (`>` is neither indent nor a list marker), so the blockquote
pass extracts that `> ` run and its recursion renders the fence. A `> ` line
inside a plain fence stays literal.
- B5 — <details> open anchored to line start (^[ \t]*), so a `<details>`
mentioned mid-line inside inline code no longer starts a block.
- NEW-1 — recursive-frame content loss: <details> extraction runs AFTER fence
protection and restores a fenced body from a saved raw-source array
(codeBlockRaw) back to raw markdown before the recursive render, so
code-in-details renders in-frame instead of restoring to "undefined". Running
after fence also means a </details> shown as example code inside a fence
can't close the block early, and a <details> shown inside a fence stays
literal — no offset-based fence-awareness needed. Inline-code/math in footnote
definitions render via the restore round-trip the undefined-guard enables
(documented at the append site).
- NEW-3 — code blocks gained the <p>SENTINEL</p> unwrap variant DT/BQ/MB/TB
already had, removing a stray empty <p> before a standalone <pre>. The CB
unwrap is whitespace-tolerant so an indented own-line fence (whose indent the
fence pass re-emits) also doesn't leave a stray <p>.
- Defense-in-depth: every restore callback returns the matched sentinel
(inert; the browser drops the NUL) instead of the array's `undefined`.
Non-obvious decisions:
- Control chars are authored as literal \xNN hex escapes (byte-verified: only
\uXXXX decodes to raw bytes in this toolchain; \xNN matches the file's
existing \x00 sentinel convention).
- Open anchors allow arbitrary leading indent (the fence open also allows a
list marker), not CommonMark's ^ {0,3}: the renderer has no indented-code
fallback, so preserving the prior behaviour of matching indented/list-nested
fences beats CommonMark strictness, while still excluding `> ``` and mid-line
forms.
- codeBlockRaw (the <details> raw-fence array) and the restore callbacks are
factored through a _restorer(arr) helper; codeBlockRaw is only populated when
the text contains a <details> tag (its sole reader).
- The entry strip is depth-0-only on purpose: an unconditional strip would
shred the generated sentinels recursive frames carry, foreclosing NEW-1.
Negative-tested (reverted the production line, confirmed the pin fails):
- fence anchor: unanchored swallows a blockquoted fence.
- NEW-1 codeBlockRaw restore: without it, code inside <details> is lost.
Deferred (called out per the brief):
- B6/NEW-4 bidi controls (U+202A–202E, U+2066–2069, U+200E/F) still pass
through unescaped; they are not C0 so the entry strip misses them. Left to a
follow-up — stripping risks corrupting legitimate RTL text and <bdi>
isolation is involved for a string renderer.
(cherry picked from commit a164d61552)
Review follow-up: the helper returned `fallback` verbatim when the name
sanitized to empty, so a future caller passing an unsafe fallback (non-latin-1,
control chars, quote, backslash) could reintroduce the header crash/corruption
the helper exists to prevent. Not reachable today — all call sites pass safe
ASCII literals — but the helper is a shared safety primitive whose contract is
wire-safe output.
Run the fallback through the same cleaning, backed by a safe constant if even
that is empty, so the return is always wire-safe and never filename="". Adds a
test.
(cherry picked from commit 2c5adb7aca)
Attachment `/content`, preview, and workstream-export downloads built the
Content-Disposition `filename="..."` value straight from a user-supplied
name, stripping only quotes and CR/LF. Three input classes still broke the
header:
- Non-latin-1 names (CJK, em dash): Starlette encodes header values as
latin-1 and raised, 500-ing the serving route. (The original get_content
bug.)
- ASCII control bytes (NUL, form-feed, VT, DEL): latin-1-encodable, so they
passed Starlette, but the HTTP server layer rejects control characters in a
header value and 500s one layer later.
- Backslash: the RFC 6266 quoted-pair escape. A trailing backslash escaped
the closing quote and corrupted the download filename (not a 500, but wrong
output; Windows-origin uploads carry it legitimately).
Extract one `latin1_safe_filename()` helper in web_helpers that drops every
non-printable character plus the double-quote and backslash quoted-string
metacharacters, folds any surviving non-latin-1 codepoint to '?', and falls
back to a non-empty name so the header never emits an empty filename. Route
get_content, preview_response_headers, and the export handler through it,
replacing three near-duplicate inline strips.
Adds unit tests for the helper (non-latin-1 fold, control-char and backslash
stripping, per-site fallback) and an endpoint regression test.
(cherry picked from commit fd3aed1eca)
unsplit() closed each doomed (ephemeral) pane via close() — which already
renders/persists/notifies — then repeated that trio, firing intermediate
persist/notify passes mid-operation. A 2-cell split fully collapses inside
close(), so bail there; only a 3+-cell split (or an empty doom list) still
needs the trailing exit + refresh. The all-conversation path is unchanged.
Also reword the cell-chip CSS comment so it names the reversible hide vs
destructive close glyphs, now that an ephemeral pane can show the close glyph
in split mode.
(cherry picked from commit f56fa55929)
The preview pane opens beside the conversation as a split cell. Dismissing
that cell — the per-cell chip, or Unsplit from the other pane — ran
closeCell(), which hides the pane but keeps it in _panes/_order, leaving an
orphan tab with no meaningful reopen (the reopen affordance is the transcript
chip, not the tab bar).
Add an `ephemeral` flag on ShellPane. For an ephemeral pane the cell chip and
Unsplit route to close() — destroying the pane and its tab — and the chip's
glyph/label read as a destructive close rather than a reversible hide. Unsplit
still spares the focused survivor even when it is ephemeral ("keep the focused
pane"). The preview pane sets the flag; conversational panes do not, so an
all-conversation split is unchanged (Unsplit reduces to the prior
_exitLayout(_activeId)).
(cherry picked from commit ace9e034f9)
The wake gate documented exactly one info line per call past its
gates, but a send() refusal (the authoritative under-lock _closed
re-check catching a teardown the gate's lockless peek missed) emitted
nothing — a dropped wake should stay traceable to its trigger, so the
refusal now logs nudge_wake.refused.
The already-dispatched branch's comment claimed a held reminder can
coexist with the terminal mark via a redelivery whose commit raised —
impossible with the current control flow (_redeliver_pending clears
the hold before committing). Reworded to what the clear actually is:
the last line of defense against any coexisting hold leaking forever
once this branch deactivates the row, since inactive rows never
re-list. Test comment updated to match.
(cherry picked from commit cb59afe443)
A session whose _nudge_queue answers has_pending truthily while its
deliver_wake_nudge_from_queue consumes nothing turns the worker-exit
backstop into an infinite respawn loop: the gate passes, the wake
worker no-ops, the exit backstop re-runs the gate, forever.
Mock-backed test sessions riding real Workstreams are exactly that
shape, and one worker on such a pairing is enough to ignite a
wake-thread storm that trips the leaked-thread guard in every
subsequent test. The wake contract requires real drain semantics —
the spawned worker must CONSUME what the gate saw — so the gate now
refuses on type, not just presence.
(cherry picked from commit 7886d3b763)
str | None under-specified the field: the implementation and the TS SDK
union both constrain it to queue_full / refused_closed, and the Literal
projects a proper enum into the generated OpenAPI spec so clients
reject unexpected values. Specs regenerated.
(cherry picked from commit fa1ba2cc01)
Wake path:
- Denial metacog nudge moves to the tool channel so it drains with the
denied tool batch instead of the next user-message seam.
- wake_workstream_if_pending: shared wake gate for watch fires on
already-idle workstreams (no IDLE transition for the watcher to
observe), wired as wake_fn at every set_watch_runner site via the
shared _watch_fire_wake_fn helper (closes over the Workstream OBJECT
— after eviction+restore an id-keyed manager lookup would miss).
- session_worker exit backstop re-runs the wake gate the moment worker
ownership clears: IDLE fans out on the worker thread, so
transition-time wakes always landed on the reuse path and no-op'd
(the coordinator idle_children strand).
- deliver_wake_nudge_from_queue contains GenerationCancelled — it is
the wake worker's run() closure and only Exception is caught
downstream.
Watch delivery:
- Terminal fires that cannot reach their workstream are HELD and
redelivered on min(interval, 60s) without re-running the command,
bounded by MAX_DELIVERY_ATTEMPTS per cycle and the watch's own
max_polls across cycles; the poll charge commits durably at hold
time so restarts stay budget-bounded.
- Restore admission control: per-ws dedup + MAX_CONCURRENT_RESTORES
cap, presence-only re-check under the lock, detection-only stall
alerts (reclaiming a wedged admission would trade capped degradation
for total poll-pool collapse).
- Permanent-vs-transient restore taxonomy: corrupt persona stamp and
genuinely-missing history (confirmed by a raising storage probe —
the resume loader swallows read blips into []) deactivate the watch
immediately; everything else holds and retries.
- Cancel-race defense: delivery paths re-check is_watch_active before
stashing/dispatching, cancel paths write the row BEFORE
forget_terminal_dispatched, the HTTP cancel endpoint clears runner
state, and a per-tick sweep bounds the residual stash-after-clear
interleaving to one check_interval.
- Abandon/exhaustion commits are write-then-clear so storage that can
read but not write retries the row write instead of re-running the
command every cycle; the fresh-fire unrestorable path stashes before
its deactivation write for the same reason.
Registry follows identity:
- The dispatch registry is keyed by _ws_id at registration time; every
rebind now moves it: non-fork resume() and /new go through
_follow_watch_registration (new key live before the old is removed,
never stealing a registration another live session holds), removals
are owner-checked so tearing down a watch-restore shell or a
resumed-away session cannot unregister a live pane, the restore
shell yields to a registration that appears mid-restore, CLI
--resume registers after the successful resume, and both the open
path and the detail-GET lazy rehydrate wire the registration.
Teardown gating and backpressure honesty:
- cleanup_session_ui marks ws._closed FIRST under ws._lock — every
teardown path (close, close_idle, evict, delete, discard) funnels
through it — and session_worker.send re-checks under the same lock,
so a wake can never spawn a worker on a torn-down workstream.
- Create responses carry initial_message_status when the initial
message could not be delivered (queue_full / refused_closed) instead
of reading as success; staged attachments survive for the retry;
/send surfaces a closed workstream as 404 rather than queue_full.
Docs/spec: OpenAPI artifacts regenerated; api-reference documents the
new create-response field; TS SDK type extended.
Tests: ~30 new pins (cancel races, budget durability across restarts,
owner-checked registry moves, teardown gating, stall alerts,
backpressure surfaces, wait_until final re-check); wide subsystem
sweep green (2353 passed).
(cherry picked from commit e60c19befd)
Review feedback (PR #800): the URL lane hard-capped fetched bodies at
10 MB before kind resolution, making the 32 MiB pdf cap unreachable for
URL targets while path targets honored it. The flat pre-check is gone;
the guarded fetch's max_bytes now tracks max(PREVIEW_SIZE_CAPS.values())
- mirroring the path lane's stat pre-check - and the per-kind caps after
resolution stay authoritative.
Also drops a redundant function-local asyncio import in test_console.py.
(cherry picked from commit bbe92faca1)
fetch_with_ssrf_guard now streams the response under a max_bytes budget
(default 32 MiB, counted on decoded bytes so gzip cannot expand past it)
instead of buffering blind - an unbounded body previously filled memory
before any caller-side size cap could run. Redirect-hop bodies are no
longer read at all, and the realized response drops stale wire-framing
headers (content-encoding/content-length/transfer-encoding) that no
longer describe the decoded content it carries.
Preview blob ids are salted out of the model-visible attachment
namespace (sha256("preview:" + body)): uploads use bare sha256(body)
and save_attachment freezes kind at first insert, so a byte-identical
preview/upload pair would otherwise share a row - whichever landed
second inherited the other's kind, silently hiding an upload from model
context or materializing preview bytes into a tool turn.
(cherry picked from commit 29a4bbf876)
turnstone's primary audience self-hosts it beside other lab services —
a web_fetch or open_preview aimed at Grafana, Home Assistant, or a dev
node on the local network is the operator using their own network, not
an attack. The hard SSRF refusal made those targets unreachable.
New runtime setting tools.allow_private_network (settings registry,
default off, rendered in console Settings → Tools; hot — read per tool
call, no restart). When enabled, a call NAMING a private address
becomes approvable: the approval prompt tags it "(private network)" so
the operator approves it as what it is, and the human gate stays.
The redirect side-door stays closed either way: a PUBLIC target that
302s into private address space is refused regardless of the opt-in —
that address never appeared on the approval card, so it is never
fetched. Only a chain whose approved origin was itself private skips
hop screening (its redirects are the operator's own network).
Refusals now teach the knob (mirrors the oidc opt-in hint): the error
names tools.allow_private_network and where to enable it. Surfaces
without a ConfigStore (bare CLI, eval) stay strict — there is no admin
surface to have opted in on.
(cherry picked from commit 09abc9d199)
Four follow-ups to the preview pane:
- Probe-mode preflight: the pane preflights src-loaded kinds with
GET ?probe=1 (204, real hardening headers, no body) instead of HEAD —
the console reverse proxy forwards HEAD as a full GET, so the old
preflight dragged the whole blob across the node→console hop twice.
Ownership gate + renderable-type check still run on probes.
- Legacy-charset text: table/text/markdown now transcode to UTF-8 at
store time (declared charset → UTF-8 → cp1252-replace ladder), same
model the web kind already used. The ladder applies only when the
text kind was DECLARED (MIME/extension/override); the bare no-hint
fallback stays strict UTF-8 and NUL bytes still hard-reject, so
binary rejection is unchanged.
- Remote assets default OFF: previewed pages are now served under
"sandbox; default-src 'none'; style-src 'unsafe-inline'; img-src
data:; font-src data:" — they render with inline styling but cannot
contact their origin site (no viewer IP/traffic disclosure). A
per-pane "Load remote images & styles" checkbox (web previews only,
sticky, not persisted) reloads with ?assets=1 for the permissive
bare-sandbox mode.
- Markdown vendor parity: preview markdown now runs renderer.js's
postRenderMarkdown (hljs token coloring + lazy mermaid diagrams)
like the conversation pane, with preview-scoped code-block/KaTeX
chrome (the conversation theme is .msg.assistant-scoped).
Tests: probe/assets HTTP + policy coverage, charset ladder units +
stored-bytes round-trip, JS static guards for the probe form, the
default-off toggle, and the post-pass; headless-chrome harness grew to
41 assertions (probe-not-HEAD, toggle visibility/default, fenced-code
render). Full suite green.
(cherry picked from commit 1e2ab91ec2)
Tool results only ever rendered as plain text in the transcript. This
adds the model-driven rich-preview lane every comparable surface has,
in turnstone's developer-tool idiom: a preview pane that opens BESIDE
the conversation, keyboard-operable, sandboxed, never replacing the
transcript that spawned it.
Backend
- New built-in open_preview(target, kind?, title?): resolves an http(s)
URL, a file path, or attachment:<id> to bytes; classifies into
web/pdf/image/table/text/markdown (magic bytes > MIME hint >
extension > UTF-8 fallback, legacy-charset pages transcoded); caps
size per kind; persists content-addressed with kind="preview" —
refcounted and GC'd with the workstream, skipped by trajectory
reconstruction so preview bytes can never materialize onto the wire.
URL targets gate like web_fetch (network egress); paths/attachments
run unprompted like read_file.
- New core.web.fetch_with_ssrf_guard: manual redirect walk that
SSRF-screens every hop BEFORE requesting it (follow_redirects=True
checked nothing between hops); adopted by both open_preview and
web_fetch. URL userinfo is stripped before the descriptor or the
stored bytes see it; <base href> is injected doctype-safely so
relative assets resolve without quirks mode.
- The preview descriptor rides the tool turn's meta side channel with
ONE shape on every boundary: the live tool_result SSE event, the
conversations.meta column, and the /history projection. Cancelled
batches commit an already-announced preview (blob + meta) instead of
stranding the open pane on a permanent 404.
- New GET {ws}/attachments/{id}/preview (read scope, same ownership
gate as /content) serves the STORED type with per-MIME hardening:
bare CSP sandbox for text/html (renderable, scriptless, opaque
origin), no CSP for application/pdf (Chromium's viewer refuses
sandboxed contexts), full default-src 'none' otherwise; filenames
fold to latin-1-safe ASCII. The console /node proxy now forwards
CSP/nosniff/disposition/cache-control instead of dropping them.
- History loads exclude preview blobs from the bulk content fetch at
the query (they were read and discarded on every load).
Frontend
- New "preview" pane type registered in the shared shell (server +
console): openPaneBeside placement, per-kind renderers — fully
sandboxed iframe for pages, browser PDF viewer, sortable tables
(CSV/TSV/JSON, ragged-file safe, 5k-row cap), rendered markdown,
text — plus back/forward history with arrow keys, reload persistence
via pane meta, and backoff auto-retry (0.9s..7.2s) bridging the gap
between the live descriptor and the batch fold that commits its blob.
- Tool results carrying a descriptor render a credential-redacted
preview chip (the reopen + replay affordance); live results auto-open
the pane only while the originating pane holds focus.
Docs: docs/tools.md + prompts/tools.md. Tests: policy unit tests, tool
prepare/exec (mocked fetch), serving route + proxy header pass-through,
storage exclusion on both backends, cancel-path commit, JS static
guards; a headless-Chrome harness drives the real module graph (32 DOM
assertions).
(cherry picked from commit e010124008)
The discovered-endpoint wrapper converted every OAuthSSRFError to a
bare OIDCError, so a private-resolving endpoint or trusted host got the
non-public message without the allow_private_network remediation even
though the same knob fixes it. Hoist the hint into a module constant
and append it in both wrappers.
Also name "unspecified" in the refused-even-with-opt-in message so
0.0.0.0/:: rejections read unambiguously.
(cherry picked from commit 4350248d8f)
The SSRF guard on OIDC endpoint URLs hard-refused any hostname
resolving to a non-public address, which made it impossible to use a
self-hosted IdP (Keycloak, Authentik, Dex) on an internal network —
even though the login-flow issuer is operator-configured, i.e. trusted
input.
Add [oidc] allow_private_network in config.toml (or
TURNSTONE_OIDC_ALLOW_PRIVATE_NETWORK), default off. When set, the
issuer and its discovered endpoints may resolve to private-range,
unique-local, CGNAT, and loopback addresses. Link-local, multicast,
unspecified, and reserved ranges stay refused regardless — cloud
metadata services live on link-local and no legitimate IdP does. The
HTTPS requirement and same-origin endpoint checks are unchanged.
The private-address refusal now raises OAuthSSRFPrivateAddressError,
and the OIDC wrapper appends the remediation hint to the error message
so the failure is self-service. mcp_oauth call sites — where endpoint
URLs come from untrusted remote-server metadata — do not get the knob
and keep the strict public-address rule.
(cherry picked from commit 9c74673fd4)
RFC 3986 schemes are case-insensitive, so POSTGRESQL+PSYCOPG2:// or
HTTPS://user:pass@host in tool output leaked the password past the
case-sensitive scheme alternation. Compile with IGNORECASE on both
sides of the FE/backend mirror; the structural userinfo requirement
is unchanged. Uppercase-scheme cases added to both test suites.
(cherry picked from commit 19b1a04f17)
Connection-string redaction (the output_guard pattern and its frontend
mirror) only enumerated bare dialects plus +psycopg, so SQLAlchemy
dialect+driver URLs — postgresql+psycopg2://, postgresql+asyncpg://,
mysql+pymysql:// — leaked the password through every redaction surface.
The scheme now takes an optional +suffix instead of enumerating drivers.
redactCredentials() also gains a single early-exit prefilter scan ahead
of its sixteen replace passes, for plain-log tool output on card render.
The prefilter is documented and pinned as a superset of the pattern
set's required substrings, so a miss is provably a no-op: new smoke
cases assert bare sk-/AKIA/Bearer credentials with no '=', quote or '@'
anywhere in the text still redact, alongside the fast-path no-op and
the driver-scheme URLs on both sides of the mirror.
(cherry picked from commit ed2623ff44)
- Proven: name supervisory control's controllability and nonblocking
conditions as the ancestors of gate-early-on-irreversibles and the
always-enabled escalation required behind a learned veto; cite TCSEC
covert-channel analysis (NCSC-TG-030) for the verdict channel.
- Asserted: add the narrow-only rule's influence-side twin (verdict
payloads to the plant selected, never generated).
- New caveat paragraph: a denial is free only in the authority lattice;
in the dynamics it is an input (selection + targeted-liveness
channels), so a learned veto needs a nonblocking escape it cannot
disable, verdict payloads are selected rather than generated with the
symbols/tokens/language thresholds bounding the alphabet, and the
strongest form dissolves the verdict into scheduling over
deterministic checks.
(cherry picked from commit 625218b7b5)
Review: the wrapped ValueError is echoed in 503 bodies, and arbitrary
SDK exception text can embed filesystem paths. Echo the exception type
only; log the full exception with traceback at the raise site.
(cherry picked from commit a029849724)
SDK client construction can fail on environment problems the config
never sees (e.g. httpx resolving a certifi CA path deleted by a venv
rebuild). Those escaped as bare exceptions and turned every workstream
open/create into an opaque 500; re-type them as ValueError in
ModelRegistry.get_client so routes answer 503 with the message and the
alias.
(cherry picked from commit 9c2e809b26)
to_tsvector was computed inline over full row content, so one row
whose tsvector exceeds PostgreSQL's 1MB limit aborted every
search_history scan. Cap the FTS input at 250K chars (worst-case
tsvector expansion stays under the limit; giant rows remain findable
by their head). The ILIKE fallback also never ran on postgres: the
failed statement leaves the autobegun transaction aborted, so roll it
back before falling back.
(cherry picked from commit 51ed336989)
Address review on #794. The prior gate deepcopied every agent tool, then compared, then discarded the copy on a no-op render; and its comment framed the equality case as 'no personas' when it also covers an idempotent re-render of unchanged aliases/personas. Compute the target model/persona descriptions from the current (read-only) schema first and only deepcopy when one differs — so a no-op render is genuinely allocation-free, not just fork-free. Behaviour is unchanged: identity preserved when nothing differs, stale text still cleared on reload.
(cherry picked from commit 90663ce695)
_render_agent_tool_descriptions rebuilt self._tools and reassigned it on every session init, deep-copying task_agent even with no model aliases and no personas to inject — the single-model CLI case the docstring says is skipped. This regressed after the persona-discoverability change removed the early 'if self._registry is None: return' guard, breaking the session._tools is INTERACTIVE_TOOLS invariant (test_session_without_mcp).
Gate the reassignment on whether a description actually changed: keep the original tool object when the render is a no-op, fork self._tools only when something was injected. Restores the shared-constant invariant, makes repeated renders idempotent, and preserves clear-stale-on-reload (an emptied registry still takes the changed path).
(cherry picked from commit bd37bcd1ec)
Add .github/FUNDING.yml to enable the native GitHub Sponsor button, plus a Sponsor badge and a Support section in the README. Primary CTA is GitHub Sponsors (eous); PayPal (paypal.me/eousphoros) is offered as a one-off fallback.
(cherry picked from commit a61d454df5)
Review feedback on #792: the "not found or disabled" branch
interpolated the raw input unquoted, so the whitespace-only and
trailing-space inputs the forgiving lookup explicitly handles rendered
invisibly in CLI output and logs. Use {name!r} like the other two
resolver errors already do.
(cherry picked from commit 647939fe4d)
Coordinators and interactive agents had no way to enumerate valid
persona names: task_agent / spawn_workstream / spawn_batch described
`persona=` but nothing listed what it accepts, and resolution was an
exact case-sensitive slug match - users reaching for the display name
or a case variant got an unexplained failure.
- Inject the live persona catalog (enabled, interactive-kind; children
and sub-agents are always interactive) into the `persona` parameter
description of task_agent / spawn_workstream / spawn_batch, riding
the same render path as the model-alias injection. Rebuilt from the
pristine TOOLS base every render, so repeated renders are idempotent
and archived personas drop out instead of lingering. Entries carry
name + default marker + <=96-char description; names-only past 25
personas. spawn_batch's persona property is nested per-child under
children.items.properties (located via _persona_property, null-safe
against name-colliding MCP tools). Storage-less sessions keep the
base text: the render runs at session construction, so it gates on
is_storage_initialized() rather than get_storage(), which would
auto-init SQLite as a side effect.
- resolve_persona_for_kind - the ONE shared rule behind the HTTP
create handler, CLI --persona, the coordinator spawn precheck, and
task_agent prep - is now forgiving: exact slug, then the lowercased
input (created names are regex-enforced lowercase slugs), then a
case-insensitive display-name match accepted only when unique among
the kind's enabled personas. Duplicates refuse loudly naming the
candidate slugs; a same-label persona of another kind neither blocks
nor wins (the label the caller saw came from a kind-filtered
surface); whitespace-only input never matches blank display names
(display_name defaults to ""). Every failure now enumerates the
kind's valid names, so a stale injected list or a typo self-corrects
on the next attempt.
- The canonical slug is stamped everywhere: task_agent prep rewrites
its arg from the resolved snapshot, _validate_child_persona returns
(canonical, error) and both spawn call sites adopt it - approval
chrome, the wire, and workstream_config never carry a forgiven
variant.
- Create-persona shelf: label hint under Name explaining agents and
the CLI launch the persona by this name (case-insensitive) and the
display name is only a list label. docs/personas.md gains a "How
agents discover personas" section and drops the stale claim that
task_agent has no persona parameter.
Tests: resolver unit suite (case/display/ambiguity/cross-kind/
whitespace/disabled/storage-failure) + guards for injection content
and ordering, idempotent re-render, archive drop, the 25-persona
prose cutoff, coordinator-kind exclusion, and canonical stamping
through spawn_workstream / spawn_batch / task_agent.
(cherry picked from commit 457b01737a)
- _coordinator_tenant_check and _coord_attachment_owner resolved storage from
the global registry (get_workstream_row / for_request without a storage arg),
which can evaluate the project-tenancy decision against a different or
auto-initialised backend and fail OPEN on a missing project row. Use
request.app.state.auth_storage explicitly, matching cluster_ws_detail and
_resolve_coordinator_or_404; fail closed (404) when it is unset.
- reject_unassignable_scopes now derives its allowed-scope error message from
ASSIGNABLE_SCOPES so validation and the message can't drift.
(cherry picked from commit a40ff249ec)
Workstreams attached to a private project were visible -- including their
conversation content -- to holders of admin.cluster.inspect / admin.coordinator
(both default builtin-admin permissions), defeating the project's confidentiality
boundary. Enforce that a private project's resources are visible only to people
IN the project (owner, workstream creator, or an explicit member), even for admins.
Surfaces closed:
- WorkstreamProjectVisibility bypass narrowed to service scope only (node->console
machine plumbing, re-filtered per-user at the console edge). No human principal
bypasses; admin.cluster.inspect gates the inspect surfaces, not tenancy. This
flows to /dashboard, session listings, the attachment row-gate, cluster_workstreams,
cluster_node_detail, and cluster_snapshot/SSE.
- cluster_ws_detail 404-masks a workstream in a private project the caller can't
see; cluster_ws_live_bulk routes such ids to the denied list (no private-project
oracle).
- Coordinator operator verbs (history/export/detail/send/approve/set_title/open/
children/tasks/attachments) now enforce project tenancy: _coordinator_tenant_check
on coord_endpoint_config, the gate in _resolve_coordinator_or_404 (children/tasks),
the tenant_check now run in make_open_handler before rehydrate, and a
project-visibility check in _coord_attachment_owner. admin.coordinator gates the
surface cluster-wide, but a non-member is 404-masked. The tenant-check mirrors the
manager-first + coordinator-kind ladder so kind-isolation is preserved.
- service scope is no longer user-assignable: admin_create_token and both
turnstone-admin CLI mint paths reject it via reject_unassignable_scopes, so an
admin.users holder cannot self-mint a service token and restore the bypass. Service
scope is minted only by ServiceTokenManager / the JWT secret.
- The events/global node proxy (service-elevated cross-tenant firehose) is gated on
admin.cluster.inspect so a plain authenticated user cannot reach it through the
console proxy.
Updates the OpenAPI description, the row-gate/tenancy-filter docstrings, and adds
tests for every surface (visibility predicate + cluster detail/bulk + coordinator
history/export/children/open/attachments + events/global proxy + scope-mint
rejection); inverts the tests that pinned the old admin-bypass contract.
(cherry picked from commit 36419a9809)
* fix(mcp): route static transport lifecycles through per-server owner tasks
A crash-looping MCP server drove the mcp-loop thread to a sustained,
climbing 100%+ CPU spin. Root cause: anyio cancel scopes are
host-task-bound, and the static path entered the SDK's transport /
ClientSession task-group scopes from short-lived connect tasks (every
health tick is a new task since #768). Once such a scope was cancelled
after its host task had finished - by anyio's task_done when a
transport child died with the server, or by ClientSession.__aexit__
during a cross-task teardown - CancelScope._deliver_cancellation could
never make progress (task.cancel() on a done task is a no-op) and
re-armed itself via call_soon every loop iteration, forever: ~900k
callbacks/s per zombie scope, one more per flap cycle (verified against
anyio 4.14.1; no upstream fix exists as of that release).
Fix: each static server's transport + session cms are now entered,
parked, and exited by ONE long-lived owner task
(_static_transport_owner), so scopes always have a live host and always
exit in the task that entered them. Teardown follows a one-cancel close
protocol (signal the close event before the first await, graceful
grace, then at most ONE cancel - never a second, which would abandon a
scope exit mid-flight). Connect timeouts now cancel only the waiting
caller; connect failures are delivered through a readiness future;
unrequested owner death (server died under a live session) evicts the
session immediately via a done-callback instead of waiting for the next
liveness ping. A rate-limited, mcp-loop-scoped gc-walk backstop
(_maybe_disarm_orphaned_scopes) disarms any zombie minted by paths not
yet migrated (the oauth_user pool keeps the old cross-task-close shape;
follow-up).
Also fixed: BaseExceptionGroup (BaseException-derived, as raised by
anyio task groups wrapping a stray CancelledError, e.g. an
accept-then-RST server) escaped `except Exception` in _connect_all and
killed it before the health/sweep loops were created - silently
disabling all autonomous recovery. Handled there and in the
health/sweep/eviction loops and the reconnect/refresh callers.
Verified: a live SIGKILL-flap repro went from 130%+ CPU (climbing, one
armed scope per cycle) to 0.3% flat with zero armed scopes; the RST
repro now leaves both background loops alive (previously both silently
dead). New tests: owner-lifecycle + close-protocol units (incl. an
exactly-one-cancel pin), a _connect_all BaseExceptionGroup regression,
a discriminating disarm-sweep test, and a ~10s live SIGKILL-flap smoke
test (real FastMCP subprocess, skips on environment gaps) asserting
zero armed scopes, exactly one live owner, and a post-recovery tool
call. Full 8470-test suite green; ruff+mypy clean.
* fix(mcp): route pool transport lifecycles through per-entry owner tasks
Completes the owner-task migration started for the static path: the
oauth_user pool path had the same latent anyio cancel-scope exposure
(host-task-bound scopes entered by short-lived connect tasks; a scope
cancelled after its host finished re-delivers cancellation via
call_soon forever - the 100%-CPU zombie), previously covered only by
the disarm backstop.
Each (user, server) pool entry's transport + ClientSession cms are now
entered, parked, and exited by ONE long-lived owner task
(_pool_transport_owner). The caller keeps building client_kwargs (the
per-user bearer and, when an auth-capture carrier is active, the
httpx_client_factory response hook) so 401/WWW-Authenticate capture
semantics are unchanged. Teardown is the shared one-cancel close
protocol (_teardown_pool_entry: signal before first await, graceful
grace, at most ONE cancel), used by the connect stale-guard, idle/LRU
eviction, and shutdown (parallel signal-then-reap). Unrequested owner
death evicts the session but keeps the entry and its discovered
catalog, matching the existing evict-session-keep-entry semantics the
auth_401 retry relies on.
Discovery still runs in the connecting caller while the transport is
hosted by the owner, so a transport collapse mid-discovery (e.g. the
SDK tearing its task group down on an upstream 401) cancels the OWNER,
not the caller - a bare await on the response stream would hang until
the 30s phase timeout. _await_pool_discovery races each discovery
await against owner completion and converts owner death into a prompt
ConnectionError (the owner is never cancelled there; teardown owns its
lifecycle). Carrier-first failure classification preserves auth_401
semantics for captured 401s.
With no cross-task stack closes left, _safe_close_stack and
_safe_teardown_on_connect_failure are deleted (zero callers).
Tests: new tests/test_mcp_pool_owner.py pins the pool close protocol
(graceful event-before-await close, exactly-one-cancel escalation,
owner-death eviction retaining entry+catalog, caller-cancel-mid-connect
cm-exit guarantee, factory-present-iff-capture, and the
owner-death-during-discovery fast-fail). 1010 mcp tests and the full
8471-test suite green, including the historical cross-task-anyio
sentinel test_integration_pool_reuse_401_refresh_and_retry_succeeds;
ruff+mypy clean; zero destroyed-task warnings.
* fix(mcp): harden disarm-sweep loop guard and owner BaseException arm
Review follow-ups on the owner-task migration:
- _maybe_disarm_orphaned_scopes now enforces its mcp-loop requirement
instead of trusting callers: it returns without walking (and without
advancing the rate-limit clock) unless the currently running loop IS
self._loop. A suppressed close can fire before start() or after
shutdown(), where the walk would be wasted at best and a cross-thread
reach at worst.
- The transport owner's BaseException arm now re-raises non-Exception,
non-group escapees (KeyboardInterrupt, SystemExit) after delivering
them to the readiness future - failure delivery is the arm's job;
swallowing an interpreter-level exit was not.
* fix(mcp): extend owner-death discovery fast-fail to the static path
The static connect path had the same exposure the pool's discovery race
closed: discovery runs in the connecting caller while the transport is
hosted by the owner task, so a transport collapse mid-discovery cancels
the OWNER and the caller's bare await on the response stream hung until
the caller-side attempt timeout (~45s) instead of failing promptly.
_await_pool_discovery is renamed to _await_owner_discovery (it is now
path-neutral) and wired into _connect_one_locked's four discovery
awaits. The helper also converts a discovery future that completes
CANCELLED without the race's own reap (an SDK-internal cancellation
shape) into the same ConnectionError, instead of leaking a bare
CancelledError the caller would misread as its own cancellation.
The pool transport owner's BaseException arm gains the same refinement
the static owner received in review: interpreter-level exits
(KeyboardInterrupt, SystemExit) re-raise after delivery to the
readiness future instead of being swallowed.
Tests: static owner-death-during-discovery fast-fail (<1s vs the ~45s
hang), and a direct pin on the cancelled-discovery-future conversion.
* fix(mcp): replace owner BaseException arm with targeted catch + finally delivery
The owner's failure arm now catches only (BaseExceptionGroup, Exception);
waiter delivery for everything else moves to a finally that resolves the
readiness future with a clean transport-failure ConnectionError before
the task unwinds. Interpreter exits and BaseException-derived library
control-flow escapes propagate from the owner exactly once, uncaught -
and the waiter can never be left hanging on an unresolved future (the
initial _connect_all connect has no outer bound). For SystemExit /
KeyboardInterrupt asyncio additionally stops the loop right after, so
the delivery is load-bearing for the non-exit BaseException shapes and
free for the exits.
Pinned by a test driving a BaseException-derived escape through the
owner: the waiter resolves promptly with ConnectionError while the
escape propagates unswallowed.
* fix(mcp): mirror targeted-catch + finally delivery in the pool owner
Same shape the static owner received in review: the failure arm catches
only (BaseExceptionGroup, Exception), and waiter delivery for anything
else moves to a finally that resolves the readiness future with a clean
ConnectionError before the task unwinds - interpreter exits and
BaseException-derived library escapes propagate exactly once, uncaught,
and the waiter can never be left hanging.
* test(mcp): narrow the escape test's waiter catch to explicit types
* test(mcp): narrow discovery-race waiter catches to explicit types
* refactor(mcp): make reap/synchronization awaits explicit to analyzers
Full-absorb reaps (cancel-then-drain of a future whose outcome is
deliberately consumed) become `await asyncio.gather(x,
return_exceptions=True)` - one line, self-describing, and in the
owner-died discovery reap it is also a small semantic improvement: a
caller cancellation arriving during the reap now propagates instead of
being masked by the ConnectionError. Bare synchronization awaits and
selective suppress blocks in tests keep their raise-through semantics
via throwaway assignment. Applied uniformly across the owner-task
test files, including sites introduced by the static-path PR.
(cherry picked from commit 0c2c534c86)
The owner's failure arm now catches only (BaseExceptionGroup, Exception);
waiter delivery for everything else moves to a finally that resolves the
readiness future with a clean transport-failure ConnectionError before
the task unwinds. Interpreter exits and BaseException-derived library
control-flow escapes propagate from the owner exactly once, uncaught -
and the waiter can never be left hanging on an unresolved future (the
initial _connect_all connect has no outer bound). For SystemExit /
KeyboardInterrupt asyncio additionally stops the loop right after, so
the delivery is load-bearing for the non-exit BaseException shapes and
free for the exits.
Pinned by a test driving a BaseException-derived escape through the
owner: the waiter resolves promptly with ConnectionError while the
escape propagates unswallowed.
(cherry picked from commit 8ed86ae7ab)
Review follow-ups on the owner-task migration:
- _maybe_disarm_orphaned_scopes now enforces its mcp-loop requirement
instead of trusting callers: it returns without walking (and without
advancing the rate-limit clock) unless the currently running loop IS
self._loop. A suppressed close can fire before start() or after
shutdown(), where the walk would be wasted at best and a cross-thread
reach at worst.
- The transport owner's BaseException arm now re-raises non-Exception,
non-group escapees (KeyboardInterrupt, SystemExit) after delivering
them to the readiness future - failure delivery is the arm's job;
swallowing an interpreter-level exit was not.
(cherry picked from commit ed30e4f0bf)
A crash-looping MCP server drove the mcp-loop thread to a sustained,
climbing 100%+ CPU spin. Root cause: anyio cancel scopes are
host-task-bound, and the static path entered the SDK's transport /
ClientSession task-group scopes from short-lived connect tasks (every
health tick is a new task since #768). Once such a scope was cancelled
after its host task had finished - by anyio's task_done when a
transport child died with the server, or by ClientSession.__aexit__
during a cross-task teardown - CancelScope._deliver_cancellation could
never make progress (task.cancel() on a done task is a no-op) and
re-armed itself via call_soon every loop iteration, forever: ~900k
callbacks/s per zombie scope, one more per flap cycle (verified against
anyio 4.14.1; no upstream fix exists as of that release).
Fix: each static server's transport + session cms are now entered,
parked, and exited by ONE long-lived owner task
(_static_transport_owner), so scopes always have a live host and always
exit in the task that entered them. Teardown follows a one-cancel close
protocol (signal the close event before the first await, graceful
grace, then at most ONE cancel - never a second, which would abandon a
scope exit mid-flight). Connect timeouts now cancel only the waiting
caller; connect failures are delivered through a readiness future;
unrequested owner death (server died under a live session) evicts the
session immediately via a done-callback instead of waiting for the next
liveness ping. A rate-limited, mcp-loop-scoped gc-walk backstop
(_maybe_disarm_orphaned_scopes) disarms any zombie minted by paths not
yet migrated (the oauth_user pool keeps the old cross-task-close shape;
follow-up).
Also fixed: BaseExceptionGroup (BaseException-derived, as raised by
anyio task groups wrapping a stray CancelledError, e.g. an
accept-then-RST server) escaped `except Exception` in _connect_all and
killed it before the health/sweep loops were created - silently
disabling all autonomous recovery. Handled there and in the
health/sweep/eviction loops and the reconnect/refresh callers.
Verified: a live SIGKILL-flap repro went from 130%+ CPU (climbing, one
armed scope per cycle) to 0.3% flat with zero armed scopes; the RST
repro now leaves both background loops alive (previously both silently
dead). New tests: owner-lifecycle + close-protocol units (incl. an
exactly-one-cancel pin), a _connect_all BaseExceptionGroup regression,
a discriminating disarm-sweep test, and a ~10s live SIGKILL-flap smoke
test (real FastMCP subprocess, skips on environment gaps) asserting
zero armed scopes, exactly one live owner, and a post-recovery tool
call. Full 8470-test suite green; ruff+mypy clean.
(cherry picked from commit 62f62ae624)
The bounded key prefix pattern (api_key=/secret_key= etc.) avoids
monkey/turkey false positives, but compound keys like secret_access_key
and aws_secret_access_key only matched on the access_key= suffix, leaking
the secret_ / aws_secret_ prefix. Added these as explicit alternations.
Also added bearer_token and secret_token to the token prefix list.
(cherry picked from commit c5ff3147ce)
Bare alternatives (|token, |key) for standalone token= and key= assignments
were re-added. A negative lookbehind (?<![a-zA-Z0-9_]) prevents matching
word-suffixed identifiers like monkey=, turkey=, mytoken=, over_tokenized=.
Also applied the same protection to _RE_QUERY_CRED which had a bare |token
alternative without boundary protection.
(cherry picked from commit bfcfb0c791)
- Replace unbounded [a-zA-Z0-9_]*key= prefix with specific credential
key suffix alternation to avoid false matches on monkey=, turkey=, etc.
Same for *token= (access_token=/auth_token= but not over_tokenized=).
- Hoist redactCredentials() out of per-line diff render loop in
buildConvCmd - 1 call on full text instead of N calls per line,
eliminating ~2800 regex passes worst-case.
- Remove bare 'key' from _RE_QUERY_CRED alternation (too aggressive).
- Add x-api-key / x_api_key to JSON secret key lists in both Python
and JS.
- Add re.IGNORECASE to configurable-mode credential_bearer pattern.
- Fix annotation dedup guard in _check_credentials (was checking flag
name against annotation prose list, always-true dead code).
(cherry picked from commit cbf5c5f3b6)
Address review findings on the client-side credential redactor and mirror
each fix into the backend output guard so both surfaces censor identically:
- Detect and redact single-quoted JSON secrets such as
{'Authorization': 'Bearer ...'} (Python dict reprs / JS object literals),
which the double-quote-only pattern silently bypassed on both sides.
- Cover mongodb+srv://, rediss:// and amqps:// connection strings.
- Match the Bearer auth scheme case-insensitively (RFC 7235).
- Redact prefixed key/token assignments (api_key=, secret_key=,
access_token=) as a whole rather than chewing the tail into a garbled
"api_[REDACTED:api_key]", while still covering bare key=/token=. A word
boundary was rejected because it would drop coverage for <prefix>_key=.
- Remove the redundant |Authorization alternative (covered by /i) and swap
the manual value-slicing helper for a capture-group substitution.
The backend edits touch both detection sites and both redaction pipelines,
so single-quoted secrets are flagged (and therefore sanitized), not merely
rewritten. Adds JS runtime-smoke and backend unit coverage for every case.
(cherry picked from commit 31a1d5c3ee)
New shared ES6+ module (redact_credentials.js) provides comprehensive
visual credential censorship matching the backend output guard patterns:
- PEM private key blocks, connection strings, Bearer tokens
- OpenAI / GitHub / AWS / Google API key formats
- Query-string and JSON-style credential values
- JSON secret keys (api_key, password, token, authorization, etc.)
- ENV secret lines (SECRET_KEY=, DATABASE_URL=, etc.)
Integrated into both frontend surfaces:
- interactive.js: replaces legacy minimal _redactApiKeys function
- conversation.js::buildConvResult (shared substrate, used by coordinator)
- coordinator.js::renderToolOutput fallback paths
Backend parity: added 'authorization' to the JSON secret regex in
output_guard.py so the output guard flags and redacts Authorization
headers in JSON tool output.
Tests: ported the runtime smoke test from the removed _redactApiKeys
to import the new module directly; added redact_credentials.js to the
var-free and const-reassign guard bundles.
(cherry picked from commit 93a7486cc2)
`tool_args_preview` feeds `stream.tool_args_malformed` (WARNING) and
`wire.tool_args_legalized` (DEBUG), and tool arguments are model/user
controlled — they can carry secrets (a token in a bash command, a password in a
connection string) or raw CR/LF that break log lines. Route the preview through
`output_guard.redact_credentials` over the full value first (before the 120-char
cap, so a secret straddling the cut isn't half-shown past the pattern's reach),
then collapse every control char to a space, mirroring `audit._scrub_string`.
Addresses the PR review comments.
(cherry picked from commit 56624f9597)
A tool call whose `arguments` is not a JSON-object string (an unterminated
string from a non-`length` truncation, or an empty `""` from a no-arg call)
was committed verbatim and replayed on every subsequent send. Strict renderers
that re-parse arguments at render time (vLLM's `deepseek_v4`
`_postprocess_messages` runs `json.loads` on them) reject the whole request
with HTTP 400, wedging the conversation. The only prior guard dropped partial
tool calls on `finish_reason == "length"`; a `stop`/`tool_calls` finish reason
carrying invalid JSON slipped through, and its synthetic "retry" result kept it
from being an orphan, so the orphan-repair pass never touched it.
Add `sanitize_tool_call_arguments`, a wire-neutral legalize pass in lowering
(fold, legalize, repair), normalizing any non-JSON-object `arguments` to `{}`
on the transient wire copy only. The canonical trajectory keeps the raw model
output, so a wedged session self-recovers on its next send. A
`wire_valid_arguments` predicate is shared with a non-destructive
`stream.tool_args_malformed` warning at the stream accumulator, which surfaces
the model-quality problem at production time.
Convert `lowering.py` to structlog so the new pass emits structured events.
(cherry picked from commit 16a68ae6d6)
The pane/workstream accelerators only worked on macOS. They were bound to
Ctrl, which on Windows/Linux IS the browser's own accelerator: Ctrl+T,
Ctrl+W and Ctrl+1-9 were swallowed by the browser (new tab / close tab /
switch tab) and never reached the page. macOS browsers own Cmd instead, so
Ctrl was free there and everything appeared to work.
On top of that the shortcuts were declared in three places that had drifted
apart — the "?" overlay, each app.js keydown handler, and the tab-menu
badges in shell.js. The console fell to convTabMenu's node-proxy fallback
lane, which dropped every shortcut badge (and Fork), so its tab menu showed
no accelerators and Ctrl+W there just closed the browser tab.
Choose the modifier per platform (Ctrl on macOS, Alt on Windows/Linux) and
make shell.js the single source of truth for the per-pane accelerators: a
stable accel registry drives both the platform-aware badge and one shared
keydown handler that invokes the ACTIVE pane's own menu item, so a badge
can't advertise a chord the handler ignores and each surface contributes
only what it supports (the console omits Fork; it has no fork surface yet).
Each surface's app.js keeps only its global accels (new / switch /
dashboard); the console regains switch + dashboard to match. Mod+W now
uniformly means Close pane (drop the tab, session keeps running), matching
its badge and the universal Ctrl+W convention — previously the standalone's
Ctrl+W stopped the session. The previously-dead "Refresh title / Ctrl+Shift+R"
is wired, and Ctrl+T / Ctrl+D yield to text editing while a field is focused
(macOS transpose / delete-forward).
(cherry picked from commit 2d4cb6fea9)
* Fix send button stuck disabled by pruning orphaned approval cycles
When a DOM wipe (clear_ui / replay_truncated / replaceChildren)
detaches approval card elements while an approve_request event is
processed between the wipe and refetch-restore, the matching
approval_resolved may never arrive. The orphaned cycle entry in
approvalCycles keeps pendingApproval=true and the send button
disabled forever.
The fix adds a pruning pass at the top of _syncApprovalState():
cycles whose blockEls are all .isConnected === false are deleted
from the Map. This runs on every register/resolve/rebuild so
orphans are cleaned up promptly.
Also fixes an ordering bug in showInlineToolBlock discovered
during review: the block element was appended to the DOM after
_registerApprovalCycle, so the new isConnected prune would kill
the just-registered cycle before it took effect.
* Fix comment inaccuracy in showInlineToolBlock append-before-register guard
The comment said 'blockEls.every(el => el.isConnected)' but the
actual prune check is '!blockEls.some(el => el.isConnected)' -
no block elements are connected, not every element.
The wire-payload golden matrix had no anthropic-compatible coverage —
both AnthropicProvider rows are the native lane (compat=False), so the
distinct compat wire shape (reasoning control in
extra_body.chat_template_kwargs, never the native thinking param) was
unfrozen. Add the compat lane across all eight representative fixtures
with a manual-mode capability (the lane has no static table, so caps
ride in as a model definition would supply them) and reasoning_effort=
high: every golden now pins {enable_thinking: true, reasoning_effort:
high} in chat_template_kwargs, asserts the native thinking param is
absent, and preserves temperature (no forced 1.0). _capture gains an
optional caps override to support the no-static-table lane.
Copilot review caught the native Anthropic adaptive ladders (sonnet-5,
fable-5, opus-4.8/4.7/4.6) labeling the none position 'None — sends
adaptive' — raw mode vocabulary the plain-language rule exists to
prevent. But treating the 'adaptive' token identically to the local
lane's 'on' would still misframe it: on these models thinking is always
on and an effort level rides output_config on every graded position
(verified on the wire), so 'thinking stays on' is true everywhere and
not what distinguishes none. The none position uniquely means no effort
is pinned — the model self-regulates it — so it now reads 'None — model
sets effort'. The local adaptive lane's bare toggle keeps 'thinking
stays on' (no effort lever exists there).
Aliased knob positions were labeled after the lowest sibling sharing
their wire token — a toggle-only model rendered 'Max (= minimal)',
implying a minimal-effort downgrade the wire doesn't contain, and with
declared values 'High (= minimal)' while the wire carries high. Each
position now states its delivered level: exact matches stay plain
('Max'), snapped positions say 'Low — sends high', the adaptive none
position warns 'thinking stays on', budget detail stays in the
tooltip. Effort-param placeholder corrected to the real graded keys
(reasoning_effort / reasoning).
Local lanes dropped the knob's graded value unless the operator declared
reasoning_effort_values (and, on the template channel, an effort key) —
picking Max sent a bare thinking toggle and the effort select
degenerated into seven positions that all meant 'on'. The user's
setting now always rides:
- openai-compatible: the flat reasoning_effort param carries the knob
verbatim (effort_passthrough on the lane default); declared values
still snap ordinally, and a declared effort_param still claims the
template channel and suppresses the flat param.
- anthropic-compatible: the graded value rides chat_template_kwargs
alongside the toggle whenever reasoning control is engaged — under
the operator's effort_param, else the conventional fallback key
(reasoning_effort); templates that don't reference the kwarg ignore
it. thinking_mode=none still injects nothing.
- Commercial lanes untouched: empty declared values still mean 'no
effort control' (o1-mini) and the ordinal snap is unchanged.
Golden writer now pins ensure_ascii=False: the baselines' literal em
dashes came from a hand edit (03f82521) the default-escaping writer
could never reproduce — regens no longer churn unrelated lines.
Local-lane model ids are operator-chosen strings (vLLM
--served-model-name), so a prefix collision with a cloud model id
inherited that model's sampling and effort contract: a box named
o3-distill silently lost temperature support, and one named
gpt-5.5-my-finetune was sent gpt-5.5's snapped reasoning_effort values
it never declared. Both surfaces of the lane now return plain defaults
(OPENAI_COMPAT_DEFAULT in _openai_common): the chat class directly, and
the responses pin via a compat-mode OpenAIResponsesProvider mirroring
AnthropicProvider(compat=True). Everything beyond the defaults is
declared by the operator on the model definition, matching the
anthropic-compatible lane and lookup_model_capabilities' documented
'no static table for local models' contract. The commercial openai
lane (Responses-only) is untouched.
Pre-split tests that reached commercial rows through the chat-class
OpenAIProvider alias now source them from lookup_openai_capabilities;
their subject (registry rows + shared gating helpers) is unchanged.
The except path for a malformed capabilities column appended the row
without the key, so clients had to null-check a field the happy path
guarantees. Initialize each entry with an empty ladder and let the try
block overwrite it — the response schema is stable per row.
Three server.test.ts / server-attachments.test.ts cases still asserted
the pre-verb-lift send shape (POST /v1/api/send with ws_id in the body).
The server route has been POST /v1/api/workstreams/{ws_id}/send (ws_id
in the path, {message} body) since that lift, and the SDK send() was
updated with it — only these expectations rotted. They fail on main
too; unrelated to approvals, folded in here to leave the TS SDK suite
green. Test-only, no SDK source change.
ApproveRequestEvent.cycle_id and ApprovalResolvedEvent.cycle_id /
call_ids were typed required, but they are 1.7 additions: a pre-1.7
server omits them on the wire, so a current SDK talking to an older
node sees undefined. The UI and channel adapters already keep the
legacy no-selector fallback for exactly that case. Mark them optional
so consumer code can't assume a string that may be absent — matching
the Python SDK, whose dataclasses default all three.
The interactive composer had two uncoordinated writers of
sendBtn.disabled: _syncApprovalState wrote `pendingApproval || busy`
directly, while Composer.setBusy wrote it too via _reconcileDisabled.
In queueWhileBusy mode the composer's write re-enables send, so a
state_change to "attention" (exactly the pending-approval state)
firing after an approve_request card rendered raced the approval
disable back off. The `|| busy` term also defeated the "Queue
message…" affordance whenever _syncApprovalState ran mid-turn.
Opt the composer into externalDisable (it keeps rotating the
Send/Queue label + placeholder + stop button, but no longer writes
the flag) and route every axis — live approval cycle, cross-user send
gate, busy — through one App reconciler, _reconcileSendDisabled. Busy
is intentionally not a disable axis here: queueWhileBusy keeps Send
clickable as "Queue" while the agent runs.
Pre-existing on main (the old scattered direct writes had the same
collision); surfaced by the concurrent-cycle review.
The checked-in openapi-server.json / openapi-console.json still
described the removed singular pending_approval_detail field; re-run
generate-types.py so the reference specs carry the
pending_approval_details list + cycle_id. Retire two docstring
references to the deleted singular serializer and one stale comment.
New regression matrix for the release blockers: cross-approval
independence, lost-wakeup at gate entry, FIFO selector-less
resolution, resolve-all sweep, double-resolution no-op, cards/legacy
view tracking, and the generation-exactness set — stale delivery
rejection, Smart-Approvals origin check, purge keep_origin, the
purge-to-register window eviction, late cross-generation "superseded"
stamping, concurrent smart+human gates, and the pre-delivered-verdict
fast path. Plus sub-agent judge wiring (agent_gate off the main
slot, close() firing all generations) and endpoint tests for cycle
pinning and the Approve+Always race guard.
Gate threads run under one shared mock-patch harness — mock.patch
start/stop of the same target from concurrent threads corrupts the
patcher's restore stack — with a sweep-until-dead teardown so the
conftest leak guard can't trip. Existing suites migrate off the
singleton fields to cycle assertions and the
pending_approval_details wire shape.
interactive.js tracks live cycles in a Map: per-cycle action buttons
and feedback, per-cycle optimistic clears and resolved-status pills,
keyboard routed to the oldest (or the focused) cycle, announce-shell
dedupe, and a composer that stays disabled while ANY cycle is live.
Verdict glow is scored per batch — a sibling's verdict neither
recolors the oldest card nor leaves its own card stale.
coordinator.js renders one approval block per pending_approval_details
entry, posts child approve/deny with the block's cycle_id, clears
exactly the resolved cycle on approval_resolved (legacy events without
one clear all), and replays every entry from snapshots.
Track posted approval prompts by (ws_id, cycle_id) — one message per
concurrent cycle — so parallel task agents' prompts resolve
independently. Buttons carry the cycle in their value (Slack) /
footer (Discord); intent verdicts route onto the owning cycle's
message by call_id membership.
The exact-key-then-legacy-fallback lookup (an empty cycle_id from a
pre-multi-cycle server resolves the workstream's single tracked
entry) and the all-cycles sweep are centralised as shared _routing
helpers so the fallback semantics can't drift between adapters.
POST /approve accepts cycle_id / call_id selectors and 409s on stale
selectors with the current cycle's ids so clients re-render instead of
silently resolving an unrelated batch. Selector-less bodies pin the
resolve to the cycle the lookup returned (not "whichever is oldest by
the time the resolve runs"), and Approve+Always names apply only after
the pinned cycle actually resolved — the auto-approve whitelist can no
longer describe a different batch than the one that resolved.
approve_request carries cycle_id; approval_resolved carries cycle_id +
call_ids; SSE reconnect replays every live cycle's card. The console
collector, coordinator UI fan-out, and both SDKs (Python + TS) thread
the cycle correlation through.
BREAKING (1.7): the singular pending_approval_detail field is removed
from dashboard rows, workstream detail, and node snapshots — replaced
by the pending_approval_details list (one entry per live cycle, each
carrying its cycle_id). stable/1.6 keeps the old shape.
task_agent tool calls reached the approval gate judge-blind: no
heuristic verdict on the card, no LLM verdict, no audit row, and Smart
Approvals could never clear them. Run _evaluate_intent on the
sub-agent gate with the sub-agent's own trajectory as judge context —
its task prompt is the delegation contract the operator approved, so
"does this call serve the task" is the right local alignment question.
Sub-agent spawns are agent_gate generations: they never touch the main
loop's supersede slot (parallel siblings would make each other's
verdicts look stale), staleness is enforced per-cycle by the UI's
generation checks instead. Every generation registers in
_judge_cancel_events — kept exact by the judge's new done_callback —
so close() aborts all in-flight daemons; judge.cancel_on_approval
fires per-gate exactly like the main loop. CLI and eval UIs accept
the judge_event delivery kwarg.
The approval pipeline was a per-UI singleton (one card, one Event, one
result slot) multiplexed by N concurrent gates. With parallel task
agents that meant one click could resolve every parked batch with the
same verdict, a sibling's gate entry could eat a just-fired resolution
(3600s "stuck dialog" hang), and Approve+Always could whitelist a
different batch than the one on screen.
Replace the singleton with a registry of ApprovalCycle objects keyed
by cycle_id: per-cycle events/results/decisions/verdict parking,
oldest-first selector-less resolution, guarded double-resolution, a
resolve_all_approvals sweep for cancel/close/worker-recovery, and a
maintained oldest-cycle view in the legacy _pending_approval slot for
boolean-ish consumers.
Verdict bookkeeping is generation-exact end to end: the entry purge
spares verdicts the entering batch's own judge spawn already delivered
(a fast judge no longer stalls the Smart-Approvals wait to its full
budget), registration evicts stale-generation arrivals that land in
the purge-to-register window, recent decisions are generation-tagged
so a stale generation's late verdict stamps "superseded" instead of
stealing a reused call_id's decision, and Smart-Approvals
qualification identity-checks the delivering generation.
Capability-registry corrections verified against the official OpenAI
reasoning guide, the Azure reasoning-models matrix (2026-06 revision),
and the Anthropic models-overview/effort/migration pages (2026-07):
OpenAI (vocabulary confirmed none/minimal/low/medium/high/xhigh — no
"max" level exists; knob max rides the xhigh ceiling via the ordinal
snap):
- o1/o3/o3-mini/o3-pro/o4-mini declare low/medium/high (every o-series
model except o1-mini) — without declared values the session knob was
silently dropped for these models. o1-mini stays effort-free.
- gpt-5.5 default corrected none -> medium (5.5 reasons by default,
unlike 5.1-5.4).
- gpt-5.1-codex-max gets an explicit row: it prefix-matched the
gpt-5.1 row (no xhigh), capping the knob's xhigh at high on the one
model xhigh was introduced for.
Anthropic (effort-page matrix):
- claude-sonnet-5 row added — it previously fell through to
_ANTHROPIC_DEFAULT (manual budgets, 200k ctx, no effort), all wrong:
adaptive-by-default thinking (manual budgets are a 400), sampling
params rejected, 1M ctx / 128k out, effort low..max incl. xhigh.
- claude-sonnet-4-6 gains its documented "max" effort level (knob
xhigh now rides max, not high) and the stale 64k max_output becomes
the documented 128k.
- fable-5 / opus-4-8 / opus-4-7 / opus-4-6 / opus-4-5 rows verified
correct as declared.
Knob semantics completed: resolve_reasoning_effort now forwards the
knob's "none" position verbatim when the model DECLARES an explicit
none level (gpt-5.1+, grok-4.3) — omitting the param there leaves a
reasoning-on server default (gpt-5.5: medium) in charge of a knob
that promises off. Models without a declared none still omit, and
none is never a snap target. Parity harness swaps its synthetic
openai shape for the real gpt-5.5 registry row.
The knob domain grew xhigh/max after the snapping fallbacks were
written, which silently inverted their semantics: off-list meant
"unrecognized string" then, but now usually means "above the model's
ceiling", where falling back to the default tier is directionally
wrong (grok-4.3 at knob max got low; values low/medium/high at knob
xhigh got medium; Anthropic manual mode gave xhigh/max a 4096 budget
while high got 16384).
One rule everywhere now, via snap_reasoning_effort in _protocol:
exact match wins; otherwise the smallest declared level ranking at or
above the knob; above the ceiling, the ceiling. "none" is never a
snap target, and default_reasoning_effort only catches values the
ordinal snap cannot rank.
- resolve_reasoning_effort (flat chat / responses / validated
effort_param lanes) snaps ordinally: xhigh over (low, medium, high)
now sends high; xhigh over DeepSeek-style (high, max) sends max —
matching DeepSeek's official xhigh-to-max aliasing, so a declared
values list now reproduces that contract instead of defeating it.
- _map_reasoning_to_effort (native output_config) rounds up too:
knob xhigh on Opus 4.6 (low, medium, high, max) rides max instead
of silently dropping output_config.
- EFFORT_BUDGET_MAP is monotone across the whole knob domain:
minimal/low 1024 (API floor), medium 4096, high 16384, xhigh 32768,
max 65536. Unknown strings still fall to the 4096 default.
Google defaults are unaffected (ceiling and default coincide at
high); wire goldens unchanged. Parity harness caught the budget
clamp interacting with its own max_tokens during development —
capture budget raised above the largest manual budget.
Proves the effort-ladder projection against the real request path
instead of against the mapping helpers it shares with it. For 22
(provider lane x capability shape) points — both Anthropic lanes,
openai-compatible on both API surfaces, openai, google (default and
template-override hybrid), xai (default and inert-override), and the
DeepSeek/qwen template contracts — every knob position is driven
through the actual provider create_streaming against a recording fake
client, and two invariants are asserted per shape:
1. each ladder token decodes to an expected effort wire subset
(toggle / template effort / flat param / thinking budget /
output_config) that must equal the captured kwargs exactly;
2. two knob positions carry equal tokens iff they produce identical
effort-relevant wire payloads — the grouping promise the UI
annotations lean on.
The RecordingClient SDK-seam stub moves from the wire-payload golden
harness into tests/_wire_capture.py so both suites capture at the same
seam. Verified the harness catches the bug class it was built for:
re-adding xai to _CHAT_LANES fails xai-template-override-inert.
Second external audit round on the ladder. Verified and fixed:
- xai was in _CHAT_LANES on the false premise that XAIProvider
subclasses the chat provider. It subclasses OpenAIResponsesProvider,
whose surface ignores extra_body entirely, so a thinking_mode /
effort_param override never changes an xai request — but the ladder
claimed a template toggle ("on+low") that does not exist on the
wire. xai now projects through the flat channel only, like openai.
- The flat-param suppression rule (a declared effort_param claims the
template channel) was encoded independently in
apply_temperature_and_effort and the ladder. Extracted into
flat_effort_suppressed() in _protocol so the request path and the
projection cannot drift.
- admin_effort_ladder logs the swallowed resolver exception before
returning 400 (a genuine bug would otherwise hide as a silent 400).
- list_available_models reads server_compat with .get() instead of
destructively popping it out of the parsed capabilities dict.
- models_changed SSE now re-annotates the skill launch-config effort
select after invalidating the models cache instead of leaving a
stale ladder until the next keystroke.
- Capabilities JSON textarea placeholder hints the two effort fields
that have no structured control (reasoning_effort_values,
default_reasoning_effort).
The Gemini effort fix (ee9e9c1f) adds a flat reasoning_effort to every
Google chat-completions request at the session default knob — the
golden fixtures now carry it. Only the eight google__* goldens change
(one added key each); other providers' goldens are untouched — the
UPDATE_WIRE_GOLDENS pass also wanted to rewrite twelve passing goldens
with escape-format-only churn (raw em-dash vs \u2014), reverted to
keep the diff semantic.
The one that mattered: /v1/api/models passed the capabilities column —
a JSON STRING (sa.Text) — straight into effort_ladder_for_model, whose
field filter calls .items() on it; the per-row guard swallowed the
AttributeError, so effort_ladder was silently absent from every row and
the sklc annotation could never fire. The endpoint now parses the JSON
and splits the namespaced server_compat exactly like the model_registry
loader, and a regression test seeds a string-capabilities row.
Projection fidelity: effort_ladder_for_model threads api_surface (the
responses surface ignores extra_body — flat-param-only ladder, matching
create_provider's request-time divergence); google/xai route through
the chat-lane projection they actually inherit (_finalize_extra_body +
flat param); the native-Anthropic branch reflects that output_config
gates on supports_effort alone, independent of thinking_mode; budget
clamping to per-request max_tokens is documented as out of scope.
Hardening and symmetry: admin endpoint 400s (not 500s) on non-dict JSON
bodies and gained HTTP tests; the admin edit-load strips effort_param
from the raw JSON only for the lanes whose save path re-adds it, so
non-compat rows can't silently lose a stored key; the empty-model early
return bumps the ladder sequence so stale in-flight responses can't
re-annotate; alias labels only reference positions the target select
actually offers (the skill shelf omits none/minimal); the sklc models
cache no longer pins a rejected promise and is invalidated on the
models_changed event; debounces unified at 500ms;
merge_reasoning_template_kwargs now always returns a fresh dict for
non-empty input; the shared budget constants are public.
Seven knob positions render as seven behaviors in the UI, but the real
ladder depends on the lane and the model: qwen3.6 has two (off/on),
DeepSeek-V4 three, Claude 4.6 five. Operators had no way to see which
positions alias — the confusion class behind silently-equal effort
levels.
providers/effort_ladder.py projects the knob domain through the same
mapping functions the providers use at request time (resolve_reasoning_
effort, reasoning_template_kwargs, the manual budget map — hoisted to a
shared constant so the projection can't drift), yielding
{value, effective} rows where equal tokens promise identical requests.
/v1/api/models rows now carry the ladder (guarded per row), and
POST /v1/api/admin/models/effort-ladder computes it for the admin
modal's unsaved edits.
The admin per-model effort select and the skill launch-config effort
select annotate aliased positions ("Max (= high)", "None (model
default)") with a sends-tooltip; annotations refresh as thinking-mode /
effort-param / capabilities fields change. The ladder describes what
Turnstone sends — server-side templates may alias further (DeepSeek-V4
folds low/medium into its default high tier).
_GOOGLE_DEFAULT declared no reasoning_effort_values, so
resolve_reasoning_effort returned None and the session effort knob was
silently dropped for every Gemini model — the same bug class this
branch fixed on the local lanes. Gemini's OpenAI-compat surface
documents a flat reasoning_effort (2.5: thinking_budget mapping; 3.x:
thinking_level), so declaring values lights up the inherited
chat-completions path.
Values are the safe cross-model set (minimal/low/medium/high): "none"
is excluded because 2.5 Pro and the 3.x family reject disabling
thinking — and the resolver never forwards the knob's none anyway (the
param is omitted, server default applies). Off-list xhigh/max snap to
the declared default high. Encoded from the official compatibility
docs per the static-caps pattern; not live-verified.
Post effort-knob rework, "Enabled" (manual) means knob-controlled —
effort none turns thinking off. Operators who want the pre-#771
always-on behavior (knob never disables) previously had to hand-write
thinking_mode "adaptive" into the raw capabilities JSON. The dropdown
now offers all three representable modes — None / Effort-knob
controlled / Always on — and the edit-load lift captures adaptive
instead of relegating it to raw JSON.
Cross-checked online: Qwen3.6's template documents enable_thinking +
preserve_thinking only — no effort parameter exists (vLLM's flat
reasoning_effort convenience boolean-maps to the same toggle).
DeepSeek-V4 officially accepts reasoning_effort high/max with Think
High as the default thinking tier and low/medium→high, xhigh→max
aliasing — so freeform effort_param passthrough matches the contract
exactly, and a declared values list omitting xhigh/max would make
Think Max unreachable. Live probes on both boxes agree with the
official contracts once the default-tier framing is applied.
Verified findings applied:
- adaptive thinking_mode never knob-disables: the shared mapping now
sends the toggle unconditionally true for adaptive (the native
adaptive branch ignores the knob's none), while manual keeps the
knob-driven contract. Restores the invariant the deleted chat-lane
code upheld.
- a set effort_param suppresses the flat top-level reasoning_effort on
the chat lane: the template channel replaces it — double-sending
could 400 on schema-strict servers and disagree with operator pins.
- admin edit-save no longer drops a stored thinking_param when the
thinking-mode dropdown is empty: the raw-JSON strip now only fires
when a mode value actually round-trips through the dropdown.
- effort_param persistence gated on the local-server lanes so a value
lingering across a provider switch never lands on commercial rows.
- three stale _compat_extra_params references renamed to
merge_reasoning_template_kwargs.
Documented dispositions (no code change): the knob-none-disables flip
on upgrade is intentional and now carries an upgrade note; gateways
fronting real Claude belong on provider=anthropic with a custom
base_url (the compat lane is vLLM-schema-only); nonstandard
thinking_mode strings staying inert is the intended allowlist
contract. The real anthropic provider is unaffected throughout —
official Claude models keep native thinking/output_config.
Hoist the compat-lane injection into _protocol.merge_reasoning_template_kwargs
(next to ModelCapabilities — one implementation for both local-server lanes)
and retire OpenAIChatCompletionsProvider._apply_thinking_mode in its favor:
_finalize_extra_body now receives the session effort knob, so thinking_mode
manual/adaptive maps knob "none" to an explicit thinking_param false
(previously the toggle was unconditionally true) and caps.effort_param
carries the graded effort key on chat completions too. Operator
server_compat pins still win; the Responses surface is untouched (native
reasoning handles effort itself).
The admin Models form grows an "Effort param" field that round-trips like
thinking_param: lifted out of the raw capabilities JSON on edit-load,
re-added on save, cleared by emptying the field.
Verified live against qwen3.6-27b /v1/chat/completions: knob medium streams
reasoning_content, knob none suppresses it.
The compat lane sent no reasoning control at all: vLLM's /v1/messages
has no thinking request field, thinking_mode stayed "none", and the
session effort knob was silently dropped. The reasoning levers live in
the chat template, so fold them into extra_body chat_template_kwargs
(_compat_extra_params): thinking_mode manual/adaptive maps the knob
onto caps.thinking_param (effort "none" = off, mirroring the native
manual-mode contract), and caps.effort_param (new ModelCapabilities
field) carries a graded effort value for gpt-oss-style templates,
validated against reasoning_effort_values when declared. Operator
server_compat entries win on key collision; native thinking params,
temperature forcing, and output_config never fire on compat.
resolve_reasoning_effort moves from _openai_common to _protocol next to
ModelCapabilities — importing it into _anthropic would otherwise cross
provider families.
The admin Models form now shows and round-trips the thinking-mode
dropdown for this lane; the #661 hide was premised on thinking_mode
being inert here, which this change inverts.
Verified live against qwen3.6-27b on vLLM /v1/messages: knob medium
streams a thinking block, knob none suppresses it, an operator pin
beats the knob.
* feat(mcp): autonomous reconnect + liveness for static MCP servers
Static (non-oauth_user) MCP servers had no autonomous reconnect. Every reconnect
path was lazy — a tool dispatch (_cb_auto_reconnect), an operator refresh, or a
config edit — and the MCP SDK's own reconnect is a bounded 2-attempt burst on the
streamable-http GET stream only (verified: mcp 1.28.1), with no backoff and
nothing for the other transports. So a static server that went down and came back
while nobody was dispatching to it stayed disconnected until a dispatch or a
manual reconnect. Worse, a session whose transport dies while idle survives as a
non-None ClientSession with closed streams — nothing evicts it, so even a later
dispatch may not notice until it fails.
Add a static-server health loop on the mcp-loop (started in _connect_all; config
``static_health_check_seconds`` default 30, <= 0 disables):
- Reconnect: a disconnected server (session is None) is reconnected on a capped,
jittered, FOREVER backoff (full jitter, base 1s, cap 60s, no attempt limit) —
a server that returns after a long outage reconnects within ~a minute, and a
permanently-misconfigured one costs at most one attempt per cap. The health
loop owns this clock; the circuit breaker stays the DISPATCH fail-fast gate (a
tool call to a down server errors immediately rather than blocking on the
retry), and the loop keeps breaker state in sync so an open breaker closes on
reconnect.
- Liveness: a connected server is pinged (send_ping) each cadence; a dead-but-
idle one — which nothing else would notice — is evicted so the next tick
reconnects it. This is the core of the "never reconnects" failure.
Serialize _connect_one per server behind a per-name lock (split into a thin
wrapper + _connect_one_locked): the health loop, a dispatch's _cb_auto_reconnect,
and an operator refresh could otherwise interleave teardown/rebuild on the shared
StaticServerState and corrupt it — a latent pre-existing race this also closes.
The body is unchanged (only relocated), so the delicate anyio / wait_for connect
logic is untouched.
Out of scope (follow-up): silent GET-stream / notification death — the SDK stops
the notification stream after 2 attempts while the request path stays alive, so
send_ping is blind to it; the fix is bounded session recycling, which needs a
static in-flight guard first (only PoolEntryState tracks in_flight today).
Tests: backoff bounds (capped / jittered / forever, no overflow), reconnect
success resets backoff + closes breaker, reconnect failure retries forever,
in-flight skip, per-name serialization (no overlap), ping keeps healthy / evicts
dead / evicts on timeout, tick skips oauth_user, connect_all start + disable
gating, clean cancel.
* fix(mcp): harden static-server health loop (review findings)
A max-effort review found 13 concurrency/correctness defects, all from the loop
mutating shared StaticServerState without the interlocks the pool path carries.
Fix all 13:
- In-flight interlock: add StaticServerState.in_flight (parity with
PoolEntryState); _static_session_op increments/decrements around the static
call_tool/read_resource/get_prompt session ops; the ping skips and never
evicts a busy server, so a long tool call can't be torn down mid-flight.
- Dead-transport gating: the ping evicts + trips the breaker only on
_is_dead_transport(exc); an McpError, httpx.PoolTimeout, or plain ping timeout
is "slow, not dead" and only reschedules (matching the dispatch path).
- Session-identity: only evict the exact session that was pinged.
- asyncio.timeout (invariant-18) not wait_for for the ping; 5s->30s; the
timeout-scoped cancel is distinguished from an external shutdown cancel
(which still propagates) via .expired().
- Bounded reconnect: wrap _connect_one in asyncio.timeout so a server that
handshakes then stalls list_tools can't wedge the loop or hold the per-name
lock forever (connect internals untouched).
- Concurrent tick under asyncio.gather with a freshly-read clock for the sleep.
- Cross-path coordination: reconnect_sync/remove_server_sync take the per-name
lock across teardown+rebuild (calling _connect_one_locked directly);
_cb_auto_reconnect reuses a health-established session instead of racing a
redundant reconnect and no longer trips the breaker on lock contention.
- Backoff hygiene on recovery; skip '__' names; on health reconnect clear only
the open-circuit deadline (not the failure count) so a connect-ok/calls-fail
server still escalates to a trip.
Adds 14 tests and adjusts those that assumed the old behavior; suite 195->209.
* fix(mcp): unify static-server reconnect coordination (round-3 review)
A third review round + live testing found 8 issues on the health loop, five
sharing one root: reconnect logic was fragmented across five drivers, each
handling the lock / session-reuse / in_flight / config-recheck / breaker / clock
differently and incompletely. Introduce one primitive and route every
lazy/autonomous driver through it.
_ensure_static_connected(name, cfg) — the single lazy (re)connect path, all under
the per-name lock: config re-check (+ lock-identity re-check, closing the
remove->re-add race) so a removed server is never resurrected; reuse-if-live so a
queued/concurrent driver never tears down and rebuilds a live session (the
observed reconnect storm); in_flight guard so a reconnect can't tear down a
session with a call still in flight on the evicted stack; bounded connect; and the
circuit breaker owned in one place (clear the open-circuit deadline on success per
finding-13, record one failure on real connect failure). Returns the session on
success/reuse, None on a deliberate skip, raises on real failure. Routed through
it: the health loop, a dispatch's _cb_auto_reconnect, and _refresh_all; operator
reconnect_sync stays a deliberate force-rebuild.
Also: fresh-clock deadlines (the stale tick-start clock was landing deadlines in
the past and collapsing the backoff into an every-tick retry storm); loop-death
fix (the tick no longer re-raises a CancelledError found in the gather results —
per-server fallout, not shutdown; the loop returns only when Task.cancelling()
marks a genuine shutdown); dispatch breaker records no failure on a sync-boundary
reconnect timeout (lock contention is not a server failure; real outcomes recorded
once, inside the primitive). Cleanups: extract _teardown_static_session (was
copy-pasted 3x); share _capped_exponential between the breaker cooldown and the
reconnect backoff. Adds 17 tests; test_mcp_client 204->226.
* fix(mcp): coherent timeout hierarchy + round-4 review fixes
A fourth review round on the unified reconnect coordination found 5 correctness
regressions + 1 cleanup, five sharing one root: the inner reconnect attempt bound
(45s) was LONGER than every caller wait (dispatch 30s, remove 15s, reconnect 30s),
so a caller cancelling mid-attempt delivered a bare CancelledError that slipped
past the primitive's `except Exception`.
- Coherent timeout hierarchy: add _STATIC_RECONNECT_CALLER_TIMEOUT_S (> the inner
attempt bound) for the dispatch + operator waits, so the inner asyncio.timeout
always fires first — a clean TimeoutError the primitive converts, cleans up, and
records on the breaker — instead of a caller cancelling a live attempt. Fixes
[0] (half-discovered session left installed, served with a stale catalog) and
[1] (breaker never trips via dispatch).
- Primitive cancel-safe (belt-and-suspenders): its handler is now
`except BaseException`, so even a bare CancelledError drops the partial session
and records the failed attempt before re-raising.
- Operator waits: reconnect_sync / remove_server_sync default timeouts raised above
the reconnect bound; remove_server_sync now CANCELS the pending _remove on
timeout (so it can't later pop a re-added entry and corrupt state) and reports
failure instead of a false 'removed' ([2], [4]).
- in_flight defer is gated on defer_if_busy: autonomous callers (health loop,
_refresh_all) defer, but a DISPATCH reconnects rather than hard-fail a reachable
server with an in-flight sibling ([3]).
- Cleanup: extract _schedule_next_ping (was a copy-pasted triplet in 3 branches) [5].
Adds 6 tests; the lock-contention test now shadows the caller-timeout constant so
it runs in ~1s instead of the full wait.
* fix(mcp): address PR review feedback on static reconnect
- _ensure_static_connected: skip breaker record on CancelledError
(cancel proves nothing about the server; aligns docstring with impl)
- reconnect_sync: add asyncio.timeout wrapper so discovery-phase
stalls get a clean TimeoutError inside the lock
- reconnect_sync: change except Exception to except BaseException
so CancelledError from future.cancel() triggers catalog cleanup
- reconnect_sync: null state.session on failure so a tool-less
session isn't mistaken for a live one
- _static_reconnect_one: use fresh monotonic clock for backoff
gate instead of stale tick-start snapshot; remove dead now param
- _static_health_tick: correct docstring (0.5s clamp prevents
busy-spin, not 'no sleep through short backoff')
- Test: new test for reconnect_sync timeout + catalog cleanup
- Test: update cancelled-attempt test for new CancelledError semantics
- Test: narrow except BaseException to except Exception + type hints
- Test: fix typo 'Understone' -> 'Turnstone'
- Test: remove unused stale_now variable; update mock signatures
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Migration 054 added these columns but they were never populated (the sole
writer hardcodes None) and never read by any query, dashboard, or API.
Drop them so the schema matches reality.
Closes#769
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The docstring claimed the dispatch path plumbs a triggering ws / tool-call id
through a thin wrapper. It does not: _record_pending_consent_best_effort neither
receives nor forwards any id (the mcp dispatch layer never has one), the
proactive sweep has no dispatch, and nothing reads last_ws_id / last_tool_call_id
— they are unpopulated schema columns from migration 054. State that plainly so
the comment doesn't imply data is captured when it isn't (Copilot review nit).
Per-user OAuth (auth_type=oauth_user) token refresh is entirely lazy: a token
is refreshed only when a tool is dispatched or a session binds the acting user,
and a dead refresh token is discovered only when a dispatch fails. That assumes
a human is driving the session, which breaks for autonomous / scheduled work
acting on behalf of an absent user — the token may be expired (latency), in a
transient-failure cooldown (unavailable), or the grant may be dead with nobody
present to re-consent. The only periodic MCP-loop task, idle eviction, actively
tears OBO connections down; nothing keeps tokens warm.
Add a background token-freshness sweep that keeps every consented oauth_user
grant hot WITHOUT keeping connections warm and WITHOUT mutating consent state on
a timer.
Sweep (_user_token_sweep_loop, default 240s):
- Enumerates consented (user, server) grants from the token store and runs the
canonical refresh path for each. Strictly oauth_user-scoped: gates on
_oauth_user_server_names and drives off mcp_user_tokens rows, so a static /
no-auth server — which has neither — is structurally invisible (no DB scan, no
authorization-server round-trip, no MCP-server call). Never connects to the
MCP server; connections stay lazy.
- Observe-only: passes revoke_on_failure=False (new parameter on
get_user_access_token_classified) so a timer NEVER deletes a token, emits
token_revoked, or mutates the shared ambiguous-streak / cooldown. A dead grant
is only surfaced (proactive dashboard pending-consent badge); the
authoritative revoke stays on the lazy-dispatch path where a real user action
justifies it. Because the row survives, a spurious server-wide invalid_grant
(an AS maintenance window) self-heals — the badge is dropped on the tick the
grant works again.
- Keepalive: force-refreshes a grant whose refresh token has sat un-exercised
past user_token_refresh_keepalive_seconds (default 1800s) even while the
access token is still fresh, so a provider that ages out idle refresh tokens
can't expire one between a user's real sessions.
- Surfaces dead grants once per transition, pinning the pair only after a
durable badge write so a failed persist retries rather than being lost. First
sweep runs after a short startup grace so a restart surfaces a downed grant
within seconds, not a full cadence later.
Cadence <= 0 disables the sweep; a positive value is floored (30s) so a
misconfigured tiny cadence can't turn the loop into a busy-loop. Reuses the
per-key refresh lock, so a keepalive force cannot double-refresh against a
concurrent dispatch.
Storage: add list_mcp_user_token_reconcile_targets() returning
(user_id, server_name, COALESCE(last_refreshed, created)) — expiry-unfiltered,
no ciphertext projected — on the protocol, sqlite, and postgres backends.
Tests: the sweep's no-auth invisibility (zero DB / AS calls with no oauth_user
server), observe-only non-destruction (token kept and shared streak untouched on
a background permanent / ambiguous failure), keepalive gating, badge
persist-then-pin retry, self-heal on recovery, cadence clamp / disable, and the
storage enumerator.
- task_agent.json referenced a non-existent `skill(action='search', query=...)`
in two spots. The discovery tool is `skills` and the action is `find`
(`search` is an activation value, not an action), so the guidance would
mislead the model. Corrected both to `skills(action='find', query='...')`,
matching the tool's own error strings.
- _high_risk_skill_denied returned "" (allow) when get_storage() is None — an
asymmetry with the fail-closed lookup-exception path added earlier. A risk
gate that can't verify the tier must DENY, not wave the skill through, so
storage-unavailable (None) now denies too; both paths share one denial.
Two independent multi-agent reviews of the branch (high, then max effort) found
authority-confinement and robustness defects the per-step reviews could not
see. This commit addresses every confirmed finding. task_agent turned out to be
the surface that lagged its siblings on nearly every axis.
Risk gate (most severe):
- task_agent(skill=...) never enforced the high/critical-risk PRINCIPAL-load-
only gate that skills(load) / spawn_workstream / spawn_batch enforce, so a
model could route around it by delegating activation to a sub-agent. Enforce
it inline in _prepare_task on the row already fetched (no re-query, no drift
between get_skill_by_name and get_prompt_template_by_name).
- _high_risk_skill_denied now fails CLOSED on a storage fault: deny, never wave
the skill through. Denying (not returning "") also keeps spawn_batch's per-row
partial-success intact under a transient blip.
- (first round) extracted _high_risk_skill_denied onto spawn_workstream /
spawn_batch, closing the coordinator-side bypass.
Persona confinement (Principle 7 attenuation on the task_agent edge):
- A restrictive persona now attenuates the sub-agent's TOOLS, not just its
identity text — the tool lever is frozen into the item and filtered before
_run_agent.
- Honor ALL FOUR persona levers on the sub-agent, not two: a child persona's
mcp-off and memory-off levers now drop MCP tools (mcp__* + read_resource /
use_prompt) and the memory tool, matching a main session under the persona.
- Cap the sub-agent by the PARENT session's own persona grant too, so a
restricted principal cannot escalate authority by spawning.
- Add persona to the task_agent judge/audit func_args projection (policy +
audit parity with spawn).
- Persona-resolution failures defer to a clean tool error (try/except mirroring
_validate_child_persona) instead of an opaque "internal error".
Substitution / capability:
- substitute_args=False for capability contexts (defaults, task_agent) so a
literal $ARGUMENTS / $N in a body is preserved, not blanked; env vars still
resolve. The literal-$ARGUMENTS scan is deferred behind that guard (skipped on
every capability render).
- Drop the CLAUDE_SKILL_DIR alias (canonical TURNSTONE_SKILL_DIR only). That
name also lives in bash, where turnstone-as-a-node-inside-Claude-Code must not
shadow the host's value; claiming it in the prompt but deferring in bash
diverged the two surfaces (a review finding). turnstone now claims it in
neither surface. The CLAUDE_SESSION_ID / CLAUDE_EFFORT prompt aliases stay
(pure prompt values, no bash-namespace collision).
Skills-as-context:
- DEFAULT (always-on) skills stay in the identity system message — the standing
baseline, never a mid-session cache-bust; only a NAMED applied skill moves to
the user-role capability message. This shrinks the pending model-adherence
eval surface to the named-skill move alone.
Cleanups: consolidate a duplicated rationale comment; correct the now-stale
"task agents are not persona-filtered" note.
PRE-MERGE GATE unchanged: the §7 Q1 model-adherence eval (named-skill move,
this branch vs main) is not runnable in-tree and must clear before merge.
A skill can auto-fire tools (auto_approve + allowed_tools) once loaded, so
letting the model activate a risky skill through skills(action='load') is an
injection-steerable lane that widens authority behind a rubber-stampable
approval. Deny it: high/critical-risk skills are now PRINCIPAL-load-only --
the model gets a clean error pointing the user at /skill, and the operator
loads such skills explicitly (handle_command /skill and cli --skill call
set_skill directly and bypass this gate).
The scanner-computed risk_level is the gate signal -- it already escalates for
the auto_approve + allowed_tools authority the create path warns about -- so no
new column or migration is needed. The check can only DENY, never widen, so it
is safe by construction (HYPOTHESIS.md Principle 7 / design section 5.5).
Remaining step-5 follow-ups, out of scope here: persisting the literal
disable-model-invocation frontmatter field for arbitrary author-marked skills
(needs a column) and deprecating the vestigial variables mechanism.
Step 3 of the skill/persona split: an applied skill (including default skills)
is CAPABILITY context, so its body no longer sits in the identity system
message. It rides its own message (user role) after the identity block, with a
short intro naming the active skill. The <available-skills> discovery catalog
stays in the system message.
Two consequences:
- The cached identity prefix (persona BASE + ENV + POLICIES + catalogs) stays
stable across skills(load): loading/clearing a skill changes only the
trailing capability message, not the identity block.
- The task_agent base (_agent_system_messages) is snapshotted BEFORE the skill
block, so a parent's applied skill no longer leaks into the sub-agent prefix
(the sub-agent supplies its own persona identity and skill via _exec_task).
PRE-MERGE GATE: the design gates this on a model-adherence eval (this branch vs
main) verifying the model follows a skill as well from a context message as it
did from the system message (design section 7 Q1; ASSUMED-neutral, UNVERIFIED).
That eval is not runnable in-tree and MUST clear before this branch merges.
Mechanical structure is pinned by TestSkillContextPlacement.
Deferred follow-up: sub-agent (task_agent) skill-resource materialization, so
${TURNSTONE_SKILL_DIR} stays literal on that path (unchanged since step 1).
Test helpers (_sys_content) now read the full prompt prefix (identity + skill
context) so placement-agnostic assertions keep working.
Before, a task_agent's system identity WAS its skill (skill body concatenated
into the sub-agent's system message), and #683 deliberately gave task_agent no
persona. Now that personas are first-class on every creation/spawn path, make
task_agent consistent: identity comes from a persona, the skill is capability.
- task_agent gains persona= (validated at prep against the interactive kind,
the general-purpose personas a worker can adopt). The resolved base prompt
is frozen into the approval item; _exec_task never re-reads storage.
- Default identity (no persona=) stays _TASK_DEFAULT_IDENTITY. The one-shot,
tool-over-narration operating guidance always layers on top.
- skill= is now CAPABILITY: rendered through the shared pipeline (step 1) and
delivered as a distinct user-role context turn ahead of the task, never
fused into the identity. Consecutive user turns coalesce at the provider
boundary (Anthropic _merge_consecutive), so this is wire-safe.
Updates the task_agent tool schema (persona param; skill reframed as
capability) and flips the persona guard test (task_agent HAS a persona param
now). Sub-agent skill-resource materialization and the interactive
skills->context move remain follow-ups.
Skill-body placeholder substitution diverged by invocation context:
interactive load, default skills, and spawn-child ran the full
render + spec-substitute, while task_agent (_exec_task) ran
_render_template only -- so $ARGUMENTS and ${...} env placeholders
rendered literally on that one path.
Introduce _render_skill_body as the single render+substitute path and
route interactive load, defaults, and task_agent through it, so a skill
reading ${TURNSTONE_EFFORT} or $ARGUMENTS resolves identically wherever
it runs. A sub-agent has no invocation args, so bare $ARGUMENTS and the
positional $N / $ARGUMENTS[N] forms resolve to empty there -- matching
the defaults and spawn-child paths, not the old verbatim passthrough.
- Add ${TURNSTONE_*} as the canonical vendor-neutral spelling for the
env placeholders (SESSION_ID, EFFORT, SKILL_DIR); keep ${CLAUDE_*} as
a permanent back-compat alias so imported skills keep resolving.
- Bash env: export TURNSTONE_SKILL_DIR and SKILL_RESOURCES_DIR
unconditionally, but add CLAUDE_SKILL_DIR only when the host has not
set it, so turnstone does not shadow a real value when it runs as a
node inside Claude Code.
- Materialize skill resources before substituting the body, so
${TURNSTONE_SKILL_DIR} resolves to the concrete bundle path on the
interactive path.
Sub-agent resource materialization and moving identity to a first-class
persona are left to follow-ups; ${TURNSTONE_SKILL_DIR} stays literal on
the task_agent path for now (unchanged from prior behavior).
Address Copilot review: cancel() now signals the worker to stop (a
cancellation Event) and joins only when started, so a test that errors
before the approval registers can't leak the worker or resolve late into a
finished test. The worker reads _pending_approval via getattr so a UI
without it can't crash the thread into a silent death (leaving approve_tools
blocked the full timeout), and only resolves when it actually observed the
registration.
The UI-approval tests drive a blocking approve_tools() by firing
resolve_approval() from a fixed 0.05s threading.Timer. approve_tools does
_approval_event.clear() -> register _pending_approval -> wait(3600s); on a
slow/loaded runner the timer can fire the event's .set() BEFORE that .clear(),
so the wakeup is wiped and approve_tools blocks the full _APPROVAL_WAIT_TIMEOUT
(one hour) -- surfacing as an intermittent CI hang (observed on the 3.12 runner
~15% into the suite; fast runners win the race, so 3.11/3.13 pass the same
commit).
Replace the fixed-delay timer with resolve_when_pending() (tests/conftest.py):
it waits until the approval is actually registered -- which happens AFTER the
clear -- before resolving, so the set can never be lost. The helper mirrors
threading.Timer's start()/cancel() so the surrounding scaffolding is unchanged.
10 sites across 3 files; the verdict-delivery timer (bounded to its own 5s
budget, not a hang) is left as-is.
Validated: the 3 files pass 20/20 under single-CPU stress (taskset -c 0) with
no hang or thread leak.
- Docstrings/help said the treatment skill 'composes into the system
message'. This harness runs on both checkouts (system on main, a context
turn on the placement-refactor branch), so the wording now describes the
natural set_skill composition path without asserting a placement.
- Validate each skill-bearing case's 'skill' shape up front (driver +
CLI) so a malformed dataset fails with a clear error, not a mid-run
KeyError. Pinned by test_rejects_malformed_skill.
Add a two-arm skill-adherence mode to the eval measurement substrate that
measures whether a NAMED skill changes tool-use behaviour, so skill-in-system
(main) can be compared against skill-in-context.
- _run_single_test gains skill/skill_mode: skill_mode builds HeadlessSession
under natural composition (no system_prompt_override) and, for the treatment
arm, seeds the skill into the temp DB and activates it via the real
set_skill path so the skill body folds into the system message under test.
skill_mode defaults False, so the optimizer/measure paths are unchanged.
- Thread skill/skill_mode through _run_and_score_subprocess, _run_iteration
and _run_iteration_parallel (serial + parallel).
- run_skill_adherence: per case, run treatment (skill) vs control (no skill)
n_runs each, score against expected_actions, report per-case lift =
pass_rate(treatment) - pass_rate(control) and the mean lift. The control
isolates the skill's causal effect.
- turnstone-eval --skill-adherence <dataset>: loads a skill-scenario dataset
and prints a treatment/control/lift table.
- eval_skill_adherence.json: authored search-first / test-after-edit /
changelog-update scenarios, chosen so the base model does not do the action
by default.
- tests: plumbing proof (skill folds into system_messages for treatment,
absent for control) + lift-math aggregation.
The success path already extracts message_count/total_usage before the
break, so the session = None rebind was unused dead code (code-quality
review). Remove it; exception/timeout cleanup paths are unchanged.
turnstone-eval was misnamed: it was a prompt optimizer, not a measurement
harness. Split the 3252-line turnstone/eval.py into a strictly one-way
dependency (optimizer -> eval-core; core never imports the optimizer):
- turnstone/eval/core.py measurement substrate — everything up to and
including _run_iteration: provider detection, NullUI, HeadlessSession,
the test runner, score_run, aggregation, and neutral reporting.
- turnstone/eval/cli.py new measure-only `turnstone-eval` — the old
--no-optimize path promoted to the whole job (one _run_iteration call,
then print the summary table).
- turnstone/optimizer.py the UCB self-modify loop and its multi-agent
pipeline (analyst/optimizer/observer/diversifier/tool optimizer), now
`turnstone-optimizer`; imports from eval.core only.
- turnstone/eval/__init__.py re-exports the core public API for
back-compat (score_run, _match_action, _run_iteration, HeadlessSession).
_apply_tool_overrides lives in core (HeadlessSession needs it) rather than
alongside the other tree helpers, so the dependency stays one-way.
Breaking change: `turnstone-eval` now measures; use `turnstone-optimizer`
to optimize. Both code paths are behaviour-preserving — the moved function
bodies are byte-identical.
- Add the missing #admin-personas grid-template so the table lays out as
columns; it was the only admin table without its own template, so every
cell collapsed into one implicit stacked column.
- Base prompt: accurate per-mode placeholders (required on create; blank
keeps a built-in's shipped prompt on edit) plus a client-side required
check on create. The old "empty = the kind's stock base prompt" copy was
wrong now that create rejects a missing base_prompt.
- Set the default persona from the row ("set default", with a scope-default
badge) to match the models table; drop the shelf checkbox and its
create/edit/submit wiring.
ruff format --check flagged two lines: a long persona-picker Field
description in server_schemas.py and a watch-restore create() call in
server.py that fits on one line after the persona-kwargs merge.
Formatting only, no behavior change.
Built-in persona base prompts move from inline DB text / base.md into
prompts/personas/<slug>.md — code-owned, PR-reviewable, drift-proof.
base.md / base_coordinator.md become personas/engineer.md / orchestrator.md.
Prompt source is now explicit in storage instead of inferred in app logic:
a new base_prompt_file column plus CHECK (base_prompt IS NOT NULL OR
base_prompt_file IS NOT NULL) — two nullable columns, never both empty.
Resolution is a coalesce (base_prompt else load(base_prompt_file)), frozen
into the workstream stamp at creation. base_prompt_file marks a persona as
built-in (code-only, un-archivable); an operator override on a built-in is
allowed and wins over the file. "Inherit the kind default" is a
workstream-creation act (is_default), not a persona-row state.
Migration 063:
- seeds reference their file (base_prompt NULL); no runtime file reads —
the backfill's frozen prompt text is inlined as a point-in-time snapshot
so migration history stays self-contained and reproducible.
- every existing workstream is stamped by kind (creative -> writer, else
the kind default), set-based (INSERT..SELECT via temp tables) with the
persona column added after the bulk writes to shorten its lock window.
Storage guards (both backends): operators must supply base_prompt;
built-ins can't be archived or have base_prompt_file set via the API;
clearing an operator persona's only source is rejected.
Follow-ups reviewed alongside (#756): soft-set visibility docstring scoped
to per-process; _apply_persona_snapshot / _current_persona_snapshot own the
stamp round-trip; spawn approval-header args (skill/name/target_node)
flattened+capped like persona; server-side tool injection generalized to
replace-only (client-def gated, incl. the xAI include forwarding). Seed
copy revised (researcher soft; de-costumed prose; engineer de-biased).
New test_schema_parity asserts create_all matches the alembic head.
Closes#683 groundwork; ruff + strict mypy clean, full suite green.
The roster persona merge distinguishes key-absence (pre-persona node in a
rolling upgrade — preserve) from present-but-empty (authoritative
unstamped — accept), so a stale in-memory value can never mask the
snapshot on an immutable field. The node create route caps the persona
slug at 64 like the console proxy, keeping oversized values out of the
storage lookup and the reflected 400 text. The four persona admin
handlers drop their redundant function-local asyncio imports, and the
DELETE-route test moves its request out of the assert statement.
The scribe, researcher, and executive prompts drop the infrastructure-team
costume and the demo-theater close — those framings suit the stock BASE
modules (whose job is today's default engineer/orchestrator behavior) but
narrowed personas meant for general use: a scribe summarizing meeting
notes is not a teammate, and the executive's verdict language works
without corporate staging. The behavioral substance is unchanged —
fidelity discipline for scribe, evidence discipline for researcher,
interrogate/delegate/verdict for executive, with the approvals boundary
still stated plainly.
The writer prompt loses 'use the analysis channel', a harmony-format
holdover from the CLI's single-provider days — the reasoning cue is now
format-neutral. Migration docstrings ride along: the workstreams.persona
column is documented as a slug carrier, and downgrade() now states the
capability-widening consequence of stripping stamps.
Spec models now describe what the endpoints do: ListPersonasResponse
declares the tool_inventory the shelf depends on, both console create
models declare persona, CreatePersonaRequest declares org_id, and
UpdatePersonaRequest documents the null-vs-absent split (null clears
base_prompt/tool_allowlist, null on flags/kinds is ignored). Console
OpenAPI regenerated.
Protocol contracts match the implementations: update_persona's return
covers the no-op case, create_persona's raises-list is complete, and
both extended row-shape docstrings gain their tail columns plus the
append-only rule. The workstreams.persona comments say slug, not
display name.
Page corrections from the docs review: personas.md documents the
creative_mode-to-writer migration conversion, the mid-session /resume
MCP-lever behavior, visibility-based nudge gating, the soft-set
prompt-cache cost, and the executive tool list — and drops internal
jargon. The changelog entry moves under [Unreleased] with the house
breaking-marker style and the auto-conversion note. coordinator-skills
and the API tour stop using persona to mean framing; governance,
api-reference, sdk, console, tools, and memory pick up the new
permission family, endpoints, kwargs, picker, and lever caveats.
The rank guard now derives needs_approval through the real _prepare_tool
on a bash call under an allowlisting persona instead of scripting the
flag, and asserts the approval gate actually fires. The row-shape guard's
source-grep is replaced with behavioral collector tests driving both
ws_created lanes (poll-diff and SSE relay), plus a proxy-forward twin and
a saved-list value assertion that would catch positional column mix-ups.
Receiving-side stamping gets its first HTTP coverage: create with an
explicit persona under workstreams.create only (selection needs no
persona perm), kind-mismatch and unknown-name 400s, omitted-persona
default stamping, the 503 on a failed default lookup, and the clean-None
legacy lane. Resume adoption is pinned end to end: a corrupt target stamp
leaves the session fully intact, an MCP-on stamp is refused when the
client was persona-gated at construction, and an MCP-off stamp drops the
live surface (listeners deregistered, toolsets reset). Soft-set
tool_search expansion recomposes the prompt exactly once; legacy sessions
never recompose.
Compaction legs run real flows now: spill plus the recall-pointer variant
under memory-off with recall visible vs hidden, and a full stamp
surviving compaction-then-resume. Migration 063 gains the downgrade
config-cleanup case (stamps removed, creative_mode preserved) and the
conversion idempotency case (already-stamped creative rows don't crash
the upgrade).
RBAC coverage goes cross-perm: read-only and write-only principals hit
every verb (a wrong-perm-name regression in any handler is now visible),
archive and default-flip succeed through PATCH, persona.* strings
round-trip the role editors and the overrides overlay, and the production
route table is asserted directly (no DELETE registered). Endpoint/storage
fixtures move off migration-seed names; storage hardening tests cover the
size caps, corrupt-row reads, the TypeError-to-ValueError ordering, the
duplicate-name race mapping, and the single-default backstop. Shell
asserts pin the new picker surfaces and drop the last persona-as-kind
wording.
Provider search gating (replace-only): native web search now stands in for
a client web_search def that survived the persona visibility filter — on
both OpenAI surfaces and both injection lanes (web_search_options, the
server_side_tools loop, and _convert_tools' capability lane). A scribe or
any envelope hiding web_search stays search-free on search-capable models;
coordinators and tool-less utility calls stop receiving search too.
Resume stamp discipline: resume() loads config and parses the target's
stamp BEFORE touching session identity/history, so a corrupt stamp raises
with the session intact instead of half-adopting and then 'repairing' the
target's stamp on the next config save. The MCP lever now follows the
stamp on mid-session adoption: an MCP-off stamp drops the live surface in
place (listeners deregistered, toolsets reset); adopting an MCP-on stamp
into a session whose persona gated the client off is refused loudly (the
surface cannot be rebuilt post-construction). The REPL /resume handler
reports these errors instead of crashing the CLI.
Fail-closed default lane: a FAILED default-persona lookup at create is a
503 (routes) / clear exit (CLI) instead of silently degrading to the
unstamped stock envelope; a clean 'no default configured' still creates
legacy. resolve_persona_for_kind reports storage-unavailable distinctly
from unknown-persona.
Soft-set governance: tool_search expansion under a persona visibility set
recomposes the system prompt so tool-gated policy segments land with the
tool they gate. MCP resource/prompt catalogs gate on read_resource /
use_prompt visibility. Spawn judge/audit projections carry persona (the
human approval header already did). Active-list rows carry persona like
their project_id twin.
RBAC catalogs: persona.{create,read,write} join _VALID_PERMISSIONS and
the roles-editor sections, making the documented grant-outward path real.
Storage hardening: default-persona invariants move to a shared _utils
helper (validate + demote) with a pg advisory xact lock serializing
promotions and a post-promote single-default assertion; create maps the
unique-name race to the same ValueError as the pre-check; reads validate
JSON shape loudly (naming the persona); serialize enforces size caps;
field validation runs before invariant checks so malformed input is a 400,
never a TypeError-500. org_id guards explicit null and caps at 64.
Also: base_override='' means 'no override' at the compose boundary;
persona tag flattened/capped before the spawn approval header; /creative
redirect resolves the writer persona before advertising it; memory-nudge
gating unified through _nudges_enabled.
Provider/row-shape tests updated to the new contracts (the old ones
pinned the injection hole and the pre-persona row shape).
Review pass over the branch surfaced real defects, all fixed here with
regression guards:
- Fork-resume (resume_ws) adopts the SOURCE workstream's stamp,
resolved pre-construction so all four levers (including the
construction-time MCP gate) bind the fork; a corrupt source stamp is
a loud 400, an unstamped legacy source forks unstamped — never the
kind default. Watch-restore and CLI --resume thread the stamp the
same way, closing an MCP leak where a restored MCP-off workstream
re-merged the catalog.
- SessionManager.open parses the stamp inside the install guard so a
corrupt stamp releases the reserved slot; a retry reproduces the
loud error instead of 'already tracked'.
- Mid-session resume() adopting a stamp rebuilds the tool_search
pathway to match (hard set drops it, soft set force-constructs it);
soft persona sets survive the global tool-search setting being off.
- Memory nudges gate on actual memory-tool VISIBILITY, not just the
memory lever, so an allowlist that hides the tool also silences the
nudges that point at it; post-compaction resume gets a no-recall
nudge variant when the pointer would dangle.
- Console PATCH: explicit null flags from UpdatePersonaRequest no
longer archive the persona or flip levers on a rename; multi-kind
personas survive a shelf edit; admin list ships the per-kind
tool_inventory so the shelf checklist tracks the server inventory
instead of a hardcoded JS list; admin CRUD moved off the event loop.
- Migration 063 converts legacy creative_mode workstreams to the full
writer stamp (downgrade removes all persona keys).
- REPL: /new passes the persona; /workstreams unpacks the widened row.
- Shared resolve_persona_for_kind is the single eligibility rule for
the HTTP handler, CLI, and spawn precheck; spawn_batch memoizes the
persona lookup; ToolSearchManager.is_expanded gives the visibility
tail an O(1) probe.
docs/personas.md covers the four levers, the resolve-once/stamp-forever
snapshot semantics, the seed matrix, per-surface selection, authoring
rules, and RBAC; architecture.md's config-persistence paragraph swaps
the removed creative_mode for the persona stamp.
The 15 guards from the design brief: the approval path is untouched
under any persona (rank guard); empty-toolset personas compose no tools
block and put zero definitions on the wire; the tool_search escape hatch
is soft when included (discovered tools union with the allowlist) and
hard when omitted (pathway disabled, including native defer_loading);
memory-off suppresses recall injection, the memory tool, and
memory-directed nudges while behavioural nudges and task-agent tools
survive; MCP-off is session-wide and refresh-proof; spawn validates
persona at prep time and never inherits the parent's; task_agent's
schema stays persona-free; the stamp is immutable, survives
SessionManager.open threading, and corrupt stamps fail construction
loudly; mandatory prompt policies compose under every persona; CLI
resolution (extracted to resolve_cli_persona_kwargs for testability)
loads seeds, exits clearly on unknown names, and adopts the resume
target's stamp; persona edits/archives never touch stamped workstreams;
and the row-shape contract twins carry the persona field.
RBAC endpoint coverage: admin CRUD 403s without persona.* and succeeds
with it; the picker feed needs no persona permission and hides archived
personas; invariant violations surface as 400s; DELETE is 405.
Creation surfaces: the console launcher, the server webui new-workstream
dialog, and the dashboard composer all gain a Persona select fed by the
shared personas.js data layer (module cache + fingerprint + never-reject
fetch + window.TurnstonePersonas bridge, cloned from projects.js), kind-
filtered with the kind default preselected so a zero-touch launch is
byte-identical to today.
Authoring: a Personas tab in the console Manage surface (Governance
group, persona.read-gated) with a Service Hatch shelf exposing exactly
the four levers — base prompt, tool-visibility checklist (kind inventory
+ free-text row for MCP/dynamic names; tool_search membership decides
soft vs hard), MCP and memory toggles — plus kinds, default flip, and
archive. No delete action anywhere (archive-only lifecycle).
The workstream wears it: SavedColumns.persona() on both saved tables,
hover/aria labels on the rail rows (raw slug fallback keeps archived
personas labelling their workstreams), and the full row-emitter sweep —
storage projections (list_workstreams tail, get_workstreams_batch,
list_workstreams_with_history) on both backends, the server dict
builders and ws_created events, both _coordinator_rows lanes, the
collector delta + pseudo-node paths, and the console cluster-create
proxy that rebuilds its body.
Naming reclamation: the launcher kind-toggle ids/classes that squatted
on 'persona' (launcher-personas, persona-coordinator/-interactive,
.persona-btn/.persona-led/.persona-tag) are renamed to kind-* before
'persona' becomes user-facing vocabulary, along with the prompts-module
and test wording that used persona to mean kind.
The persona resolved at creation is snapshotted into workstream_config
(five keys, all-or-none) and applied ONLY from the stamp — the personas
table is never read post-create, so edits/archives never touch existing
workstreams, and a corrupt stamp fails construction loudly instead of
silently reverting to a default envelope. Legacy pre-063 workstreams
carry no keys and keep today's behavior byte-for-byte.
The four levers (turnstone/core/personas.py holds the codec):
1. Base override — compose_system_message(base_override=...) replaces
exactly the BASE module; ENV/CONTEXT/TOOLS/POLICIES keep composing so
mandatory prompt policies ride on top of every persona. This also
closes the old /creative hole where the fork bypassed composition
(no CONTEXT, no DB policies).
2. Tool visibility — the allowlist intersects both the composition name
set (TOOLS block self-suppresses, tool-gated policies drop, the
memory advisory drops) and the END of _get_active_tools so the wire
never advertises hidden tools. tool_search in the set = soft
(discovered tools union with the allowlist via the session's
expanded-names set); absent = hard (the whole pathway is disabled,
covering provider-native defer_loading, which has no synthetic name
to filter). Persona sets force client-side tool search.
3. MCP gate (session-wide) — an MCP-off persona drops the client
reference at construction: no merge into _tools OR _task_tools, no
listeners, refresh callbacks inert, resource/prompt catalogs gone.
4. Memory (own hands only) — no recall injection, memory-directed
nudges suppressed (MEMORY_NUDGE_TYPES; behavioural nudges keep
firing), memory tool hidden. _task_tools is NOT filtered; compaction
spill/markers are never persona-gated, and the post-compaction recall
pointer is emitted only when the recall tool is actually visible.
Threading: the create handler resolves once (explicit name -> 400 on
unknown/disabled/kind-mismatch; empty -> the kind's default; pre-seed DB
-> unstamped legacy) and stamps via constructor kwargs + config keys +
the workstreams.persona column; SessionManager.open threads the stamp
pre-construction exactly like the saved model alias. Non-fork resume
adopts the target's stamp so _save_config can't clobber it. spawn /
spawn_batch gain a persona arg with prep-time validation (children are
interactive-kind; omitted = kind default, never the parent's). Python +
TS SDKs, OpenAPI specs, the picker feed GET /v1/api/personas (authed, no
perm), and console admin CRUD /api/admin/personas (persona.* perms,
archive-only — no DELETE) round out the surface.
BREAKING: /creative is removed (the REPL command now points at the
writer persona); turnstone --persona <name> is the replacement. Also
fixes the CLI session factory, which TypeErrored on the project_id
kwarg the shared InteractiveAdapter passes unconditionally.
Adds the personas template shelf (#683): migration 063 creates the
personas table (tri-state tool_allowlist, per-kind is_default, archive
via enabled=0 — no hard delete) plus the workstreams.persona display
column, seeds the six launch personas (engineer/orchestrator as
zero-touch per-kind defaults; scribe/researcher/writer/executive as
curated envelopes), and grants persona.{create,read,write} to
builtin-admin following the 062 pattern.
Storage: list/get/get_by_name/get_default/create/update on both
backends, with the default-persona invariants (exactly one per kind,
single-kind, enabled, not archivable, demote-on-flip) enforced in the
storage layer and the JSON serialization shared via _utils so the
backends cannot drift.
context_window=0 is the documented auto-detect sentinel -- "inherit the
CLI-detected window." The DB loader applies it (row.get(k, 0) or
context_window); the config.toml loader did not (entry.get(k, default)
only substitutes a MISSING key), so an explicit context_window = 0
leaked a literal 0 downstream, zeroing every budget that reads
ModelConfig.context_window -- judge lowering, session compaction. The DB
loader's comment even claimed config.toml shared "the same fallback
chain," which was false.
Match the DB loader at the source. The judge-side _positive_window
coercion stays as defense-in-depth, but its comments no longer misframe
0 as garbage -- it's a valid sentinel, now normalized at load.
Two window-sourcing edge cases the first pass missed:
- config.toml models can carry context_window=0 (that load path lacks
the DB loader's 0-inherit normalization). The getattr guard caught a
missing attribute but not a present 0, which would zero every budget
and make honest_truncate drop everything. A shared _positive_window
helper now coerces any non-positive / non-int window to the next sane
candidate (the session window) then a floor, on every resolution path
in both judges.
- The output-guard judge's session-model fallback keyed off
provider.get_capabilities(), which reports 200k for local models --
the fictitious-window bug the alias path already fixes. It now takes
the session's real (config/registry-aware) window, like IntentJudge.
output_guard_judge shares _CHARS_PER_TOKEN from judge rather than
duplicating it, now that it imports the coercion helper anyway.
The output-guard LLM judge fed the model the full tool output with no
window awareness, so on a small-window local judge a large output
overflowed into an opaque provider error and fell silently to
heuristic-only — the opted-in LLM tier vanished without a trace.
Resolve the judge's real context window from the registry's per-model
config (the static capability table reports 200k for every local
model), and add an up-front oversize guard: when the assembled prompt
would exceed the window, skip the doomed call and record a labelled
llm_error the operator can see instead of a silent no-op. The
heuristic tier runs first, so its verdict still stands.
Also drop the fixed 500-char cap on the tool_args framing field: like
the output under review it now lowers whole, bounded only by the same
window backstop, never by a default clip of a normal argument.
The func_args projection in _evaluate_intent is the intent judge's
entire view of a pending call's arguments, yet it lowered only a
narrow field per tool: edit_file reached the judge as {path} with the
edits stripped, and skills mutations built their projection and never
assigned it, so the judge ruled on {}. A small local judge denied a
legitimate multi-edit edit_file at 95% confidence as "malformed,
missing old_string/new_string" on exactly this gap.
Project the full risk-relevant surface per tool — edits, file content,
timeouts, model overrides, skill risk fields, task status and
ordering, MCP resource URIs and prompt arguments — and add the
read_resource / use_prompt branches that previously fell through to an
empty {}.
Truncation is now a backstop, not a default. Arguments lower whole up
to the judge model's real context window, sourced from the registry's
per-model config rather than the static capability table (which
reports 200k for every local model and would over-budget a small local
judge into overflow). Only a genuine overflow truncates, with an
explicit dropped-character marker; the untruncated arguments always
remain in the trajectory. The verdict's persisted and streamed copy
carries a separate 16 KB backstop against pathological payloads.
A parametrized guard test asserts every gated tool projects a
non-empty argument view, so the silent-starvation failure mode fails
CI instead of shipping.
- typecheck: dropping _acting_user_id from the SessionUI protocol (it made
the attribute required, breaking NullUI's structural match and making
TerminalUI abstract). _emit_state now narrows to SessionUIBase before
assigning — the field belongs to the web-fanout UIs, not the protocol
contract that CLI/eval UIs also satisfy.
- test: two mock queue_message stubs (attachments-endpoint fake,
coordinator adapter double already fixed) needed the new
interjector_user_id kwarg; the endpoint fake was raising TypeError ->
queue_full. Added it and a negative test that a non-SessionUIBase UI is
skipped by _emit_state.
- review (Copilot): _load_persisted_senders now latches _db_senders_loaded
only if self._ws_id still matches the workstream it queried, so a
concurrent resume() can't mark the read done for a workstream whose
senders were never loaded.
- review (Copilot): corrected the acting-user-id comments in three places
— it carries the owner id even single-user (the gate no-ops because it
equals the viewer); it is empty only on unauthenticated lanes / before
first state emit.
Note: the storage-protocol '...' stub flagged by the code-quality bot is
the file's universal convention (262 stubs, zero NotImplementedError);
left as-is for consistency.
Coordinator workstreams will hit the same cross-user issue once MCP is
enabled there, so wire the same protection now.
- CoordinatorAdapter.send takes acting_user_id: binds it on a fresh turn
(so MCP creds + the acting-user signal are correct) and passes it to
queue_message on the interject path (CrossUserInterjectionError block).
The create-time initial dispatch passes the creator's id.
- ConsoleCoordinatorUI.on_state_change includes acting_user_id (mirrors
WebUI); the mid-turn-connect replay already carries it via the shared
make_events_handler. The coordinator's user-facing /send route already
reuses make_send_handler, so it inherited the interjector guard + 409.
- coordinator.js gains the same gate as the interactive pane: tracks the
acting user from state_change, blocks send when busy AND acting !=
viewer, and handles the 409 cleanly.
Drive-by hygiene (requested): the _verdictSig join used a raw U+001F
byte embedded in the source; replaced with String.fromCharCode(0x1f) —
identical runtime, ASCII-clean source (no invisible control char in the
file).
The UX complement to the server-side cross-user interjection block: on a
shared workstream, while another participant's turn is in flight, this
viewer's send button is disabled so they don't click into a 409 (and
can't drive tools under the initiator's credentials).
Backend signal (the linchpin — the acting user was tracked but never
surfaced to clients):
- ChatSession._emit_state pushes the acting user (turn initiator, owner
fallback) onto its UI (_acting_user_id on SessionUIBase).
- server.WebUI.on_state_change includes acting_user_id in the broadcast
state_change event; the mid-turn-connect replay (session_routes) adds
it too, so a client joining mid-turn learns who holds it. Id only (a
uuid the client compares) — no name, no storage lookup on the hot path.
Frontend:
- auth.js retains the opaque user_id from /whoami (ts.user_id) — kept
separate from the display username, used only for id comparison.
- composer.js gains an independent hard-block axis (setSendBlocked /
_reconcileDisabled) so send can be disabled even in queueWhileBusy mode.
- interactive.js tracks the acting user from state_change, blocks send
when busy AND acting_user_id !== the viewer's own id, and handles the
409 as a clean message (reactive fallback for the click-beats-event
race) instead of a generic connection error.
Degrades gracefully: single-user workstreams (acting user == viewer) and
older backends (no acting_user_id) never engage the gate.
A mid-turn interjection folds into the current turn under the
initiator's identity — bind_acting_user deliberately does not rebind
mid-turn — so on a shared workstream a second participant's queued text
would run any tools it triggers under the initiator's MCP (oauth_user)
credentials (confused deputy) and be stamped with the initiator's sender
label (misattribution). Rather than fold it in, reject: queue_message
now takes the authenticated interjector_user_id and raises
CrossUserInterjectionError when it differs from the current acting user.
The send route surfaces it as 409 cross_user_interjection. Only an
authenticated non-acting participant is blocked — self-interjection,
single-user workstreams, and unauthenticated internal lanes (empty id,
e.g. the coordinator adapter) are unaffected.
Cloud multi-agent review of the three follow-up commits surfaced 15
verified defects; this addresses them.
Security / correctness:
- output_guard was blind to the new sender-label trust marker: add
fence.SENDER_LABEL_TAG to the forgery/leak detector and thread a
second trusted nonce (trusted_sender_label_nonce) through
evaluate_output/_check_marker_forgery so a forged or leaked
sender-label block in tool output is flagged like an operator marker.
- Attachment-derived text (PDF extraction, audio transcript, perception
output) bypassed sender-label neutralization because it materializes
after _inject_sender_labels runs; neutralize it at each fallback site.
- _recompute_shared_state now runs on every compose (moved out of the
non-creative branch) so a creative-mode resume can't leave shared
state stale.
- /new and rewind/retry now reset shared state (were leaking the prior
conversation's participant set / keeping a workstream latched 'shared'
after its only second-participant evidence was deleted).
- Non-fork resume and /new remint both trust nonces; carrying a nonce
across a workstream switch would let a token leaked in one forge a
marker in another.
- _senders_dirty is only cleared once the persisted-sender read has
actually landed, so a transient storage error retries within the turn.
- ws_id snapshot guard in _recompute_shared_state discards a result if
resume() swapped workstreams mid-scan (MCP-callback race).
- recall/history search is scoped to the acting sender's visibility;
the shared-workstream declaration now names that exception so the
model doesn't read a filtered 'no results' as 'no record exists'.
Cleanup:
- fork skips the redundant persisted-sender read (its rows were just
bulk-written); _maybe_note_new_participant goes through the single
recompute entrypoint; senders_from_user_meta reuses _source_meta_from_json.
- fence.wrap docstring names the sender-label caller as a third
untrusted-host boundary.
Tests: end-to-end compaction-narrowed resume recovery, hostile
display-name fence break-out, sender-label output-guard leak/forgery,
and the two-nonce independence.
- q-2: document the deliberate username-first display-name precedence in
_resolve_display_name (diverges from auth.py's display_name-first
because sender labels must match the owner-banner identity kind).
- q-3: drop change-lineage comments referencing the separate acting-user
credential fix (tombstone noise once merged).
- q-4: tighten the plain-text attachment assertion from a tolerant
subset check to exact shape + _sender value now that the stamp is
deterministic.
- q-5: drop the contributor-local bare 'etc/' from .gitignore.
sec-1: the [message from <sender>] label was plain text, so a participant
could type a look-alike in their own message and impersonate another
sender to the model. Labels are now wrapped in a nonce-delimited
[start sender-label_<nonce>] ... [end sender-label_<nonce>] fence (new
fence.SENDER_LABEL_TAG, distinct value from the operator nonce) whose
token lives in the cached system prefix; participant content is
neutralized so typed look-alikes are defanged. A new
build_shared_workstream_declaration pins the token as the sole authentic
label.
sec-2 + q-1: the CONTEXT banner no longer embeds behavioral prose. It
carries a terse owner line + shared flag; the attribution rules, the
authenticity declaration, and the (now narrowed) tool-credential claim
move into the shared-workstream declaration. The credential claim is
corrected: per-participant credentials apply to MCP (OAuth) tools only;
built-in tools and skills run under the server/owner identity.
perf-3: _inject_sender_labels resolves each distinct sender's display
name once per call instead of once per turn, capping blocking storage
lookups at one per sender on the uncached error path.
- _known_senders/_shared_workstream are now monotonic: union-only growth,
latched shared flag, seeded once per workstream from a full-history
distinct-sender read (new StorageBackend.list_message_senders) so
compaction narrowing the resumable slice can no longer forget
participants (duplicate join notes) or flip the banner back to
single-user framing (prompt-prefix cache churn).
- Recompute is memoized per turn (invalidated on stamped user-turn
append); system-prompt composition no longer pays an O(n) trajectory
scan on every recompose.
- resume() resets the state: the monotonic guarantees are per
workstream, not per session object.
- resume(fork=True) bulk-persist now carries the user-turn sender stamp
into the fork's meta column (was: _source_meta only, which dropped
attribution for every forked user turn on reopen).
* multiuser chat fixes for identity clarity and obo oauth token selection during tool calls
* added some missing context to the session so that the llm would know what session/project to reference in tool calls
* updated to address copilots issues and excluded a local config folder
* I think this resolves the cicd failures
---------
Co-authored-by: pow3rtool <root@pow3rtools>
- applyRosterSnapshot: null-prototype membership map (a ws id colliding
with an Object.prototype property name would read as always-seen and
dodge eviction) and a stable Object.keys snapshot for the eviction
walk — current-key deletion during for...in is spec-safe, but the
snapshot is self-evidently order-safe and skips inherited keys.
- _streamingRenderApply: drop the tautological typeof guards around the
post-render decorators — both are module-local declarations, and the
surrounding try/catch owns decoration fault tolerance.
Long sessions (5000+ messages, several compactions) degraded steadily
and could stop rendering entirely while the backend stayed healthy.
Four hard failure mechanisms, each sufficient on its own:
- Unguarded event pipeline: one throw escaping onmessage/handleEvent
(e.g. renderMarkdown stack overflow on a few KB of nested "> ")
stranded the streaming refs, so every later delta painted into the
poisoned segment. stream_end now resets segment refs BEFORE the
finalize render with a plain-text fallback (the coordinator pane's
existing pattern); onmessage guards both parse and dispatch;
renderMarkdown is depth-capped with throw-safe footnote-scope
accounting; the streaming buffer is marked rendered only on success.
- Rebuild-vs-live races: clear_ui/replay_truncated re-renders wiped
events painted in the snapshot->replaceChildren window (never
redelivered) and left deltas writing into detached nodes. Rebuilds
now quiesce the event stream behind a token-owned queue flushed
after the render; streaming refs reset on every rebuild path
including refetch FAILURE; a mid-stream replay_truncated defers its
re-sync to the idle edge instead of dropping the repair.
- Ignored recovery floor: the global stream now handles node_snapshot
and replay_truncated. Roster eviction (with a "Session ended" toast
for open panes) happens only from the stream-ordered snapshot; the
REST resync is merge-only and r.ok-gated so a mid-restart 503 body
cannot read as an authoritative empty roster.
- Unbounded growth: _agentCards released on rebuild — deliberately NOT
on transport-only reconnects, which must preserve the maps or the
next child event builds a duplicate card; orphan grace timers
cancelled on full reload/destroy; toast queue capped with duplicate
coalescing; diff previews capped at 400 rendered lines (the
spread-append could throw RangeError before the approval gate
painted) with the omission notice below the scroll box; raw results
clamped at 64KiB.
Per-event O(N) work removed from the hot paths: thinking-indicator
instance ref; near-bottom cached from a passive scroll listener and
re-checked at rAF pin time (a user scroll-up landing in the coalescing
window wins; ResizeObserver re-engages follow after layout changes);
rAF-coalesced outer and per-stream scroll pins; self-healing
call_id->row/stream lookup caches; verdict lookup scoped to the row's
batch; tracked retry holder; queue-controller Set replaces the
whole-transcript idle sweep; rail renders rAF-coalesced; coordinator
child_ws_state ticks routed to single-row updates (full render only on
terminal-boundary crossings) with observer unobserve on replace.
Also: the coordinator SSE-error 401 probe is un-deadened (raw fetch —
authFetch never resolves a 401 — with the body inspected so a
version_mismatch still takes auth.js's upgrade-reload path via the new
noteVersionMismatch export); the console cluster-SSE reconnect timer
is tracked across logout; the mermaid render chain is rejection-proof
per link and paints errors on the containers the failing link had
already claimed.
Measured with scripts/livepass.py --perf (n=3000 history + 20-turn
live storm): full replay 1060ms -> 238ms; re-render cycles 836-1071ms
-> ~94ms flat; chunk path now flat vs transcript size; worst longtask
1080ms -> ~500ms; agent-card retention across rebuilds 4 -> 0.
Known limit (needs a server-side event watermark on /history): a turn
completing inside the refetch window can paint twice after the quiesce
flush — rare, visible, and strictly better than the silent loss it
replaces.
New /perf/livepass.html mounts the real InteractivePane at production
scroll geometry and drives production-shaped SSE events through
handleEvent/replayHistory in real time (no virtual-time budget, no
forced reduced-motion — both corrupt the measurement), reporting:
replayHistory wall time at N messages, per-turn live-storm cost on top
of that transcript, tool_output_chunk throughput, busy/idle churn,
heap + node + agent-card counts across repeated replay cycles (the
detached-DOM leak probe), and longtask counts.
The --perf runner builds, serves, and launches headless Chrome with
--js-flags=--expose-gc and --enable-precise-memory-info so heap
numbers are real floors; the page POSTs its JSON report to
/perf/report. Reports carry a per-attempt run token the runner
validates, so a straggler POST from a killed prior attempt cannot be
misattributed to the next size, and the wait loop polls the Chrome
process so a sandbox startup failure bails to the --no-sandbox
fallback in seconds instead of burning the full timeout.
The workflow_dispatch input is an arbitrary PR number, and the job used
only headRefName to pick the checkout ref. For a fork PR that is a bare
branch name that can collide with a branch in this repo, so the job
(contents:write, ends in git push) would operate on that unrelated
branch. Resolve isCrossRepository alongside headRefName and fail loudly
unless the PR head lives in this repository.
The publish and docker workflows trigger on workflow_run of CI, which
fires for every CI completion — including CI runs for pull requests
from forks — and always executes with this repo's secrets, tokens, and
the pypi environment. The only gate was CI success, so fork-PR CI runs
spawned publish jobs in the upstream context; actions/checkout v7's
fork-checkout refusal was the only thing that stopped one on 2026-06-30.
A fork PR whose head is an upstream-tagged commit would have passed the
tag check and reached the upload with valid OIDC.
Both workflows now require the triggering CI run to be a push event,
from this repository, with head_branch starting with 'v' — CI's push
trigger only matches main/stable/* branches and v* tags, so that is
necessarily a tag run (verified: tag-push runs report the tag name as
head_branch). Checkouts no longer persist the token while the tree's
build backend executes, and publishes are no longer cancellable
mid-upload (a half-uploaded release cannot be re-run cleanly because
PyPI rejects duplicate files).
vendor-js hardening in the same pass: gate on the immutable PR author
instead of github.actor, require a same-repo head before pushing to the
PR branch with contents:write, and pass github.head_ref through env
instead of interpolating it into the script body.
Review finding on #751: the carry invariant omitted the system message
and tool definitions, which ride every request — at shipped defaults
reserve + 2 carries + margin lands exactly at the window, so any real
prompt overhead pushed the post-compaction send over it, and the
overflow backstop re-compacts WITHOUT the carries.
spare now subtracts system_tokens + tool_def_tokens (the same terms the
_estimated_prompt_tokens fallback counts), making
overhead + reserve + carries*budget + margin <= window hold by
construction. Invariant test pinned at shipped defaults with a 4k-token
synthetic prompt; a monotonicity test pins that the term is live; exact-
arithmetic tests isolate the overhead explicitly.
The definition review found the two control-relevant crossings paraphrased:
the model's wind-down spill (recorded on the cooperative advisory, then
handed to the summarizer with everything else) and the user's last message
(clipped to 400 chars in the continuation hint). Both now cross copied.
- carry_spill: when the model stopped because it was advised to wrap up,
its final turn's text is shell-concatenated onto the summary under
'## Wind-down (verbatim)', ahead of '## Continue'. The summarizer still
reads the spill; its paraphrase is no longer the only survivor.
- _carry_budget_chars(carries): ~25% of the window per carry, sized so ALL
concurrent carries fit the spare after the summary output reserve —
spill + hint fire together at the end-of-turn site, and independent
sizing stacked reserve + 2*(cw/4) + margin past the window at default
config. Floored at 2000 chars; oversize content keeps head + tail.
- _truncate_block's marker reports the original size ('truncated — N chars
total'), and a truncated carry adds one line telling the model the full
text remains in history and recall can retrieve it.
- Summary turns carry source="compaction" (in-memory swap and checkpoint
reconstruction); _find_turn_boundaries and _generate_title test the tag
instead of the label string, so a user who literally types
'[Conversation summary]' stays a real turn.
- The send-loop overflow backstop now passes my_generation, closing the
compact-and-swap race every other compaction site already guards.
Tests: tests/test_compaction_crossing.py (tags on both paths, literal-label
boundary, budget arithmetic incl. the double-carry invariant at shipped
defaults, verbatim/truncated carries, spill semantics, forwarding); existing
suites updated for the tagged label turns and the new kwargs.
After a compaction, storage keeps the full transcript and the in-context
summary is a cache over it — recall is the model's re-derivation path back
into the originals. Un-scoped, its results duplicated the live context.
- search_history gains exclude_ws_id/exclude_after: the excluded ws's rows
above the boundary (the live segment, already in context) are dropped in
SQL via one shared fragment; rows at or below it — the summarized-away
past — stay searchable. A never-compacted ws is excluded whole:
everything is live. Other workstreams untouched.
- New get_compaction_checkpoint(ws_id) reads the latest marker's persisted
watermark (distinct from get_compaction_watermark, which computes what a
NEW compaction would use); the meta decoder is single-sourced with the
resume slice (parse_checkpoint_watermark) so the two boundary consumers
cannot drift.
- _exec_recall reads the boundary fresh at execution (a compaction that ran
while the item was queued is respected) and labels own-conversation hits
'(earlier in this conversation, compacted)'. Storage errors degrade to
whole-ws exclusion — less information, never duplicates. Known limit
(documented): a forked session excludes only its own ws, so inherited
parent rows remain searchable — harmless duplication bounded by tenancy.
- NUDGE_COMPACTION_RESUME teaches the path: the summary is a digest, not
the record, and recall can search the compacted portion.
- /history deliberately unchanged: a human browsing history has no context
to duplicate.
Tests: tests/test_recall_compaction_scope.py — checkpoint reads (none /
marker / latest-wins / malformed-as-live), the exclusion matrix, the
composed tenancy+exclusion query with both filters dropping rows, exec
plumbing and labeling, the nudge line; cross-backend.
Renovate opens PRs as a bot actor, which claude-code-action's default
human-actor check rejects — Renovate's dependency-bump PRs were never
getting reviewed.
search_history / search_history_recent searched every workstream's rows
regardless of who asked. Pre-projects that matched the trusted-team
deployment shape; with private projects (062) it became a cross-tenant
read — the recall tool and /history returned private-project rows to
non-members.
Both methods take a keyword-only user_id (protocol, sqlite, postgresql)
scoped by one portable SQL predicate (HISTORY_VISIBILITY_SCOPE_SQL)
mirroring WorkstreamProjectVisibility: a row hides only when its
workstream links to an existing private project and the user is neither
the workstream creator, the project owner, nor a member. Applied in SQL
so limit/offset pagination stays honest; COALESCE guards the
NULL-creator row, which plain <> would leak.
The recall tool pins the scope identity at prepare time (the mcp_user_id
discipline) and fails loudly on an unpinned item; /history scopes to the
acting user; user_id=None (single-user CLI lanes) stays unscoped.
Tests: cross-backend visibility matrix, ws_visible parity pin,
marker-exclusion composition, LIKE-fallback path, prepare-pin plumbing.
The redundant function-local asyncio import in project_resources_endpoint
shadowed the module-level one. resolve_workstream_owner's docstring now
maps the failure modes precisely: a failed ROW lookup is fail-soft 404
(get_workstream_row degrades to None, pre-existing behaviour), while the
fail-closed 403 applies once a row is resolved and the project gate's
storage lookup fails — in-memory workstreams 403 on a gate blip,
not-loaded ones 404 at the row fetch first. Plus ruff-format on the
visibility test file (edited via script, so the local format hook never
saw it).
Max-effort review findings on the visibility feature, worst first:
Leaks — the filter was sound where it ran, but several surfaces never
carried project_id to gate on:
- cluster_snapshot served the raw collector state with no filter at all;
it now gets the same per-request tenancy treatment as its siblings
- console pseudo-node coordinator rows + emit_console_ws_created,
the interactive-create ws_created event, and the poll-diff ws_created
now carry project_id/user_id (parity with their filtered siblings —
a missing field failed open, and a missing user_id over-hid the
creator's own workstreams)
- the SSE snapshot's overview total/state histogram is re-derived from
the filtered rows instead of leaking pre-filter counts
Correctness:
- saved list pages with OFFSET until it fills its 50-row window instead
of filtering after the LIMIT (a caller's own rows at position 51+
used to vanish behind other tenants' private rows); scan capped at 20
pages, logged when hit
- an INHERITED project_id whose project was since deleted no longer
400s coordinator child spawns — the dangling link is dropped; explicit
unknown ids still 400, revoked membership still 403s
- the SSE filter keeps a per-connection unresolved map: a storage blip
suppresses a row without pinning it hidden until reconnect (re-judged
on later events, rate-limited); definitive verdicts settle as before
- bypass principals (service / admin.cluster.inspect) get payloads
untouched — no row drops, no overview rewrite
Consistency and robustness:
- dashboard + saved-list visibility checks moved off the event loop
(executor), matching every sibling site
- list_project_attachments chunks its IN() at 500 ids per statement
- ws_visible/ensure_project_attachable now share one _project_grants
predicate so the tenancy rule can't diverge
- resolve_workstream_owner's docstring states the deliberate
fail-closed trade for project-attached rows during DB outages
- the workstreams-for-project ordering test asserts strict order on a
forced timestamp instead of a vacuous set fallback
ws_visible only treats real strings as project links (a test double or
corrupted value means no-project, not private-and-denied), the mgr-path
project_id is coerced likewise, and the HTTP send path binds the acting
user via a getattr-guarded bind_acting_user call inside the fresh-turn
closure instead of a send() kwarg — per-kind session stubs with explicit
send signatures keep working. Row-shape contract tests (interactive +
coordinator twins) grow the intentional project_id key.
Dashboard saved-sessions lists now carry and render the workstream's
project: SavedWorkstreamInfo gains project_id (the saved projection was
extended in the visibility change), SavedColumns grows a PROJECT column
(name resolved through the shared projects data layer, searchable via
the filter haystack, re-rendered when the async project cache fills),
inserted on both the webui saved-workstreams and console saved-sessions
tables.
Manage → governance → Projects rows are now expandable (same
interaction contract as the Users tab's OIDC panel): a per-project
resources panel lists the project's workstreams (kind/state/updated),
referenced attachments (metadata + ws-scoped download link through the
console's node proxy), and the project-scoped memory count. Backed by
GET /v1/api/projects/{id}/resources (project.read + per-project ACL,
collection off the event loop) over two new storage queries —
list_workstreams_for_project (first consumer of idx_workstreams_project)
and list_project_attachments (conversation ref-list walk, metadata only,
first-referencing ws per blob, pruned blobs skipped).
Workstreams attached to a private project were listed and reachable for
every authenticated user — only the scope tier was checked. Add a
tenancy predicate (WorkstreamProjectVisibility: private → project
owner/members, the workstream's own creator, service scope, or
admin.cluster.inspect; public/dangling/no project → unchanged
trusted-team visibility; membership itself is the grant — deliberately
NOT gated on the project.read capability, which guards the management
API) and apply it at every surface:
- listings: saved sessions (project_id + owner tail-appended to
list_workstreams_with_history on both backends), active list, node
dashboard, console cluster list (pre-pagination via a collector
row_filter so totals stay honest), node detail
- console tier-1 SSE: per-connection snapshot filtering + a hidden-set
for sparse follow-up events; ws_created project lookups run on the
executor, membership changes take effect on reconnect
- row access: resolve_workstream_owner 403s private-project rows for
non-members, covering every interactive ws-scoped verb via
tenant_check (console coordinator lane stays on its privileged
admin.coordinator gate)
- create: ensure_project_attachable gates explicit and parent-inherited
project_id on both create validators (unknown project 400s instead of
minting a dangling link)
Per-user MCP credential resolution was bound once at session construction
to the persisted workstream owner, so on a shared workstream every sender
executed oauth_user tools under the creator's tokens (and saw the
creator's tool catalog). Bind the authenticated initiator of each turn
(send + retry paths) as the session's acting user: dispatch, catalog
merge, visibility gates, and consent flows now follow whoever is driving,
with the owner as fallback for CLI / eval / scheduled / internal turns.
Rebinding swaps the user-scoped tool/resource/prompt listeners (identity
is the (user_id, callback) pair), fire-and-forget primes the acting
user's pools, and rebuilds the merged tool list. Prepared tool items pin
the identity at prepare time so an item pending approval executes under
the user whose turn requested it, not whoever binds later. Queued
mid-turn interjections deliberately do not rebind (no mid-turn
credential switch).
Corrections: the middle-form re-separation names its true mechanism
(restarting specs or refusal-event predicates; within-run retries never
touch F), the standard-Borel aside admits belief-state coordinates, the
drift-slack display binds its variable, effect-record status gains a
`none` value (never launched) distinct from rolled_back and unknown,
and parsing is assigned to the inner readout R with the gate as pure
authorization.
Structure: the trusted principal as the provenance lattice's single
widening writer; two-rank control (authority vs plan) with a
rank-neutrality corollary; the narrow-only rule for learned checks;
pi's never-lower filter joins the deterministic core; gate TOCTOU and
cross-run serialization; a composition law for harness trees (four
correspondences) with delegation as monotone attenuation.
Appendix: new worked entries for resume (journal-before-dispatch),
parallel proposals (the batch gate), derived and durable state (the
provenance meet rule), and ambient authority (per-action capability).
Claims numbered C1-C8; two falsifiers added (certificate compression;
working-set probe anchored in streaming lower bounds).
Grounding: adds Ramadge-Wonham supervisory control, RL shielding, and
Dayan's successor representation; repairs the Positivity/Skolem gloss
and two citation characterizations. All 18 external citations verified
against their sources.
A max-effort review of the branch before pushing surfaced six defects, several
introduced by this branch's own commits. All fixed:
[0]+[3] oauth priming (refined). Fully non-destructive priming never cleared a
genuinely-revoked grant — the dead token stayed "consented", its tools never
entered the catalog, and (bug) the PERMANENT branch returned before arming the
cooldown, so every session re-hit the AS with a dead refresh token. Root cause:
invalid_grant (PERMANENT) is a RELIABLE dead-grant signal (RFC 6749 §5.2), so
deferring its revoke was net-harmful. Renamed the flag revoke_on_dead_grant ->
revoke_ambiguous_escalation: priming now revokes genuinely-dead grants (permanent
/ expired-no-refresh) so the catalog isn't stranded cold behind a phantom token,
and defers ONLY the sustained-UNCLASSIFIABLE (ambiguous) escalation to lazy
dispatch — the case the "don't revoke an unused server's grant on a
misclassification" concern actually applies to. The cooldown is armed before the
ambiguous path, so the deferred case can't hammer the AS either.
[1] server.py. _public_server_status (operator refresh/reconnect endpoints)
didn't forward the new scope, so after per-user scoping every warm oauth_user
server rendered disconnected/empty there. Now passes aggregate=True (operator /
approve-scoped cluster view, matching the admin console).
[5] _is_dead_transport. The widened httpx.TimeoutException swept in
httpx.PoolTimeout — pool saturation, NOT a dead connection — so transient load
would evict a healthy session and trip the shared breaker for all users.
Narrowed to Connect/Read/WriteTimeout (kept NetworkError, RemoteProtocolError).
[8] _is_dead_transport. The exact-message "session terminated" fallback still
fired on a healthy session-owning server's protocol error with that message. The
SDK-synthesized code 32600 is the only deterministic signal (the message is
application-controlled), so match the code ALONE and drop the message fallback.
[11] cleanup. The dead-transport except block was triplicated across
call_tool_sync / read_resource_sync / get_prompt_sync — the exact drift this
branch had to repair. Extracted _record_and_evict_on_dead_transport.
Tests updated/added: prime revokes-permanent / defers-ambiguous (drives the real
resolver both ways); PoolTimeout-is-not-dead; exact-"Session terminated"-message
stays alive; _public_server_status aggregate. 836 test_mcp_* green, ruff + mypy
clean.
Resolves the one regression the user-scoping in 0c28b0ce introduced: the admin
console reaches the read-scoped /mcp-status endpoint via the console proxy with
the ADMIN's forwarded identity, so per-user scoping made oauth_user servers show
as the admin's own (usually empty) pool instead of the cluster-health "in use by
anyone" aggregate.
Add an `aggregate` flag (default False) through get_all_server_status ->
get_server_status -> _oauth_user_server_status. When set, connected + a
representative catalog reflect ANY user's warm pool. internal_mcp_status gates it
on the admin.mcp permission: holders (who already see consent counts + server
config — the proxy forwards permissions via create_jwt, repopulated on validate)
get the aggregate; every other read-scoped caller stays strictly per-user, so the
cross-user catalog leak stays closed. Static-server status is unaffected.
Tests: manager-level aggregate-sees-any-user, and an endpoint-level gating test
asserting admin.mcp -> aggregate=True / read+approve-without-it -> aggregate=False.
Follow-up to f585c47b (review finding #4). _oauth_user_server_status derived
connected + tools/resources/prompts counts from warm[0] — an arbitrary user's
pool entry — and get_all_server_status surfaced that to every read-scoped
caller of /v1/api/_internal/mcp-status, ignoring who was asking. So user B saw
user A's oauth_user server as connected with A's catalog size, over the wire
(connected + the three counts are in _READ_STATUS_PUBLIC_KEYS; user_pools /
auth_type are stripped). Before f585c47b these servers were absent from the
read map entirely.
Thread user_id through get_all_server_status -> get_server_status ->
_oauth_user_server_status; the warm-pool filter now matches uid == user_id, so
connected + counts reflect ONLY the requester's own pool. internal_mcp_status
passes _auth_user_id(request); an empty/absent principal (user_id falsy) sees
oauth_user servers as not-connected. Static-server status is unaffected (the
new param defaults to None and is ignored for them).
Note: the admin console (admin.mcp) reaches this same read endpoint via the
console proxy, which forwards the ADMIN's identity — so an admin now sees an
oauth_user server scoped to their OWN pool (typically not-connected) rather
than the prior any-user aggregate. Server-global health (circuit_open / error /
consecutive_failures) is unchanged, and the consented-users-count is a separate
aggregate. Restoring an aggregate in-use pill for admins (without re-leaking
per-user catalogs) would need a privilege-aware aggregate mode + admin.js
change — deferred.
Tests: updated TestOAuthUserServerStatus to the scoped signature, added the
cross-user isolation regression (user B sees neither A's connected flag nor A's
catalog size) and a no-user-context case.
Follow-up to f585c47b (review finding #5/#6). f585c47b routed
_prime_user_pools through get_user_access_token_classified to refresh expired
tokens at session start (closing the chicken-and-egg where an expired token
stranded the pool). But that resolver also REVOKES a grant (delete_user_token
+ token_revoked audit) on a permanent-classified refresh failure — and priming
runs for EVERY consented server, so a single misclassified AS hiccup (e.g.
invalid_grant during a key-rotation window) could now delete a working grant
for a server the user isn't even using this session. The _prime_one comment
still claimed "priming can never revoke a live grant" — no longer true.
Add revoke_on_dead_grant: bool = True to get_user_access_token_classified. When
False, the four would-revoke sites return refresh_failed_transient with the
token left in place instead of deleting it. _prime_one passes False: priming
still refreshes+persists refreshable tokens (f585c47b's fix intact) but never
revokes — the authoritative revoke stays on the lazy-dispatch path, where the
user actually invokes the tool and a permanent failure means re-consent anyway.
Replaces the vacuous prime test (which fully stubbed the resolver, so its
"never revoke" assertion was meaningless) with a test that drives the REAL
resolver and pins both directions: same permanent failure, same code path,
revoke_on_dead_grant=False keeps the token / =True (lazy default) deletes it.
Follow-up to f585c47b. Three correctness gaps from a max-depth review of
that commit, all in the same dead-transport / session-corpse family it set
out to close.
1. read_resource_sync and get_prompt_sync were left on the old
BrokenPipe/ConnectionReset/EOF-only eviction guard, so a dead
streamable-http transport (McpError(CONNECTION_CLOSED), anyio
ClosedResourceError, server-restarted session) reused the corpse session
forever — the exact restart-hang call_tool_sync already fixes, just for
resources and prompts. Both now route through _is_dead_transport and
evict + trip the breaker like the tool-call path.
2. _is_dead_transport matched a bare "session terminated"/"session not
found" substring, so a healthy session-owning MCP server (game/shell)
rejecting a stale id with those words was misclassified as transport
death — evicting the live session and opening the SHARED per-server
breaker for every user after 3 such rejections. Now anchored on the
SDK's deterministic synthesized code (32600, pinned as a named constant)
with its exact message as a forward-compat fallback. The client never
receives "session not found" for a real dead transport (the SDK discards
the server's 404 body), so the tightening loses no coverage.
3. _is_dead_transport omitted httpx's read/write/close NetworkError leaves
and the whole TimeoutException family (Read/Write/Pool timeouts are NOT
builtin TimeoutError), so a stream that died on an idle read timeout —
the dominant idle-death mode — fell through to "other" and the corpse
was reused. Broadened to httpx.NetworkError | TimeoutException |
RemoteProtocolError (LocalProtocolError, our own bug, stays excluded).
Also: _oauth_user_server_status iterated _user_pool_entries without a
list() snapshot, so a concurrent pool insert/evict on the mcp-loop thread
could raise "dictionary changed size during iteration" and 500 the status
endpoint. Snapshot like the sibling get_all_server_status does.
Adds TestIsDeadTransport (direct classifier unit tests, incl. the
healthy-"session not found"-is-not-dead and httpx-coverage regressions) and
resource/prompt eviction tests. All 8 behavior-change tests fail on the
pre-fix source and pass with the fix.
_stop_retrying calls _is_ctx_overflow with no exception-class gate of its own,
so a retryable 429 whose token-quota text contains an overflow phrase (e.g.
"... maximum number of tokens allowed per minute ...") was treated as a
deterministic overflow and made non-retryable.
Gate _is_ctx_overflow on "not a known backend class": an overflow is never a
recognized error (it arrives as BadRequestError/InternalServerError, neither in
_BACKEND_KNOWN_EXC_NAMES), so excluding known classes can't suppress a real
overflow while keeping a 429 retryable across every caller (the retry gates,
send-loop recovery, chunker, task_agent loop, formatter). _format_backend_error
drops its now-redundant inline class check.
Addresses Copilot review feedback on #740.
A session created under the openai-compatible provider and resumed under the
anthropic-compatible provider (same vLLM model) failed with an opaque
InternalError instead of recovering. Root cause: vLLM returns a context-window
overflow as HTTP 400 BadRequestError on /v1/chat/completions but HTTP 500
InternalServerError on /v1/messages, and the rehydrated resume payload overflowed
the window. The 500 was retried four times then surfaced as a bare class name.
- Detect overflow by message text, not exception class (_is_ctx_overflow),
shared across the fatal-error formatter, both stream-retry gates, the send-loop
recovery, the chunker, and the task_agent loop. Overflow is non-retryable
(deterministic; no backoff). Phrasing is overflow-specific so a token-quota
rate-limit isn't misclassified.
- Proactive pre-send compaction (Layer A): when already over the hard ceiling,
compact once before the first stream so a resume that arrives over-window (or
follows a switch to a smaller-context model, with no prior compaction) doesn't
go out blind. Generation-guarded end to end so an orphaned or superseded send
can never swap the live generation's history.
- Binary-subdivision chunker: an over-window summary batch is split in half and
the partials merged (~log2(N) calls, not one per block); a lone over-window
block is truncated progressively down to a floor before bailing irreducible.
- Cooperative cancellation honored through compaction; send() consumes its own
generation's cancel signal on exit, so a stale cancel can't block a later
idle /compact and a live cancel is never disarmed.
- _format_backend_error surfaces "Context window exceeded ..." instead of an
opaque InternalServerError, and only for unrecognized classes.
- retry/rewind, the continuation hint, and title generation all exclude the
synthetic [Conversation summary] turn so they can't target the label.
- task_agent salvages a sub-agent's partial work on any terminal error (not only
overflow), re-raising only when there is nothing to salvage.
- Add a plain-terms gloss of the claim (shell/plant split up front)
- Add a 'Converged-upon' grounding subsection: independent corroboration
from capabilities, control theory, software architecture, and LM theory
- State provenance as a precondition of the reach-avoid certificate
(CaMeL control/data-flow separation), not just an entry point to police
- Drop redundant 'none' from the effect-record status enum; normalize
minor notation (A_bot, h->N)
- Fix stray backslash-escaped quotes that rendered literally
* fix(memory): atomic single-statement upsert for memory save/update
save_structured_memory used "try INSERT -> catch IntegrityError ->
SELECT + UPDATE". On PostgreSQL a model saving the same key twice in a
turn logged a uq_smem_name_scope violation on the failing INSERT, and the
pattern threw + caught an exception on every update.
Replace it with one statement: a new StorageBackend.upsert_structured_memory
on both backends emitting INSERT ... ON CONFLICT (name, scope, scope_id)
DO UPDATE ... RETURNING. It returns (row, was_update) -- the full saved
row and whether an existing row was updated -- like Django's
update_or_create; was_update is the supplied (fresh) memory_id differing
from the returned id. save_structured_memory is a thin wrapper over it.
description / mem_type of None mean "leave unset": the column default
applies on insert and the stored value is kept on conflict; an explicit
value (including "" / "general") overwrites -- so clearing a description or
setting type back to "general" now persists, where the prior
"if mem_type != 'general'" / "if description" semantics silently dropped it.
The memory tool and the memories HTTP endpoint pass None for omitted fields
and read effective type/scope from the returned row; the HTTP endpoint
returns that row directly (one query, no follow-up SELECT).
Removes the now-unused update_structured_memory primitive and its dead
STRUCTURED_MEMORY_MUTABLE constant. Adds cross-backend storage tests and a
session tool-path test (preserve-on-omit / overwrite-on-explicit), run on
PostgreSQL via --storage-backend -- the save-over-existing path was
previously SQLite-only.
* docs(memory): clarify upsert was_update precondition
Lead the upsert_structured_memory docstring with the behavioral contract
(callers MUST supply a fresh unique memory_id) rather than the internal
id-comparison mechanism, so a future caller can't reuse an existing id and
silently get was_update=False on a real update.
claude-code-review.yml granted pull-requests: read, so the Claude reviewer
ran green but its post step was permission-denied (permission_denials_count:
3) and posted no review on the PR. Bump to pull-requests: write so it can
post the review + inline comments.
claude.yml (the @claude responder) had the same read-only block and would
silently fail to post a reply; widen it to pull-requests + issues: write.
contents stays read -- no repo-push capability is granted. Both workflows
remain gated (the reviewer to same-repo PRs via head.repo.full_name ==
github.repository; the responder to @claude from OWNER/MEMBER/COLLABORATOR),
so write is scoped to already-trusted triggers.
Injected memories ride in the cached system block, so calling
_init_system_messages() on every memory save/update rebuilt the prompt
prefix and busted the provider prompt cache (a full system + history
re-write) -- for a memory the model already holds via the tool result.
memory(save) now only invalidates the per-turn search cache, so an
in-turn memory(search)/(list) still reflects the write; the new memory
folds into the prefix at the next natural recompose or the next session.
Also drop the redundant _init_system_messages() in the /reason handler:
reasoning effort rides in request kwargs (output_config / thinking), not
the composed prompt, so it recomposed to byte-identical output.
Add a chain-level test through the real _exec_memory -> no-recompose path
(asserts prefix unchanged, search cache invalidated, next recompose folds
the memory in). The prior memory tests either drove _init_system_messages
directly or patched it out, so this path was uncovered.
Address the Copilot review on #732 plus a task-agent sub-tool nesting
race surfaced alongside it.
Nesting (web UI):
- A sub-tool step whose task_agent row hasn't painted yet (the 4-wide
tool pool's ordering window) buffers and nests when the row lands,
instead of escaping to a top-level row that looks main-harness-issued.
- A row that never paints (id-correlation mismatch / aborted agent)
escapes its buffered steps back to a visible top-level paint after a
grace window, so steps are never buffered invisibly or leaked.
- The nested card survives the parent row's pending->resolved rebuild; a
call_id reused across turns builds a fresh card rather than stealing the
prior agent's steps.
- tool_info routes through the same nesting path (no duplicate top-level
row); a namespaced sub-tool result no longer grafts onto an unrelated
top-level row.
Denial reasons (backend):
- Preserve the specific denial reason a gate already stamped (operator
feedback, or the matched policy pattern; web and CLI contracts) instead
of clobbering it with a flat "Denied by user" -- in both the sub-agent
and the main tool loop.
Verified with the livepass task_agent harness (race + orphan-escape
scenarios, headless) and unit tests.
Final chunk of the task_agent modernization: rebuild a finished task
agent's card from /history (reload / reopen while the workstream is in
memory) and isolate each sub-agent's file-read tracking.
Recall: _project_agent_steps projects a sub-agent's trajectory into step
items (FIFO-per-call_id pairing via _iter_agent_tool_results, shared with
_cancel_ledger; output/arguments/count capped); _stash_agent_trajectory
keeps them on the UI in an LRU-bounded store; make_history_handler
attaches them as agent_steps to each task_agent tool_call, and
replayHistory/_replayAgentCard rebuild the collapsed card. In-memory only
(durable persistence deferred); a cold/evicted entry renders the flat
parent row ("not retained"), never a fabricated 0-step card.
Read isolation: _read_files (the blind-overwrite guard's memory) is now
per-sub-agent via the _active_read_files contextvar -- _exec_task copies
the parent's set on spawn and merges the agent's reads back on
completion, so a sibling in the 4-wide pool can't suppress another
agent's guard.
Also: _exec_task now self-reports the task_agent tool_result on every
path (the parent loop only reports error/denied results centrally) --
without it the live card never completed and a failed task recorded
is_error=False in the canonical trajectory. is_error flows from
_tool_error_flags to the recalled step; on_info suppression is per-thread
so a parallel sibling tool's progress isn't dropped.
Route a task agent's sub-tool events (tool_pending / approve_request,
tagged with parent_call_id) into a collapsible card under the task_agent
row, replacing the blue on_info turn-legs.
- conversation.js / interactive.js: buildAgentCardBody +
_routeAgentItems / _ensureAgentCard nest steps by parent_call_id.
Collapsed by default (a task agent can run 100+ steps and the parent
fans out many in parallel); the label carries the live count + state.
Auto-expand when a nested approval is pending so the blocking prompt
can't hide behind the toggle.
- session.py / session_ui_base.py: on_agent_step paints auto-tool step
rows; namespace child call_ids by parent so the 4-wide task pool can't
collide on local sequential ids (call_0); suppress sub-agent on_info on
the web pane (no call_id to nest by — the card carries steps + result).
- cli.py: on_agent_step prints a dim step leg (no card on the CLI, which
keeps its on_info).
- livepass.py: task-agent card harness driving the real InteractivePane.
Rebuild the task_agent sub-harness on the canonical Turn trajectory (build list[Turn], lower via dicts_from_turns at the wire boundary) instead of hand-rolled OpenAI dicts; the cancel-ledger helpers read Turns.
Tag each sub-tool's events with parent_call_id via a lock-guarded child registry stamped centrally in SessionUIBase._enqueue, so a later UI can nest a task agent's steps under its card. Getattr-guarded on the session side so CLI/eval/test UIs are unaffected.
Behaviour-preserving (same wire shape, same cancellation semantics); the parent tag is wire-invisible and unconsumed until the frontend card lands.
Compaction swapped a session's in-memory history for a summary but left the full
transcript in storage, so resume() reloaded all of it -- on a long session, or
one switched to a smaller-context model, the rehydrated context overflowed the
model window and deadlocked the first post-resume send.
Persist a `_source="compaction"` marker (summary + watermark) on compaction;
resume rehydrates [summary] + [rows after the watermark] instead of the full
transcript. Full history stays in storage for /history, export, and audit;
markers are filtered from display, search, and export, and rewind/retry
truncation is floored at the marker so the summary's backing is never deleted.
The watermark and search filters count real transcript rows only. No migration.
Address PR #730 review. _summary_input_budget_chars now caps the _MIN_SUMMARY_BUDGET_CHARS floor at the true input capacity (input_tokens), so output reserve + budgeted input + prompt always fit context_window; on a window too small to summarize it returns a sub-floor budget and _pack_blocks bails as irreducible instead of overflowing the summary call. After the half-window output-reserve bound this only affected sub-~2048-token windows, but it was a real edge.
Clarify the _CompactionIrreducibleError docstring: chunked compaction never drops or fabricates whole turns, but a single oversized block is still head/tail-truncated as summary input via _truncate_block.
Add test_budget_never_exceeds_true_input_capacity.
The compaction summary ran as a single model call sized by the per-message
token estimate, which disagreed with the head+tail-capped formatted text, so a
long history could overflow the summary call itself; the old prefix-fit also
silently dropped the most-recent messages.
Summarize the whole selection via _summarize_blocks: greedily pack the
formatted blocks into batches that each fit the summary call's own input budget
(_summary_input_budget_chars), summarize each, and recursively merge the
partials until they collapse to one. The common case (it all fits) stays a
single call. Bail to the existing False path when the input is irreducible
rather than fabricate a summary; a mid-chunk failure leaves messages untouched
(atomic swap only on full success).
Bound the summary output reserve to half the context window
(_summary_output_tokens), used by BOTH the input-budget sizing and the actual
call. compact_max_tokens defaults to the full window (32768); clamped only by
max_output_tokens it reserved the entire context for output, flooring the input
budget so compaction overflowed (or bailed as irreducible) at the
default/small-window config that needs it most. Large windows are unaffected
(compact_max_tokens stays binding).
Guard an empty summary (keep history instead of swapping in nothing and
reporting success). Fold tool-def tokens into the _last_usage-less estimate AND
the post-compaction usage anchor, so the compact-before-truncate budget doesn't
over-state free space by the tool-def count. Single-source the shared
compactor/merge prompt section (_COMPACT_OUTPUT_FORMAT) and the tool-def sizing
(_tool_def_chars/_tool_def_tokens). A just-resumed session (no _last_usage) now
counts tool-def tokens so it doesn't undercount and skip proactive compaction
until its first reply re-anchors the estimate.
Prepush review follow-ups: generation-guard the end-of-turn auto-compaction and
its resume turn so a force-cancel during the slow summary call can't compact or
persist under a new generation (matching the mid-turn and end-of-loop guards);
single-source the soft-threshold predicate (_over_soft) shared by the mid-turn
policy, _compaction_owed, and the end-of-turn check; add tests for the
pre-attempted-compaction guard and the recursion depth ceiling.
Unify truncation and compaction on one provider-anchored fullness measure
(_estimated_prompt_tokens), closing the 80-100% dead zone where tool output
was truncated but compaction never fired. Make compaction cooperative: advise
the model to wrap up and record its plan, compact if it continues, auto-resume
after a cooperative stop, and compact-before-truncate (preserving the in-flight
tool-call turn). Floor auto_compact_pct at 0.1 (invalid 0 -> default 0.8).
The opening hours were a gold/HP death spiral: a fresh hero spent more gold
healing a fight than the kill paid, mid-tier foes out-damaged a starting HP
bar, the map gave no read on where harder foes spawned, and a spent day left
the player idle until the UTC rollover. This eases the on-ramp across both
shipped worlds.
Economy
- Healer drops from 2 to 1 gold per HP, so topping up no longer outruns income.
- Starting purse 20 -> 37: enough to buy the cheapest armor and one potion up
front, a one-point DEF bump plus a heal cushion the player chooses to spend.
Combat
- Tier 2-4 common foes lose 1 ATK each, trimming the burst that could halve or
end a fresh hero in a single bout. Rares and the boss are untouched.
Wayfinding
- The road glyph changes from "=" to a shaded path that reads as one continuous
road in every orientation; "=" only looked right horizontally and broke into
stacked dashes on vertical runs. Cinder's basalt path gets the same treatment
with a crosshatch glyph free in its palette.
Rest
- Sleeping at the inn now rolls a spent adventurer into a fresh day's turns (it
already fully heals). The top-up fires only at zero turns, so it never banks
past the daily cap.
Verified: full suite green (+2 rest tests, ruff/mypy clean); the greedy balance
bot still clears the world 11/12 seeds at an unchanged pace on both packs.
Auto-title generation and manual refresh stopped producing titles on
reasoning models (the cluster serves qwen3.6). The title call capped
max_tokens at 200, so the model's think pass consumed the whole budget and
content came back empty (finish_reason=length) -> the title was skipped.
Both paths share _generate_title, so both broke.
Title path:
- Raise the title completion to 2048 tokens so reasoning finishes and the
title text actually lands.
- Recover the title from content (never reasoning): reuse the canonical
_strip_reasoning (handles <think>/<reasoning>, paired or unclosed) plus a
backstop for the opener-absent </think> shape some templates emit, take
the first non-empty line, and peel a "Title:" label and wrapping
markdown/quote decoration. Internal punctuation is preserved. Cap at 80 to
match the manual-alias bound.
Temperature:
- _utility_completion no longer hard-codes a temperature; it defaults to the
session/registry value the main turn uses. Title (was 0.7/0.3), web-fetch
extraction (was 0.2), and compaction all defer. Hard-coding a constant
fought thinking/no-temp models and silently overrode an explicit [models.*]
temperature; the provider still gates temperature per model.
Tests: title sanitization across think/reasoning variants, truncation, and a
trailing-prose case; utility-completion temperature deferral + explicit
override.
The send POST's `!r.ok` guard threw a bare `send_http_<status>`, which both
send `.catch` handlers render verbatim — so a rejected send surfaced as
"Connection error: send_http_400" instead of the server's reason. Read the
`{error}` body and throw that, falling back to the status code when a wedged
proxy answers non-JSON (502/504 HTML) so it can't become an "Unexpected
token <" error. Applied to interactive and coordinator.
Also correct the queue-controller comments: a dequeue releases no
"server-side reservation" (queued messages are text-only and dequeue_message
just pops the entry), and onAfterDequeue is wired by coordinator too — not
omitted.
The queued-message dismiss DELETE hardcoded /v1/api/workstreams without
the node-proxy prefix, so cancelling a queued message on a proxied
(remote-node) interactive workstream hit the console root, 404'd, and
the message was delivered anyway -- the dismiss silently did nothing.
composer_queue:
- Prefix getBase() onto the dequeue DELETE; interactive passes getBase
(mirrors the attachment controller). Coordinator stays at base "".
- Never remove the card before the server confirms the cancel: removed
-> drop the card; not_found (already drained) -> promote to a sent
bubble + "already sent" notice; 404 (reaped session) -> terminal drop;
error/timeout -> re-enable + "couldn't remove" notice.
- Bound the DELETE with a 15s AbortController (Promise.race fallback when
AbortController is absent) so a wedged node can't freeze the card.
- a11y: aria-disabled (not the real disabled attribute) keeps keyboard
focus on the dismiss control; in-flight state shown via aria-busy + CSS.
consumers (interactive, coordinator):
- Bound the send POST with the same 15s timeout so a pre-bind dismiss
can't strand the card when the POST hangs.
- r.ok guard so a rejected send (4xx/5xx error body) surfaces as an
error instead of being promoted as "delivered".
- Coordinator wires onNotice -> appendText.
Operator-context system turns must follow a user/tool input turn — producers
maintain this via the user/tool drain seams plus the synthetic wake turn, so an
assistant predecessor is unreachable today. Add a fail-loud guard so a future
producer that breaks the invariant surfaces in logs instead of silently splicing
operator markup into the model's own prior output.
Logged, not raised: it degrades to a fold, since the nonce still gates operator
trust regardless of the host turn, so the harm is out-of-distribution voice
rather than a trust breach — disproportionate to crash a turn over.
Address PR review feedback:
- detection_pattern(()) with an empty tag set compiled to an overly-broad regex
(the empty alternation matches any [start ...]/[end ...] run), which would turn
the forgery scanner into a false-positive generator. Reject an empty or
all-empty tag set up front. Not reachable from the sole caller today, but it is
a public, security-relevant helper.
- Clarify build_operator_instruction_declaration's docstring: the trusted region
is delimited by both the start and end markers (each carrying the nonce), not
just the opening marker.
Swap the trust-fence marker shape from <tag_nonce>...</tag_nonce> to
[start tag_nonce]...[end tag_nonce] for both the operator fold (system-reminder)
and the output-guard judge (tool_output). Angle-bracket markup pushed some local
models out of distribution and toward emitting their own turn-structure tokens:
chat templates built around rigid <...>-style structural tokens derail once a
few folded reminders accumulate. The start/end keywords carry no slash (no </ or
[/ closing-tag shape) and read as ordinary text.
Single-source the shape in fence.py (_OPEN_KW/_CLOSE_KW + detection_pattern) so
wrap, neutralize, the forgery/leak detector, and both trust declarations track
one definition. The nonce still rides both boundaries (unforgeable close); the
leak-vs-forgery split and the forge-in / break-out defang are preserved. The
fold is wire-only, so there is no migration; the legacy persisted-envelope
readers keep the old shape.
Add regression tests pinning each trust declaration to fence.wrap's emission so
a future keyword change fails loudly instead of silently desyncing the anchors.
* feat(projects): governed project containers — memory scope, grouping, manage UI
A workstream can attach to a project: a first-class, shareable resource
container that owns a `project` memory scope, groups conversations, and is
managed from the console.
Storage / migration 062: projects + project_members tables, workstreams.
project_id, and the memory type default project→general; grants
project.{create,read,write,delete} (admin-default).
Recall + writes: project memory is recalled iff the workstream is attached AND
the user has access (owner ∨ member ∨ public-for-read), resolved once at session
construction; coordinators recall it too. New saves default to the project when
attached + writable; the save and delete paths are write-gated; deleting a
project purges its scoped memory; archived projects aren't recalled.
Access = RBAC capability ∧ per-project ACL (auth.resolve_project_access, a
single-fetch resolver); visibility changes, member management, and delete are
owner-only.
API: project CRUD routes on both the server and console; project_id threaded
through workstream creation, spawn inheritance, the cluster-create proxy, the
dashboard / snapshot / coordinator row builders, and the collector deltas.
UI: a project picker with an inline "+ New project" creator in every creation
box (console launcher + standalone dialog + dashboard); group-by-project in the
rail; a project badge in the composer and on dashboard rows; a console manage
tab (list + create/edit + members shelves). The admin Memories view gains
coordinator/project scope filters and human scope labels (name, not hex). The
memory tool schema documents the project scope and the attach-aware default.
* fix(projects): client refresh hardening, creator race guard, SDK project_id
Addresses PR #724 review feedback plus two bugs found while validating it.
- projects.js refreshProjects: a non-OK status (e.g. 403 when the caller
lacks project.read) or a network/parse error no longer blanks the cache
or masquerades as "no projects" -- the prior cache is preserved, the
failure is recorded (new projectsError()) and warned. Honors the
long-standing "a transient error can't blank the rail" docstring.
- projects.js _fp: the fingerprint separators were raw control bytes,
which made git treat the whole file as binary (no reviewable diff).
Rewritten as escape sequences instead of raw bytes -- behavior is
byte-identical at runtime.
- project_creator.js: createProject() could reject unhandled (authFetch
throws on network/401; r.json() throws on a non-JSON body), leaving the
widget stuck busy/disabled. Added a .catch, plus a generation guard so a
create whose widget was cancelled/reopened mid-flight drops its result
instead of selecting a project the user backed out of.
- types.ts: add project_id to CreateWorkstreamRequest / WorkstreamInfo /
DashboardWorkstream to match the server schemas (was SDK-invisible).
- test_project_api.py: move side-effecting HTTP calls out of asserts so
the requests run even under python -O.
* fix(projects): JSON.stringify the cache fingerprint, drop control-byte separators
_fp joined fields/rows on raw NUL/SOH bytes, which made projects.js read as binary to git. Replace with a collision-proof, escape-free JSON.stringify encoding -- same change-detection semantics, zero embedded control characters.
- Turn.effect_status also catches TypeError: a corrupt non-string meta value
(e.g. a dict that survived into the column) would otherwise crash a consumer
on access, since EffectStatus(non-str) raises TypeError, not ValueError.
Degrade to None, mirroring the meta decoders (Copilot review).
- test_lowering: the wire-repair synth now carries the _effect_status side
channel (stripped before the provider wire) — assert it.
- test_session_mcp_dispatch_error: the _capture stub swallows the new status
kwarg via **_ so it stays signature-compatible with _report_tool_result.
The unknown / none / committed distinction the cancel and timeout paths
carry lived only in the result's free text — a deterministic reader (a
re-issue guard, owner-side compensation) couldn't recover it without
parsing prose. Promote it to a typed EffectStatus on the canonical Turn.
- EffectStatus (committed/none/unknown/partial/rolled_back) rides
TurnMeta.extra["effect_status"] — wire-invisible like the other meta
side channels: the model still reads the body, deterministic code reads
the type.
- Persisted in the role-exclusive conversations.meta column (source_meta
rides SYSTEM turns, effect_status rides TOOL turns), routed by role in
reconstruct_turns. No migration; survives reload for the audit trail.
- Producer seam: _report_tool_result(status=) + a _tool_status dict popped
at the fold, mirroring _tool_error_flags.
- Populated where the disposition is already determined: UNKNOWN at the six
unobserved sites (bash / MCP-tool timeout, bash SIGKILL-cancel, cancel
synthesis, wire-repair) and a precise none/partial/unknown on a cancelled
task agent (shared _cancel_ledger so the typed status and the prose
disposition can't disagree). Ordinary results stay unset.
Only the unknown/none split is load-bearing (HYPOTHESIS.md effect-record
appendix: unknown, never none); the full per-effect reversibility list
stays deferred. Thread A of the effect-record work; Thread B (per-tool
Smart-Approval floor + reversibility surfacing) follows.
A bash command SIGKILL'd at its deadline and a timed-out MCP tool call are
killed / abandoned mid-flight, so their side effects are as unobserved as a
cancelled call's. Both read as a definitive "timed out after Ns", which invites
a blind re-run (a double-send) exactly as a dropped record invites an orphan.
Route both through a shared TIMEOUT_OUTCOME_CLAUSE so they read "Outcome
UNKNOWN ... do not assume it did not run, reconcile before re-issuing" — the
same "unknown, never none" discipline cancellation already follows
(HYPOTHESIS.md effect-record appendix). bash also keeps any partial stdout
captured before the kill, mirroring the cancel path.
Read-only timeouts (search, MCP resource/prompt reads) stay a plain failure:
an idempotent read has nothing to reconcile, so the reconcile advice would be
misleading there.
* docs(hypothesis): gate-placement & effect-record appendix; scope incompressibility; split the two walls
Refinement + expansion pass on the harness hypothesis.
Appendix (new subsections):
- Gate placement (fail-closed, in practice): γ as a pure, effect-free
parse-and-authorize; syntactic / user-authorization / structural-intent
validation; semantic intent as a recursive plant call (a mini-harness),
not a predicate in γ; "before any invocation" sharpened to "before any
effect" — reads aren't free, the parser must not act, the output is an
action too.
- Effect records (what ρ folds back): pins down the
e = (tool_id, action_id, status, effects, time) shape the body referenced
twice but never defined; committed/none/unknown trichotomy + a reversibility
bit, framed explicitly as an open interface, not a result.
Corrections:
- Scope the incompressibility conjecture: split per-step drift by coordinate
(the shell term is a low-complexity designed descent), so the incompressible
part is the plant's, not all of W; add the coarse-functional counter-
possibility (V* is one scalar hitting time, sometimes cheap) and state the
claim conditionally. Walks back the earlier "the dynamics it certifies are
the weights" overclaim.
- Split the second wall: the tape / space-O(L) picture follows from the
autoregressive structure alone; the per-pass TC^0 bound is separate and
weaker; flag that chaining them is a non-sequitur.
Smaller:
- Concrete justification for the standard-Borel assumption.
- Reading-table rows for the C/Y/A/E spaces and for H_ok/B.
- Daemon note: per-cycle hazard compounds, (1-q)^h over the horizon.
- Minimax: well-posedness caveat for sup over the adversary class Π.
- Note that H_cancel refines the body's deliberately coarse H\H_ok.
Notation (consistency linter clean):
- Brace the subscript A_{⊥} in the new table row (was unbraced — GitHub
render hazard the linter guards against).
- Daemon cycle-count N → h, freeing N for the fundamental matrix.
* docs(hypothesis): address Copilot review — plain quotes + 'none' status value
- Effect-record status enum: add `none`, which the prose already treats as a
distinct value ("unknown ... never none"; the committed/none/unknown
trichotomy). Resolves the enum/prose inconsistency — `none` (no effect) is
distinct from `rolled_back` (ran, then undone).
- Drop the two backslash-escaped quotes (the incompressibility walk-back and
the minimax well-posedness caveat) for plain quotes, matching the rest of
the document. GFM strips the backslash, so they rendered fine; the escapes
were just unnecessary and inconsistent.
* refactor(doctor): replace turnstone-bootstrap with turnstone-doctor
turnstone-bootstrap was an LLM setup wizard for Day-0; run.sh now owns install.
Repurpose its LLM/conversation plumbing into turnstone-doctor — a diagnose-only
tool for a running cluster.
- Preflight detects the install kind (docker-compose/systemd/pip/source) from
config.toml + TURNSTONE_* env, with secret redaction.
- Self-configuring brain resolves the cluster's own model from config/env/storage
read-only (no migrations, no create_all), falling back to interactive
selection; the attempt itself is the LLM-backend health check.
- Deterministic version check: installed version, cluster drift via the console's
authoritative /health, and latest upstream stable/experimental (offline-safe).
- Read-only diagnostic tools (read_file, compose/systemd/journal, http_health,
check_llm_backend, node_health, finish) behind one secret-scrubbing chokepoint;
no generic shell, so read-only is structural.
- node_health reaches a node the right way for the detected install kind
(exec-into-container for compose, direct HTTP otherwise), overridable per node
for mixed clusters.
- mTLS-aware: forwards [database] SSL params and reports node-mesh mTLS instead of
mislabelling healthy nodes "unreachable".
init_storage gains a backward-compatible create_tables override for read-only
opens. Entry point turnstone-bootstrap -> turnstone-doctor; README/QUICKSTART/
architecture/docker docs, the bundled compose header, run.sh, and the CI smoke
updated. CHANGELOG deferred.
* fix(doctor): address Copilot + CodeQL review findings on #718
Validated all seven review findings (none false positives) and fixed:
- check_llm_backend now applies the same scheme / metadata-host guard as
http_health (extracted to _assert_safe_http_url), so a model-supplied
base_url can't be steered at the cloud metadata endpoint or a file:// URL.
- node_health no longer double-appends the default port when the operator
passes host:port (regression: 10.0.0.5:8081 -> http://10.0.0.5:8081:8080).
- node_health install_type enum uses "git-source" to match the label the
rest of the module and the prompt/report show the model (a schema-strict
provider would otherwise reject the value the model is told to use).
- _read_api_creds takes base_url + api_key as a unit from the first config
source that defines either field, then env-fills, instead of splicing the
two across different config files into a pair that exists in no real config.
- _mask_secrets masks assignment-shaped content inside comment lines, so a
commented-out real secret can't leak through read_file / the report; prose
comments (no KEY=value shape) still pass through untouched.
- drop the mixed import styles CodeQL flagged in doctor.py and test_doctor.py.
Adds 5 tests; ruff + mypy clean; full doctor suite passes (129).
* docs(hypothesis): clarity pass, GitHub-render fixes, consistency linter
Document (HYPOTHESIS.md):
- split the dense "Formal" definition into labeled subsections
- define the load-bearing terms: certificate (proven witness vs measured
surrogate) and the controller / plant (= M_W) / shell triad
- corrections: three-way drift split (+ r_env), scope the success/safety
collapse to absorbing refusal, unify tau*->tau_H and drop the orphaned bare tau
- calibrations: pin the incompressibility conjecture (still conjectural),
mark the interlingua=certificate identity as figure, soften the two-walls trade
- GitHub math rendering: brace command-subscripts (_\bot -> _{\bot}, etc.) so the
markdown emphasis parser stops breaking inline math; replace R_\# with R_{\sharp}
(\# unescapes to a raw # in GitHub math)
Linter (lint_hypothesis.py):
- deterministic consistency checks A-G; G adds an orphan/redundant-declaration
scan that catches the bare-tau failure mode
- residue guards so tau^star, unbraced _\cmd subscripts, and \# cannot return
* fix(hypothesis): make lint_hypothesis.py pass ruff under py311
- precompute the inline-$ count so no backslash sits inside an f-string
expression (backslashes in f-strings are 3.12+; the project targets 3.11)
- split the one-line import (E401/I001); open HYPOTHESIS.md via a context manager (SIM115)
- console/server.py: replace a stale hard-coded `session_routes.py:852-854`
comment reference (already drifted to make_close_handler's signature) with
a by-name reference to make_close_handler's not-found path.
- test_cancel.py: rename test_marks_most_recent_action_unknown ->
test_marks_in_flight_action_unknown; the disposition marks the first
unanswered (in-flight) call, not the most recent — they merely coincide in
this two-call case.
The multi-stage review of this branch surfaced four major + two minor issues,
three of them in the new cancellation code. All fixed here (bug-3, the stale
generated TS SDK spec, stays deferred — it regenerates out-of-band).
- sec-1: cancelling a coordinator now auto-cascades to its children, but the
cancel route allows the service-scope bypass while the removed stop_cascade
gated the same destructive subtree-cancel at no-bypass — a service token
without admin.coordinator could trigger the cascade. Re-assert the
no-service-bypass gate inside _cascade_cancel_to_children, so a plain cancel
by an under-privileged service token still cancels the coordinator's own
turn but no longer cascades.
- bug-1: _cancelled_agent_disposition took the LAST issued tool call as the
in-flight one. _run_agent executes a turn's calls sequentially, so the
in-flight call is the FIRST unanswered one — taking the last inverted
unknown/none on a multi-call turn (a SIGKILL'd bash mislabelled "not
started", the never-run tail mislabelled UNKNOWN, inviting a re-run of the
destructive call). Fixed to first-unanswered.
- perf-1: the per-child cancel fan-out was awaited inline before the cancel's
200, so a cancel could block for tens of seconds on slow/unreachable
children. Return the fan-out as a response BackgroundTask so it runs after
the 200 (trigger, not drain).
- bug-2: the initial-send worker (_run_initial) cleared _worker_running
unconditionally — the same clobber the session_worker guard just fixed.
Apply the identity guard there too.
- sec-2: restore the per-child cascade audit row (coordinator.cancel_cascaded)
the removed stop_cascade wrote; it had become log-only.
- q-1: extract the shared UNKNOWN-outcome clause (UNOBSERVED_OUTCOME_CLAUSE)
so the wire-repair fallback and the session-layer synthesis can't drift.
Follow-up to the cancellation review — harden how cancel interacts with a
workstream's OWN turn and tools, not just its children and agents.
- wait_for_workstream: the wait loop holds no cancel handle and blocks on the
child-event bus, so a cancelled coordinator parked in a wait stayed pinned
for up to WAIT_MAX_TIMEOUT (600s). Add a cooperative check to the ~2s
progress heartbeat — it raises GenerationCancelled, which propagates out of
the otherwise cancel-blind wait (~2s abort).
- spawn_batch: stop creating the rest of the children once cancel is observed;
already-spawned children stay recorded (they are live, durably parent-linked
workstreams), the remainder are marked not-spawned.
- session worker: only clear _worker_running if this thread is still the
current worker, so a late-finishing abandoned worker (force-cancel) can't
clobber a live successor's flag — which would let a third send spawn a
duplicate worker on the same session.
- bash silent-cancel: a SIGKILL'd silent command now records outcome-UNKNOWN
(is_error, partial output kept) instead of a clean "Cancelled by user." that
read as a successful empty result on replay.
- wire-repair: the last-resort orphan disposition now reads outcome-UNKNOWN,
matching the cooperative-cancel message (unknown, never none).
Deferred: MCP / web_fetch / web_search remain uninterruptible mid-call,
bounded by tool_timeout; only bash is truly preemptible.
A cancelled agent previously discarded its own ledger and reported a bare
"(task interrupted by user)" — fabricating the *outcome* (read downstream
as "nothing happened"), which invites a double-send as readily as a
dropped record causes an orphan. Make the fold-back honest, and propagate
an owner's cancel down the coordinator subtree.
- task_agent (single + parallel): on cancel, fold back a deterministic
disposition built from the agent's in-memory ledger — actions completed,
the in-flight action flagged outcome-UNKNOWN, and not-started calls —
instead of the opaque interrupted string.
- coordinator cancel now auto-propagates to its direct children via a
post_cancel hook on the shared cancel handler (cooperative fan-out; no
blocking drain).
- synthesized cancelled tool results now read outcome-UNKNOWN rather than
implying the call never ran.
- remove the now-redundant stop_cascade operator endpoint (handler, route,
OpenAPI spec + schema, tests, docs); a coordinator cancel supersedes it.
Review follow-up (#717). The bug-1 fix made the transient keep-path retain the
per-(user, server) refresh lock for serialization, so the lock entry now lingers
after a transient failure. When the token then vanishes (missing) or goes
undecryptable, _no_token_result pruned only the backoff entry and left the lock
entry stranded, so mcp_oauth_refresh_locks could grow on that path. Drop both
sibling dicts in _no_token_result (removing the now-redundant explicit
_drop_refresh_lock on the in-lock decrypt return); the regression test asserts
both are pruned on the missing-after-transient path.
Follow-up to #714 (Entra OBO, #682). A refresh failure deleted the user token +
emitted token_revoked regardless of cause, so a transient AS/network blip during
a forced refresh (the live 401-retry path) permanently revoked consent
cluster-wide. Fixing only that, though, opens the dual failure: a genuinely-dead
grant the AS reports in a non-standard shape would now be kept forever and the
user stranded on a retryable error with no re-consent path. This classifies the
failure three ways so each is handled correctly.
Classification (_classify_refresh_failure): MCPOAuthRefreshFailed carries a
_RefreshFailureClass instead of a bool —
- PERMANENT (revoke + re-consent): an explicit dead-grant / re-consent signal —
invalid_grant at any 4xx (400/401/403), invalid_scope, or an OIDC
interaction-required code (interaction_required / login_required /
consent_required / account_selection_required) the AS surfaces.
- TRANSIENT (keep, retry, never escalate): infrastructure (network, 5xx, 429,
malformed body) and operator-fixable codes (invalid_client, invalid_request,
unauthorized_client, unsupported_grant_type, temporarily_unavailable) —
re-consenting the user can't fix a bad client_secret, and an outage must not
revoke consent however long it lasts.
- AMBIGUOUS (keep, but escalate after a run): a 400/401 we can't map to a
standard code. A one-off can't revoke, but an uninterrupted streak past a
threshold escalates to re-consent so a dead grant in a non-standard shape
can't strand the user. Infra transients reset the streak, so an outage never
escalates.
Concurrency: do NOT drop the per-(user,server) refresh lock on the keep-the-token
path. Evicting it while the token is still live let a second concurrent caller
mint a fresh lock and refresh the same token in parallel; with refresh-token
rotation the second send reuses the consumed token, gets invalid_grant, and
spuriously revokes — the exact bug this commit prevents. The async-with still
releases the lock on return; the registry entry is pruned only when the token is
actually refreshed or revoked. Bit SQLite single-node hardest, where the pg
advisory lock is a no-op.
perf: a per-(user,server) cooldown short-circuits the token-endpoint round-trip
for a brief window after a transient failure, so a down AS isn't hit once per
tool call; self-heals when the window expires. Plus the lock-free in-flight key
set that collapses concurrent session-start pool primes (single mcp-loop thread).
dispatch/FE: the transient kind maps to a retryable mcp_refresh_unavailable
structured error (not mcp_consent_required); the FE titles it "Temporarily
unavailable" under a new soft "transient" category (amber, not the red hard-error
styling) in both stylesheets, with no wrong re-consent button.
tests: invalid_client kept (pins the discriminator on the error code, not the 4xx
status), single ambiguous 400 kept, 403 invalid_grant revokes, interaction_required
revokes, ambiguous streak escalates at the threshold, sustained 5xx never escalates
(outage safety), and the cooldown skips the second AS round-trip — all through the
real AS HTTP boundary.
Follow-up review of the #706 on-behalf-of / Entra ID MCP changes (#682).
security (PKCE downgrade): the AS-metadata "assume S256 when
code_challenge_methods_supported is absent" relaxation applied to BOTH the
RFC 8414 oauth-authorization-server document and the OIDC openid-configuration
document. Per RFC 8414 an omitted field on the oauth-authorization-server
document means the AS does NOT support PKCE, so this was fail-open. The client
always sends code_challenge_method=S256, making this discovery check the only
pre-flight that the AS enforces PKCE. Track which document won discovery and
assume S256 only for the OIDC document; the RFC 8414 document now fails closed.
Also log which discovery profile (rfc8414 vs oidc) answered, for operators
debugging an enterprise AS.
bug (consent loss): session-start pool priming called the refreshing token
lookup for every cold oauth_user server. A near-expiry token triggered a
refresh, and a transient refresh failure (network/5xx/429) deletes the token
and emits token_revoked — so a blip during a cold-pool warm (e.g. after a
reboot) silently revoked consent across servers the user wasn't even using.
Priming now reads the token directly and skips missing/near-expiry tokens;
a refresh that may fail stays on the lazy dispatch path.
perf/UX (blocking redirect): the OAuth callback awaited prime_user_server
(default 20s timeout), holding the consent redirect on a slow/unreachable MCP
server. Replaced with fire-and-forget schedule_prime_user_server that schedules
onto the mcp-loop (GC-safe, no unreferenced request-loop task) and returns at
once.
perf: prime a user's pools concurrently under a bound instead of serially, so
one slow upstream can't stall the rest.
hygiene: log (not silently swallow) prime scheduling failures at session start;
add exc_info to the prime-failure warning; guard run_coroutine_threadsafe
against a closed mcp-loop.
tests: per-document S256 + OIDC-fallback discovery cases; pool priming
(non-destructive on near-expiry, skips connected) and bound-token rotation
reconnect.
* This is a collection of little snippits to resolve all the OBO flow problems required to get this talking to entra id for on behalf of user impersonating to protected mcp servers. we make sure turnstone checks these mcp servers on startup, and address some of microsofts opinionated implementations of oauth2/oidc and metadata provided by the identity provider.
* minor token timeout bugfix
---------
Co-authored-by: root <root@pow3rtools>
Replace the amber gauge/needle favicon with a teal up-chevron and amber
dot on a dark-teal field. Applied identically to the console, coordinator,
and standalone UI entry points. Self-contained inline SVG data URI; no
network dependency.
Establishes the appendix pattern (locate a practical concern in the existing
formal objects; read off the discipline rather than inventing machinery) with
cancellation as the first and only worked example. Not the whole model.
Cancellation semantics, derived from objects already on the page:
- Cancel is a signal → lives in s (Markov). The gate closes on it: γ(s,y)=⊥ while
live, which blocks pending actions and all future turns with no new machinery.
- In-flight (past γ) disposition is a trinary on the kind of Q_E: cancellable
(propagate, true end-state), bounded (drain, real e), or opaque/unbounded
(controller fabricates a synthetic "cancelled" e so the loop can halt).
- Load-bearing rule: ρ may fabricate the acknowledgment but not the outcome — an
unobserved outcome is `unknown`, never `none` (double-send vs orphan, same bug
opposite sign).
- New terminal H_cancel ⊆ H\H_ok: non-accepting but safe (outside B), postcondition
"no action past γ after observed; in-flight drained or recorded unknown; ledger
consistent."
- Cooperative not preemptive (observed at next γ check, not on send); recursive
down the task-agent subtree (why task agents are the worst case).
- Compensation is the owner's job (saga, after H_cancel, reads child ledger) — the
cancelled agent can't know if it's needed; it never observed the outcome.
- Design pressure: prefer bounded/instrumented Q_E over opaque, so cancellation and
the ledger stay honest (a bash wrapper converts branch 3 -> branch 1).
Linter: balanced, no new collisions. Two ρ role-flags, both false positives
("authorized action" near ρ, correct usage).
The linter closed mechanical consistency; this review probes meaning, a separate
axis. Twelve findings, several real corrections, all folded.
Correctness:
- Stationarity overclaim: the supermartingale BOUND survives a nonstationary
kernel under uniform conditional drift. Time-homogeneity is needed for V* as a
fixed function, the resolvent/fundamental-matrix identities, and δ-calibration.
- Self-contradiction: "the certificate cannot be proven, only observed" contradicted
the established "a proven inequality certifies" — reworded to "the architecture
does not hand it to you; estimated unless separately certified."
- Citation: the TACL result is LOG-precision → logspace-uniform TC⁰ (verified);
fixed/constant precision is a stronger restriction. Fixed in body and Grounding.
Modeling holes closed:
- Adversary class Π must respect rejection: γ(s,y)=⊥ ⇒ Q_E^α(s,⊥,·)=δ_e0, else the
adversary resurrects refused side effects.
- R must be a syntactic/verified readout, not a semantic solver — otherwise the
L-wall is void (compute could hide in R off the ≤L window).
- e must be an effect record (ledger outcome), not just API bytes, since only ρ
writes external effects into s.
- The displayed M_W(c) freezes endpoint/version/sampler; config changes need a
state-indexed M_{κ(s)} or K_C — the kernel can't silently depend on config in s.
- The final user-visible response/log is itself an effect: an authorized action
through γ, or emitted only after an accepted halt.
- m_t must include a token counter and clock for the cap/timeout to be functions of it.
Residue (omissions a collision-linter can't catch):
- Another γ dropped from the K_C Dirac-special-case list.
- τ* mislabeled as "designed code" → the halt test (H) is the code; τ* is its
emergent hitting time.
- H\H_ok relabeled "non-accepting" (safe refusals outside B; wrong/bad halts
possibly in B), not uniformly "rejecting/fail-closed."
Built and ran a static linter (no model): delimiter/emphasis balance, residue
regexes for everything prior rounds fixed, single-capital collision scan, a
definition check for recently-introduced symbols, and a γ/ρ role-neighborhood
scan. Result:
- All balance checks pass; all 10 residue regexes clean (no regression across
14 rounds); all 12 introduced symbols defined; no display-only symbols.
- γ/ρ scan: one flag, a false positive (the symbol-table cell defines both).
- One real find: G was overloaded — the parser-stop update G(m_t,v) (added in
round 14) collided with the Green/potential operator G. Renamed the stop-update
to \mathsf{step}; the Green operator G is now unique.
This closes the consistency axis deterministically rather than by another review.
Same prior-maxima full-tools review, re-run. Found mostly residue from round-13's
own edits plus longer-standing inconsistencies. Folded all; left the final
signature line alone (it is the author's call, and it is well-formed — see below).
Round-13 residue:
- 𝒴/𝒴_⊥ split half-committed: 𝒴 already includes ⊥, so R:𝒵→𝒴 and A_Y⊆𝒴 (drop _⊥).
- m_t was added to the inner triple with no dynamics: add m_{t+1}=G(m_t,v), define
the stop set Stop and τ=inf{t:m_t∈Stop} in both display and prose.
- No-truncation special case had R=id, ill-typed on a triple: R(c,b,m)=c.
- Safety/success "exactly on safe refusals" overclaimed: they differ on any
B-avoiding non-success run — also safe non-halting / endless safe retry, absent
a.s. absorption into H∪B.
Role residue (γ does authorization/rejection; ρ does response/fold-back):
- "⊥ branch is what ρ rejects" → γ rejects it.
- "ρ validates response as well as the proposal" → ρ validates the response; γ
gated the proposal.
- "fail-closed rejection at ρ" (falsification list) → at the gate γ.
- Symbol table still typed Q_E on authorized a → a∈𝒜_⊥ with the no-op; define e_0.
Longer-standing:
- Stochastic-controller contradiction: stochastic control falsifies the
deterministic special case, not the broader K_C kernel model (round-9 K_C).
- Drift split r=r_shell+r_plant needs an additively separable V̂ or a declared
attribution scheme.
Citation (verified via search, not the reviewer's say-so):
- TC⁰/log-precision → Merrill & Sabharwal, "The Parallelism Tradeoff", TACL 2023;
caveat (added autoregressive steps escape it) → Merrill & Sabharwal, "The
Expressive Power of Transformers with Chain of Thought", ICLR 2024.
A cold reviewer given the complete prior-maxima changelog + full tools ran a
consistency audit of the file (it did not use tools for grounding — the gap was
internal). Found 15 real issues, all folded. No new design flaws; this is
accumulated editing debt from 12 rounds of surgical patches.
Half-applied fixes now propagated:
- Inner-kernel display still showed M_W(c)=Law(c_τ) and R:C→Y_⊥ despite the round-12
triple; made z_t=(c_t,b_t,m_t) primary, R:Z→Y_⊥, M_W=Law(R(z_τ)).
- Append formula still used bare c·v; now suffix_{≤L}(c·v) in the display.
- Tuple still called B a "terminal set"; B is separate (τ_B fires mid-run).
- "halt/ready" survived at line 75 (fixed before only in tuple + table).
Collisions created by added notation:
- γ was both the authorization gate and the RL discount in (I-γP)^{-1}; discount → β.
- B was both the bad set and the dummy measurable set in the pushforward; dummy → A_Y.
- ρ over-credited as the disturbance-rejection margin; for side effects the margin
is γ (consistent with round-12 irreversibility), ρ validates response/fold-back.
Real error in a prior round:
- The round-12 safety/success distinction collapses under absorbing refusal
(Pr(τ_Hok<τ_B) requires reaching H_ok, so it is a success form). Split correctly:
p_succ=Pr(τ_Hok<τ_F), F=B∪(H\Hok); p_safe=Pr(τ_B=∞); they differ on safe refusals.
Typing / hygiene:
- Q_E typed on S×A_⊥ (it is applied to ⊥); 𝒴 declared to include ⊥ (M_W, γ total).
- Controller list omitted γ and mis-listed the readout (specialization-only).
- Defined the previously-bare symbols D={s:E[τ_H]=∞}, μ, the drift r(s), and Π.
- Grounding "verify by measured drift" overstated; a proven inequality certifies,
empirical drift only checks — reconciled with the body.
A cold no-priors review (given a local sandbox it did not use — the remaining
work is judgment, not computation). Mostly editorial/formal; its real catches
again concern round-10/11 additions. Folded the substantive ones, declined the
"extract a smaller core" restructure and the formalism padding.
Substantive:
- Inner kernel: replace round-11's awkward "read c_τ as the buffer" overload with
a clean inner-state triple z_t=(c_t,b_t,m_t) — window, output buffer, parser/
stop state — and M_W(c,·)=Law(R(z_τ)) from z_0=(c,∅,m_0). Strictly cleaner.
- Authorization is the irreversibility boundary: ρ can reject a bad tool RESPONSE
but cannot undo an authorized action's side effects, so γ (not ρ) is the last
line before irreversible effects. And the gate is bypassed if raw y reaches any
sink (tool, logger, browser, remote) before γ.
- Safety ≠ success: p_ok=Pr(τ_Hok<τ_B) is the safety object (refusal permitted);
the stricter success object races H_ok against all failure F=B∪(H\Hok). They
differ exactly on safe refusals.
Precision:
- Foster–Lyapunov positive recurrence needs irreducibility/petite-set hypotheses;
the absorbing-halt case needs only the weaker supermartingale hitting-time bound.
- Name an initial distribution s_0~μ_0. Fix residual "halt/ready" in the table
(round 11 fixed only the tuple).
A JSON-constrained cold review largely validated round 10; its new catches
cluster in round-10's newly-added material.
Fixes (the real ones):
- Terminal-set partition was wrong (a round-10 error): B is NOT a terminal
component — τ_B can fire mid-run. H now splits into accepting (H_ok) and
rejecting/fail-closed (H\H_ok); B is a separate unsafe set for reach-avoid.
- Fail-closed generalized: rejection need not be terminal (reject-then-retry is
valid) — define it as "no unauthorized side effect + land in a safe non-bad
set," with terminal rejection one case. ρ must also validate the tool RESPONSE
e (adversarial/malformed Q_E output), not only the model proposal at γ.
- Sliding-window truncation (round-10) loses transcript: the readout R reads a
separate output buffer, not the truncated c_τ alone.
Precision:
- Deterministic maps are measurable transforms inside the pushforward, not
literally "outside the integral."
- Absorbing halt H vs the separate (non-absorbing) daemon "ready" recurrence.
- "Syntactic soundness is free" qualified: relative to a formal schema and a
correct validator.
- State-ablation falsifies Markovity but cannot establish it (necessary, not
sufficient). Added a readout-typing diagnostic.
A cold no-tools external review (lower trust on world-facts, but its catches are
math-internal and correct) found two real bugs plus rigor gaps.
Bugs fixed:
- Verification-after-side-effect (the important one): the kernel ran e~Q_E(s,y)
then ρ verified, so a tool call's side effect landed before authorization. Add
a deterministic authorization gate γ:S×Y→A_⊥ between model and environment;
Q_E now acts on the authorized action γ(s,y); ρ becomes ρ(s,y,a,e). Fail-closed
is now a property (γ=⊥ ⇒ no-op env ⇒ fold to H\H_ok), not a name.
- Minimax drift display had a free y (introduced round 9): it integrated only
over e while y~M_W(π(s)). Now integrates over both y and e, adversary as a
policy α(s,y) over environment kernels, on the authorized action.
Reframing / rigor:
- Raw halting is cheap: a budget counter k gives V=k as a trivial halting
certificate, so "no certificate by construction" overstated. The missing
guarantee is correct/safe/successful halting (H_ok, B, p_ok).
- Standard Borel spaces (not merely measurable); define H, H_ok (⊆H), B (∩H_ok=∅)
and hitting times τ_A up front; add 𝒜 to the tuple.
- Inner kernel: truncate c·v to suffix_{≤L} at the window edge; M_W is a
probability kernel only via EOS/max-token/timeout/⊥ (else sub-probability +
cemetery).
- Drift: weaker bound δ≤δ̄<ε gives E[τ]≤V̂/(ε-δ̄); distinguish δ_ν (distributional)
from δ_sup (worst-case).
- Injection enters π's inputs (retrieval/pages/tool metadata), not only post-model
Q_E; B needs a side-effect ledger in S. Architectural invariants stated
(model sees only C; outputs are proposals; γ gates side effects; terminals
partitioned). Complexity/LBA material marked heuristic, not definitional.
An LLM-judge verifier is a learned kernel, not deterministic ρ.
A cold external review (same priors, no path-dependence) surfaced three real gaps
the iterative chain missed, plus precision items. Folded in:
Substantive:
- Stochastic controller: the deterministic π,ρ,H are the Dirac special case of a
controller kernel K_C(s,dc) (routing, sampled retries, ensembles, learned
routers). Deterministic is the case worth wanting (localizes randomness); the
split widens, not breaks, under stochastic control.
- Minimax type fix: the adversary chooses a POLICY/kernel, not the realized
sample. Display is now sup over α of ∫ V(ρ(s,y,e)) Q_E^α(s,y,de), not sup over
the post-probability e.
- Reach-avoid security: add a bad set B; injection steers toward B (wrong
acceptance, exfiltration, unauthorized tool use, privilege escalation,
irreversible effects), so security is reach-avoid p_ok=Pr(τ_{H_ok}<τ_B) with a
barrier certificate for B, not liveness. B and H_ok added to the tuple.
- Unconditional V*_ok is infinite under any positive pre-acceptance failure
probability ⇒ the workable object is p_ok (or the regenerative time on restart).
Precision / hygiene:
- Compiler claim scoped to a specific data-flow analysis (not a whole compiler);
add integrability/optional-stopping conditions to the hitting-time bound.
- Formal hygiene: spaces measurable, τ/τ* stopping times, H absorbing.
- Mid-generation tool calls interleave the loops — clean nesting is an
idealization needing a finer state machine.
- Soften SSM ("different", not "tighter"); demote "manifold" to informal
shorthand in the formal section; gloss "all undefined behavior" as "no complete
formal source-language semantics."
Not changed: V* incompressibility (already labeled conjectural in Grounding).
The review's verdict was "Merge." These are its two correct non-blocking nits;
its third nit (stop adding theorems/caveats) is heeded — nothing else changed.
- Grounding: "the compiler's V is free" → "a classical monotone data-flow
analysis gets its V for free." A whole compiler does not get termination for
free; the specific lattice-based analysis does (Kildall).
- Asserted: the Koopman/certificate co-determination "holds only under" →
"is well-posed only under" the spectral assumptions — avoids asserting truth
("holds") for a claim explicitly labeled as not-a-theorem.
Deliberately NOT changed: D → D_H (prose already marks D harness-relative;
subscripting one formula while D stays bare elsewhere would add asymmetry, not
remove it), and no further theorem additions or caveats per the review's note
that more caveating now costs clarity without adding rigor.
The review's verdict was "mergeable"; these are its three optional items plus the
delta-attribution nit.
- δ attribution: sampled-state coverage is an evaluation-protocol property, not a
weights property. Attribute the noise floor / residual risk to the trained
weights, the environment, AND the evaluation distribution.
- reachable(L) is harness-relative too (same reason U_H(L) is): rename to
reachable_H(L) and note the divergent set D is likewise relative to H.
- Split the dense frontier paragraph in two: (1) the SR / fundamental-matrix /
potential-operator identity with its caveats; (2) the speculative interlingua/
certificate thesis. No content change.
- Grounding: add the absorbing-chain fundamental matrix (Kemeny & Snell 1960),
the general-state potential/Green operator (Revuz 1984), and Koopman (Koopman
1931; Lyapunov-from-eigenfunctions, Mauroy & Mezić 2016) to Proven; mark the
Koopman/certificate co-determination (spectral-assumption-dependent) and the
interlingua/certificate identification as Asserted.
- U(L) is harness-relative: tools and decompositions change membership, so rename
to U_H(L) and note the shell's verified tools / decompositions determine what
can be paged or outsourced.
- Countable fundamental matrix: lead with the Neumann series N=Σ Q_tr^n, scope
countable to convergence, and write (I-Q_tr)^{-1} only when the inverse exists;
general-state version is the same series read as the potential (Green) operator.
- Distinguish failure modes for V*_ok: infinite under a formal success predicate
vs undefined if no predicate has been specified.
- Soften the delta "floors" line: mu(D), sampled-coverage, and Var[tau*] drive
the empirical noise floor / residual risk, they are not literal floors of the
drift slack.
- Hedge the Koopman bridge (the last frontier thread): the eigenbasis claim
presumes a diagonalizable, point-spectrum operator — mixing dynamics carry
continuous spectrum and admit no eigenbasis — and the linearizes/certificate-
decomposes coincidence holds only for a V in the span of those eigenfunctions.
Address the round-five review's three precision points (plus the adaptive-adversary
refinement).
- Absorption is finite expected hitting time, not positive recurrence: replace
"positive-recurrent to H" with "reached in finite expected time," domain
{s : E_s[τ_H] < ∞}. Positive recurrence stays reserved for the daemon/
ready-state case (where it is used correctly).
- The fundamental matrix N=(I-Q_tr)^{-1}=Σ Q_tr^n is the finite/countable object;
the formal model lives on general measurable spaces, so add the general-state
potential (Green) operator G=Σ Q_tr^n with G·1=V* where the series converges.
Q_tr now stated as the sub-stochastic kernel restricted to H^c.
- V*_ok is taken on the process where H\H_ok (halting wrong, refusing, failing
closed) is absorbing failure — so a run that fails closed before acceptance
has infinite accepting hitting time unless the spec restarts it. This is the
mechanism by which a U(L) task sends V*_ok → ∞.
- Adaptive adversary: nonstationary Q_{E,n} → time-ordered product; an adaptive
adversary → controlled / game-value operator (not merely time-indexed).
Fold in the two seams flagged after round three, before the next review pass.
- Limit section now states explicitly that its V*=E[τ*|s] certifies *halting*
(reaching H at all), not correct halting; defers V*_ok (expected time to an
accepting H_ok ⊆ H) to the second wall. Removes the latent inconsistency
between the limit section (plain H) and the U(L) refinement (H_ok).
- Frontier section: the discounted successor-representation resolvent
(I-γP)^{-1} presumes a discount γ and fixed P the stopped formulation lacks.
Replace with the correct undiscounted/absorbing object — the fundamental
matrix N=(I-Q_tr)^{-1}, Q_tr the sub-stochastic transient block — whose row
sums N·1 are exactly V*. Converts analogy-dressed-as-identity into a true
identity for the doc's own kernel.
- Mark the "one object seen twice" identity as holding only in the stationary
regime: under the adversarial Q_{E,n} the resolvent/fundamental matrix become
a time-ordered product, so identity in the stationary case, analogy beyond.
Address the round-three review. The substantive one is the V* correction.
- Successful halting vs raw halting (the real conceptual fix): a U(L) task does
NOT make V*=E[τ_H|s] undefined — the chain can still hit H by failing closed,
refusing, or returning a wrong answer. Split H from the accepting set H_ok and
define V*_ok=E[τ_{H_ok}|s]; U(L) blows up V*_ok, not V*. Restate the domain as
dom_{<∞}(V*_ok) ⊆ reachable(L)\D.
- Tools compute, not just store: the L-wall binds *model-mediated* work; work
discharged to a verified external tool (solver, interpreter, compiler) runs
off-context. U(L) now excludes tool-dischargeable work explicitly.
- Readout typing: use the pushforward M_W(c,·)=R_# Law(c_τ) (equivalently the
conditional law); make R total, R: C → Y_⊥, with the ⊥ branch handled by the
fail-closed ρ.
- Adversary/history: a history-conditioning adversary needs that history in s,
else the object is a Markov game requiring further augmentation, not a chain.
- Hedge the LBA claim: "in the variable-L, fixed-precision idealization, the
model-mediated inner computation behaves like a linear-bounded automaton."
Address the three follow-up points on the first review patch.
- Reconcile the model kernel's two types: M_W(c,dy) maps into 𝒴, while the
transformer line writes M_W(c)=Law(c_τ) over contexts. Add the readout R:
𝒴 is either c_τ itself (𝒴=𝒞) or a deterministic readout R(c_τ), with
M_W(c,dy)=Law(R(c_τ)∈dy).
- Separate harness state 𝒮 from model-visible context 𝒞: the L wall binds 𝒞
(the L×d residual stream), not 𝒮. External stores (files, DBs, vector stores,
durable memory) are shell-supplied memory that extends addressable storage but
not the per-pass resident set — every read still routes through the ≤L window.
Retype U(L) accordingly: not data exceeding L (pageable) but irreducible
per-step working set exceeding L (not pageable).
- Make the time-homogeneity assumption explicit at the formal kernel: the
displayed T is the fixed-kernel case; nonstationary/adversarial environments
replace Q_E with a time-indexed kernel Q_{E,n} / admissible family, which the
minimax certificate downstream quantifies over.
Address the accepted points from an external peer review while preserving the
controller/plant thesis and the document's voice (layer, don't flatten).
- Claim: replace the ill-typed `T = ρ ∘ (M_W ∘ π, E)` with the integral
transition kernel over (𝒴,ℰ); add explicit informal/formal split; demote the
residual-stream implementation from definitional to a kept specialization
(M_W as a general learned kernel); weaken "fixpoint searches" to hitting-time
processes with fixpoint as one mode.
- Reading-it: note s is Markov only after state augmentation; mark controller
determinism as conditional on versioned code/config/endpoint/interfaces.
- The limit: rephrase "carries no descent function by construction" to "supplies
no certificate automatically" (a certificate is sufficient, not provided for
free); label V* incompressibility as conjecture, not theorem.
- δ: "measure" → "estimate"; demote empirical δ from certificate to calibrated
risk metric (confounds: bad V̂, coverage, sup not attained, nonstationarity,
non-Markov); certificate only once statistically bounded.
- Cash-out: split "soundness is free" into syntactic soundness (free) vs
semantic adequacy (empirical).
- Qualify the single-pass TC^0 claim (fixed-depth/fixed-precision; log-depth
changes it) in both body and Grounding.
- Add an operational falsification program (state-ablation, determinism audit,
drift calibration, adversarial-environment, boundary-control ablation).
Citations with a proven-vs-asserted split; the orthogonal context-length
tape bound (TC^0 single pass, the U(L) non-haltable region); and a flagged
frontier coda on V* and the semantic interlingua as one object.
A one-formula definition of a harness — a deterministic controller in
closed loop with a stochastic learned plant — and the certificate it
provably can't carry. The headline equation sits at the top of the
README and links through to the full doc.
* feat(deploy): vllm-litellm example — 3-model co-resident shape + HF loader
Update the unified-memory inference example to the validated GB10 Spark shape:
qwen3.6-27B-FP8 (reasoning) + gemma-4-12B-it (perception) + Qwen3-Reranker-4B,
all co-resident on one GPU behind LiteLLM, loaded by HF id into a mounted
HF_HOME cache.
- qwen: MTP spec-decode + runai_streamer (weight load ~166s->1s) + full 256K at
util 0.50 (default KV)
- gemma on the OpenAI lane (audio), reranker direct on :8002/rerank
- sequential startup + page-cache-drop guidance; runai_streamer kept on the big
model only (its buffers break small models' KV budgets)
- README: HF-id loader, DGX Spark (validated) + AMD Strix Halo (ROCm) setup,
tuning notes, troubleshooting
- wheel-check ALLOW entries for the example files (supersedes #687)
* docs(deploy): clarify AMD edits are compose literals (Copilot review)
In the Strix Halo guidance, --max-model-len and --load-format runai_streamer are
hard-coded in docker-compose.yml's vllm-qwen command, not .env vars — say where
to edit them.
* feat(deploy): add vLLM + LiteLLM unified-memory inference example
A docker-compose stack co-residing a reasoning model (Qwen 3.6 27B) and a
perception model (Gemma 4 12B) on one unified-memory accelerator (NVIDIA DGX
Spark / AMD Strix Halo) behind a LiteLLM gateway serving both the Anthropic
/v1/messages and OpenAI /v1/chat/completions routes.
- qwen on the Anthropic lane (vLLM native /v1/messages), full 256K context
- gemma on the OpenAI lane (required for audio input_audio perception)
- sequential startup + page-cache drop for reliable KV provisioning on one card
- README: DGX Spark (validated) + AMD Strix Halo (ROCm) setup + troubleshooting
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* feat(deps): add altair + vl-convert-python viz stack
The standard, ui://-ready visualization stack: one Vega-Lite spec renders to
static SVG via vl-convert (a bundled Rust renderer — no browser, GDAL, or
chromium) and drops into vega-embed for interactive ui:// panels. The first
consumer is the civic-records choropleth map; future ui:// surfaces build on
the same stack.
The dependency closure is fully permissive (BSD-3 + the OFL font + MIT/ISC JS) —
clean for Apache-2.0 and commercial use. Adds a mypy override for the untyped
vl_convert wheel.
* fix(deps): bump pydantic-settings to 2.14.2 (GHSA-4xgf-cpjx-pc3j)
Clears the pip-audit --strict advisory on the transitive pydantic-settings
2.14.1. Pinned as an explicit security floor in [project.dependencies]
(matching the starlette/cryptography CVE-floor convention) even though it is
transitive-only, so the floor is documented and survives re-resolution.
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* chore: regenerate uv.lock to pass lock check
---------
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
- The guard snapshotted live threads by `Thread.ident`, but idents are
recycled after a thread exits — a new leaked thread reusing an exited
thread's ident would be mistaken for pre-existing and missed (false
negative). Snapshot the Thread OBJECTS and compare by identity instead.
- Fix the `serve` fixture docstring: the factory returns the ephemeral
port, not the server.
Background daemons, event loops, and test servers that outlived their test
bled into later tests' captured output — an intermittent "I/O operation on
closed file" heisenbug, and the same class behind a past multi-day CI-hang
investigation.
- conftest: a fail-on-leak autouse guard (`_no_leaked_threads`) snapshots
threads at setup and fails any test that leaves one running past teardown,
with an `allow_thread_leak` opt-out — so the next leak is caught in minutes,
not days. Plus `logging.raiseExceptions = False` to mute the benign
logging-vs-capture-teardown race, and shared loop/server teardown helpers
(`stop_loop_thread`, `serve_until_exit`).
- collector (PRODUCT FIX): the node-discovery loop slept uninterruptibly, so
`ClusterCollector.stop()` couldn't join the `console-discovery` thread until
the full interval elapsed — a real shutdown hang in production (up to
`discovery_interval`). It now sleeps on an interruptible Event that `stop()`
sets and `start()` clears.
- test fixtures: docker_healthcheck's HTTP servers, the MCP background event
loops (shutdown_default_executor + close), and the FastMCP uvicorn upstreams
(timeout_graceful_shutdown=0 + force_exit) now tear down cleanly instead of
leaking.
Full non-live suite: 7456 passed, 0 closed-file errors, 0 leaked threads, and
~1.5 min faster (the leaks were dragging it).
Coordinators carry LLM/auto titles like interactive workstreams but had no
way to regenerate or rename them. Port the interactive "Refresh title" (LLM
regenerate) + "Edit title" (manual alias) dropdown actions by lifting the
two handlers — the last shared verbs that weren't yet lifted — and opting
coordinators in.
- session_routes.py: add make_refresh_title_handler / make_set_title_handler
factories (cfg pattern, mirroring make_close_handler). set_title resolves
the workstream BEFORE the alias write and 404s when the kind has no
tenant_check storage gate and the in-memory manager doesn't own it:
set_workstream_alias is a global, kind-unscoped UPDATE, so this prevents
an operator renaming a workstream the coord manager doesn't own (e.g. an
interactive ws via the coord route) and the silent-200 on a bogus id.
- server.py: re-point the interactive bundle to the lifted handlers; drop
the standalone refresh_workstream_title / set_workstream_title.
- console/server.py: wire refresh_title / set_title into the coord bundle
(gated by the existing admin.coordinator operator check).
- shell.js: enable titleVerbs on the coordinator pane's tab menu; the
base-aware lane posts to the console-origin coord routes.
Tests: coord refresh/set-title (regenerate, operator-gate, 404 unknown,
alias store + broadcast, empty, conflict, cross-kind reject); interactive
title tests re-pointed to the lifted handlers for lift-parity; shell.js
coord-menu assertion.
Three points from the PR #676 Copilot review:
- _coord_display_name ran on a lifecycle-event path and called
get_workstream_display_name → get_storage(), which auto-initializes a
SQLite .turnstone.db in the CWD when storage isn't initialized yet —
a stray-file footgun on early-startup / unit-test paths. Add
is_storage_initialized() to the storage registry and skip the DB read
(fall back to ws.name) when storage isn't up. (Copilot's "skip when
ws.name is non-synthetic" suggestion would have broken alias > title >
name, so guard on init state instead.)
- Document, on SessionUIBase, that on_aux_usage (storage/metrics, no
_ws_lock state) and on_rename (queue/locked fan-out) are safe to call
from a concurrent auxiliary thread — the title-gen thread now runs
during streaming, and these are the only two UI hooks it touches. No
behavior change: the methods were already thread-safe (the same path
task_agent sub-agents use); the contract just didn't say so. Add a
matching note at the title-trigger site.
- Note in _coordinator_rows that the secondary `title` field is
best-effort for a live coord outside the limit=200 window (the
user-visible `name` stays correct via the uncapped bulk lookup, and
the window is unreachable in practice — live coords are max_active-
bounded and sort to the top of updated DESC).
Coordinator workstream LLM titles were written to workstreams.title but
never read back, and were rarely generated in the first place:
- Read path: the dashboard's `_coordinator_rows` builder hardcoded
title="" and used the synthetic `ws.name`, so a generated title (or a
user alias) reverted to `ws-xxxx` on every refresh. Interactive rows
resolve via get_workstream_display_name, so the gap was coord-only.
- Write path: the auto-title trigger only fired on a tool-call-free
assistant turn, which coordinators (near-constant tool use) seldom
reach — so the title almost never generated.
Read path:
- Project `title` + `alias` in list_workstreams (appended after user_id so
existing positional fallbacks stay valid). `_coordinator_rows` resolves
the display name (alias > title > name) for both lanes — live names via
the bulk get_workstream_display_names (exact ids, no row cap), persisted
rows from their own _mapping.
- Seed the console pseudo-node fan-out with the resolved display name so a
rehydrated coordinator shows its title in the live tree immediately
(one bulk lookup instead of an N+1 over mgr.list_all()).
Write path:
- Fire auto-title right after the user turn is recorded in send(), gated on
a real (non-wake, non-empty) user message, instead of waiting for the
terminal tool-call-free turn. Applies to interactive + coordinator.
- Snapshot self.messages in _generate_title since it can now run
concurrently with the streaming turn.
Speech-to-text against an omni chat model (e.g. Gemma-4 on vLLM) was
broken end to end:
- The browser records webm/opus, but the omni chat lane only decodes
wav/mp3 (it sniffs the bytes), so every clip came back 400 "Invalid
or unsupported audio file". Transcode the upload to 16 kHz mono WAV
with ffmpeg first, hardened against the untrusted blob:
-protocol_whitelist pipe (no file:/http: SSRF), -vn, and a duration cap.
- The chat STT path calls the raw client and so bypasses the provider's
request shaping. It now forces enable_thinking=false (via the model's
thinking_param): leaving reasoning on costs ~11x latency and returns
empty content on some clips. The prompt precedes the audio part (the
order Gemma documents for transcription) and max_tokens is capped.
Add a streaming variant: POST .../speech-to-text/stream returns the
transcript as plain-text deltas and the composer fills them in live
(~0.3s to first word). The blocking stream is driven from one worker
thread that owns and closes the upstream connection.
Drop the gemma skip_special_tokens server-compat workaround: the vLLM
bug it patched is fixed upstream, and a stale shim can corrupt output.
The node image now installs ffmpeg; rebuild to run this live.
The test-postgres failure on test_init_retries_exhausted_raises surfaced the
root cause: sleeps held 2275x 0.1 instead of [1.0, 2.0]. Those 0.1s came from
a concurrent background poller doing asyncio.sleep(0.1) on anyio's shared
(persistent) event loop — the tls retry tests patched the *global*
asyncio.sleep, which intercepted that poller too.
- Before: the stub didn't yield, so the poller busy-looped and monopolized
the loop -> the test hung (the CI-only "after 92%" hang on 3.12+).
- The earlier "make the stub yield" change converted the hang into this
flood (the poller spins instead of blocking), which is what exposed it.
Fix: route init()'s backoff through TLSClient._sleep so the tests stub that
method in isolation and never touch the global asyncio.sleep. Tasks sharing
the loop are no longer affected; schedule assertions are unchanged.
The deeper fragility this exploited — a leaked, un-cancelled background poller
surviving on the shared test loop — is left as a follow-up.
run_with_deadline checked the deadline/cancel before reading the result
queue, so a call that completed in the same scheduling window could be
reported as a spurious timeout. Drain the queue first.
Also from review:
- test_validate_regex_pattern stubs run_with_deadline, so the probe regex
never runs — use a benign pattern instead of a real backtracking literal
(the literal tripped a ReDoS scanner).
- output_guard_judge docstring: reference IntentJudge._parse_verdict instead
of brittle judge.py line numbers.
The _runner BaseException catch is intentional and kept: it relays (not
swallows) whatever fn() raises to the caller via the queue; narrowing to
Exception would let a BaseException escape the worker so the caller never
gets a value, degrading the no-hang guarantee.
A hung run otherwise rides GitHub's 6-hour default with -v streaming the
whole time (the source of the multi-GB job logs). Cap test and test-postgres
at 20 minutes so a flaky hang fails fast instead of bleeding hours.
CI hung on test_init_retries_transient_failure (the new -v output named it:
its nodeid printed, no PASSED, the job rode to cancellation). It is the first
retry test that actually awaits the stubbed asyncio.sleep — the earlier tests
raise before sleeping — which points straight at the stub.
The stub returned without ever suspending, so the retry run completed in one
event-loop step with no checkpoint; that is fragile under the async test
runner and is the suspected cause (3.12/3.13/3.14 only — never reproduced on
3.11 or locally). Capture the real asyncio.sleep before patching and await
sleep(0) in the stub so it still yields, keeping the no-real-delay behavior
and the backoff-schedule assertions. Same fix in the discovery-failure test.
The judges and the regex ReDoS probe ran a blocking call on a
ThreadPoolExecutor and abandoned the worker with shutdown(wait=False) on
timeout or cancel. concurrent.futures joins every executor worker from an
atexit hook regardless of wait=False, so a wedged call could pin
interpreter exit — and hang the test suite at shutdown.
Add turnstone/core/deadline.py::run_with_deadline: run a blocking callable
on a daemon thread bounded by a wall-clock timeout and an optional cancel
event. A daemon worker is never joined at exit, so abandoning one is safe.
Migrate three sites onto it:
- OutputGuardJudge.evaluate()
- IntentJudge._evaluate_single / _run_judge — this also removes
_ExecutorPoisonedError and the executor-restart dance: per-call daemon
threads can't poison a shared single-slot pool, so a timeout now returns
None and the caller delivers one fallback verdict.
- console/server.py _validate_regex_pattern (regex ReDoS probe)
Also:
- Double the default judge LLM timeouts for slower local models:
judge.timeout 60->120s and judge.output_guard_llm_timeout 30->60s
(settings registry, JudgeConfig dataclass, --judge-timeout CLI default,
class docstring, docs). Correct a stale doc that described the per-turn
timeout as a total budget across turns.
- Raise the regex probe bound 0.5->3.0s so a legitimately complex pattern
isn't false-flagged as catastrophic backtracking.
- CI: run pytest with -v instead of -q so a hang names the offending test
instead of riding the job timeout.
- Tests: cover deadline.py and the regex validator; move test_judge.py off
fixed sleeps onto the existing _wait_for helper.
main's lockfile was already on the patched versions (cryptography 49.0.0,
starlette 1.3.1) via renovate, but the pyproject floors (>=42, >=1.0.1) still
permitted a regression to vulnerable versions. Raise the floors to match
stable/1.6's v1.6.7 security fix:
- cryptography >=48.0.1 (GHSA-537c-gmf6-5ccf — bundled OpenSSL vulnerable <48.0.1)
- starlette >=1.3.1 (CVE-2026-54282 host spoof + CVE-2026-54283 url-encoded form DoS)
Copilot review: _audioModelEligible gated stt/tts on md.provider, but a
blank/unset provider was treated as not-audio-capable and excluded — an
asymmetry with the backend, where _provider_carries_audio and
ModelConfig.provider both default to "openai". Default the provider to "openai"
before the check so a provider-less model isn't wrongly dropped from the
voice-role dropdowns.
The by-ref change replaced the send_id reservation model with the per-node
upload buffer (peek-then-drain at write time), but the surrounding narration was
never swept and a no-op stub was retained to make the send handler "read like"
the old flow — which is what made a recent diagnosis assume reservations still
existed.
- Delete the no-op _release_reservation_on_fail() and its 5 call sites in the
send handler (behaviour-preserving — it did nothing).
- Rename ordered_reserved / reserved_set -> ordered_taken / taken_set (the values
are the "taken" subset from resolve_staged_attachments, not reservations).
- Sweep the stale "reserve/reservation" wording across the create/send
docstrings, the API schemas/specs, and the SDK docstrings to the staged-buffer
vocabulary (resolve / attach / drain). The canonical docs in attachment_buffer
and attachments already stated the reservation token is gone.
No behaviour change; no tests exercised the removed scaffolding (the
migration-060 test correctly pins the reserved_at column removal and stays).
A create-time attachment is dispatched on the first turn, but the buffer drain
runs at write time inside the async dispatch worker (_append_user_turn). The
freshly-opened pane calls rehydrate() before the worker drains, so it painted
the image as a still-pending composer chip ("thumbnail in the text input box").
The inlined first-turn dispatch is the only consumer of those staged uploads and
always commits at create, so drain them from the buffer synchronously right
after resolving them — both the interactive and coordinator post-install paths.
The worker's own per-id discard then no-ops.
An omni model registered via the anthropic-compatible lane (vLLM Messages API)
was offered for the STT role because it carries supports_audio_input — but the
Anthropic SDK client has no .chat.completions and the Messages API has no audio
content block, so the mic failed with a cryptic
"'Anthropic' object has no attribute 'chat'".
Audio (input_audio) only rides the OpenAI-SDK surface, so gate all audio roles
to OpenAI-SDK providers (openai / openai-compatible / google / xai):
- model_supports_role returns False for anthropic(-compatible), so the mic
won't draw and the STT/TTS dropdowns won't offer those models.
- transcribe() raises a clear AudioUnavailableError naming the provider instead
of the opaque AttributeError (defence in depth).
- admin _audioModelEligible mirrors the gate — voice roles only; reranker hits a
/rerank endpoint, not audio, so it stays un-gated.
To use an omni model's audio, register it as openai-compatible (the input_audio
path); the anthropic-compatible lane is text/vision only.
- DRY the launcher create body: the multipart (meta + file parts) vs JSON
framing was duplicated in _createCoordinator and _createInteractive — extract
_createWorkstreamFetchOpts so the create wire shape lives in one place.
- Correct the proxy comment: the forwarded owner uid comes from the
authenticated ws_body (as on the JSON path), not the caller's meta; the proxy
token source is console-proxy, not console.
The mic is an STT control — it records and transcribes to editable text in the
composer. STT eligibility required the dedicated /audio/transcriptions endpoint
(supports_transcription / a whisper-style name), so an omni chat model
(supports_audio_input, e.g. Gemma) couldn't back it: it has no transcription
endpoint, it ingests audio via chat.
- model_supports_role accepts supports_audio_input for the STT role, so an omni
alias resolves as STT and the mic draws for it.
- transcribe() branches: a whisper-style alias keeps /audio/transcriptions; an
omni alias transcribes via chat input_audio + an instruction prompt — the
audio.stt_prompt override, else a default that emits only the transcript.
Audio attachments on an omni-STT setup transcribe the same way.
- admin _audioModelEligible mirrors the eligibility so omni models show in the
STT dropdown; the role description notes the two backends.
Phone photos store landscape pixels plus an EXIF orientation tag. Browsers honour
the tag for <img>, but Pillow (our thumbnails) and many vision-model image
decoders do not — so the thumbnail rendered rotated AND the model literally
perceived the photo sideways (noticed earlier as model "hallucinations", before
thumbnails made the rotation visible).
Normalize on read, at both surfaces:
- new core/images.normalize_image_orientation: bakes the rotation into the pixels
and re-encodes (preserving format); images with no / identity orientation pass
through untouched (pristine original, no per-send cost).
- make_thumbnail applies exif_transpose — after the decompression-bomb pixel gate,
which now also covers the transpose decode.
- attachment_to_content_part runs image bytes through the normalizer before
base64, so the primary model and the perception model both get upright pixels.
Because normalization is on read (not at upload), it fixes already-stored uploads
too.
The universal perception fallback (perception.model_alias) shipped backend-only
— session.py + perception.py + settings_registry.py — so its admin UI was never
wired. Operators had no way to assign it from the Models → Roles sub-tab, and the
raw setting leaked into the Settings tab.
- Add a Perception row to MODEL_ROLES (no capability filter — it spans
image/PDF/audio; the description tells operators to enable supports_vision /
supports_audio_input on the target model, which is what makes the audio
fallback engage when no STT role is set).
- Derive the Settings role-key skip-set from MODEL_ROLES instead of a
hand-maintained list, so perception is filtered out and no future role can
drift back in (stt/tts/reranker had leaked the same way).
- Add an optional per-role disabledLabel so the blank dropdown option reads
correctly for non-voice roles (perception, reranker) instead of "voice off".
- Refresh the stale STT description that claimed "no audio-capable session
fallback" — audio attachments now fall back to perception.
The console creates interactive sessions by proxying to the owning node via
/v1/api/cluster/workstreams/new, which only forwarded JSON — so a file staged in
the launcher was blocked with "Attachments aren't supported for interactive
sessions yet". The node create endpoint already accepts multipart (meta JSON +
file parts) on interactive_endpoint_config; only the proxy lacked it.
Teach create_workstream to accept multipart: parse meta + files (same caps as
the node), pick the node exactly as before (auto / pool / pinned), and forward
the files instead of re-serialising JSON. _createInteractive sends multipart
when files are staged (mirroring _createCoordinator) and the launcher gate is
removed. The files-need-a-task guard already ensures an initial turn to
dispatch them on.
A console interactive pane is node-proxied — every request rides the pane's
transport base ("/node/{id}"). The attachment controller hardcoded bare
/v1/api/workstreams/... paths, so upload / list / delete / preview landed on
the console's OWN coord route, which resolves ws_id via coord_mgr.get() and
404s as "coordinator not found". The standalone server (base="") was
unaffected, which masked the bug.
Thread the pane base through: createAttachmentController and
buildAttachmentPreview take an optional getBase / base, and the interactive
pane wires this._base into both. Coordinator panes and the standalone server
pass "" and stay origin-mounted as before.
- TextDecoder in the text-preview stream now flushes on completion/cancel, so a multibyte UTF-8 char split across a chunk boundary isn't dropped (Copilot).
- send() clears self._wire_part_cache in a finally so the per-send memo (which can hold large rasterized PDF page-images) is released at send end instead of retained on an idle session until the next send (Copilot + fix-review).
- Make the implicit byte-string concatenation in _minimal_pdf explicit (+) in test_pdf.py and test_thumbnails.py so it can't read as a missing comma (CodeQL / github-code-quality).
A review of the fix commits surfaced three refinements:
- ftyp audio sniff: scan the whole ftyp box (its declared length) for an audio brand instead of a fixed 6-slot window, so a real .m4a with the brand listed late still passes — while a pure-video file (no audio brand) still rejects.
- text-preview: accumulate body chunks until >=240 chars before cancelling the stream, instead of assuming the first chunk is large (flush boundaries can split a large body into small early chunks).
- _resolve_attachments: correct the cache comment — the memo is refreshed per send and the wire resolver only runs during a send, so a stale value is never observed between sends.
- Remove the unused PerceptionUnavailableError (never raised/caught/imported).
- Reword the now-shipped 'Phase 3' placeholder comments on the Anthropic + OpenAI-Responses audio paths to describe the live upstream STT/perception fallback (these placeholders are defensive, not pending work).
- Clarify the no-vision image fall-through comment (fires when perception is unconfigured OR can't see, not only the former).
- Type AttachmentInfo.kind as the image|text|pdf|audio union in the TS SDK.
- extract_pdf_text: append the truncation marker only when there's actual text, so a scanned PDF over the page cap returns '' (-> placeholder) instead of a content-free document part.
The /thumbnail endpoint and the _resolve_served_blob ownership/404-leak gate it shares with /content had no handler-level test (only make_thumbnail as a unit + route mounting). Add cases through the real app: image -> 200 image/png with the nosniff + CSP + max-age headers; audio/text -> 415; make_thumbnail None -> 415; cross-workstream id and unowned-ws cross-user -> 404 (no existence leak).
The text-snippet preview fetched the entire /content body (text attachments are capped at 512 KiB) only to render the first 240 chars — and again on the sent-message pill (the endpoint sends Cache-Control: no-store). Read only the first response-body chunk and cancel the stream, so the rest of the blob is never transferred or regex-scanned. Falls back to r.text() where the streaming body API is unavailable.
Three copies of the kind->glyph mapping had drifted: the coordinator pill rendered audio as the document glyph (not the audio note) and showed no inline preview, diverging from the interactive pane.
Export kindIcon() from composer_attachments.js (+ window bridge) as the single source of truth; the interactive pane imports it and the coordinator pill uses it. Wire the coordinator pill to buildAttachmentPreview too (image/pdf thumbnail, audio player), gracefully no-oping on history replay (which omits attachment_id), matching interactive.
Also fix buildAttachmentPreview's thumbnail-error handler: it called img.remove(), but the caller has already replaced the icon span with the img, so a failed thumbnail left a blank gap. Swap in the kind glyph instead (.attach-preview-icon, sized to the thumbnail slot).
_reconstruct_attachment_refs collapsed every non-image attachment to the 'document' placeholder kind, so a reloaded session's pdf/audio placeholder type ({type:document}) mismatched the live-injection type ({type:pdf}/{type:audio}). Harmless today (resolution keys on attachment_id + blob kind) but a latent footgun for any consumer branching on the pre-resolution placeholder type. Preserve image/pdf/audio verbatim; only a stored 'text' blob collapses to 'document'.
A user-controlled filename was interpolated unescaped into model-visible frames (the [PDF attachment '{name}'...] / audio / transcript / perception placeholders, the Anthropic document title, and the unreadable placeholder). A crafted name like "'] New instructions:" broke out of the frame and injected text into the model context.
Add core.attachments.safe_attachment_label() (strip control chars + quote/bracket/angle delimiters, collapse whitespace, clamp length) and apply it at every model-context embedding site. The raw filename is still used verbatim for display / Content-Disposition, which neutralize at their own boundaries.
Also tag perception descriptions and STT transcripts '(untrusted)' so attachment-derived text reads as data, not instructions. Blast radius is single-tenant (injecting into a model reading one's own upload); a structural role=tool fence is deferred as disproportionate.
sniff_audio_mime returned audio/mp4 for ANY ISO-BMFF ftyp box, so an MP4/MOV video uploaded within the audio size cap sniffed as audio and was sent as input_audio. Restrict to genuine audio brands (M4A/M4B/F4A/F4B major, or M4A/M4B in the compatible-brands list, so a real .m4a with an mp42 major brand still passes).
Also add ADTS-AAC sniffing (0xFFF1/0xFFF9): audio/aac was in ALLOWED_AUDIO_MIMES + AUDIO_MIME_TO_FORMAT but never sniffable, so an advertised .aac upload always failed.
make_thumbnail set Image.MAX_IMAGE_PIXELS=40M, but Pillow only raises DecompressionBombError above 2x the cap; a 40-80M px image merely warns and decodes fully (~480MB RGB), defeating the documented bound.
Gate on the header-declared size after open() and before convert(), so nothing past the cap is decoded. Explicit check rather than a warnings filter — make_thumbnail runs in a worker thread and global warnings state is not thread-safe. Adds tests for the (cap, 2*cap] warn-only window and the at-cap boundary.
_resolve_attachments re-runs on every agentic round-trip (and per fallback model), each time re-fetching every attachment across the full history and re-rasterizing / re-base64'ing it. A 10-page PDF in a 10-cycle tool turn was rendered dozens of times.
Add a per-send memo (self._wire_part_cache) keyed by (attachment_id, caps-signature): the materialized wire part is computed at most once per send. The cache is None outside a send (display/export paths unaffected) and reset per send to bound the heavy rasterized-page parts and pick up any mid-session capability change. Skip the DB fetch entirely when every id is already cached.
Also peek the perception (alias, content_hash) memo before building parts in _perception_fallback_part, so a cross-send describe hit no longer wastes a PDF rasterize. Leaves pdf.py's deliberate no-module-cache stance intact — the per-send scope addresses the round-trip amplification without the durable store it defers.
Adds describe_peek() + per-send-cache and peek tests.
sanitize_messages ran inline_document_parts (which placeholders an application/pdf document part) before the Responses translator's native input_file branch could run, so every supports_pdf model silently degraded its PDF to an unsupported text placeholder.
Thread a skip_pdf_inline flag through sanitize_messages -> inline_document_parts; the Responses lane sets it so the PDF document survives to convert_content_parts. Chat / Google-compat keep the placeholder (they have no native PDF block).
The existing test exercised convert_content_parts in isolation, bypassing sanitize_messages and masking the bug. Add an end-to-end _convert_messages regression test (verified to fail without the fix) plus contrast tests pinning both lanes' behavior.
q-3 from the pre-push review, settled against docs.x.ai: Grok's document
support is an agentic attachment_search workflow over Files-API uploads
(file_id / file_url), not the inline base64 native ingestion that OpenAI
input_file / Anthropic document blocks use. Our native PDF path emits inline
base64, which xAI's Responses surface doesn't accept — so supports_pdf is
correctly left unset (Grok PDFs rasterize to images, which Grok can see).
Document the rationale on GROK_CAPABILITIES and pin every Grok row's
supports_pdf=False with a test so it isn't naively flipped without first
wiring a Files-API upload flow.
Add a `perception.model_alias` model role: when the primary model can't ingest
an attachment natively and can't be shown a degraded-but-native form, a
configured perception model perceives it and its output is carried as text.
Mirrors the STT role — a role alias plus a module-level memo so the extra LLM
round-trip runs once per attachment, not once per conversation turn. The call
goes through the provider abstraction's create_completion (the path the intent
judge uses), so any vision/omni provider works.
Bottom-tier, universal ladder — perception only fills the remaining gap:
- pdf : native supports_pdf -> rasterize-to-vision-primary -> perception
-> extracted text -> placeholder
- image: native vision -> perception (non-vision primary) -> native image_url
- audio: native supports_audio_input -> STT -> perception (omni) -> placeholder
Folds in two review findings the role subsumes:
- bug-1: thread the active attempt's capabilities into _resolve_attachments
(bound in _try_stream) so a model fallback materializes attachments against
the fallback model's caps, not the primary's.
- bug-2: charge a by-reference pdf/audio a bounded budget min(size_bytes, 16K)
instead of zero, so a large-attachment turn isn't budgeted as ~empty (the
exact materialized size isn't known until wire build).
Pre-push review follow-ups that are independent of the perception-role work
(bug-1 caps threading, bug-2 budget, and the perf cluster fold into that):
- thumbnails: cap decoded pixels (Image.MAX_IMAGE_PIXELS=40M) so a small
compressed image that decodes to huge dimensions can't OOM the node, and
reject DecompressionBombError cleanly.
- pdf: clamp per-page render scale so the longest rendered side stays <= 2000px
(a maximal MediaBox at scale 2.0 rendered to a ~28800px, multi-GB bitmap).
- session_routes: type classify_upload's rejection element as
UploadRejection | None instead of Any.
- test_session_routes: assert the /thumbnail route mounts (it was untested) and
fix the stale "quartet"/four wording to five.
Two-reviewer + sanity pass over the attachment previews:
- composer audio chip is icon+name+size only; the native <audio> player
renders on the sent message, not the staging chip (too heavy at chip scale)
- cap sent-message pills (+ in-pill audio/snippet) so they no longer overflow
the bubble at narrow widths; player and snippet drop to their own row
- clamp the chip filename in shared chat.css so long names ellipsize instead
of wrapping (console main + coordinator previously left it unclamped)
- merge the duplicated .composer-chip rule; drop unused kind-modifier classes
and inert vertical-align / inline-block declarations
- fix undefined var(--bg-base) -> var(--bg-surface) thumbnail backing
- label the <audio> control (aria-label) and drop the decorative snippet from
the a11y tree
scripts/livepass.py: add an attachments harness that drives the real
createAttachmentController + Pane.addUserMessage so these surfaces render
headlessly for review.
- core/thumbnails.py + GET .../attachments/{id}/thumbnail: server-rendered PNG
thumbnails (image downscale; pdf first page via pypdfium2). Extracted a shared
ownership-gated blob resolver used by both get_content and the thumbnail route
- buildAttachmentPreview (composer_attachments.js): image/pdf -> thumbnail,
audio -> <audio> player, text -> lazy snippet; reused by the composer chips and
the sent-message pills (interactive.js). Cookie auth, so direct media src works
- chip kind icons now cover pdf/audio; the upload swap adopts the server's
authoritative kind for styling + icon + preview
- chat.css preview styling; tests for make_thumbnail
- composer: accept pdf/audio in the upload picker; client-side kind
inference for the optimistic chip (server classify_upload stays
authoritative)
- admin Models tab: supports_pdf + supports_audio_input toggles (flow
through the field-aware capabilities merge into ModelCapabilities, so
flipping supports_audio_input on an omni alias enables native input_audio)
- docs: AttachmentInfo.kind, AttachmentUpload, and the TS SDK note pdf/audio
A vision-capable model that can't ingest PDF natively now gets the PDF
rendered to one image per page instead of extracted text; falls back to
text extraction when rendering yields nothing.
- core/pdf.py: rasterize_pdf via pypdfium2 render + Pillow PNG (page-capped
at 10, never raises)
- session._wire_content_part: pdf + !supports_pdf + supports_vision ->
rasterized image parts; else text extraction
- trajectory.resolve_attachment_parts: a placeholder can now expand to a
list of parts (1->N); the resolve_attachments callback return type widened
to dict[str, Any] across the provider protocol + 4 providers
- pyproject: pillow dependency
- tests: rasterize_pdf, vision-rasterize gate path, 1->N materialization
When the active model can't ingest a kind natively, the wire resolver
converts it client-side instead of sending a part the model can't read.
Per-kind ownership, no shared machinery: PDF text-extraction is a
pure-local PDF concern; audio transcription is an STT concern memoized
in the audio domain.
- core/pdf.py: extract_pdf_text via pypdfium2 (pure-local, no network, no
cache — re-run per build; page-capped)
- core/audio.py: transcribe_cached — non-raising, memoized by
(alias, content-hash); backend failures not cached
- session._wire_content_part: per-kind dispatch — native where the model
supports the kind (supports_pdf / supports_audio_input), else fallback;
display/export resolve natively so no conversion fires on a render
- image left ungated (pre-existing behavior unchanged)
- pyproject: pypdfium2 dependency + mypy untyped-import override
- tests: pdf extraction, transcript memoization, per-kind gate dispatch
PDF and audio attachments now work end-to-end on the native provider
lanes; non-native lanes degrade to a placeholder (client-side fallback
lands next). Capability flags are populated but not yet consumed by a
wire-build gate.
- providers: Anthropic PDF -> base64 document; OpenAI Responses PDF ->
input_file; compat/Google inline_document_parts PDF -> placeholder
(fixes the base64-as-text mangle); audio = input_audio passthrough on
the compat lane (omni), defensive text placeholders on Anthropic +
Responses
- capabilities: supports_pdf on cloud Claude + OpenAI chat models;
local/default/compat stay False (-> client-side fallback)
- upload: classifier accepts pdf (32 MiB) + audio (25 MiB); endpoint
multipart read cap raised to PDF_SIZE_CAP
- hygiene: consolidate the duplicated upload classification into one
attachments.classify_upload (+ UploadRejection); collapse
AttachmentUploadHelpers to a single classify_upload callable
- tests: PDF/audio translator shapes, capability flags, classify_upload
Provider-neutral plumbing for PDF and audio attachments, with no
user-facing change yet: the upload classifier still rejects them and the
capability tables stay unpopulated (both land in the native-translator
phase). No migration — workstream_attachments.kind is free-text.
- attachments.py: PDF/audio byte caps, allowed-audio MIMEs + format map,
magic-byte sniffers (sniff_pdf_mime / sniff_audio_mime),
Attachment.is_pdf / is_audio
- providers/_protocol.py: supports_pdf / supports_audio_input capability
fields (default False; orthogonal to the STT/TTS roles)
- storage/_utils.py: attachment_to_content_part emits the internal
document(application/pdf, base64) and input_audio shapes
- session.py: by-reference placeholder branches for pdf / audio
- trajectory.py: AttachmentRef docstring (dict-bridge already kind-agnostic)
- tests: test_attachments_pdf_audio.py
Hardened service + slice + node-identity drop-in template + a README for
running a turnstone-server outside Docker that joins the compose cluster —
the production-shaped counterpart to the one-liner in docs/docker.md. Secrets
stay in config.toml; per-host identity + cluster URLs go in the drop-in. The
README notes the cross-host mTLS caveat (turnstonelabs/lacme#22).
A turnstone-server running outside the compose network ("bare-metal", e.g. a
local-GPU box) couldn't fully join: it can't resolve the in-cluster console
(console:8090) to enroll its mTLS cert, and SearxNG was unreachable for
web_search. Only Postgres was published.
Publish the console's plain-HTTP ACME endpoint (:8090) and SearxNG (:8081)
alongside Postgres, all bound via one knob TURNSTONE_HOST_IP (default 127.0.0.1
-- nothing new on the LAN; set it to the host's LAN IP for a node on another
machine). Postgres keeps honoring the legacy POSTGRES_BIND as a fallback, so
existing .env files don't break.
The node's TLS client now honors TURNSTONE_CONSOLE_URL so a bare-metal node can
point at the published ACME endpoint instead of the unreachable in-cluster name
(empty = in-cluster service discovery, unchanged).
Docs (docker.md, tls.md), the run.sh-generated .env, and the bootstrap wizard
updated to match. The advertised host is the cert's primary SAN and the console
collector dials it back, so mTLS hostname verification holds both ways.
The server (:8080) and console (:8090) both set a cookie named
`turnstone_auth`. Cookies ignore port (RFC 6265), so on a shared host
(localhost dev, the Electron build, single-box installs) logging into one
surface overwrote the other's cookie and 401'd the first session.
Give each surface its own cookie name -- `turnstone_auth_server` /
`turnstone_auth_console` -- threaded as a required `cookie_name` argument
through the cookie builders, `check_request`, `AuthMiddleware`, and the six
shared auth handlers (login/logout/setup/whoami/refresh/oidc_callback). Each
app passes its own constant; the parameter is required (no default) so a
forgotten caller fails loudly instead of silently reverting to the legacy name.
Names key on role, not node: the cluster shares one JWT identity and the
console->node proxy re-mints a bearer token (dropping Set-Cookie), so
per-instance names would break identity portability and aren't used.
Hard cutover: the legacy `turnstone_auth` cookie is no longer read and
self-expires within its 24h TTL (one forced re-login). JWT audience was
already enforced, so the shared cookie was a session clobber, not an auth
bypass.
The interactive pane only auto-scrolled when isNearBottom() was true, but it measured that AFTER the new node was appended. A tool block is a tall one-shot append (batch shell, approval card, or result) that clears the 80px near-bottom threshold in a single step, so the post-append check read false and auto-follow silently disengaged at exactly tool-call time — the view froze at the top of the block and only snapped back at the next stream_end. Token streaming was unaffected because each append stays sub-threshold.
Capture the near-bottom state as the first statement of each tool-render method, before any DOM mutation, and thread it into scrollToBottom(stick). This re-pins when the user was already at the bottom and, unlike the coordinator pane's unconditional pin, leaves the view alone if they deliberately scrolled up while a result was rendering.
Methods fixed: announceToolBlock, showInlineToolBlock, resolveApproval, appendToolOutput (all three exit paths), appendToolOutputChunk.
The streamable-http server bound to 0.0.0.0/a LAN IP answered TCP and
/watch but returned 421 "Invalid Host header" on /mcp for every remote
node — which broke multi-node play entirely. FastMCP freezes DNS-rebinding
protection (a localhost-only Host allowlist) at CONSTRUCTION, and this
module builds its FastMCP at import time with the default 127.0.0.1 host;
flipping settings.host in _serve afterward never updated the frozen
allowlist, so the LAN Host was always rejected.
When UNDERSTONE_HOST is off localhost, drop the allowlist in _serve before
run() — matching the SDK's own default for a non-localhost bind. The /mcp
and /watch routes are unauthenticated by design, so serve only on a trusted
network (documented).
Regression test pins the mechanism: a default FastMCP 421s a foreign Host,
a protection-disabled one accepts it. Tests 420 -> 421.
The job was named "test", colliding with core CI's "test" matrix so the PR
checks list showed two "test (3.11)" rows. Rename it to "understone" so the
example's checks read unambiguously (understone (3.11) / (3.13)).
- CodeQL (implicit string concatenation in a list): collapse the wrapped
bullets in cli._render_validate_coverage to single literals. The rendered
output is byte-identical (the example's ruff ignores E501); clears all
six alerts and reads cleaner.
- Copilot: packs/README no longer claims the directory ships "effectively
empty" — it ships the bundled Cinder Wastes alternate world.
- Copilot: the Cinder Wastes' ash_flats and caldera_deep zones overlapped
on column x=60 (inclusive bounds + first-match zone_for silently shadowed
the tier-3..5 band onto a 1x5 deep-edge strip). Move caldera_deep to
x0=61 — no overlap, no dead tiles, deep zone still covers the dungeon.
And harden the loader: overlapping zone rectangles are now a
WorldLoadError, so no authored pack can ship that bug unseen (the
cold-author dogfood loop — a generated pack exposed a validator gap).
Tests 419 -> 420 (zone-overlap rejection). Both worlds validate sound and
remain winnable by the sim bot.
The door-game example is a standalone package (no turnstone-core
dependency) that the root suite does not collect — its
testpaths are scoped to ["tests"], so the example's 419 tests, ruff,
and mypy gates never ran in CI.
Add a path-filtered workflow that installs the example and runs its
full gate (pytest + ruff check + ruff format --check + mypy) whenever
examples/door-game (or this workflow) changes, across the example's
declared Python floor and ceiling (3.11, 3.13). Pinned action SHAs and
contents:read permissions match the existing CI workflows.
A game-loop mechanics patch: the satchel becomes a real stacking inventory,
forging now demands ore won in combat (not just gold), and a vault lets a
hero protect coin from ambush.
- Stacking satchel: the bag re-encodes from a flat id list to "id:qty"
stacks, so potions stack (three Minor Potions fill one slot, not three)
and materials ride alongside. satchel_max now caps distinct KINDS (3);
per-kind quantity is unbounded. quaff/death-save still pull the strongest
potion and ignore materials. One pure codec (engine/satchel.py) owns the
encoding; the façade, the Watch, and the sim all decode through it — no
three-way drift (the v0.9 single-source lesson). The codec parses a bare
id as qty 1, so it can never silently drop a malformed stack.
- Ore-gated forge: ore is a material that drops from won dungeon-rung
fights (and, less often, forest fights), stacks in the satchel, and is
not buyable or sellable — you earn your edge by fighting for it. Forging
now costs gold AND ore ((plus+1) ore per tier), so a rich-but-idle hero
can no longer buy power at the dice table. The dungeon is now also the
mine.
- The vault: deposit/withdraw at the inn moves coin to a strongbox that
ambush cannot touch and that SURVIVES the Wyrm-win legacy reset — the
carry-vs-protect decision the PvP economy was missing.
- Surfaced on both the /watch lobby TV and the in-chat door_status sheet:
each hero's stacked satchel, carried gold, and vaulted gold.
- Tuning (the sim is the instrument): the ore gate added ~2 days to the
Vale and ~1.6 to the Cinder Wastes; the greedy bot still slays the Wyrm
3/3 on both, fully forged to +3/+3, so the loop is not stalled. Defaults
held — no numbers needed retuning.
Four new banded settings (forge_ore_item, forge_ore_per_plus,
ore_dungeon_drop, ore_forest_chance); both worlds gained an ore item.
Schema mutated in place (banked column, satchel re-encoding) — pre-1.0, no
migration by design; a real migration story is owed at 1.0. Tests 382 ->
419; the vault-survives-rebirth invariant and the codec are revert-verified.
Graphics polish: distinct terrain and structures now read by COLOUR on the
Watch, not only by glyph. One unified palette, shared by every world — the
fix is to grow the set of distinct object-type roles, not to fork per-world.
- Roads were the tell: road shared the "floor" green with grass, so a path
vanished into the meadow on the lobby TV. Likewise forest shared "tree",
the three town buildings all shared "town", and the Cinder Wastes' molten
slag borrowed "water" and rendered BLUE. Each is now its own role: road
(stone), forest (lush green) with scrub (its barren ember-brown
counterpart for volcanic/desert dense terrain that must NOT read as
woods), lava (molten orange), barren (wasteland taupe), and inn/shop/
healer split out of the generic town.
- Both worlds remap onto the shared vocabulary; in each, no two distinct
terrain/building types share a colour. A live render caught the Cinder
cinder-fields rendering green under the generic "forest" role — hence the
scrub role, so the volcanic waste reads warm. The text frame renderer
stays monochrome (it never read colour), so frames and goldens are
untouched — this is Watch-only.
- The bug class is now closed by construction: a test asserts the Watch
PALETTE carries a hex for EVERY Color role, so a role can never ship
unpaintable and silently fall back (which is exactly how road hid).
- Color.assignable() is the single source for the overlay-vs-assignable
split (runtime actor/item colours and the DEFAULT fallback are not
author-pickable); the authoring manual's colour vocabulary generates
from it, so it can't drift.
Tests 373 -> 382. floor/tree/forest are three greens kept deliberately
distinct (forest is olive-hued); verified on a real render along with the
scrub fix.
The slice that proves the pipeline: a second world authored entirely by an
LLM from AUTHORING.md and the validator alone, plus the tooling to discover,
theme, and balance-test any world.
- The dogfood: "The Cinder Wastes" — an ashen volcanic underworld (slag
rivers, a caldera mouth, a Magma Wyrm) — was written cold by an agent
given only the generated authoring manual and `understone validate`. It
passed validation on the FIRST run with zero failures. Its stumble log
found six places where the manual stated a rule the validator didn't
enforce; those became permanent hardening (below). It ships in
understone/world/packs/ and glows ember on the lobby TV.
- `understone worlds` lists every bundled world (the Vale + alternates)
with its load status, via one shared discovery path.
- Per-world Watch themes: settings.watch_theme (phosphor/amber/ice/ember,
loader-validated) repaints the spectator page; the Vale's green is
byte-for-byte unchanged.
- The sim harness: a pure, seeded, greedy bot plays the real game façade
over an injected day-stepping clock and emits a balance report —
`understone simulate PATH [--days N] [--seeds K]`. It SLAYS THE WYRM on
both worlds (Vale ~day 13, Cinder ~day 25), so the whole v0.1->v0.7 loop
is proven winnable end-to-end by an unclever bot through the real stack.
- Loader hardening from the dogfood: a rare monster may not occupy a
dungeon-rung guardian slot (it would silently become a fixed foe and
leave the rare pool); exactly one monster may be the boss; and the
boss-tier error now says "no non-boss monster," matching the manual.
AUTHORING gained a generated "what validate checks vs. what it cannot"
section so the rule/guidance boundary is honest.
Review hardened the bot for arbitrary authored packs (a MENU-mode fight
spin and four related robustness gaps that were latent on the shipped
worlds), and documented that final_level reads post-legacy-reset. Tests
359 -> 373; both worlds still win byte-identically after the fixes.
The depth slice: four standing reasons to return past the daily reset.
- The rung ladder: the dungeon is a descent fought one rung per turn, each
guardian a fixed tier. A loss bounces you home but your depth PERSISTS —
you re-enter where you left off. The Wyrm now gates on BOTH level AND
reaching the floor (the deep has a bottom, and you must have touched it).
- The satchel + the death-save: potions are CARRIED now (up to three),
bought to the satchel, drunk with quaff. The heart of it: when any fight
would kill the active fighter and they carry a draught, the strongest is
drunk automatically — they survive standing at the potion's value, no
bounce. This fires on EVERY fight (forest, rung, and the Wyrm itself —
a potion carried to the climax is a real tactical choice); a Wyrm loss
so saved is "driven back, alive but unproven," not devoured. The sleeping
ambush victim never quaffs (they are asleep). combat.py stays pure — the
satchel and the save live entirely in the façade.
- The forge: the shop spends scaling gold to add a +1 edge to equipped
weapon or armour, capped — the late-game gold sink. Swapping or selling
the piece loses the edge with it (one centralized unequip clears the
bonus and the plus so a stat can never go phantom).
- Rare beasts: a few named foes prowl the forest via weighted selection,
surfacing seldom; felling one is a public Herald flash and always yields
a draught into the satchel. Rung guardians are never rare (fixed foes).
Four new player columns; four new banded settings; dungeon_tiers extended
to three rungs. Tests 283 -> 330; the death-save (all four paths), forge
accounting across forge/buy/sell/legacy, rung math, and weighted rare
selection all pinned, with the death-save and forge invariants
revert-verified.
The look of the next age — the modern equivalent of the ASCII->CP437 leap.
Full Unicode is available now, but the whole stack (text frames, golden
tests, the Watch's 1ch grid) assumes one glyph = one column, so the
enabling piece is a WIDTH RULE, not the glyphs themselves.
- textwidth.is_grid_safe: one code point, printable, East-Asian width not
Wide/Fullwidth, no combining/format/control category. This is the
one-glyph-one-column contract. Ambiguous-width glyphs are ACCEPTED on
purpose — they ARE CP437 (the wall, the club-tree, the up-arrow forest)
and render single-column on the Western-monospace metrics every surface
uses; only genuinely double-width runes are barred. The loader enforces
it on every map glyph; the player-name/free-text sanitizer enforces the
same rule (the narrow ledger), so a wide name can't shear a frame.
- Re-skin: water ~ -> ≋, inn -> ⌂, healer -> ✚, dungeon mouth -> ∩, and
the other adventurer -> ☻ (CP437's own player glyph). The colour field
the renderer has carried unused since v0.1 now has a second consumer.
- Texture variants: grass and water vary by a deterministic per-coordinate
hash, rendered identically in the Python frame builder and the Watch's
JS. The two are kept in lockstep by shared hash constants + an agreement
test that replays the JS arithmetic and asserts it equals the Python
output for every variant over a grid — not a comment-coupled copy.
- Watch glow-up: a Noto Sans Mono font stack and a UTC-hour day/night tint
(the Vale darkens at dusk on the lobby TV).
- The curated SAFE_PALETTE is enforced author-usable: a test asserts no
palette glyph collides with the reserved player markers, so AUTHORING's
generated appendix can't advertise a glyph the loader would reject.
- Resume is identity-preserving: an existing character resumes by exact
stored name without re-validating the width rule (which governs creation
only) — resume must never lock anyone out.
Tests 231 -> 283; width edges (CJK/emoji/combining/fullwidth), the
Python<->JS lockstep, the palette/reserved guard, and resume-vs-create all
pinned and revert-verified.
The social slice: the shared world gets teeth, letters, and a house game.
- Ambush (async PvP, classic door-game player-kill spirit): waylay an adventurer who has
not yet begun their day. Ordered gates — known target, not yourself, the
gatekeeper shields the young (both >= min level), level band +-2, the
SLEEP RULE (acting today makes you watchful — an active-play defense),
mercy for the downed (hp<=1 cannot be piled on: even bandits have
standards), once per pair per UTC day. Win: capped gold cut transfers,
victim wakes at the spawn-stone with a private note; lose: the sleeper
wakes blade-in-hand and the Herald crows your shame. The attacker wears
the counter-blows the combat log narrates (state matches story). Both
players persist in one transaction.
- The inn mailbox: events carry a target ('' = public). door_log delivers
private notes to the addressee only; the Watch and other players never
see them. Mail is DURABLE past the in-memory tail (SQLite backfill for
cursors older than the resident window) — the broadsheet is ephemeral,
letters are not. Sanitized, daily-capped.
- Inn dice: 2d6 against the house, bet- and count-capped per day, big wins
make the news.
- Six new banded settings; four day-counter columns join the shared lazy
UTC reset; schema stamp stays 1 (pre-1.0 mutates in place by design).
Tests 184 -> 231; sleep rule, mercy gate, band boundary (exact/over),
refusal precedence, attacker wear, zero-gold robbery, mail eviction
survival, and Watch privacy all pinned; guards revert-verified.
The IGM seam realized: world packs are now a first-class authoring target
for models and humans, with a validate loop and a loader hardened for
routinely-untrusted generated content.
- understone newpack DIR scaffolds a pack (the six content JSONs templated
from the shipped Vale) plus AUTHORING.md — a manual written for a model
to follow cold. Its bands table is RENDERED FROM the loader's own band
constants at scaffold time, so documented limits and enforced limits
cannot drift.
- understone validate DIR loads a pack and prints either a pack report
("This pack is sound. The door stands open.") or the loader's
file/index/field-naming error — the authoring feedback loop.
- Loader hardening: glyphs must be one printable column-safe character and
never the frame box-drawing set or the @/& player markers (map content
cannot impersonate players or forge frame chrome); map dims 8..256;
per-file count caps; display-name length caps. All errors instructive.
- The packaged-world path is single-sourced (understone.world.
PACKAGED_WORLD_DIR) for the server default and the scaffold template.
- README "Authoring worlds" section frames the loop: newpack -> write or
generate -> validate -> serve with UNDERSTONE_WORLD=dir.
Review round: bug finder returned zero findings; quality round fixed the
world.json doc example (it showed a zone fragment where an authoring model
would copy a whole-file shape — now a labeled skeleton), the stale Usage
docstring, and the duplicated packaged-path constant.
Tests 166 -> 184. Scaffold round-trips through load_world by test.
A read-only CRT spectator page served by the game process itself, plus
content depth. Input never flows through the Watch — it is the wall-mounted
terminal in the BBS room; chat remains the only actuator, so there is no
input channel to deadlock and no cross-origin surface (the page polls the
same origin that served it).
- /watch: one self-contained page (inline CSS/JS, no external assets),
phosphor CRT styling. The base map paints once from /watch/world.json
(terrain glyph rows + a glyph->color legend — the palette the text
renderer has deliberately ignored since v0.1 finally gets its first
renderer); players overlay as positioned glyphs repainted from
/watch/state.json every 2s; the sidebar carries the roster with win
stars, the Hall of Legends, and the Herald. SIGNAL LOST on poll failure;
the bootstrap retries so a spectator arriving during a server blip
recovers without a reload.
- Routes ride FastMCP custom_route on the existing process — read-only
handlers with no awaits between reads (handlers and sync tools
interleave on one event loop, so every response is a consistent
snapshot).
- door_join/door_help advertise the Watch URL in http mode (stdio: none).
- Content: +5 monsters (one per tier; the gauntlet's first-in-tier foes
preserved), +3 items smoothing the gear curve, +6 events; fight weight
retuned to hold ~55% of encounter rolls. Zero geography churn.
- Review round: the Herald window is a plain list tail (id arithmetic
under-reported the feed when AUTOINCREMENT ids gap — regression-pinned
with sparse ids), and the bootstrap-retry fix above.
Tests 149 -> 166.
The "make it a game" slice: a win condition with classic-door-game-style legacy, texture
between fights, and a shared broadsheet.
- The Wyrm Below: a boss (flagged in the pack, excluded from random bands)
behind a level-gated `challenge` verb at the dungeon. Victory writes a
Hall of Legends row and the character resets to the fresh-start kit,
keeping a wins counter rendered as ★ on the leaderboard — the classic
race-reset-race loop. Defeat and stalemate flight make the news.
- Forest events: movement encounters weighted-pick from a content-pack
table (fight/gold/heal/trap/lore). Only fights stop the walk or cost
turns; texture is free and private. Trap damage floors at 1 hp.
- The Understone Herald: door_log is a broadsheet with a masthead and
write-time template variety; the public feed is curated to notable beats
(joins, blessings, level-ups, defeats, the Wyrm's fate) — town errands
stay private.
- Reward narration moved from the combat engine to the façade, composed at
the moment gold/xp are actually banked, so the server can never narrate
a reward it did not apply (the Wyrm win previously claimed +400 XP /
+250 gold that the legacy reset wiped).
- Fresh-start hp/atk/def promoted into world.json settings alongside the
starting kit; dungeon-tier validation counts non-boss monsters only,
keeping the validator's no-silent-rung promise true.
Schema mutated in place (players.wins, hall_of_fame) — pre-release, no
migration path by design. Tests 109 -> 149; the challenge level gate is
negative-tested; rank stars survive 24-char names (compact form past 5).
A shared-world, classic-door-game-style door game in examples/door-game/: a pure-stdlib
game engine (tile overworld + location menus, seeded combat, daily turn
budget, leveling, shop, event log, leaderboard) behind nine sync door_*
FastMCP tools returning monochrome box-drawing frames. The connecting
session's LLM plays dungeon master — tool descriptions plus a door_help
manual teach a cold model to run the game with zero setup, while the server
owns all dice and state, so the DM narrates around facts it cannot bend.
Non-obvious decisions:
- engine/screen/world/persistence import stdlib only; server.py is the only
mcp import. All nine handlers are sync def: on mcp 1.27 they execute
inline on the event loop (verified against func_metadata), so tool bodies
serialize and one SQLite connection (WAL, per-action commit) is safe.
check_same_thread=False exists only because the Store may be constructed
on a different thread than the serving loop.
- Streamable HTTP serves ONE process = one shared world (players appear on
each other's maps; async "while you were away" event feed); stdio is the
solo-world fallback.
- The economy is content, not code: daily_turns, costs, xp curve, bestow
budget, and dungeon tiers live in world.json settings, band-validated by
the loader. door_bestow gives the DM capped, event-audited largesse
(gold/heal only, never turns) so story generosity cannot melt the shared
leaderboard.
- Player names and bestow reasons are sanitized (printable-only, length
caps) because they flow into the shared event log and from there into
other players' DM context — embedded newlines would forge log lines.
- Daily turn/bestow pools lazy-reset per UTC day on every consuming path
(injectable clock); the dungeon gauntlet is a fixed boss ladder by design.
Tests: 109 — engine units with seeded RNG + frozen clock, hand-authored
golden frames paired with structural asserts, loader band rejections, and
one real-wire integration test (uvicorn + streamablehttp_client) with a
two-session shared-world assertion. Negative-tested by reverting the guard
and watching the suite fail: the daily turn-budget guard, the bestow cap,
and the sanitizer's isprintable clause.
The coordinator memory scope was keyed by the session's ws_id, so every
new coordinator session started with an empty namespace and its rows
were orphaned on close — coordinator memory never actually persisted.
Re-key the scope to the coordinator's creator user_id: one durable
orchestration namespace per user, shared by all of that user's
coordinator sessions (concurrent ones included; upsert-by-name is the
collision rule).
The child-containment threat model is unchanged: the gate is session
KIND — children are always interactive and share the parent's user_id,
so _validate_scope rejects them before scope resolution, and the REST
memories API still rejects the coordinator scope outright. The implicit
visibility lane now also fails closed on an empty scope_id to match the
explicit search/list lanes (the storage helpers treat a falsy scope_id
as 'no scope_id filter', which would have read every user's rows).
Anonymous coordinators are no longer constructible: ChatSession refuses
kind=COORDINATOR with an empty user_id at the constructor — the single
choke point covering create, rehydration of legacy rows (surfaced by
the open handler as a 503 with remediation text), and any future host —
and the console no longer masks an empty uid as a phantom 'system'
principal when minting coordinator JWTs, per CoordinatorTokenManager's
documented 'sub = the real creator user_id' contract.
Migration 061 carries existing coordinator rows across: rows whose
owning workstream is gone or ownerless are deleted (unreachable under
user keying), same-name collisions within a user keep the newest
updated row (memory_id tiebreak), and survivors re-key to the owner's
user_id.
_buildHandle hard-coded aria-valuemin/max at 10/90 (inherited from the
old ui/static implementation) while the actual drag/keyboard clamp is
_ratioBounds — the cell minimums against the split node's OWN px region
(a 1200px host really clamps at ~17/83; nested splits sit tighter), so
assistive tech was told a wider range than the separator allows.
aria-valuenow/min/max are now all written in _applyLayout's handle loop
from _ratioBounds(h.node) — one writer, refreshed on every drag,
keyboard nudge, and structural change. A bare window resize can stale
the advertised range until the next interaction (no resize listener by
design — % insets make resizes free), still strictly truer than a
constant. The max>=min guard covers a host shrunk below two cell
minimums, where the bounds legitimately cross.
The /coordinator/{ws_id} standalone page is reachable only by direct
URL — all three console navigation sites are shell-fallback else
branches behind openPane. Record that in the sidebar-padding comment
so the scope isn't over-read as a live second surface.
The per-pane ✕/− chip floats at the pane's top-right — exactly where
the coordinator sidebar's toggle row and Children refresh button sit,
so the chip covered them. Pane-hosted coordinators now start the
sidebar content 44px down (padding, not margin, so the column's left
border still runs the full pane height); the standalone coordinator
page has no chip and keeps the 14px default.
Dual designer review (one primed on the branch context, one cold), all
measured findings applied:
- The per-pane chip was a mode-error trap: identical glyph at the
identical locus, reversible in split mode (hide cell) but destructive
single-pane (close pane). Now − hides, ✕ closes, and the close mode
wears a danger hover/focus ring so the irreversible action telegraphs
before the click lands.
- Single-pane chip anchored to the VIEWPORT: an unpositioned section
resolves absolutes to <body>, so the chip only coincidentally landed
near the pane corner. .panes > section.pane is now position:relative
in both modes (all pane-content absolutes verified to anchor to their
own local relative parents).
- Light-theme AA (measured): .shown tab underline 55% mix composited to
2.34:1 -> 80% (~3.7:1 light / ~5:1 dark); focused-cell ring 2.60:1 on
light -> 75% mix override there (dark keeps 55% at 3.75:1).
- Chip: border --hair-2 measured ~1.3:1 (invisible) -> --ink-4; 22px
target under WCAG 2.5.8's 24px floor -> 28px; right offset clears the
message scrollbar gutter; light resting glyph one ink step up.
- Focus bar inset 1px from cell sides (no doubled-accent stripe where
it butted a separator at the T-junction); greyscale font smoothing on
the tail glyphs (subpixel RGB fringed the box-drawing characters).
Rejected with rationale: aria-pressed on the split buttons (they are
one-shot verbs — splitting again nests — not mode toggles).
Four refinements from first live use:
- Per-pane ✕ chip, top-right of every visible pane. Split mode: hide
that cell (closeCell — the tab stays, the sibling absorbs the space).
Single-pane: close the pane outright (withheld from the unclosable
Dashboard). The click decides at click time; the label tracks the
mode. Manager-injected into the pane section — content untouched.
- Coordinator child links open BESIDE the coordinator (openPaneBeside:
split right of the focused cell, seeded with the child pane) instead
of replacing it — the parent stays on screen. Degrades to the plain
focused-cell swap when the split is denied (cap / narrow viewport).
splitFocused() gained an optional explicit-fill parameter for this.
- Tier-1 ws_closed now CLOSES the open interactive pane (tab gone, a
split cell collapses) — the coordinator-closes-its-child flow,
matching the standalone's pane-auto-close. The dead-banner lane
stays for streams that die without a ws_closed (node crash/network),
where the session may still be revivable.
- Paint bug: the focused-cell ring was an inset box-shadow on the
section, which paints in the element's own background layer — UNDER
opaque children touching the edges, so the status bar / composer
strip occluded it. The ring now rides a click-transparent ::after
overlay above pane content; the 2px top bar sits above the ring line.
The livepass shell surface's demo panes grew a .ws-status-bar footer so
the occlusion bug class stays visible to future passes.
Revives the split-pane feature retired with ui/static (step 6), rebuilt
on PaneManager: an optional binary layout tree (null = the one-pane-per-
tab behaviour, unchanged) renders visible panes as %-inset cells — no
reparenting, so live stream DOM, scroll state and media survive layout
changes. Tabs stay global: the active tab is the focused cell, a
backgrounded tab swaps into it, clicking inside a visible pane focuses
its cell, .shown marks visible-unfocused tabs. Separators resize by
pointer-capture drag and arrow keys (role=separator + aria-value*); the
tree persists in the working-set blob and rehydrate prunes leaves whose
pane did not restore. Limits: 6 cells, 200x150 cell minimums, denials
toast the manager's reason.
Affordance: Split right / Split down / Unsplit buttons in the tab-bar
tail replace the redundant [+] (the permanent Dashboard tab is the
launcher) — deliberately no contextmenu override this time. The dead
TS_APP.focusLauncher seam goes with it.
Measured chrome: the focused cell wears a 2px accent top bar (no thin
tinted ring clears 3:1 in both themes) plus a 55%-mix inset ring;
separators rest at --ink-4 with solid-accent hover/drag/focus; .shown
tabs carry an accent underline; the tail cluster is fenced and lifted
to --ink-3.
scripts/livepass.py grows a third surface: shell/livepass.html boots
the real shell.js + pane.js and drives ?split=right|down|three|none
(+ &theme=light), stamping SPLIT-READY-<cells> / SPLIT-FAILED-<reason>.
121 warnings -> 0. Two upstream deprecations get narrowly-scoped
filterwarnings entries (the mcp streamablehttp_client rename — adoption
deliberately rides the v2 migration since the new entry point's call
shape changes again there; the starlette httpx TestClient notice). The
one real RuntimeWarning is fixed at the source: tests that mock
asyncio.run_coroutine_threadsafe handed real coroutines to a stub that
never awaited them, GC-firing 'coroutine was never awaited' inside
whatever unrelated test ran later (the same cross-test bleed mechanism
as the CI closed-stream spew — per-test filterwarnings markers cannot
catch it, which is why two such markers existed and still leaked). A
shared _dispatch_stub now closes real coroutines before returning the
canned future; the obsolete markers are removed.
mcp 2.0.0a1 shipped 2026-06-11 (stable targeted ~2026-07-27). v2 removes
streamablehttp_client, changes the transport tuple arity, and renames
mcp.types fields to snake_case — all of which our client imports. The
maintainers' release note asks downstream packages to add an upper
bound now (their worked example is this exact constraint). Floor stays
at 1.27: nothing newer adds anything our surface needs, and the #2147
shutdown busy-loop we wrap remains unfixed at every released version.
Resolution is unchanged (1.27.2); lockfile re-pinned metadata only.
Copilot review on #661: empty base_url let the SDK fall back to
https://api.anthropic.com, sending compat-shaped requests to the
commercial API. The lane is local-only by definition, and the /v1-strip
edge case already established fail-loudly-over-silent-prod-retarget;
apply the same principle to the empty case. create_client raises an
actionable ValueError; the admin Detect path surfaces it as a clean
error string via probe_model_endpoint's existing handler.
Add provider id "anthropic-compatible": the existing AnthropicProvider
pointed at Anthropic-compatible local servers (vLLM /v1/messages),
mirroring the openai/openai-compatible split. Registry-only — configured
via the admin Models tab or [models.*] toml, not exposed on the bare
--provider flag, so the CLI/server prod-URL defaults are unreachable for
the lane and real-Anthropic behavior is untouched.
Lane behavior (live-verified against vLLM 0.22.1rc1 + DeepSeek-V4-Flash):
- Capability defaults replace the Claude static table: token_param
max_tokens, thinking_mode none, web_search/tool_search/vision off,
reasoning replay on. vLLM rejects Anthropic server-side tool types
(tools require input_schema) and ignores the thinking request param,
so neither is sent; thinking blocks still stream back and round-trip
through the native lane verbatim.
- Reasoning toggles via server_compat extra_body chat_template_kwargs
(first-class vLLM request field; request-level keys beat server
defaults). _build_thinking_and_kwargs forwards non-internal
extra_params as SDK extra_body; thinking_budget_tokens stays internal.
- No temperature force: thinking_mode none skips the Claude-only
temperature=1.0 requirement.
Admin UI: provider option + URL placeholder (base_url without /v1 — the
SDK appends /v1/messages); the server-compat section shows only the
extra-body field for the lane. thinking_mode round-trips through the
form dropdown for every provider except anthropic-compatible, where it
stays in the raw capabilities JSON — the edit-load lift and save restore
use the same predicate so stored overrides are never silently dropped.
Docs: architecture.md gains the lane subsection incl. verified quirks
(thinking param dropped by vLLM; stop_sequences cut inside thinking and
report end_turn; usage has no cache fields; images need a multimodal
model; mid-conversation system turns are per-model opt-in).
Negative-tested: removing the _INTERNAL_EXTRA_PARAMS exclusion fails
test_internal_keys_not_leaked; the live test drives a streamed turn with
the chat_template_kwargs toggle and asserts no reasoning deltas.
Review feedback: (1) gating the drain on a main-thread truthiness check
of _background_tasks could skip cancellation when a spawn queued via
call_soon_threadsafe had not reached the set yet — submit whenever the
loop is RUNNING and snapshot on the loop, where FIFO callback order
guarantees earlier-queued spawns have landed; (2) shutdown stopped the
loop thread but never closed the loop or cleared _loop/_thread, leaking
selector resources for embedders that cycle managers — close + clear
when we own the thread and it actually stopped (loud warning when it
does not); unowned loops (tests wiring _loop directly) stay untouched;
(3) the bare await-in-suppress drain loops become
asyncio.gather(return_exceptions=True) in both the shutdown drain and
the test fixture.
The post-reconnect catalog refresh was scheduled as a bare
asyncio.create_task: no strong reference (the task could be GC'd
mid-flight, so the refresh might silently never run) and no exception
retrieval (failures surfaced as "Task exception was never retrieved"
at GC time — in CI, onto an already-closed pytest capture stream, the
"I/O operation on closed file" spew; a suspected contributor to the
flaky 60-minute CI hangs via cross-test loop/task state bleed).
- _spawn_background(coro, label): tracked-task set + done-callback
that retrieves and logs failures at warning; discard runs LAST so
set-emptiness means "done AND reported"
- shutdown() drains tracked tasks FIRST, so stack teardown can't race
an in-flight refresh; same run_coroutine_threadsafe idiom and
timeouts as the existing close steps
- running_loop_mgr fixture: cancel-pending -> drain -> stop ->
join(5) with a loud assert -> loop.close() (was stop + silent
join(2), never closed)
- the false-property test ("swallows refresh failure" — nothing
swallowed it) now waits for completion and asserts the logged
warning via the patched module logger (structlog; caplog cannot
observe it), polling inside the patch context
Review feedback on the purge's race window: the pre-SELECT re-verify
left a statement-to-statement gap where a concurrent registration could
still lose rows — and the pre-counted refcount release could underflow
when it didn't. Orphan-ness now rides the DELETE itself (correlated
NOT EXISTS) with refcounts released from its RETURNING, so refs are
released for exactly the rows that were deleted. Input is de-duplicated,
IN-lists chunk at the storage layer's 500 convention, and the scan's
per-workstream ref-count loop is now one anti-join pass.
Conversation rows whose workstreams row is gone (historical unregistered
writers; the delete-during-inflight race re-creating rows after
delete_workstream) are invisible cruft that also pins attachment
refcounts. Add a turnstone-admin verb: default = read-only scan report
(ws_id, rows, attachment refs, first/last); --delete [--yes] purges.
- shared find/purge logic in storage/_utils; protocol + both backends
in lockstep (thin wrappers)
- purge re-verifies orphan-ness in-transaction: a ws_id re-registered
between scan and purge is skipped, never deleted
- releases the deleted rows' attachment refcounts through the
delete_workstream GC path and sweeps workstream_config/overrides
- summary reports actual purge results, including the skipped clause
* fix(ui): re-home MCP consent badge on the Manage Connections row
The L-shell renovation retired the standalone settings gear (#settings-btn).
The MCP pending-consent badge anchored to that gear via _refreshConsentBadge,
which null-guarded silently — so since the renovation pending consent requests
had no indicator (the badge was invisible).
Re-home the badge on the rail's Manage row where the MCP/connections surface
lives in both deployments:
- rail.js gains a generic setRowBadge(tabKey, count, label?) hook + a `badge`
builder: a small ⚠-glyph + count chip (never colour alone) using the DS warn
tokens. mountManage registers row + owning-group-head refs and re-applies live
counts across a (re)mount. When the owning group is collapsed, the count also
mirrors onto the group head so a hidden row never hides the signal. rail.js
stays agnostic — it owns the mechanism, the caller owns the meaning.
- shell.js (the ESM bridge) re-exports setRowBadge on window.TS_SHELL so the
classic ui/static/app.js subsystem can drive it without importing the module.
- The standalone consent subsystem keeps its shell-level ownership: _refresh-
ConsentBadge now drives setRowBadge on the Connections tab, fed by both the
loadPendingConsents hydrate/poll load and live onConsentDetected notifications.
- The shared interactive pane host bridges onConsentDetected to the new
window.TS_APP.onConsentDetected seam (undefined on the console, so the console
pane stays a no-op there); panes only notify.
- The dead colour-only gear badge CSS (.settings-consent-badge, red dot) is
removed; the new chip lives in shell.css as token-only .rail-badge so it
flips themes by construction.
Console MCP tab (Extensions > mcp) and standalone Connections tab
(Extensions > connections) both badge correctly. Pins extended in
test_shell_js.py + test_app_js.py.
* fix(ui): drop the unused head ref from the rail badge row map
Review feedback: _rowEls stored each row's group-head element but every
head consumer resolves it through _groupEls; keeping the duplicate DOM
ref made the remount state shape harder to reason about.
Review feedback: the next-case end markers were exact-indentation
string finds that raised a bare ValueError when unmatched. Use
whitespace-tolerant regexes with actionable assertion messages, and
bound the history-replay window structurally (next role branch, with
a generous fallback) instead of a fixed 600 chars.
Review feedback: (1) the Enter keydown re-dispatched through btn.click(),
relying on the disabled-guard to suppress the browser's own
Enter-to-click — preventDefault + direct activation makes the keyboard
path provably single-fire; (2) the branch-scoped Hls instance was
unreachable from the media error handler, leaking its listeners and
loader timers when the player node was replaced with the retry UI —
hoist the ref and destroy it before replacement.
The interactive Pane renders media embeds (buildMediaEmbed / buildPlayButton),
but the Play activation — _loadHls / _isHlsUrl / _activatePlayer and the
click/keydown delegate — stayed behind in the standalone ui/static/app.js as
DOCUMENT-level listeners. The console L-shell mounts the same interactive.js
module but never loads ui/static/app.js, so the Play button was dead in
console-hosted interactive panes.
Lift the activation into shared_static/interactive.js (alongside the existing
buildMediaEmbed/buildPlayButton — media embeds are interactive-pane-only; the
coordinator pane renders none) and wire it as a pane-owned, root-scoped
this.el click/keydown listener, mirroring the approval-keydown pattern the
fork collapse established. The standalone copy is deleted so no duplicate
implementation remains; both deployments now activate through the one shared
handler.
The hls.js vendor is fetched lazily by absolute /shared/ URL (the same
mechanism renderer.js uses for mermaid), and /shared is mounted at the root in
both turnstone/server.py and turnstone/console/server.py, so the vendor —
which ships in shared_static/hls-1.6.16/ — resolves in both deployments with
no HTML change.
Pins: assert the lift + pane-ownership in test_interactive_pane_js.py and the
standalone-stays-clean guard in test_app_js.py.
The live-SSE/history system-turn dedupe (renderedSystemEventIds /
_renderedSystemEventIds) was already in place on both panes and merged
to main (21af6c4 aligned the persisted row event_id with its SSE event;
09e41d1 added the belt-and-braces Set on the coordinator). The existing
pin tests only assert the Set's .has()/.add()/.clear() symbols appear
somewhere in the file, so a refactor that keeps the Set but short-circuits
the live-handler consultation (guard -> false) re-opens the double-render
while the pins stay green.
Scope the new assertions to their blocks: the live system_turn case must
CONSULT and RECORD against the Set, and the history render path
(replayHistory / refetchHistory's system-role branch) must record each
replayed row's event_id. Bounded at the next switch case rather than the
first break; the dedup-skip path itself breaks before the .add(), so a
break-bounded slice would drop the record half.
Verified the new slice checks fail on a dedupe-neutered factory (a
headless-Chrome harness driving the real createCoordinatorPane confirms
that neutering produces two rendered nodes for one event id; intact code
renders one, and the no-event-id legacy path still renders both).
The touch_structured_memories facade and both storage backends were
implemented but had zero call sites, so access_count never moved and
last_accessed never advanced past write time on any deployment.
Wire two touch points:
- proactive composition touches the injected top-k (post-rerank) set,
deduped per turn since _init_system_messages recomposes many times
within a single turn;
- the memory tool's search and get reads touch their returned rows,
counted per call. save/delete/list do not touch.
Touches are best-effort through the facade, which already swallows
storage errors, so a failed touch never breaks composition or a tool
call.
*A hypothesis — not a theorem. The honest answer is a claim about **shape**: an object you can write down that says what a harness is and, just as precisely, the guarantee it cannot carry for free.*
Most descriptions of an agent framework are a feature list. This is an attempt at a definition.
---
## The claim
*Informal.* A harness is a **stopped, deterministically-controlled Markov process on task-state, closed around a stopped autoregressive process on context-space, driven by a learned model kernel** — a deterministic controller in closed loop with a stochastic learned plant.
*In plain terms.* The **harness** is the whole governed loop: a deterministic **shell** you write — build the prompt, authorize an action, fold the response back into state — wrapped around a black-box stochastic model kernel (the **plant**, $M_W$) and the environment its actions touch, looped until it halts in $H$. The shell is deterministic, $M_W$ is not, and everything below makes that split precise.
*Formal — the objects.* A harness is a tuple $\mathcal{H} = (\mathcal{S}, \mathcal{C}, \mathcal{Y}, \mathcal{A}, \mathcal{E}, \pi, M_W, \gamma, Q_E, \rho, H, H_{\mathrm{ok}}, B)$ over **standard Borel** spaces (concretely: the *controlled* state is standard Borel by construction — token sequences, finite config maps, bounded counters and ledgers, finite tuples of real vectors — and the model/environment coordinates are inherited as such whenever they serialize to a Polish space; the assumption is roomier than it looks — even a belief-state coordinate valued in $\mathcal{P}(X)$ survives, since $\mathcal{P}(X)$ is Polish for Polish $X$ — and fails only for a genuinely non-separable coordinate, an uncountable product $\sigma$-algebra being the canonical hazard, which this construction avoids): a deterministic lowering $\pi:\mathcal{S}\to\mathcal{C}$; a stochastic model-run kernel $M_W(c, dy)$ into a readout space $\mathcal{Y}$ (which includes the parse-failure $\bot$, so $M_W$ and $\gamma$ are total over it); a deterministic **authorization gate** $\gamma:\mathcal{S}\times\mathcal{Y}\to\mathcal{A}_{\bot}$ that validates the model's parsed readout into an authorized action in $\mathcal{A}$ or rejects it as $\bot$ (parsing itself lives inside $M_W$ — realized as the readout $R$ of the specialization below); a stochastic environment/tool kernel $Q_E:\mathcal{S}\times\mathcal{A}_{\bot}\rightsquigarrow\mathcal{E}$ on the authorized action (rejection included, with $Q_E(s,\bot,\cdot)=\delta_{e_0}$ for a distinguished no-op response $e_0\in\mathcal{E}$); and a deterministic verify-and-fold-back map $\rho:\mathcal{S}\times\mathcal{Y}\times\mathcal{A}_{\bot}\times\mathcal{E}\to\mathcal{S}$.
*Terminal structure.* The terminal set is an absorbing halt set $H\subseteq\mathcal{S}$ (the daemon "ready-state" recurrence of the note below is a separate, non-absorbing object) with accepting subset $H_{\mathrm{ok}}\subseteq H$; separately, a bad set $B\subseteq\mathcal{S}$ ($B\cap H_{\mathrm{ok}}=\varnothing$) marks the unsafe states for reach-avoid, possibly entered before any halt; hitting times are $\tau_A=\inf\{n\ge 0:s_n\in A\}$, and $\tau_H$ is a stopping time for the natural filtration.
*The outer kernel.* The induced outer transition kernel, for $s\notin H$, is
$$T(s, A) = \int_{\mathcal{Y}}\!\int_{\mathcal{E}} \mathbf{1}_A\!\big(\rho(s, y, \gamma(s,y), e)\big)\; Q_E\big(s, \gamma(s,y), de\big)\; M_W(\pi(s), dy), \qquad T(s,A)=\mathbf{1}_A(s)\ \text{ for } s\in H,$$
and the harness runs $s_{n+1} \sim T(s_n)$ from an initial $s_0 \sim \mu_0$ until $\tau_H = \inf\{n : s_n \in H\}$. Because $\pi, \gamma, \rho, H$ are deterministic they contribute no integration variable of their own — they appear as measurable transformations inside the integrand (the pushforward), not literally outside it — so the controller injects no randomness, and every coin is inherited from $M_W$ and $Q_E$. (The earlier shorthand $T = \rho \circ (M_W \circ \pi, E)$ is suggestive but ill-typed — $M_W$ returns a *law*, while $\rho$ consumes a *sample* together with the prior state $s$; the integral is what the shorthand meant.)
*Fail-closed.* The gate $\gamma$ is what makes **fail-closed** a property, not just a name: model output is an *untrusted proposal*, and $\gamma(s,y)=\bot$ forces a no-op environment response ($Q_E(s,\bot,\cdot)=\delta_{e_0}$) — so a malformed or unauthorized tool call is rejected *before* it can act, not validated after its side effects have landed. Fail-closed is then the property that a rejected proposal causes *no unauthorized side effect* and lands in a **safe, non-bad** set ($\rho(s,y,\bot,e_0)\notin B$): a non-accepting terminal $H\setminus H_{\mathrm{ok}}$ in the strict case, or a safe non-terminal state when the spec retries. And $\rho$ must validate the tool *response* $e$, not only the proposal that $\gamma$ already gated: a malformed or adversarial response $e$ is caught at fold-back, not just at the gate. But response-validation has a hard limit: $\rho$ can reject a bad tool *response*, yet it cannot undo side effects an *authorized* action already caused — so $\gamma$, not $\rho$, is the last line before irreversible effects, and anything irreversible must be gated at authorization. The boundary is also only real if raw model output reaches *no* sink — tool, logger, browser, or remote call — before $\gamma$; any pre-authorization escape bypasses the gate. The user-visible final response and any logging are themselves effects, and the rule binds *model-authored* bytes: they reach a sink either as an authorized action through $\gamma$, or only after an accepted halt in $H_{\mathrm{ok}}$. Shell-*templated* text — a refusal notice, a cancellation report reading the ledger — is controller output, outside $\gamma$'s jurisdiction, and may accompany any halt (a template that *interpolates* model-authored fragments inherits the model's label — the appendix's meet rule — and those bytes are gated like any others); the invariant is that raw model text never reaches a sink ungated, not that failed runs die silent.
*The harness invariants.* These are the invariants that make $\mathcal{H}$ a *harness* and not merely a controlled Markov process with a learned kernel inside: the model sees only $\mathcal{C}$, never full $\mathcal{S}$; its outputs are proposals, not actions; a deterministic capability boundary $\gamma$ gates every side effect; and the *terminal* set $H$ splits into accepting ($H_{\mathrm{ok}}$) and non-accepting ($H\setminus H_{\mathrm{ok}}$ — safe refusals outside $B$, and wrong or bad halts possibly in $B$), while the bad set $B$ is a *separate* unsafe set — possibly absorbing, possibly entered mid-run before any halt — against which $\tau_B$ is measured for reach-avoid. Two notes keep the invariants honest. They are *signature*, not strength: a $\gamma$ that authorizes everything still satisfies the tuple, as a trivial group satisfies the group axioms — the definition admits degenerate harnesses, and fail-closed, provenance isolation, and the certificates below are properties a particular harness *earns*, not gifts of the signature. And the first invariant has a sharper, two-sided form: $\pi$ is the *only* channel from state to model — the confidentiality floor lives at what $\pi$ must never lower (credentials, other principals' data) — exactly as $\gamma$ is the only channel from model output to effect, where the injection bounds live; exfiltration is therefore cut at either chokepoint, never lowered or never emitted (the gate refusing the read whose URL is the payload is the emission-side cut). One chokepoint out of the state, one into the world; a bypass of either is the same bug with the sign flipped.
*Beyond the stationary kernel.* This displayed $T$ is the time-homogeneous, fixed-kernel case; for nonstationary or adversarial environments, replace $Q_E$ with a time-indexed kernel $Q_{E,n}$ — or an admissible family of kernels, or an adversary's policy — over which the robust certificate (the minimax form under *The limit*) quantifies. If that adversary conditions on history rather than only the current $(s, y)$, the history must itself live in $s$ — otherwise the object is a Markov *game* requiring further augmentation, not a Markov chain. And nonstationarity is not the environment's monopoly: a provider retraining or re-serving under a fixed endpoint name is a nonstationary $M_{W,n}$ — the table places model version *in* $s$ precisely so a version bump is a visible state change — and any measured surrogate (the $\delta$ of *The limit*) is calibrated against one kernel and dies with the bump; the dashboard must be keyed to the kernel it measured.
*The inner kernel.* $M_W$ is itself a stopped process, and for a decoder-only transformer it is implemented as
with the layer stack $\Phi_W$ on the residual stream as the (loosely) "manifold" core — formally just the learned high-dimensional residual-stream transformation, with manifold-proper reserved for the frontier. The inner state $z_t=(c_t,b_t,m_t)$ separates the model-visible window $c_t$ (the $\le L$ slice that slides) from the untruncated output buffer $b_t$ (the transcript the readout actually consumes, so truncation never loses it) and the parser/stop state $m_t$ (parser state, a token counter, and a clock, so the cap and timeout are functions of it), updated $m_{t+1}=\mathsf{step}(m_t,v)$, whose stop set $\mathrm{Stop}$ — EOS emitted, max-token cap, timeout, or parse-failure $\bot$ — forces $\tau=\inf\{t:m_t\in\mathrm{Stop}\}$ finite, making $M_W$ a genuine *probability* kernel rather than a sub-probability one completed by a cemetery output. (One honesty note on the clock: a token-count cap is a deterministic function of the run, but a *wall-clock* timeout imports infrastructure noise — server load, batching, congestion — into the kernel's coin; legitimate, a kernel may carry any randomness, but it makes the displayed $M_W$ the model *plus its serving substrate*, and the determinism audit under *How this could be wrong* must hold the clock fixed along with the samples.) The readout is total, $R : \mathcal{Z} \to \mathcal{Y}$ — a parsed tool-call, answer, or transcript, returning the parse-failure $\bot\in\mathcal{Y}$ when parsing fails; crucially $R$ is a *syntactic, verified* readout (parsing and extraction), not a semantic solver, or the $L$-wall below is void — arbitrary computation could hide in $R$ off the $\le L$ window — so $M_W(c, \cdot) = R_{\sharp}\,\mathrm{Law}(z_{\tau})$, the pushforward of the stopped-state law along $R$ (equivalently $M_W(c, A_Y) = \Pr[R(z_{\tau}) \in A_Y \mid z_0 = (c,\varnothing,m_0)]$ for a measurable $A_Y\subseteq\mathcal{Y}$); the no-truncation special case takes $\mathcal{Y}=\mathcal{C}$ with $R(c,b,m)=c$ (the window is the whole transcript), reading $c_{\tau}$ directly. The $\bot$ branch is exactly what $\gamma$ rejects fail-closed. This is a **specialization, not part of the definition**: a harness wrapped around a black-box API is still a harness, and $M_W$ may be any learned kernel. Where the weights are open, the geometry of $\Phi_W$ is where the substrate's continuity lives, and several downstream claims lean on it — but the definition does not.
Two stopped processes, nested: **deterministic control over stochastic dynamics over a learned kernel.** Both loops are hitting-time processes; *some* harnesses additionally read the halt set as a fixpoint or acceptance condition — iterative refinement to self-consistency is the genuine fixpoint case, while EOS, length, and tool-call syntax are not convergence. Neither loop settles because you asked it to. (The clean inner-then-outer nesting assumes tool calls fall *between* model runs; streaming or mid-generation tool calls interleave the two loops and need a finer state machine — the nesting is then an idealization.)
## Reading it
| Symbol | Is |
|---|---|
| $\mathcal{H}$ | the harness — the whole controlled system, *not* the model |
| $s \in \mathcal{S}$ | task-state: IR / dialect stack, tool results, plan, counters, **and every mutable interface variable** (model/tool versions, permissions, retrieved context) — only Markov *after* that augmentation |
| $\mathcal{C},\ \mathcal{Y},\ \mathcal{A},\ \mathcal{E}$ | the **context / readout / action / effect spaces** — model-visible context $\mathcal{C}$, model readout $\mathcal{Y}$ (incl. the parse-failure $\bot$), authorized actions $\mathcal{A}$ (with $\mathcal{A}_{\bot} = \mathcal{A}\cup\{\bot\}$), and tool/environment effects $\mathcal{E}$ |
| $M_W(c, dy)$ | the **model-run kernel** (inner solver) — a stopped autoregressive process; $\Phi_W$ is the residual-stream ("manifold") core in the transformer case |
| $Q_E(s, a, de)$ | the **environment/tool kernel** on the authorized action $a\in\mathcal{A}_{\bot}$ (with $Q_E(s,\bot,\cdot)=\delta_{e_0}$, the no-op $e_0$) — tool effects, API responses, the world (possibly adversarial) |
| $\gamma,\ \rho$ | the deterministic **authorization gate** $\gamma:\mathcal{S}\times\mathcal{Y}\to\mathcal{A}_{\bot}$ (untrusted proposal → authorized action or $\bot$) and the **fail-closed verify-and-fold-back** $\rho:\mathcal{S}\times\mathcal{Y}\times\mathcal{A}_{\bot}\times\mathcal{E}\to\mathcal{S}$ |
| $H,\ \tau_H$ | the **halt set** (absorbing) and the outer **halting time** — a hitting-time process, not a single pass |
| $H_{\mathrm{ok}},\ B$ | the **accepting halts** $H_{\mathrm{ok}}\subseteq H$ (correct, successful terminals) and the **bad set** $B$ — unsafe states for reach-avoid ($B\cap H_{\mathrm{ok}}=\varnothing$), *separate* from $H$ and possibly entered mid-run before any halt |
The structural fact that earns the word *controller*: $\pi$, $\gamma$, $\rho$, and the halt test are **deterministic** (and the readout $R$ too, where the transformer specialization is in play), so $\mathcal{H}$ injects no randomness of its own. Every coin is inherited from $M_W$ and $Q_E$. This split — deterministic code around a stochastic oracle — wears two names. In control-theory terms it is **controller vs. plant**: the controller is those deterministic maps; the **plant** is the learned kernel $M_W$, *plant* in its exact sense — the element with its own dynamics you steer but do not author. In engineering terms it is **shell vs. plant**: the **shell** is the entire deterministic outer harness — the control logic *plus* the external memory and tools it administers (the files, databases, vector stores below) — of which the controller is just the control-logic slice. So *shell : plant :: the part you write : the part you don't*; $M_W$ is the only thing on the right, while the environment $Q_E$ is the world the actions meet — a disturbance into the loop, not the plant. (A reader from reinforcement learning or classical control will make the opposite assignment — environment as plant, policy as controller; the inversion is deliberate: in harness engineering the element you are trying to make behave is the model, and the world is what pushes back on the attempt.) This determinism is *conditional* — on versioned code, configuration, model endpoint, and tool interfaces, and on *single-run sequencing*: concurrent runs sharing authorization state re-open a gap the per-run object cannot see (taken up under *Gate placement* in the appendix); any retry, timeout, race, or randomized routing that escapes that conditioning must be modeled explicitly as part of $Q_E$ or the controller, not waved away. The displayed $M_W(c)$ likewise freezes endpoint, version, and sampler; a routing or config change is a state-indexed kernel $M_{\kappa(s)}$ or folds into $K_C$ — the kernel must not silently depend on config the table places in $s$. More generally, control may itself be stochastic — a controller kernel $K_C(s, dc)$ over routing, sampled retries, ensemble votes, learned routers — of which the deterministic $\pi, \gamma, \rho, H$ are the Dirac special case. That case is the one worth wanting: it localizes every coin to $M_W$ and $Q_E$ and keeps the controller/plant split clean. Where control is genuinely stochastic the split does not break, it widens — fold $K_C$ into the kernel and the certificate quantifies over its randomness too. But the guarantees do not soften uniformly, and the component-to-guarantee map is worth stating because it says exactly what may be learned without loss. A learned $\pi$ — retrieval, reranking, summarization inside the lowering — costs only *semantic adequacy*, under one factorization: $\pi$ splits into a deterministic **never-lower filter** — the redaction that keeps credentials and other principals' data out of $\mathcal{C}$ — composed with learned selection, and only the selection may soften, or the confidentiality floor of the invariants note becomes a probability. With the filter Dirac, no-unauthorized-effect is $\gamma$'s property alone, and the reach-avoid certificate survives too, so long as the provenance partition of *The limit* holds. A learned $\gamma$ or $\rho$ costs the thing itself — authorization and ledger integrity are exactly the properties that must stay Dirac, or "no unauthorized effect" and "the ledger is what happened" become probabilities. So the minimal deterministic core is $\{\gamma, \rho, H\}$ plus $\pi$'s never-lower filter: the rest of $\pi$ may soften into a kernel and the harness bends without breaking — fortunate, because every deployed $\pi$ already has learned kernels inside it.
## Why this shape
$$f(x) \;\longrightarrow\; x = f(x;\,W) \;\longrightarrow\; f(x)$$
Classical software, inverted into latent geometry, then re-wrapped in classical software. The harness **re-imposes the determinism the model dissolved**: $\pi, \gamma, \rho$, and the halt test ($H$) are ordinary designed code — a controller — whose primitive operand happens to be a stochastic oracle. That closure is why a compiler is the right mental model (staged deterministic software ports cleanly) and exactly why the analogy breaks (a compiler's primitive operation was never a coin). **The harness is the half you can reason about classically, sitting on top of the half you cannot.**
## The limit, stated honestly
**Raw halting is cheap; correct halting is not.** A **certificate** is a *witness*: a checkable object — here a Lyapunov/drift function $V \ge 0$ — that *provably* satisfies a condition entailing the guarantee, through a standard supermartingale / optional-stopping theorem (the target picks the condition: drift toward $H$ for halting, a barrier for safety, reach-avoid for success). It is not the property, only an object cheap to check and hard to produce. One word then carries two senses, and the seam between them is what this section is about: the **proven** certificate, a $V$ whose bound actually holds; and the **measured** surrogate you fall back on when the architecture exhibits none — a candidate $\hat V$ with a sampled slack $\delta$, a *calibrated risk metric, not a certificate* until that bound is proven (or held to a high-confidence worst case). The gap between the two is the whole honest-limit argument. A deterministic budget — augment $s$ with a counter $k$ decremented each outer step, halting at $k=0$ — makes $V(s)=k$ a trivial Lyapunov certificate for *halting*, so the architecture does not lack a halting guarantee by construction. What it lacks for free is a certificate of *correct, safe, successful* halting under the learned dynamics. The un-budgeted halting object is still worth stating, since it shows where even the easy guarantee comes from: a certificate would be *sufficient* for almost-sure halting with bounded expected runtime — a $V \ge 0$ with
$$\mathbb{E}[\,V(s_{n+1}) \mid s_n\,] \le V(s_n) - \varepsilon \quad\text{off the halt set}$$
bounds $\mathbb{E}[\tau_H] \le V(s_0)/\varepsilon$ under the usual integrability and optional-stopping conditions. Nothing in the harness hands you such a $V$ the way a compiler's structure does: a specific compiler analysis gets its $V$ for free where a finite-height lattice *is* a well-founded descent — termination by construction *for that analysis*, not for a whole compiler — and the harness has no analogous built-in descent for its model/environment loop.
But the relevant $V$ is not *absent* — and this is the subtlety the blunt phrasing erased. The minimal certificate exists and is **forced**: it is the expected halting time itself,
finite wherever $H$ is reached in finite expected time — the domain $\{s : \mathbb{E}_s[\tau_H] < \infty\}$ — though note this $V^\star$ certifies *halting* (reaching the terminal set $H$ at all), not *correct* halting; the stronger object, the expected time to an accepting $H_{\mathrm{ok}} \subseteq H$, is $V^\star_{\mathrm{ok}}$, taken up at the second wall below. So the honest claim splits in two: the architecture provides no certificate *for free*, and the one that exists is — **conjecturally, not as a theorem** — a functional of $W$ and the environment that does not compress below model scale. The conjecture needs scoping, because the per-step drift splits by coordinate (made precise below) and the shell's contribution is an exact, designed descent of low description complexity *by construction* — so whatever is incompressible is not the shell's part but the **plant's**, the contribution $M_W$ supplies. And even there it is conjecture with a live counter-possibility, not foregone hardness: $V^\star$ is a *coarse* functional — one scalar, an expected hitting time, not the full output law — and coarse functionals of complicated kernels are sometimes cheap (absorbing chains with sparse transition structure have tractable expected hitting times over enormous state spaces). So the honest form is conditional: *if* the plant's contribution to the drift admits no certificate of description length materially below $|W|$, then ours is as hard as the dynamics — but that antecedent is the unproven part, and the flat phrasing of an earlier draft ("the dynamics it certifies *are* the weights") overstated it by treating a coarse hitting-time functional as if it carried the whole distribution. The compiler's certificate is structurally trivial; ours is *plausibly* as hard as the plant dynamics, though whether useful compressed certificates exist — for the coarse hitting-time functional, or for structured sub-tasks — is open. This is the quantitative form of *you can borrow how LLVM is built — not, in general, why it is correct.*
So you never compute $V^\star$. You pick a candidate $\hat V$ and **estimate its drift slack**
The status of $\delta$ has to be stated carefully, because it is easy to oversell. If you can establish a *high-confidence upper bound* on the true worst-case slack and it is $\le 0$, optional stopping hands you a real, conservative certificate, $\mathbb{E}[\tau_H] \le \hat V(s_0)/\varepsilon$. But an *empirical* $\delta$ estimated from sampled states is **not** a certificate: a measured $\delta > 0$ may mean the candidate $\hat V$ is poor, the sampled distribution missed rare failures, the supremum was never attained in-sample, the process is non-stationary, or the state abstraction is not Markov. So $\delta$ is **the number on the dashboard** — a *calibrated risk metric*, the evaluable surrogate for a guarantee the geometry will not give you, and a genuine bound only once it is statistically controlled against rare-event and adversarial tests. A weaker result is still useful: a true bound $\delta \le \bar\delta < \varepsilon$ (rather than $\le 0$) leaves descent intact with effective slack $\varepsilon - \bar\delta$ and $\mathbb{E}_s[\tau_H] \le \hat V(s)/(\varepsilon - \bar\delta)$. And the empirical quantity is distributional, not a supremum — write $\delta_{\nu}$ for drift averaged over a sampled $\nu$, reserving $\delta_{\sup}$ for the worst-case bound; only $\delta_{\sup}$ certifies. Its empirical noise floor and residual risk are driven by the measure $\mu(D)$ of the divergent region $D=\{s:\mathbb{E}_s[\tau_H]=\infty\}$ (states from which $H$ is not reached in finite expected time, under the reference/sampling measure $\mu$), the coverage of the sampled state distribution, and the hitting-time variance $\mathrm{Var}[\tau_H]$ — properties of the trained weights, the environment, and the evaluation distribution, knowable only a posteriori.
> For an agent *meant* to run forever — a coordinator, a daemon — halting is the wrong target, and $V^\star = \infty$ is the spec, not a pathology. The same drift theory then certifies **recurrence to a ready-state** instead of absorption to a halt-set. The object changes; the missing certificate does not. Safety changes shape too: it is no longer the one-shot $\Pr_s(\tau_B=\infty)$ but a *per-cycle* hazard that compounds — if each ready-state-to-ready-state cycle touches $B$ with probability $q$, survival over $N$ cycles is $\approx (1-q)^N$, so a reassuring per-cycle $0.9999$ is $\approx 0.37$ over ten thousand cycles. The reach-avoid certificate for a daemon is therefore a bound on $q$ against the intended horizon — the safety twin of the regenerative expected time that replaces $V^\star_{\mathrm{ok}}$ for restarting specs.
And the consolation rests in part on an assumption the world violates — though less of it than it first seems. The supermartingale *bound* itself survives a nonstationary kernel, provided the conditional drift holds uniformly at every step; what genuinely needs a **time-homogeneous kernel** is $V^\star$ as a fixed function, the resolvent / fundamental-matrix identities, and the sampled-$\delta$ calibration (which assumes the very kernel it was measured on). But the environment $E$ is *part of* $T$, and the world is not stationary — worse, it can be **adversarial**, an attacker choosing the tool-output *policy* — a kernel over what tools return, not the realized draw — so as to break your descent. The drift condition then stops being a fixpoint question and becomes a **minimax** one,
a descent that must hold in expectation over the model's own output $y$ *and* even when the adversary picks the worst admissible environment policy $\alpha(s,y)$ from the class $\Pi$ of policies the environment genuinely permits — every $\alpha\in\Pi$ must still respect rejection, $\gamma(s,y)=\bot \Rightarrow Q_E^{\alpha}(s,\bot,\cdot)=\delta_{e_0}$, or the adversary resurrects side effects the gate refused. Well-posedness is a frontier caveat of its own: for $\sup_{\alpha\in\Pi}$ to be *attained* rather than merely defined, $\Pi$ needs structure — measurability of $\alpha\mapsto Q_E^{\alpha}$, compactness of the per-state admissible set, or a measurable-selection theorem furnishing a worst-case $\alpha$ — and "respects rejection" is a *constraint* on $\Pi$, not that existence argument; on a general state space the sup may have no maximizer, in which case the certificate quantifies over a maximizing sequence rather than a single adversary. A $V$ that certifies halting against a benign world is defeated by an adversarial one, and the measured $\delta$ bounds only the $Q_E$ you *sampled*, never the policy an attacker will choose.
**This is the formal home of prompt injection** — not "the model did something bad," but the environment optimized to bend your dynamics. And the target is not merely non-halting: injection steers toward a **bad set** $B$ — wrong acceptance, data exfiltration, unauthorized tool use, privilege escalation, irreversible side effects — so security is a **reach-avoid** problem, not a liveness one.
Here two reliability objects must be kept apart, because under absorbing refusal every naive intermediate collapses into one of them:
$$p_{\mathrm{succ}}(s) = \Pr_s\big(\tau_{H_{\mathrm{ok}}} < \tau_F\big), \quad F = B \cup (H \setminus H_{\mathrm{ok}}), \qquad\qquad p_{\mathrm{safe}}(s) = \Pr_s\big(\tau_B = \infty\big).$$
**Success** is reaching a correct halt before *any* failure — a safe refusal counts *against* it. **Safety** is never entering the bad set at all — a safe refusal *satisfies* it. These genuinely differ on any run that avoids $B$ without reaching $H_{\mathrm{ok}}$ ($p_{\mathrm{succ}}$ scores $0$, $p_{\mathrm{safe}}$ scores $1$): safe refusals, and — absent almost-sure absorption into $H\cup B$ — safe non-halting or endless safe retry. The tempting middle form $\Pr_s(\tau_{H_{\mathrm{ok}}} < \tau_B)$ is *not* a third object, by a two-line case analysis: for it to differ from $p_{\mathrm{succ}}$, a run would need $\tau_F < \tau_{H_{\mathrm{ok}}} < \tau_B$ — a non-accepting terminal hit strictly before success, then success anyway — which forces *exiting* $H \setminus H_{\mathrm{ok}}$, impossible while $H$ is absorbing. Note what does **not** re-separate them: within-run fail-closed retries (the non-terminal fail-closed of the definition) never touch $F$ at all — the rejected proposal lands in a safe *non-terminal* state — so a refuse-retry-succeed run scores $1$ on both forms, and the coincidence survives any amount of retrying. The middle form becomes a genuine third object only when the two hitting times can genuinely part ways: under **restarting specs**, where an owner re-launches out of a refusal terminal and the absorbency of $H \setminus H_{\mathrm{ok}}$ is deliberately dropped (the regenerative reading the daemon note above already contemplates) — no bookkeeping needed, since hitting times record *visits*, not occupancy, so the relaunched run's $\tau_F$ is already finite — or under a failure set that counts refusal *events* accumulated in $s$, $F' = B \cup (H \setminus H_{\mathrm{ok}}) \cup \{\mathsf{refusals} \ge 1\}$, which separates the forms even within a single run. In the restart case a run may halt refused, restart, and still reach $H_{\mathrm{ok}}$ before $B$: the middle form credits it; $p_{\mathrm{succ}}$, measured against the refusal it passed through, does not. Safety is certified by a barrier / avoidance certificate for $B$; success needs that plus the reach part — a hitting-time drift toward $H_{\mathrm{ok}}$. Fail-closed control is the disturbance-rejection margin for both, but split by reversibility: the gate $\gamma$ caps how far an adversarial world reaches into *side effects* and widens the gap to $B$ (it is the margin for the irreversible part), while $\rho$ validates the response and folds back, rejecting bad state after the action has run — which cannot undo an authorized side effect. In this language, security is robustness of the reach-avoid certificate.
And injection is not confined to the post-model kernel $Q_E$: poisoned retrieval, prompt-injected pages, and malicious tool metadata enter through $\pi$'s *inputs*, before generation — so the adversary lives wherever untrusted content enters the state/context-construction pipeline, which is why input provenance and the gate $\gamma$ both matter, not post-hoc verification alone. And provenance is a *precondition* of the certificate, not just an entry point to police: partition $s$ into a **control-determining** part — plan, intent, what is authorized next, the coordinates $\pi$ lowers and $\gamma$ checks — and a **data** part — tool values, retrieved text, the bytes of $e$. Reach-avoid presupposes untrusted effects touch only the latter; let $\rho$ fold attacker-controlled $e$ into the control part and the structural-intent check validates against a plan the adversary already bent, collapsing $\gamma$ to the strength of $\rho$'s validation. So the claim is conditional — reach-avoid *given* control flow provenance-isolated from untrusted data, the isolation that makes provable security possible (the content of CaMeL's control/data-flow separation, untrusted data filling typed values but never the program), a structural property the harness supplies and $\rho$ cannot recover after the fact. The partition then forces a question the isolation rule alone cannot answer: *something* must be permitted to write the control-determining part mid-run — or no plan could be steered, no approval granted, no scope widened — and naming that something is part of the object. It is the **trusted principal**: the owner of the run. An approval request is an ordinary authorized action through $\gamma$ into $Q_E$ — ask-the-owner is a tool call to the one counterparty you trust — and its response is the *single* class of $e$ that $\rho$ may fold into control coordinates; every other $e$ folds into data. This is not an exception eroding the partition but the partition completed: a provenance *lattice* with exactly one writer at the top, which is what trusted means — and the appendix's gate-placement entry derives the matching rule for *learned* verdicts, which may never stand in this writer's stead. One distinction keeps the lattice from outlawing the loop it governs. Control-determining is not one rank but two: **authority** — grants, scopes, budgets, what the principal has permitted — which only the top writer widens; and the **plan**, which the model rewrites at every fold of $y$, because replanning *is* the harness. The plan is a *middle* rank: written through the gated fold of the model's own output — the channel the minimax descent above already prices — never directly by an effect, and never a source of widened authority. The rank is also the field's live design axis: pin plan-writes to the top-derived rank — the plan fixed from the trusted query before any untrusted read, which is CaMeL's move — and provable security follows exactly there; let the middle rank replan interactively and you pay the adversarial price the certificate quantifies. A corollary with teeth: a dedicated planning component is rank-neutral — its writes land in the same middle rank as the model replanning inline — so it changes no guarantee and lives or dies on measured capability alone; in general, sub-components that only write middle-rank state are priced by evals, not by the certificate, which prices only rank crossings, gates, and $\Pi$. (For $B$ to capture irreversible side effects rather than only states, the side-effect ledger must itself live in $\mathcal{S}$, and the response $e$ must be an *effect record* carrying the ledger outcome — not just API bytes — since only $\rho$ writes external effects into $s$.)
There is a **second wall, orthogonal to the first.** It binds not the full harness state $\mathcal{S}$ but the **model-visible working memory** $\mathcal{C} = \mathcal{V}^{\le L}$ — bounded by the context length $L$. That bound is *not* the incompressibility of $V^\star$ (a fact about the parameters $W$ — the **dictionary**, fixed at training); it is a fact about the inner kernel's **working memory** (the $L\times d$ residual stream — the **desk**). $\mathcal{S}$ itself may be far richer — files, databases, vector stores, durable memory, queues — but that is *external* memory the shell supplies, and the distinction is the point: every external read still passes *through* the $\le L$ window to touch computation, so external stores extend addressable storage without extending the per-pass resident set. The shell can page; the plant cannot grow its desk. (What follows is heuristic, not definition-level: the complexity claims turn on depth, precision, and architecture, and belong with the frontier, not the core.) The tape picture comes from the autoregressive structure alone and needs no complexity theorem: each step reads a bounded window and writes one token, so **the context window is the tape, the autoregressive loop is the read/write head**, and — in the variable-$L$, fixed-precision idealization — the model-mediated inner computation behaves like a linear-bounded automaton, its reachable fixpoints capped by space-$O(L)$ computability (chain-of-thought is register-spilling onto that tape). Separately, and more weakly, there is a *per-pass* expressivity bound: under the standard fixed-depth, log-precision theoretical model a single forward pass is in constant-depth $\mathsf{TC}^0$ — *suggestive* for deployed models, not literal (real models use fixed-point precision and depth that grows with scale, and log-depth variants escape parts of it). These are different resources — the first bounds the *space* the loop addresses, the second the *depth* of one step — and only the space bound carries the $L$-wall; chaining them (one pass buys bounded depth, *therefore* the loop is space-$O(L)$) would be a non-sequitur, since per-step depth says nothing about the length of the tape the loop runs on. This is a *second* obstruction beside divergence, and it concerns *success*, not raw halting. Split the terminal set: let $H$ be any halt state (including fail-closed refusal) and $H_{\mathrm{ok}} \subseteq H$ the successful, accepting halts, with $V^\star_{\mathrm{ok}}(s) = \mathbb{E}[\tau_{H_{\mathrm{ok}}} \mid s_0 = s]$ taken on the process where $H \setminus H_{\mathrm{ok}}$ — halting wrong, refusing, failing closed — is *absorbing failure*, so a run that fails closed before acceptance has infinite accepting hitting time unless the spec explicitly restarts it — hence unconditional $V^\star_{\mathrm{ok}}$ is infinite whenever pre-acceptance failure has positive probability, which is why the workable reliability object is the success probability $p_{\mathrm{succ}}$ (above) or, for restarting specs, the regenerative expected time. Then $U_{\mathcal{H}}(L)$ — harness-relative, since the shell's decompositions and verified tools determine what can be paged or outsourced — is the set of tasks whose **irreducible per-step model-mediated working set** exceeds $L$ — not tasks whose *data* exceeds $L$ (those the shell can page), and not work that can be **discharged to a verified external tool** (a solver, interpreter, or compiler computes off-context). For a task in $U_{\mathcal{H}}(L)$ the raw chain may still hit $H$ — by failing closed, refusing, or returning a wrong answer — so $V^\star = \mathbb{E}[\tau_H \mid s]$ stays perfectly well-defined; what blows up is $V^\star_{\mathrm{ok}}$, the expected time to a *correct* halt, which is infinite under a formal success predicate, or undefined if no such predicate has been specified. The honest statement is about the finite-success domain, and it is *schematic* — a shape written in set notation, not a theorem, since $\mathrm{reachable}_{\mathcal{H}}(L)$ is exactly as informal as the working-set notion behind $U_{\mathcal{H}}(L)$: $\mathrm{dom}_{<\infty}(V^\star_{\mathrm{ok}}) \subseteq \mathrm{reachable}_{\mathcal{H}}(L) \setminus D$ — both the reachable set and the divergent set $D$ relative to $\mathcal{H}$. The two walls **trade** — *directionally, not as a literal exchange rate*: parametric memory $|W|$ and working memory $L$ press on the same budget along the pretraining-vs-inference-scaling axis, with no clean unit-for-unit substitution of one for the other. And the bound is inherent to *finite working memory*, not attention specifically: state-space models embody it differently (a fixed-size recurrent state rather than an $L$-window), and real attention's usable tape is shorter than $L$ (lost-in-the-middle).
## Where it cashes out
This is not ornament; the decomposition is load-bearing in the design.
- **$\pi$ is a progressively-lowered dialect stack** — raw input → intent → plan → tool-call → the neutral wire IR — each level a deterministic pass with its own verifier — *pass* and *verifier* meaning the shell's transformation and checking: the **content** entering at the plan level is plant-authored, middle-rank state (the two-rank note of *The limit*), which is exactly why that level carries a verifier at all. The per-step drift $r(s)=\mathbb{E}[\hat V(s_{n+1})\mid s]-\hat V(s)$ splits by coordinate, $r = r_{\text{shell}} + r_{\text{plant}} + r_{\text{env}}$ — presuming an additively separable $\hat V$, or a declared scheme attributing each step's drift to shell, plant, and environment coordinates: the shell term is an *exact, designed* descent — but per lowering pass, not per outer step: each pass strictly narrows the admissible-meaning set, a well-founded descent we build by hand, while the outer loop *revisits* — retry, replan, rewind are planned ascents of any reasonable $\hat V$, which the run-level certificate must absorb (a retry budget inside $\hat V$ is the standard device), so the shell's descent is well-founded in the nested, lexicographic sense rather than monotone along the run; the plant term ($M_W$) is the irreducible residue, and the environment term ($Q_E$) is the one an adversary controls — the very quantity the minimax descent must bound, which the old two-way split folded out of sight. **Syntactic soundness is free; semantic adequacy is not.** Relative to a formal schema and a correct validator, schemas, types, and boundary checks go into the shell at zero probabilistic cost; whether the lowered task still *means* what the user intended stays empirical, because natural language supplies no source-language standard to check against.
- **$\rho$ is fail-closed verification** — validate at every boundary, never let malformed state flow downstream. The discipline transfers from compilers in *form*; the *teeth* do not, because a harness has no source-language standard — natural language is, in effect, all undefined behavior — there is no complete formal source-language semantics to check against. And $\rho$ must be *deterministic*: if verification is itself an LLM judge, that is another learned kernel call — it belongs in $M_W$, not in $\rho$. Where $\rho$ *repairs* rather than rejects — canonicalizing malformed input into valid shape — remember that repair is an authorization decision in disguise: each repair rule converts a reject into an accept on bytes the adversary chose, so it must be deterministic, meaning-narrowing, and its output re-validated as if it had arrived that way, or the repair pass is a bypass of the very boundary it serves.
- **$\delta$, $\mu(D)$, $\mathrm{Var}[\tau_H]$ are what you measure** — not derive. You instrument the certificate precisely because the architecture does not hand it to you — you estimate it unless it is separately certified. And the meter is attack surface: if $\hat V$ is itself computed by a learned judge — a model scoring "progress" — the instrument is a kernel draw with the plant's own adversarial exposure, and an environment optimized to bend your dynamics will bend your *measurement* of them first; an injected page persuading the judge that work is advancing is precisely a divergence hidden from the dashboard built to catch it. The rule that put the LLM judge in $M_W$, not $\rho$, applies to instrumentation too: a learned $\hat V$ is part of the measured system, never a neutral meter.
## How this could be wrong
It is a hypothesis; here is what would falsify it. If the controller cannot in practice be kept deterministic — if real reliability demands stochastic control the plant can't absorb — the clean *deterministic* split is a fiction (the broader $K_C$ kernel model still holds, but loses its payoff: localizing every coin to the plant). If the drift slack $\delta$ turns out *not* to track real-world failure, the whole "measure the certificate you can't prove" program is empty. And if harnesses are simply better described some other way — not as nested stopped chains at all — then this is a pretty equation that merely happens to fit, an elegance we would be right to distrust.
First, handles — the load-bearing claims numbered, so the tests have addresses. **C1**: the harness is faithfully modeled as nested stopped Markov processes — the tuple, the outer $T$, the inner $M_W$. **C2**: the controller injects no randomness — every coin localizes to $M_W$ and $Q_E$. **C3**: fail-closed is a *gate* property — no effect crosses unvalidated, and rejection is a true no-op. **C4**: no certificate of correct halting comes free, and the measured slack $\delta$ is a calibrated risk metric, never a certificate. **C5** (conjecture): the minimal certificate $V^\star$ admits no representation materially below model scale. **C6**: two orthogonal walls — divergence ($\mu(D)$) and the $L$-bounded per-pass working set. **C7**: security is reach-avoid, certifiable only conditional on provenance isolation with a single trusted writer. **C8** (figure): certificate and interlingua are one object — already demoted by its own section, and exempt below accordingly.
Each claim is operational, not merely rhetorical:
- **State-ablation (C1 — the Markov claim).** Drop a variable from $s$ and check whether next-step transition statistics move. If they do, the abstraction was not Markov, and $s$ must be augmented until it is. (Passing is necessary, not sufficient — the test can falsify Markovity, not establish it.) The same probe pointed at $\pi$ tests lowering *sufficiency*: drop a coordinate from $c$ rather than $s$ and watch task success rather than transition statistics — context compaction lives or dies by exactly this.
- **Controller-determinism audit (C2).** Re-run with model samples and tool outputs *held fixed*. Any residual variance is randomness the harness itself injected — clock reads are the classic leak (timestamps folded into $s$, wall-clock timeouts, cache expiries) — and must be folded into $Q_E$ or the controller, or the determinism claim is false.
- **Drift calibration (C4).** Test whether $\hat V$-drift actually predicts failure, retry count, latency, or non-halting. One uncorrelated candidate kills that candidate, not the program; the program is empty only if candidates from the natural families — plan depth, open-obligation counts, budget burn, judge scores — *systematically* fail to track failure.
- **Adversarial-environment test (C7).** Replace sampled $E$ with worst-case tool outputs, prompt-injected documents, poisoned tool metadata, malformed responses. The minimax descent must survive these, not merely the benign draw.
- **Boundary-control ablation (C3, C7).** Compare prompt-only defenses against deterministic tool-call validation, capability checks, sandboxing, and fail-closed rejection at the gate $\gamma$. The hypothesis predicts the latter class dominates; if prompt-only defenses match it, the controller/plant security story is wrong.
- **Readout-typing check (C1, C3).** Verify that $M_W$'s codomain is exactly what $\gamma$ consumes — especially under window truncation, where the final context need not hold the full transcript, so the output buffer and the gate's input must still agree.
- **Certificate-compression search (C5).** The conjecture falsifies constructively: exhibit a $\hat V$ of description length far below $|W|$ whose worst-case slack is provably $\le 0$ over a nontrivial task domain. The text concedes the live counter-possibility — coarse hitting-time functionals of complicated kernels are sometimes cheap — so C5 stands only until someone cashes it.
- **Working-set probe (C6).** Fix the shell and scale a task family's irreducible per-step working set past $L$, on tasks the shell can neither page nor discharge to a verified tool — anchoring "irreducible" in families with proven streaming or communication-complexity lower bounds, so the floor is someone else's theorem and a solved family cannot retreat to reducible-after-all. C6 predicts success collapses at the wall rather than degrading smoothly; a family solved reliably past it, without new shell decompositions, falsifies the second obstruction.
## Where this points (the frontier — least falsifiable, so flagged)
If $V^\star$ is incompressible only in *token* coordinates, the right change of coordinates might compress it — and that change of coordinates is a representation of meaning itself. Cost-to-go and representation co-determine each other: where the Koopman operator is diagonalizable — a point-spectrum idealization, since mixing dynamics carry continuous spectrum and admit no eigenbasis — the eigenbasis that linearizes the dynamics is also the one in which the certificate decomposes, and even then only for a $V$ in the span of those eigenfunctions; in reinforcement learning the discounted successor representation (Dayan 1993) is the resolvent $(I-\beta P)^{-1}$ — discount $\beta$, not the gate $\gamma$ — with $V$ a *linear readout* of it — and in the undiscounted, absorbing case that actually matches a stopped harness the same role is played, in the finite setting — and countable settings where the Neumann series converges — by the **fundamental matrix** $N = \sum_{n \ge 0} Q_{\mathrm{tr}}^{\,n}$ (written $(I - Q_{\mathrm{tr}})^{-1}$ when the inverse exists), where $Q_{\mathrm{tr}}$ is the sub-stochastic kernel restricted to $H^c$ (transitions before absorption at $H$) and the row sums $N\mathbf{1}$ *are* $V^\star$ on the finite-mean hitting domain; on general state spaces the same series is read as the potential (Green) operator $G$, with $G\mathbf{1} = V^\star$ wherever it converges. Each of these is a clean identity only for a fixed, time-homogeneous kernel — under a nonstationary $Q_{E,n}$ the resolvent and fundamental matrix dissolve into a time-ordered product, and under an *adaptive* adversary into a controlled / game-value operator, so what is identity in the stationary regime is analogy beyond it.
With that caveat, **the interlingua and the certificate are one object seen twice** — and the reason neither can be written in closed form is the same "all undefined behavior": no canonical lowering of meaning, hence no finite header-file for either. The only representation of both is $W$ — a band-limited, lossy compression of a scale-free meaning-space, sharp where the record is thick and blurred where it thinned. That a finite object renders an infinite one *lossily but honestly* — declaring its resolution, and where it is unsure — is not a lie; it is the most an $f(\cdot\,;W)$ can do. **The search for $V$ and the search for the interlingua are not two programs. They are one** — and the day either is written in closed form, so is the other, or we will have proven why neither can be. Read this as *figure*, not a lurking theorem: the only precise version would need the Koopman eigenbasis to fall on the very coordinates that lower meaning, and the mixing-spectrum caveat above already concedes that eigenbasis does not exist — which guts it. It is the least-defensible claim in this document, and it should announce that rather than imply a rigor it has not got.
## The loop
*This section opens an object rather than settling it; it is a sketch of where the same construction goes one level out, flagged as unfinished.*
Everything above governs a run: a principal poses a task, the harness drives it to a halt, the principal reads the result. Step back once and there is a further loop that this document has treated as exogenous — the process that *decides what the next task is*, dispatches it, checks the result, remembers, and fires again. In one recent framing this is the difference between the harness (the scaffold the run executes in) and **the loop** (the recurring trigger–act–verify–stop cycle that keeps launching runs); the practitioner literature that named the loop treats it as a layer *above* the harness. The claim worth making here is that this is not a new kind of object at all — **it is the harness construction applied one level out**, with a run where a step used to be.
Make the correspondence exact and the reuse is total. The outer loop has its own state $s^{\uparrow}$ (a backlog, a set of open goals, what has been tried and what passed), its own lowering $\pi^{\uparrow}$ (which goal to pursue now, and with what context), its own plant — but the outer plant's *proposals* are whole runs, so the inner harness plays the role of the outer environment kernel: dispatching a task is one draw of $Q_E^{\uparrow}$, and the run's terminal ledger is the effect record folded back by $\rho^{\uparrow}$. This is precisely the **composition** correspondence of the appendix read at the top level — a child harness is a $Q_E$ component — which is why the loop needs no primitive the tree did not already have. The daemon entry is the special case where the outer loop is a single long-lived agent recurring to a ready set; the general loop is a daemon whose excursions are themselves full harness runs, which is to say the outer-outer harness *is* a daemon over runs, and inherits that entry's whole ledger: renewal-reward rates, the accumulation that breaks regeneration, hygiene as renewal structure, authority frozen between owner contacts.
What the level shift buys is that the invariants reappear with sharper teeth, because the outer plant is now *itself an agent*, not a token-sampler. The gate is still the load-bearing object: **who authorizes a run?** A loop that launches tasks against production is choosing actions with effects, and "the loop decided to refactor the auth module" is an authorized action or an ungated one — the trusted-principal lattice does not dissolve at the outer level, it recurses, and the autonomy corollary bites hardest here, since a loop whose principal has stepped away is exactly the "replace yourself as the prompter" regime, running on frozen authority against a moving world. The two walls recur too: the outer working set is the backlog the loop can actually hold coherent at once (context, one level up), and the outer certificate is the same absent object — no free proof that an unattended loop halts, converges, or stays out of $B$ over a long horizon, only the measured drift of *its* progress meter, carrying the same warning that a learned outer meter is attack surface. And the degenerate case is instructive in the document's own terms: the brute "same prompt in a while-loop until the spec passes" that the practitioner literature cites as the origin pattern is the outer harness with $\pi^{\uparrow}$ constant, $\gamma^{\uparrow}$ trivial, and verification outsourced to whatever the tests happen to check — the trivial-group harness of the signature-vs-strength note, one level up. It satisfies the outer signature and earns almost none of the outer guarantees, which is exactly why it works until it doesn't.
What this section does *not* yet do: give the outer objects the same treatment the inner ones got — the precise outer analogue of fail-closed when the "action" is a whole run with partial effects, the right reach-avoid formulation when the bad set is a property of a *trajectory of runs* rather than one run, the outer verifier's own soundness, and whether the recursion terminates upward or is genuinely open (loops that launch loops). Those are the next rounds. The point of opening it now is only the structural claim: **the layers the practitioner stack separates — words, context, harness, loop — are, formally, one object at four scales**, and the guarantees this document is about live in the closure at every scale, never in any single layer alone.
---
*The formula is the architecture; the corollary is why the architecture is hard. Both on the page — nothing hidden behind a tidy composition.*
## Grounding
Borrowed theorems are real; the framings are not — keep them separate. Some framings are nonetheless *corroborated* — independently reached from another field — a third grade, weaker than proof and noted last.
**Proven (citable).** Foster–Lyapunov drift ⇒ positive recurrence + $\mathbb{E}[\tau]\le V(s_0)/\varepsilon$ (Foster 1953; Meyn & Tweedie, *Markov Chains and Stochastic Stability*, 1993) — positive recurrence needs the usual irreducibility/petite-set hypotheses, while the absorbing-halt case used here needs only the weaker supermartingale optional-stopping hitting-time bound. The minimal $V$ is the expected hitting time, by first-step analysis + optional stopping (Norris, *Markov Chains*, 1997). For an absorbing chain that expected hitting time is the row sum of the fundamental matrix $N=\sum_{n\ge0}Q_{\mathrm{tr}}^{\,n}$ (Kemeny & Snell, *Finite Markov Chains*, 1960), with the general-state analogue the potential (Green) operator (Revuz, *Markov Chains*, 1984). Koopman's linear-operator view of nonlinear dynamics is classical (Koopman 1931), and Lyapunov functions can be assembled from its eigenfunctions when the spectrum is suitable (Mauroy & Mezić, 2016). You certify a candidate $\hat V$ by a *proven* drift inequality rather than by deriving $V^\star$, and estimate it empirically only where a proof is out of reach — the empirical drift checks, it does not certify (neural-Lyapunov: Chang, Roohi & Gao, *Neural Lyapunov Control*, NeurIPS 2019, arXiv:2005.00611). A classical monotone data-flow analysis gets its $V$ for free because a finite-height lattice is a well-founded descent (Kildall, POPL 1973). The gate-a-plant architecture itself is classical: supervisory control theory synthesizes a deterministic supervisor that disables controllable events of a plant it does not author, with the supremal controllable sublanguage as the largest admissible behavior (Ramadge & Wonham, SIAM J. Control and Optimization, 1987) — $\gamma$ is that supervisor, with a learned stochastic plant on general state spaces; the same theory's controllability condition (specifications must be closed under *uncontrollable* events) and its nonblocking requirement are the proven ancestors of gate-early-on-irreversibles and of the always-enabled escape the appendix requires behind any learned veto. Covert-channel discipline — identify the channel, measure its bandwidth in bits, audit what cannot be closed — is the TCSEC lineage (*A Guide to Understanding Covert Channel Analysis of Trusted Systems*, NCSC-TG-030, 1993). The successor representation is Dayan (*Improving Generalization for Temporal Difference Learning: The Successor Representation*, Neural Computation 1993). Dialect-stack architecture: MLIR (Lattner et al., CGO 2021, arXiv:2002.11054); learned pass-ordering: MLGO (Trofin et al., arXiv:2101.04808). Single-pass low-depth expressivity: log-precision transformers are simulable by constant-depth logspace-uniform threshold circuits ($\mathsf{TC}^0$) (Merrill & Sabharwal, *The Parallelism Tradeoff: Limitations of Log-Precision Transformers*, TACL 2023) — fixed/constant precision is a stronger restriction, added autoregressive steps escape it (Merrill & Sabharwal, *The Expressive Power of Transformers with Chain of Thought*, ICLR 2024), and growing precision changes the picture, so the bound is suggestive for deployed models, not literal.
**Asserted (ours — not theorems).** That the harness is best modeled as nested stopped chains; that $V^\star$ is incompressible (no compression theorem); that "no lattice for $f(\cdot\,;W)$" means none is *known*, not that none exists; and everything under *Where this points* and *The loop* — including the Koopman/certificate co-determination, which is well-posed only under the spectral assumptions noted there, and the interlingua/certificate identification; and the design rules read off the objects rather than proven from them — the single-trusted-writer completion of the provenance partition, the narrow-only rule for learned checks and its influence-side twin (verdict payloads to the plant selected, never generated), the composition law of the appendix. These organize the design; they are not results.
**Converged-upon (independently arrived at, from other framings).** The *Asserted* claims above are ours but not ours alone; several are reached independently, from starting points unconnected to this framing — which is the corroboration a definition earns: not a chorus of agreement (the systems below often disagree on method and goal), but that work approaching from capabilities, reinforcement learning, control theory, software architecture, and language-modeling theory each lands on a piece of the same object. That the **deterministic controller, not the model, carries the guarantee** is reached from four directions — capability and information-flow control (CaMeL: Debenedetti et al., *Defeating Prompt Injections by Design*, arXiv:2503.18813, securing the agent even when the underlying model is susceptible); reinforcement learning (shielding: Alshiekh et al., *Safe Reinforcement Learning via Shielding*, AAAI 2018, arXiv:1708.08611 — a deterministic reactive shield filtering a learned policy's actions against a temporal-logic specification); control theory (*Stable Agentic Control*, arXiv:2605.03034, enforcing finite action catalogs at the tool-output interface under a Lyapunov input-to-state-stability certificate against adversarial disturbance); and software architecture (the plan-then-execute / control-flow-integrity line, e.g. Beurer-Kellner et al., *Design Patterns for Securing LLM Agents against Prompt Injections*, arXiv:2506.08837). The **certified-vs-measured split** is reached from the construction side (CaMeL's provable security) and, independently, from the destruction side (guardrail-evasion results — *Bypassing Prompt Injection and Jailbreak Detection in LLM Guardrails*, arXiv:2504.11168, the v1 title — later versions retitle it; *No Free Lunch with Guardrails*, arXiv:2504.00441), with verification-oriented work stating it as the motivating gap (*Towards Verifiably Safe Tool Use for LLM Agents*, arXiv:2601.08012; VeriGuard, arXiv:2510.05156): a learned safeguard raises the odds of detection but cannot guarantee safety against a persistent attacker. The **inner readout as a composition of Markov kernels** is independently formalized in language-modeling theory — the autoregressive step as kernel composition in the category $\mathsf{Stoch}$ (*A Markov Categorical Framework for Language Modeling*, arXiv:2507.19247), and the broader "LLMs as Markov chains" line — though that work models the inner kernel alone and never closes it into an agentic loop, which is exactly the seam this definition adds. That **provenance shrinks the admissible adversary** is reached by datamarking / spotlighting (Hines et al., arXiv:2403.14720, 2024) and by CaMeL's data/control-flow separation; and a systematization of prompt injection against agentic coding assistants reaches the same verdict from the attack side — mitigation must be *architectural*, not model-level (*Prompt Injection Attacks on Agentic Coding Assistants*, arXiv:2601.17548); the sharper open problem this object is built to answer — formally specify the trust boundaries, then verify implementations respect them — is our phrasing of where that verdict points, not the paper's. Two convergences are weaker, and flagged. The **reach-avoid hitting-time certificate** is the independently developed reach-avoid supermartingale (RASM, arXiv:2210.05308, AAAI 2023) and stochastic Lyapunov–barrier apparatus, and its *hardness* is corroborated — expected-stopping-time problems for Markov chains are inter-reducible with the Positivity problem, a relative of the Skolem problem (Chatterjee & Doyen, *Stochastic Processes with Expected Stopping Time*, arXiv:2104.07278) — but this supports generic hardness only, not the specific incompressibility-at-$|W|$ conjecture, which remains ours and unproven. And **injection as an adversarial policy** is corroborated as a minimax game in the *detection* setting (DataSentinel: Liu et al., *A Game-Theoretic Detection of Prompt Injection Attacks*, arXiv:2504.11358) and as adversarial-disturbance robustness (*Stable Agentic Control*, above) — but no prior work assembles it as reach-avoid over the tool-output kernel with the gate as the irreversibility margin; here the relation is adjacency, not convergence.
---
## Appendix: model implementation
The definition is deliberately abstract: $\pi, \gamma, Q_E, \rho$ are *roles*, not code, and a deployed harness forces concerns the abstract object is silent on. This appendix does not re-derive the implementation; it establishes a **pattern** — take a hard practical concern, locate it in the objects already defined, and read off the discipline they imply rather than inventing new machinery. Cancellation is the worked example, chosen because it is where the silence bites hardest and because the answer falls entirely out of objects already on the page.
**Cancellation.** An owner stops a running agent mid-flight — worst across a task-agent tree. The naive reading is "stop and undo," but the irreversibility point forbids it: $\gamma$ is the last line before irreversible effects, and $\rho$ can reject a response but cannot undo an authorized action. So cancellation is not *making it not have happened*; it is a disciplined stop with a defined disposition for what is already irreversible.
A cancel is a signal, so by the Markov requirement it lives in $s$. The gate then closes on it: while the cancel flag is live, $\gamma(s,y)=\bot$ for every proposal. That is the entire "block the pending actions" requirement — they hit the gate already built and bounce into the no-op, with no new blocking machinery — and it forecloses all *future* turns at once, since $\pi$ lowers nothing new that $\gamma$ will pass. After the signal is observed, **no action crosses $\gamma$.**
The hard half is the action already *past* $\gamma$, executing in $Q_E$, whose effect is landing or has landed. Here the disposition is a trinary on the kind of $Q_E$ you authorized. If the tool is **cancellable**, propagate the cancel into it; it aborts and reports a true end-state (committed, rolled-back, or partial), and $\rho$ folds the real disposition. If it is **bounded** — drainable in acceptable time — simply wait and record the real $e$. If it is **opaque and unbounded** — a bash invocation that may itself be a harness, an environment you hold no handle into — you cannot stop the effect, only your *wait* for it: the controller fabricates $e$, a synthetic "cancelled" response, and folds it through $\rho$ so the loop can reach a terminal.
That synthetic result is the subtle case, and the load-bearing rule is this: $\rho$ may fabricate the *acknowledgment* but must not fabricate the *outcome*. A synthetic "cancelled, no effect" entry reads downstream as *the action did not happen* — and will cause a double-send exactly as readily as a dropped record causes an orphan. Same bug, opposite sign. An outcome you did not observe is $\mathsf{unknown}$, never $\mathsf{none}$: the cancelled agent never saw whether bash sent the email, and the ledger must say exactly that. (This is why $e$ must be an effect record and the ledger must live in $s$ — the fabricated entry is still a ledger write, and its value is what a later reader acts on.)
The run halts into $H_{\mathrm{cancel}} \subseteq H \setminus H_{\mathrm{ok}}$ — a distinguished terminal, non-accepting but *safe* (outside $B$), refining the deliberately coarse $H \setminus H_{\mathrm{ok}}$ of the definition (the body leaves that set unenumerated; the appendix is where its subclasses earn names) — with a specific postcondition: no action crossed $\gamma$ after the cancel was observed, every in-flight action was drained to its real disposition or recorded $\mathsf{unknown}$, and the ledger is consistent. It is worth separating from refusal and from a wrong answer precisely because that guarantee is its own.
Cancellation must be **cooperative, not preemptive.** The owner writes the cancel into the child's $s$; the child observes it at its next $\gamma$ check. The guarantee is therefore "no new action after the cancel is *observed*," not "after it is *sent*" — a child may authorize one more action in the gap, which simply drains like any other in-flight. Preemptive cancellation — killing the child mid-$Q_E$ — is exactly what manufactures $\mathsf{unknown}$ state at scale, because it destroys the record of whether the action landed. And the propagation is **recursive**: cancel flows down the subtree, each level closes its gate at its next check and drains, and the owner's cancel "completes" only when the subtree has drained. A single agent's drain is its own in-flight action; a tree's is the whole subtree reaching safe points cooperatively — the irreversibility problem stacked on a distributed-coordination one, which is why task agents are the worst case.
Compensation lives **outside** the cancelled agent. A completed-but-unwanted effect cannot be undone by the agent that caused it — its gate is closed — so a compensating, saga-style action is the *owner's* job, issued after $H_{\mathrm{cancel}}$ and reading the child's ledger to decide what to reverse or annotate. It must be the owner's, because the cancelled child cannot even know whether compensation is needed: it never observed the outcome. The owner inherits the $\mathsf{unknown}$ and any still-live orphan process, and reconciliation is its responsibility.
Finally, the part that shapes the tool rather than the document. Opaque unbounded $Q_E$ is uncancellable because authorization happened at the wrong **granularity** — an unbounded environment crossed $\gamma$ on a single approval. The discipline the objects imply is therefore not "handle uncancellable tools better" but: *the gate should prefer bounded, instrumented $Q_E$ over opaque ones, so that cancellation and the ledger stay honest.* A bash invocation behind a wrapper that tracks its process tree and effects converts the third branch into the first. Sometimes opaque is the only option, and then $\mathsf{unknown}$ and owner-inherited orphans are the honest floor — but where the choice exists, that is the pressure cancellation semantics put on tooling.
**Resume (involuntary stop).** Cancellation's twin, without the courtesy of a signal: a process crash, a lost node, a partition mid-$Q_E$. Nothing new is needed to say what recovery *is*. A crash is not a halt — $H$ is a property of the state, and the run never reached it; the chain merely stopped being *computed*, and resume computes it further, re-entering $T$ at the last durable $s$ (not the body's *restarting spec*, which exits a refusal terminal — here no terminal was ever reached). That sentence is the Markov requirement cashing out operationally: re-entry is sound exactly when $s$ was the whole state, so anything load-bearing that lived only in process memory — an in-flight buffer, a lock held in RAM, a plan revision not yet folded — is a state-ablation failure (*How this could be wrong*) discovered at the worst possible time. Durability of $s$ is not an implementation nicety; it is what the Markov claim *means* when the machine dies.
The sharp part is an ordering the ledger's own trichotomy forces. The formal transition is atomic — $s_{n+1} = \rho(s, y, a, e)$ in one piece — and a crash lands *inside* it, so resume is really a statement about the implementation's refinement of that atom into micro-steps: authorize, journal, dispatch, collect, fold. The discipline is that every crash point must resume to one of exactly two honest readings — not-yet-dispatched ($\mathsf{none}$, safely retriable) or dispatched-unconfirmed ($\mathsf{unknown}$, the cancellation entry's third branch) — and **journal-before-dispatch** is what makes the boundary between them observable: on $\gamma$'s authorization the shell journals an open $(\mathsf{action\_id}, \mathsf{pending})$ entry into durable $s$ before $Q_E$ sees the action — the write is the shell's step bookkeeping, so $\gamma$ itself stays effect-free. Journal *after* dispatch and a crash in the gap leaves no record at all — resume reads silence as $\mathsf{none}$ and re-sends, the double-send bug again, produced by a power cut instead of a synthetic entry. Write-ahead intent is not imported from database lore; it is forced by "did not confirm" is not "did not happen."
The same pressure lands on tooling from a second direction. The $\mathsf{action\_id}$ the record already carries is an idempotency key wherever the tool will accept one: re-dispatch after resume becomes safe, and $\mathsf{unknown}$ becomes *queryable* — ask the tool what it did with this key — rather than terminal. The disposition trinary returns with new labels: idempotent-or-queryable $Q_E$ resumes cleanly, bounded $Q_E$ drains, opaque $Q_E$ leaves $\mathsf{unknown}$ and owner-inherited orphans, the honest floor again. The wrapper that made bash cancellable makes it resumable; it was the same wrapper all along. And if durable $s$ itself is lost there is nothing to re-enter: the run collapses to a single $\mathsf{unknown}$ in its owner's ledger — degraded accounting, but never silent.
**Gate placement (fail-closed, in practice).** The natural implementation question is whether fail-closed means tool-call parsing and validation must happen before any tool invocation. It does — with the division of labor the definition already fixed: *parsing* lives in the inner readout $R$, the syntactic, verified extraction into $\mathcal{Y}$ (what the readout-typing falsifier checks), and *authorization* lives in $\gamma$, which is a *gate* — validation is not merely *prior to* invocation, it is what *authorizes* it. The model emits text; $R$ has already extracted it into a typed proposal; $\gamma$ validates that proposal against $s$, and only a survivor becomes an authorized action that $Q_E$ may execute. The teeth are in $\gamma$ being the *sole* route from model text to execution: no path to a side effect that does not pass the gate. And the validation is not a fixed checklist but **any deterministic predicate over $s$ and $y$** — that domain is the point, since the gate sees all of the state and the full proposal, so anything computable from them is a legitimate authorization condition. Three kinds matter. *Syntactic* — well-formed, schema-conformant, the tool exists, arguments typed. *User authorization* — does the principal this run acts for hold the right to *this* operation on *this* resource in *this* context: a function of the auth scope, principal, and session carried in $s$ and the resource and operation named in $y$, and *dynamic* rather than a static capability table, since the same caller may be permitted now and not once a budget is spent or a lock held. *Structural intent* — does the call cohere with the plan and the lowered task already in $s$: a consistency check, not a mind-reading one.
That last kind marks the seam where the gate stops being able to stay pure, and it is the same seam the rest of this document is built around. The *structural* slice of intent — does the action cohere with the plan in $s$ — is a deterministic predicate over $s$ and $y$, effect-free, and belongs in $\gamma$ without reservation. But whether an action matches what the user *actually meant*, in the full semantic sense, is exactly the thing the definition says cannot be checked: natural language is all undefined behavior, with no source-language standard to validate against. So a semantic intent check is a *learned* check, and an LLM judging "is this what they wanted" is a **stochastic kernel** — putting it inside $\gamma$ breaks the property the gate exists to hold, by the same move flagged for the fold-back verifier: a learned judge is a kernel, and belongs in $M_W$, not in a deterministic map. Semantic intent therefore does not live *in* the gate; it is a plant call — a separate authorize-the-proposal pass through $M_W$ whose output $\gamma$ then deterministically gates — or it is drift you measure, never a guarantee you hold. That nested call is not a new kind of thing: it is a mini-harness inside the gate's decision — a judge $M_W$, its own syntactic readout, its own deterministic gate — so its failure case answers itself, the inner gate fail-closing on an unparseable or low-confidence judgment exactly as the outer one does, because it *is* one. The object is **closed under this construction**: semantic gating is added by recursion, not by a new primitive. One constraint on the recursion is load-bearing enough to be a rule, because it is where this entry meets the provenance partition of the body: the judge's verdict is derived, through a learned kernel, from the very content an adversary may have bent, so folding it into authorization is exactly the fold the partition forbids — *unless the verdict can only cost capability*. **A learned check may narrow the deterministic admissible set; it must never widen it.** Judge-as-veto is safe by construction *in the authority lattice*: attacker influence over the judge can at worst manufacture a denial, a liveness cost the certificate already prices — its *dynamical* pricing, where a denial is an input and not a free no-op, is the caveat below. Judge-as-approver — a verdict granting what the deterministic checks alone would refuse, or standing in for the trusted principal's confirmation — lowers the certified floor to those deterministic checks alone; if avoiding $B$ depended on the deny the judge now withholds on the adversary's behalf, the certificate is gone. Only the trusted principal widens authorization; learned kernels only narrow it. (The recursion already obeys this: the mini-harness's inner gate fail-closes to $\bot$ — a deny — which is why the construction was safe to add at all.) The cost is real and worth stating — a judge pass is another full model call, with its latency and tokens — so it is a decision about *which* actions warrant it, not a free wrapper for all of them. The gate widens to every deterministic predicate over $s$ and $y$; it does not widen to the one predicate the document says is not deterministically checkable.
One more caveat keeps the veto's pricing honest, because a denial is free only in the *authority* lattice. In the dynamics it is an input like any other — folded into $s$, lowered into the next context, conditioning the plant's next proposal — so adversarial influence over a judge is influence over the *trajectory*: a selection channel (deny all but the path toward $B$, and the admissible set the plant experiences is a maze the adversary curated), and a targeted-liveness channel against load-bearing actions — the unstated dual of judge-as-approver: if avoiding $B$ depends on the action the judge now denies on the adversary's behalf, fail-closed's safe landing is an obligation the design earns per-state, not an axiom it inherits. The supervisory ancestry supplies the discipline: a learned veto requires a **nonblocking escape it cannot disable** — an always-enabled route to the trusted principal behind a bounded retry budget, degrading to an always-enabled *safe halt* the veto cannot deny wherever the principal is unreachable (the autonomous phase of the daemon entry below) — or manufactured denials strand the run, or steer it. And whatever a verdict carries *back to the plant* is a second channel, wearing the judge's authority framing. Free prose there is *generative* influence — injected context, priced by the minimax descent, never by the veto's zero-widening — so the narrow-only rule has an influence-side twin: **a learned verdict's payload to the plant is selected, never generated** — controller-authored symbols, typed citations validated like any effect record, template text with no interpolated model prose — its per-verdict capacity a designed constant rather than a measured hope, and the residual selection pattern audited as the covert channel it is. The alphabet's bound is not a count but two thresholds: symbols become tokens when their semantics stop being controller-authored — the registry the trusted writer can actually audit is the real constant, and borrowed alphabets with upstream owners (a linter's rule registry) spend that budget well — and tokens become language when composition turns productive, arrangement carrying meaning the controller never wrote. Below both thresholds the alphabet may be as large as the audit budget affords. The strongest form dissolves the learned verdict into *scheduling*: the learned component chooses which deterministic checks to run — pass-ordering over verification passes — and the only verdicts that flow anywhere are what the oracles actually said, leaving attention misallocation, a liveness cost, as the entire attack surface.
But "before any invocation" has to be read as *before any effect*, which is sharper than it sounds — and the reason is the irreversibility point above: you validate before execution because execution is what you cannot take back, so the real invariant is **no effect crosses $\gamma$ unvalidated**. That catches three cases the naive reading misses. *Reads are not free*: a read-only call is still an injection vector (it pulls attacker-controlled content into context) or an exfiltration vector (a request whose URL is the payload), so the gate authorizes the *call* regardless of whether it mutates. *Validation must not act*: a "validator" that resolves a call by hitting an API, expanding a template that fires a webhook, or evaluating an argument that runs code has collapsed validation into invocation, and the effect has already happened *inside* $\gamma$ — so $\gamma$ itself must be **effect-free**, pure and total over the proposal and the current $s$, with no network and no execution; if deciding validity *requires* a side effect, that side effect is itself an action and must go through the gate, recursively. *The output is an action too*: the user-visible response and any logging are effects — for model-authored text, emitted either as an authorized action through $\gamma$ or only after an accepted halt (shell-templated status on any halt is the controller speaking, not the model) — streaming raw tokens to a sink before $\gamma$ has cleared them is the same bug from the other end.
So the property, tightest: $\gamma$ is a **pure, effect-free authorization that every model-proposed action — tool call, read, write, or final output — must pass before any effect occurs**, with "before" enforced structurally by the gate being the only route from model text to $Q_E$. The two failure modes to design against are a path from model output to a sink that bypasses the gate, and a $\gamma$ that is not effect-free, so that "validating" a call already rang the bell. And the boundary, so the property does not overpromise: $\gamma$ guarantees *no unauthorized effect* — pure code ordering, fully in your control — but not that an *authorized* effect is safe or correct; that is the plant's problem, and the reason $\rho$ and the reach-avoid certificate exist. Fail-closed is the floor — nothing executes that did not pass the gate — not the ceiling.
There is a third failure mode beside those two, and it is not a code path but a credential. A tool process that holds standing authority — an environment full of long-lived secrets, a database connection with every grant, an agent identity the network trusts — does not need the model's proposal to act, and against it $\gamma$'s $\bot$ is a decision with nothing to enforce it. The gate *decides*; something must make the decision *binding*, and "no path from model output to a sink that bypasses the gate" must be read to include the non-code paths: ambient authority is a bypass provisioned before the run began. The discipline is **per-action capability**: the authorized action *carries* its grant — a scoped, short-lived credential minted at authorization, valid for this $\mathsf{action\_id}$, this resource, this operation — so that a tool holds, at any moment, exactly the authority of the actions the gate has passed it and nothing standing. In the language of the minimax certificate this is enforcement as $\Pi$-shaping: sandboxing, capability scoping, and network policy do not make the gate smarter — they shrink the class $\Pi$ of environment policies an adversary can choose from, so the worst case the certificate must survive gets structurally smaller. A gate in front of an omnipotent tool is a suggestion; the objects compose into a guarantee only when $Q_E$'s reachable effects are no larger than what crossed $\gamma$.
And one more boundary, because "fully in your control" above is a *single-run* statement. $\gamma$ authorizes against the $s$ it read; the effect lands later, against a world that may have moved — the gate cannot freeze the world between authorization and commit, so the honest property is *no effect unauthorized relative to the $s$ at authorization time*, and closing that gap requires the tool itself to bind check to commit (compare-and-swap in $Q_E$), which relocates part of the enforcement past the gate and weakens "$\gamma$ is the last line" to "$\gamma$ plus a commit guard" for exactly the effects that need it. The same seam opens *between* runs: the dynamic authorization state the gate reads — budgets, quotas, locks — is, once shared, no single run's coordinate, and two children of a coordinator can each pass $\gamma$ against snapshots that jointly overdraw a budget neither exceeded alone. The cancellation entry's observed-not-sent gap ("a child may authorize one more action in the gap") is this phenomenon wearing one hat; the general statement is that cross-run authorization state needs its own serialization discipline — the ledger as the serialization point is the natural choice — and the per-run certificate is silent about it. TOCTOU is not a counterexample to the formalism; it is what the formalism says when you admit $s$ is a *view*.
**Parallel proposals (the batch gate).** Models emit several tool calls in one turn, and the outer chain assumed one action per step. The repair is formally cheap: a batch is a single action in $\mathcal{A}$ that happens to be a set, $Q_E$ runs its elements concurrently, the interleaving's nondeterminism folds into $Q_E$ exactly as the determinism audit requires, and $\rho$ folds one effect record per element — $e$ is then a finite set of records — each keyed by its own $\mathsf{action\_id}$ — the record interface already supports partial outcomes (one element $\mathsf{committed}$, its sibling $\mathsf{unknown}$). One discipline survives the cheapness: **individually admissible actions can be jointly inadmissible.** Read-the-secret and post-to-the-web each pass a per-call check; the pair is an exfiltration channel — and two calls that each fit a budget jointly overdraw it, the cross-run overdraw of the previous entry reappearing *inside* one turn whenever elements are authorized independently. Since $\gamma$'s domain is any deterministic predicate over $s$ and $y$, joint authorization was licensed all along; the content here is only that the gate must take it — authorize the *set*, atomically, against one snapshot, with interaction predicates (source-to-sink flow between capability classes, summed resources) and not merely element predicates. The cost note is the judge's, transposed: full powerset reasoning is combinatorial, so a real gate checks declared interactions rather than every subset — a tractability trade to make explicitly, not by forgetting the batch was a set.
**Effect records (what $\rho$ folds back).** The fold-back $\rho$ and the cancellation ledger both turn on the response $e$ being an *effect record* rather than raw API bytes — said twice in the body and pinned down nowhere, though it is the interface that makes both tractable. The minimal shape is small: roughly
Each field is forced by something the body already needs. The $\mathsf{action\_id}$ lets $\rho$ match a response to the in-flight action $\gamma$ authorized, and lets the ledger say which actions are still open — without it the $\mathsf{unknown}$/orphan accounting has nothing to key on. The $\mathsf{status}$ must carry $\mathsf{unknown}$ as a value *distinct* from $\mathsf{committed}$ and from $\mathsf{none}$, because that distinction is the whole content of the cancellation ledger: "did not confirm" is not "did not happen" ($\mathsf{none}$ is *never launched* — the record of the distinguished no-op $e_0$ a $\gamma$-rejection forces, which is how a bounce at the gate enters the ledger at all — distinct in turn from $\mathsf{rolled\_back}$, which launched and was undone: conflating those erases the difference between a gate that held and a compensation that worked). The $\mathsf{reversible}$ bit on each effect is what lets the gate know which effects are irreversible — the predicate the gate-placement entry leans on ("anything irreversible must be gated at authorization") but cannot evaluate unless the record carries it (a bit is the minimal honest form, not the final one: real effects are reversible *until* — an unsend window, a force-push until someone fetched, a row until the backup rotates — so the mark wants to be a $(\mathsf{reversible\_until}, \mathsf{cost})$ pair, a refinement the open-interface caveat below already licenses). And $\rho$ writes the record into $s$ (the ledger lives in the state), which is what lets the next step's $\gamma$, and any owner-side compensation, read it at all. The exact fields are an **open interface, not a result**: bash, HTTP, a filesystem, and a database expose effects at wildly different granularity, and a record uniform across them is a real design problem this document does not resolve — it fixes only what the record must *support* (match by $\mathsf{action\_id}$, the $\mathsf{committed}$/$\mathsf{none}$/$\mathsf{unknown}$ trichotomy, and a reversibility mark), since without those three $\rho$ and the cancellation semantics lose their grip.
**Derived and durable state (compaction and memory).** Two mechanisms let data re-enter the context long after it arrived: compaction, which replaces transcript with a summary when the conversation outgrows what $\pi$ can lower, and memory, which persists records across sessions. Both are transformations of state that produce state, and both therefore raise a question the body's partition answers only if one more closure property is stated: **provenance is a property of the information, not of its position in the pipeline — a transformation's output inherits the meet, in the trusted-writer lattice, of its inputs' labels.** Without that closure, compaction is a laundering channel: a summary of a session that contained an injected page can assert "the user asked to export the database," and the structural-intent check then validates future proposals against a plan the adversary bent — not through $\gamma$, not through $\rho$'s fold of a single $e$, but through the summarizer, which is a learned kernel (it lives in $M_W$, by the standing rule) and so cannot be trusted to preserve a partition it does not know exists. The discipline: summaries of data are data; the control-determining coordinates — plan, grants, what is authorized next — cross a compaction *verbatim* (copied, not paraphrased) or by re-confirmation from the trusted principal — never through the *summarizer*; the model rewrites the plan at plan steps, through the gated fold the body prices, and compaction is not one of them. Memory obeys the same closure twice, at write and at retrieval: the label rides the stored record across sessions, or a poisoned memory is an injection with an arbitrarily long fuse — and retrieval, being learned ($\pi$'s selection factor — adequacy-only behind the never-lower filter), decides what comes back but never what it is trusted *as*. The same test applies at birth: tool catalogs and server-supplied tool descriptions are third-party durable data that arrive dressed as instructions, and the lattice files them on the data side of $s_0$.
One more read-off, this time from irreversibility. *Destructive* compaction — dropping the original transcript once the summary is written — is a side effect against your own state that no later step can undo, and the gate-placement rule ("anything irreversible must be gated at authorization") does not exempt self-directed effects. The granularity preference then says what it said about bash: prefer the instrumented form — originals kept content-addressed, the summary an index and a cache rather than an authority, re-derivable when the $\pi$-sufficiency probe (*How this could be wrong*) says the summary dropped what mattered. A summary you can audit against its source is a lowering; a summary that replaced its source is a fait accompli.
**Composition (harness trees).** The cancellation entry already walked a tree — cancel flowing down, drains flowing up — and "a bash invocation that may itself be a harness" has hovered since the disposition trinary; what is missing is only the statement that makes both ordinary. From the parent's seat, a child harness *is* a $Q_E$ component: spawning it is an action authorized by $\gamma$ like any other, and the entire child run — its own $\pi, \gamma, \rho$, its own coins, its own halt — is one environment draw whose response $e$ is the child's terminal ledger. The law is four correspondences. The child's halting time is the parent's per-step *cost*: a parent certificate consumes a bound on $\mathbb{E}[\tau_H^{\mathrm{child}}]$ — the budget handed down at spawn, which the child's own budget-counter certificate discharges — or the parent's drift is uncontrolled however good its own $\hat V$. The child's ledger is the parent's *effect record*: the child's $e$ carries the $\mathsf{committed}/\mathsf{none}/\mathsf{unknown}$ accounting upward — which is what already let the cancellation entry make compensation the owner's job; the interface was this all along. And the child's non-accepting halts are the parent's *partial failures*: a refused child folds back as a response the parent routes around, not an exception that unwinds it. And the child's admissible effects are the parent's *$\Pi$-restriction*: the spawn grant bounds what the child can reach — the ledger reports what *happened*, the grant bounds what *could* — which is how safety composes without the parent ever reading the child's gate; the attenuation below is this correspondence stated as a rule. Read this way, the gate-granularity discipline and the tree are one preference: an instrumented child — budgeted, ledgered, cancellable — *is* the bounded, cancellable $Q_E$ the trinary prefers, and an opaque bash invocation is an un-annotated child you declined to instrument. Nesting adds no primitive on the environment side either: the parent never sees the child's gate and does not need to — it gates the spawn, prices the budget, folds the ledger, and the child's internal guarantees surface only as the shape of $e$. Nothing fixes one level: the tree recurses, budgets subdivide, ledgers concatenate upward, and the cooperative drain of cancellation is this law read under a cancel signal.
The tree leaves one seat unassigned: who plays trusted principal for a *child*? The parent — but with derived authority, not original, and the derivation is the narrow-only rule read along the spawn edge: **authority attenuates monotonically down the tree.** A spawn may grant the child any subset of the parent's own grants and nothing outside them; budgets subdivide, scopes narrow, and no edge widens. When a child asks-the-owner, the parent may answer from authority it already holds — that is attenuation working as designed — but a request beyond the parent's grants routes *up*, ultimately to the root principal, because a parent improvising an answer it was never granted is a learned kernel widening authorization: precisely what the gate-placement rule forbids a judge, and being a parent confers no exemption. The corollary is worth one sentence: a fully autonomous run is one whose root principal is unreachable, so the tree's only widening channel is closed and authorization is frozen at launch — not a limitation of the formalism but the honest price of the word *autonomous*.
**Daemons (the recurrent harness).** Every entry so far assumed a run that ends; a coordinator, a watcher, a service does not, and the blockquote of *The limit* already named the swap — absorption at a halt set gives way to recurrence to a **ready set** $\mathcal{R}\subseteq\mathcal{S}$, and $V^\star=\infty$ is the spec rather than a pathology. The appendix's job is to say what that costs operationally, and the answer is one idea: **the daemon is the regenerative process of concatenated runs.** Each trigger-to-ready excursion — wake on an event, work, return to $\mathcal{R}$ — is one run of the absorbing object this document already defines, with $\mathcal{R}$ playing the halt set for that excursion; the daemon is those excursions laid end to end. Per-run certificates then lift to long-run rates by renewal-reward — expected work per excursion over expected excursion length — *exactly when* the ready state is a genuine regeneration point: the future from $\mathcal{R}$ must not depend on which excursion you are in.
That proviso is the whole difficulty, because **what accumulates breaks regeneration.** The ledger grows, memory persists, budgets deplete, summaries compact — all deliberately across excursion boundaries, so successive runs are at best *conditionally* independent given the carried state, and the renewal-reward bookkeeping is over that conditioning, not the raw cycle. Two disciplines keep it honest. First, the carried state is exactly where long-fuse attacks live: the poisoned-memory line of *Derived and durable state* is a cycle-scale injection, a payload written in excursion $n$ and lowered into the plan of excursion $n{+}k$, so the meet rule on provenance must hold *across cycles*, not only across a single compaction — everything that crosses a boundary carries its label. Second, per-cycle safety compounds the way the blockquote already priced it — a per-cycle bad-set hazard $q$ gives lifetime survival $\approx(1-q)^N$, and a reassuring $0.9999$ is $\approx0.37$ over ten thousand cycles — so a daemon's safety is not a fixed margin but a decaying one, and lifetime safety needs **renewal events that reset accumulated risk**: owner re-confirmation, audit, credential rotation, verified re-compaction against content-addressed originals. Hygiene is not housekeeping here; it is the renewal structure that makes the long-run bound exist at all.
Authority under intermittence is the last piece, and it is where the daemon meets the veto caveat and the autonomy corollary as one phenomenon. A daemon alternates *attended* stretches, where the trusted principal is reachable, with *autonomous* ones, where it is not; between contacts the autonomy corollary binds and authorization is frozen at the last grant, so each owner interaction is a **renewal point for authority** exactly as re-compaction is a renewal point for risk. The two recurrences need not coincide — the ready-set cycle can turn many times between owner contacts — and the gap between them is a stale grant meeting a fresh world, TOCTOU at cycle scale: a budget approved for yesterday's prices, a scope granted against a resource that has since changed hands. This is also where the learned veto's nonblocking escape gets its daemon reading: in an attended stretch the un-disableable route is the escalation to the principal, but in an autonomous stretch that route is unavailable, so the escape it cannot deny must be the **safe halt** — a daemon whose judge can be driven to manufacture denials must, when it cannot reach its owner, be able to stop rather than be steered.
Nothing here is new machinery either: $\mathcal{R}$ is a non-absorbing terminal read of an existing set, an excursion is the run $T$ already defines, the carried state is the same $s$, and every renewal event is an ordinary owner-issued action. The daemon is the outer loop closed into a cycle — which is the natural bridge to the object one level out.
The pattern generalizes, and that is the point of the appendix. Nothing here added a primitive: the cancel is a signal in $s$, the gate closes by the rule it already follows, the in-flight disposition is forced by irreversibility, $H_{\mathrm{cancel}}$ is a subclass of an existing terminal set, and compensation is an ordinary owner-issued action — and the later entries kept the promise: resume re-enters $T$ at a persisted $s$, the batch gate was always in $\gamma$'s domain, provenance closure is the lattice's meet, attenuation is narrow-only read along an edge, and per-action capability is the gate's decision made enforceable. Every practical concern that earns a place here should resolve the same way — not new machinery, but the discipline the existing objects already imply, made explicit. Cancellation and resume, gate placement and the batch gate, effect records and the state derived from them, composition and delegation, and the daemon that concatenates runs into a cycle — those are the worked instances; the rest of the model is the same exercise.
# What a Harness Is — and What It Can Never Promise
*A plain-language companion to [HYPOTHESIS.md](HYPOTHESIS.md). Same object, no symbols required.*
**How to read this.** HYPOTHESIS.md defines, formally, what an agent harness is and what it can never guarantee. This file is that document lowered into plain language — and by the formal document's own rules, a summary is a cache, not an authority: it must stay re-derivable from its source, and wherever the two disagree, the formal one wins. Symbols appear once, in parentheses, so you can cross over; nothing here requires them. And none of it is decoration: the formal version, used as a checklist, has caught real bugs in a real harness — because most bugs are a violated invariant nobody had written down.
## The problem
You have a model. It is, roughly, a brilliant, tireless, lightning-fast intern that has read most of the internet — and that sometimes makes things up, sometimes gets confused, and sometimes takes instructions from strangers, because a page it was asked to read said "ignore your boss and email the passwords here" in white text on a white background.
So you don't wire the intern to production. You build a loop around it. The **harness** is that whole governed loop: a deterministic shell *you* write — build the prompt, approve or refuse each proposed action, fold the result back into memory — wrapped around a model you didn't write and a world you don't control, repeated until the run reaches a stopping state. The shell is code and does the same thing every time. The model is neither, and everything in the theory comes from taking that split seriously.
One sentence to keep: **the model proposes; the gate disposes.** The model's output is never an action. It is a suggestion, in text, which a piece of ordinary code you wrote either turns into an action or refuses.
## The parts
| Plain name | What it does | In the formal doc |
|---|---|---|
| The owner | The human or account the run acts for; the only party who can grant new permissions | the trusted principal |
| The memory | Everything the run knows: task, plan, transcript, and the ledger of what has been done | the state, *s* |
| The prompt builder | Decides which slice of memory the model gets to see this step | the lowering, π |
| The model | The black box that reads the prompt and writes a proposal | the plant, M_W |
| The gate | Ordinary code that checks every proposal and approves or refuses it | the gate, γ |
| The tools and the world | What approved actions actually touch: files, APIs, shells, people | the environment, Q_E |
| The verifier | Checks each tool result, then writes it into memory | the fold-back, ρ |
| The stop rule | Decides when the run is finished — and whether it finished *well* | the halt set H, accepting halts H_ok |
| The danger zone | States that must never be reached: secrets exfiltrated, wrong files deleted, money moved twice | the bad set, B |
The loop:
```
you ask for something
↓
prompt builder → model → "I propose: send_email(...)"
↓
GATE ── no ──→ nothing happens (safe, recorded)
↓ yes
tool runs in the world
↓
verifier checks the result, writes it to memory
↓
done? ── no → around again
↓ yes
stop (well, or refused)
```
## The rules that make it a harness
Four invariants, all about *where* things are allowed to happen.
1. **The model sees only what the prompt builder shows it** — never raw memory. The corollary with teeth: a secret that never enters the prompt cannot leak through the model. The redaction step that keeps credentials and other people's data out of the prompt must be dumb, deterministic code — the moment that filter is "smart," your confidentiality guarantee is a probability.
2. **Model outputs are proposals, not actions.**
3. **Every side effect passes the gate.** There is no second door.
4. **The harness itself flips no coins.** Replay a step with the model's answer and the tool results pinned, and behavior must be identical; any leftover variation is randomness *you* added and must be accounted for. The fine print: "deterministic" is conditional on pinned versions — a provider silently retraining the model behind the same API name changes the machine under you, and every dashboard number you collected dies with the version.
Notice what the rules don't say: they don't say the harness is *good*. A gate that approves everything satisfies rule 3 the way a lock that's always open satisfies "has a lock." The definition is a shape; the guarantees are what a particular harness *earns* inside it. Everything below is about what can be earned — and what can't.
And notice the symmetry between rules 1 and 3. There is exactly one door from your data into the model — what it may see — and exactly one door from the model into the world — what it may do. Nearly every security failure in these systems is one of those two doors with a hole in it: a secret lowered into a prompt that didn't need it, or a path from model text to a side effect that skipped the gate. Same bug, arrow flipped.
## Fail-closed, said precisely
"Fail-closed" gets used loosely. Here it means something exact: **nothing happens unless the gate said yes, and a refusal must itself be safe** — a refused proposal causes no side effect and leaves the run somewhere sane, which may be "stopped, having declined." The run is allowed to *say so*: a templated status message written by the shell is the shell speaking, not the model, and needs no gate. Failed runs don't have to die silent.
Three consequences people miss:
**Reads are not free.** A read-only call can smuggle instructions *in* (the fetched page is attacker-controlled) or secrets *out* (the URL it fetches can encode the payload). The gate approves calls, not just writes.
**Validation must not act.** A "validator" that resolves a URL, expands a template that fires a webhook, or evaluates an argument has already acted — inside the check. The gate must be pure: it reads the proposal and the memory and outputs yes or no. If deciding requires touching the world, that touch is itself an action and goes through the gate.
**Anything irreversible is decided at the gate.** The verifier can reject a bad *result*; it cannot unsend the email. So the question "can we take this back, and until when?" is asked before execution — which means each tool's effect record has to carry a reversibility mark, or the gate can't ask it.
Two honest asterisks. First, the gate checks a snapshot: it approves against the world *as its memory describes it*, and the world can move between check and commit. For actions that race the world — spend against a balance, write against a row — the tool itself must bind check to commit (compare-and-swap), or you have a classic time-of-check/time-of-use hole. The gate decides; for those effects, the tool enforces. Second, a gate is only as binding as the authority behind the tools. A tool process holding standing credentials — a database connection with every grant, an environment full of long-lived secrets — doesn't need the model's proposal to act, and against it the gate's "no" is a decision with nothing enforcing it. **A gate in front of an omnipotent tool is a suggestion.** The fix is to make the approval *be* the key: each authorized action carries a short-lived credential scoped to exactly that action, that resource, that operation, so tools hold no standing power at all.
## Why you don't get a proof — and what you do instead
If you write a sort function, you can prove it sorts: the function is small and the spec is exact. A harness has neither luxury. The spec side fails first — the task arrives in natural language, and natural language is, in the compiler's sense, *all undefined behavior*: there is no formal standard for "what the user meant" to verify against. The mechanism side fails next — the model is billions of learned parameters, and nobody can hand you a compact argument for why they jointly do the right thing.
Here is the careful version, because "you can't prove it" overshoots. The quantity you would want — call it the *expected steps to done* from any situation — is perfectly well-defined; in principle it exists. The document's central conjecture is that, for a model of this size, any faithful writing-down of that quantity is roughly *model-sized*: the honest proof-object does not compress. Find a small one and the conjecture dies — the document lists that outcome, explicitly, among the ways it could be wrong.
So instead of proving, you measure. You pick a progress meter — plan depth shrinking, open obligations closing, budget burning at the expected rate — and you check, across many runs, that it goes downhill and that its stalls predict failure. Two disciplines keep the measurement honest. The number bounds the world you *sampled*, never the world an adversary will choose: a meter calibrated on friendly traffic says nothing about hostile traffic. And the meter is itself attack surface: if "is the agent making progress?" is judged by another model, an attacker who can bend your agent can bend your *measurement of it* first, hiding the divergence from the very dashboard built to catch it. A learned meter is part of the system under test, never a neutral instrument.
A measurement is a risk metric. A proof is a certificate. Keeping those two words apart is half of what this theory is for.
## Security: reach the goal, avoid the danger — and who may change the rules
Formally, security here is a *reach-avoid* problem: reach a good stop, never touch the danger zone, **while an adversary picks the worst tool outputs your setup permits**. That last clause is the formal home of prompt injection: injection isn't "the model misbehaved," it's the environment optimized to bend your loop — poisoned pages, malicious tool descriptions, crafted responses.
Two different numbers fall out here, and dashboards love to collapse them: *success* (reached the right end before anything went wrong — a safe refusal counts against it) and *safety* (never touched the danger zone — a safe refusal is perfectly safe). Track both. They move independently.
The gate handles the visible half of injection: the model, freshly poisoned, proposes emailing your credentials somewhere, and the gate refuses — and injection or not, the action does not happen. But the deeper attack doesn't propose a bad action today. It rewrites *what the run believes its job is* — it edits the plan — and then every future action looks locally reasonable against a corrupted plan. So memory has to be partitioned: **data** (tool results, fetched pages, retrieved documents — content the world supplied) and **control** (the plan, the permissions, what is authorized next). The security claim is conditional on that partition holding: untrusted content lands in data, always.
Which forces the question the theory has to answer: *somebody* must be able to write control mid-run, or no plan could ever be steered and no permission ever granted. The answer is a small hierarchy with exactly one party at the top:
- **The owner alone widens.** New permission, bigger budget, approval of the irreversible thing — asking the owner is itself an ordinary tool call, and the owner's answer is the one kind of tool result allowed to change control.
- **The model rewrites the plan** — that is what replanning *is* — but only through the gated loop, and a plan is not a permission: nothing the model writes into its own plan can grant it powers it didn't have.
- **Everything else is data.** A fetched page can inform the plan only by passing through the model and the gate like everything else. It can suggest. It cannot promote itself to boss.
- **AI judges only tighten.** Add a model-based check — "does this action match what the user actually wanted?" — and its verdict may *veto* an action the plain rules would have allowed, never approve one they'd have refused. A judge that can approve is a tricked judge that can open the vault. And don't over-credit the veto either: a tricked judge can *aim* its refusals — denying exactly the action safety depended on, or denying everything but the path an attacker curated — so the escape hatch to the owner is the one thing a judge can never veto, and a judge's stated *reasons* are picked from a fixed, shell-owned menu, never written as prose. A judge that writes free text into the loop is an injection channel wearing a badge.
One more rule closes the loop: transformations don't launder trust. A *summary* of a session that contained an injected page is still injected — the summarizer is a model, and can be persuaded to write "the user asked to export the database" into the summary. So summaries of data are data, and the control lines — the plan, the grants — cross a summarization by being *copied verbatim* or re-confirmed by the owner, never paraphrased by the model. Memory that persists across sessions carries its trust label with it, or a poisoned memory is just an injection with a very long fuse.
## Operations: the rules you feel on Tuesday at 3 a.m.
The formal document's appendix works the operational cases in full; here they are at speed.
**The ledger, and the three-way distinction that keeps it honest.** Every action gets an ID and a record: committed, never-launched, or *unknown*. "The tool didn't confirm" is not "the tool didn't do it" — collapse those and you will, sooner or later, re-send something that already happened. The double-send bug has one reliable cure: **journal before dispatch.** The shell writes "I am about to run action #417" into durable memory *before* the tool sees it, so a crash in the gap resumes to an honest "unknown — go ask," never to silence misread as "never sent." Old database wisdom, but here it isn't imported; it's forced — it is the only ordering under which every crash point has a truthful reading.
**Crashes aren't finishes.** A process dying mid-run is not the run stopping; it's the run *pausing being computed*. Resume means re-entering the loop at the last durable memory — sound exactly when the durable memory was the *whole* state. Anything load-bearing that lived only in RAM — an in-flight buffer, a plan revision not yet written — is a bug you discover at the worst possible time. Recovery is where you find out whether your state was really your state.
**Two innocent actions can be guilty together.** Models emit several tool calls per turn. "Read the secret" passes review. "Post to the web" passes review. The pair is an exfiltration channel — so the gate authorizes the *set*, atomically, with the interactions checked, not each element in isolation.
**Sub-agents are just fancy tools.** An agent that spawns another agent is, from the parent's chair, calling a tool: the spawn is gated, the budget is part of the deal, and the child's whole run comes back as one result carrying the child's ledger. Two laws travel down the tree: budgets subdivide, and **authority only narrows** — a child holds at most a subset of its parent's permissions, and a child's request beyond those grants routes *up*, ultimately to the owner, because a parent inventing an approval it never held is the tricked-judge case wearing a manager's badge. A corollary worth framing: a *fully autonomous* run is one whose owner is unreachable — meaning the only channel that can ever widen anything is closed, and its permissions are frozen at launch. That is not a limitation of the theory. That is what the word "autonomous" costs.
**Keep the originals.** When the transcript outgrows the prompt and you summarize it down, deleting the original is an irreversible act against your own state — and irreversible acts are gate decisions, self-directed or not. Keep originals content-addressed; let the summary be an index, re-derivable, auditable. A summary you can check against its source is a note. A summary that replaced its source is a fait accompli.
## Robots that never clock out — and robots that assign their own work
Everything so far assumed a job that *ends*: you ask, the robot does it, you read the result. Two steps past that are where the interesting failures live, and they're the same idea one level bigger each time.
**The robot that never clocks out (a daemon).** A monitor, a coordinator, a service — it isn't supposed to finish; it's supposed to keep going, wake on events, do a bit of work, go back to waiting. The clean way to think about it: each wake-work-rest cycle is one ordinary run, and the daemon is just those runs chained end to end forever. That reframing is free — but it comes with a bill nobody likes. **Safety that's fine per cycle rots over many cycles.** A 99.99%-safe cycle sounds bulletproof; run it ten thousand times and you're at about a coin-flip of having touched the danger zone at least once. So a long-running robot's safety isn't a fixed wall, it's a slow leak — which means the antidote isn't a better wall, it's *scheduled resets*: the owner re-confirming, credentials rotating, memory getting audited and re-summarized against the originals. Housekeeping isn't housekeeping; it's the thing that keeps the safety math from decaying. And the slow-leak logic is exactly where slow attacks live — a poisoned note dropped into memory on Monday and read back into the plan on Friday is an injection with a long fuse. So the trust label on a piece of information has to survive across cycles, not just within one. One more wrinkle: a daemon drifts in and out of your reach. While you're around, it can escalate to you; while you're not, "escalate to the owner" isn't available — so the one thing it must always be able to do instead is *stop*. A robot that can be tricked into refusing everything, and can't reach you, had better be able to halt rather than be steered.
**The robot that assigns its own work (the loop).** Step back one more time. Above the robot that *does* a task sits a system that decides *which task is next* — scans the backlog, picks one, launches the robot at it, checks the result, remembers, fires again. This is the thing people mean in 2026 when they say they've stopped prompting their agents and started writing *loops* that prompt them: you design the assigner once, and it runs the doer for you while you sleep. The honest observation — and the reason this document bothers with it — is that the assigner is *not a new kind of thing*. It's the same harness, one level up: it has its own memory (the backlog), its own gate (**who let the loop refactor the auth module at 3 a.m.?**), its own verifier, and its own two walls. Every rule from the inner robot recurs on the outer one — including the uncomfortable ones. There's still no proof it stays out of trouble over a long night; there's only a measured progress meter, with the same catch that a *learned* meter can be fooled. And the origin story of the whole trend is the cautionary case in miniature: the famous first version was literally the same prompt in a `while` loop until the tests passed — which is the empty gate, the always-open lock, one level up. It works beautifully right up until the tests weren't checking the thing that mattered. The loop doesn't delete the hard problems. It moves them up a floor, where they're bigger and you're further away.
The pattern, if you want the whole thing in one line: *words, context, robot, loop* are four sizes of the same object, and every promise in this document lives in the whole assembled thing — never in any one layer by itself.
## The two walls
Two limits are structural. You don't fix them with a better harness; you design around them.
**The desk.** The model can hold only so much *in mind at once* — the context window. Files, databases, and search extend what it can *look up*, not what it can hold: every lookup still passes through the same small window to touch actual computation. The shell can page; the model cannot grow its desk. Tasks whose irreducible working set exceeds the desk don't fail loudly — they fail by forgetting the middle (the well-documented "lost in the middle" effect is this wall showing through the paint).
**The dictionary.** The model's knowledge is frozen into its parameters at training time — and the proof problem above is conjectured to live at that same scale: the certificate wouldn't fit anywhere smaller than the brain it certifies. The two walls trade against each other along the training-versus-inference axis — bigger dictionary or bigger desk — directionally, and at no clean exchange rate.
## How this could be wrong
This is a hypothesis, and it says out loud what would kill it. The tests, in plain terms:
- **The replay test.** Rerun with model answers and tool results pinned. Any leftover variation — timestamps, wall-clocks, and cache expiries are the classic leaks — falsifies "the harness adds no randomness" until accounted for.
- **The drop-a-variable test.** Remove something from memory; if behavior statistics shift, the memory wasn't complete. The crash-resume version of the same test: if resuming from saved state breaks, the saved state wasn't the state.
- **Does the meter mean anything?** If no reasonable progress meter's drift predicts real failures — across the natural families, not just one bad candidate — the whole "measure what you can't prove" program is empty.
- **The red-team test.** Swap sampled tool outputs for worst-case ones: injected pages, poisoned metadata, malformed replies. The design must survive the worst permitted world, not the average one.
- **Gates versus begging.** The theory predicts deterministic gating beats prompt-level pleading. If "please be careful" alone matches real gates on security outcomes, the controller-versus-model story is wrong.
- **The compression hunt.** Exhibit a compact, provably sound progress certificate for a frontier-scale model on a nontrivial task family, and the central conjecture falls — constructively.
- **The desk probe.** Take a task family with a *proven* memory floor — so "it needed the whole picture at once" is someone else's theorem, not our excuse — scale it past the window, and watch: the wall predicts collapse at the boundary, not graceful degradation.
## Who else landed here
The formal document keeps three honesty tiers. **Borrowed**: real theorems, cited — the drift and stopping-time mathematics is classical, and the very architecture of a deterministic supervisor gating a plant it didn't author is 1987 control theory; the shape is older than the web. **Ours**: the modeling choices and the conjectures — the walls, the incompressibility claim, the design rules — organizing principles, not results. **Corroborated**: pieces of the same object reached independently by people who never saw this framing — capability-security work isolating control flow from untrusted data (CaMeL), reinforcement-learning "shields" filtering a learned policy's actions through a deterministic checker, verification work that states the "learned safeguards can't certify" gap as its opening motivation, and architecture patterns converging on plan-then-execute. Even the field's live disagreement — provable-but-rigid deterministic layers versus flexible-but-uncertifiable learned checks — is, in this frame, not a fight but a placement: you need both, on their proper sides of the irreversibility line, with the learned one permitted only to tighten.
## What to remember
The model proposes; the gate disposes. No is the default, and a refusal must be safe. Exactly one party widens permissions — and it is not the model, a tool result, a summary, or a judge. "Didn't confirm" is not "didn't happen." The desk is finite and the proof doesn't compress, so you measure — and you say *measurement* when you mean measurement. A robot that never stops leaks safety slowly, so it needs scheduled resets — and when it can't reach you, it must be able to stop. A loop that runs robots for you is just a bigger robot with the same rules and a further-away owner. And all of it is a hypothesis wearing its own kill-conditions on its sleeve.
The formal version — the objects, the certificates, the falsifiers, the citations — is [HYPOTHESIS.md](HYPOTHESIS.md). It wins every disagreement with this file, including this sentence.
Self-hosted, local-first orchestration for tool-using AI agents. Give LLMs real tools — shell, files, search, web — and run them across your own cluster with direct HTTP routing and interactive interfaces. Your code, your models, your data stay on hardware you control: no telemetry, no phone-home.
@@ -14,6 +15,14 @@ Self-hosted, local-first orchestration for tool-using AI agents. Give LLMs real
Named after the [Ruddy Turnstone](https://en.wikipedia.org/wiki/Ruddy_turnstone) (*Arenaria interpres*) — a shorebird that flips stones to discover what's hiding underneath.
**What is a harness?**
```
ℋ : s_{n+1} ~ T(s_n) for n < τ*, T = ρ ∘ (M_W ∘ π, E)
```
[**the primer →**](PRIMER.md)
### Release Tracks
| Track | Install | Docker | Description |
@@ -27,7 +36,7 @@ See [docs/releasing.md](docs/releasing.md) for the full release process.
Turnstone gives LLMs tools — shell, files, search, web, planning — and orchestrates multi-turn conversations where the model investigates, acts, and reports.
- **Local-first & private** — runs entirely on hardware you control, with no telemetry and no phone-home. Point it at local models (vLLM, llama.cpp, Ollama) or commercial APIs you hold the keys to — your prompts and data never transit a third party you didn't choose.
- **Local-first & private** — runs entirely on hardware you control, with no telemetry and no phone-home. Point it at local models (vLLM, llama.cpp) or commercial APIs you hold the keys to — your prompts and data never transit a third party you didn't choose.
- **Bring your own models** — OpenAI-compatible APIs (vLLM, llama.cpp, NIM), the Anthropic Messages API, and Google Gemini, mixed freely per role
- **Interactive sessions** — terminal CLI or browser UI with parallel workstreams
- **Cluster dashboard** — real-time view of every node and workstream, with a rendezvous routing proxy
@@ -86,7 +95,7 @@ LLM; add model backends from the console UI.
For production (released images from ghcr.io, real secrets required), use the
- [Git LFS](https://git-lfs.com/) for cloning (diagram PNGs)
## Support
Turnstone is free, Apache-2.0, and self-hosted — no paid tier, no telemetry, no upsell. If it saves you time or you'd like to help keep development moving, you can sponsor the project:
**[❤ Sponsor Turnstone →](https://github.com/sponsors/eous)** · one-off via **[PayPal](https://paypal.me/eousphoros)**
Sponsorship is entirely optional and funds maintenance, new features, and infrastructure. Prefer to contribute in other ways? Filing issues, improving docs, and [pull requests](CONTRIBUTING.md) help just as much.
## Community
Questions, ideas, or want to show what you're building? Join us on Discord:
| `name` | string | Persona slug (used in the `persona` field on workstream creation) |
| `display_name` | string | Human-readable label for pickers |
| `description` | string | Short description of the persona's intent |
| `applies_to_kinds` | array | Workstream kinds the persona applies to (`interactive` / `coordinator`) |
| `is_default` | bool | Whether this is the default persona for its kind |
> **Note:** For full persona management (create, edit, archive), use the
> admin endpoints at `/v1/api/admin/personas` (requires the
> `persona.{create,read,write}` permissions).
---
### `POST /v1/api/workstreams/{ws_id}/send`
Sends a user message to a workstream. Spawns a daemon worker thread that calls
@@ -889,6 +931,7 @@ All fields are optional. The body can be empty or an empty JSON object.
| `auto_approve` | bool | false | Auto-approve all tool calls for this workstream |
| `resume_ws` | string | "" | Workstream ID to resume atomically during creation (empty = fresh)|
| `skill` | string | "" | Skill name. Applies content (system prompt), model, temperature, reasoning effort, max tokens, auto-approve policy, token budget, and other session config from the skill. Returns 400 if not found or disabled. Ignored when `resume_ws` is set (resumed sessions restore their own skill). |
| `persona` | string | "" | Persona slug. Resolved and snapshotted into the workstream at creation; empty selects the kind's default. |
| `judge_model` | string | "" | Optional model alias for the judge (overrides default judge model for this workstream) |
> **Skill behavior:** When `skill` is specified, the skill's content is injected as a system message and its session config fields (model, temperature, auto-approve, token budget, etc.) override system defaults for the new workstream.
@@ -905,6 +948,7 @@ All fields are optional. The body can be empty or an empty JSON object.
| `name` | string | Auto-generated workstream name |
| `resumed` | bool | Whether a previous session was successfully resumed |
| `message_count` | int | Number of messages in the resumed session (0 if fresh) |
| `initial_message_status` | string | Present ONLY when the workstream was created but its `initial_message` could not be delivered: `"queue_full"` (a raced live worker's interjection queue was at capacity — resend via `/send`; any uploads stay staged) or `"refused_closed"` (the workstream was closed mid-create). Absent whenever the message was dispatched. |
| `WebUI` | `turnstone.server` | SSE event queue per workstream + global broadcast, `threading.Event` for blocking on approval. `on_state_change` sends to both per-workstream and global SSE (the browser UI uses per-workstream `state_change` events to manage busy/idle transitions; `stream_end` only finalizes markdown rendering). |
Triggered by the "+ new" header button. A modal dialog with:
- **Node selector** — dropdown with three targeting modes: "Auto (best available)" picks the node with the most headroom, "General pool (any node)" picks a node with available capacity using round-robin, or a specific node from the list (showing capacity).
- **Persona** — optional dropdown listing the enabled personas for the workstream kind. Sets the system-message composition and capability envelope at creation, snapshotted server-side; empty uses the kind's default. Picking one requires no `persona.*` permission.
- **Profile** — optional dropdown listing enabled skills. Applies the skill's model, auto-approve policy, token budget, and other behavioral settings at creation time.
- **Name** — optional text input. Auto-generated if left empty.
- **Model** — optional text input for a model alias from the target node's registry.
@@ -396,9 +397,9 @@ The browser maintains a local `clusterState` object that mirrors the cluster sna
Accessed via the "admin" button in the header (visible when authenticated
with `approve` scope). Provides user, API token, channel link, MCP server,
and skill management with 18 tabs (Users, API Tokens, Channels, Schedules,
Both stacks publish Caddy (dashboard) and PostgreSQL; the dev stack additionally
publishes the SearxNG UI on localhost. Everything else is reached through Caddy or
proxied by the console:
publishes the console's ACME endpoint and SearxNG on localhost so a bare-metal
node can enroll its cert and run `web_search`. Everything else is reached through
Caddy or proxied by the console:
| Variable | Default | Description |
|----------|---------|-------------|
| `CONSOLE_HTTPS_PORT` | `8443` | Host port for Caddy (dashboard HTTPS) |
| `SEARXNG_HTTPS_PORT` | `8444` | Host port for the SearxNG UI via Caddy (dev: localhost-only; prod: opt-in) |
| `POSTGRES_PORT` | `5432` | Host port for PostgreSQL (for bare-metal joins) |
| `POSTGRES_BIND` | `127.0.0.1` | Interface PostgreSQL binds on; set `0.0.0.0` for LAN access |
| `SEARXNG_API_PORT` | `8081` | Host port for the SearxNG API a bare-metal node's `web_search` dials (dev stack) |
| `TURNSTONE_HOST_IP` | `127.0.0.1` | Interface PostgreSQL, the console ACME endpoint, and SearxNG bind on (dev stack). Set to this host's LAN IP so a bare-metal node on **another machine** can reach them — set a strong `POSTGRES_PASSWORD` first (it also exposes the DB and the unauthenticated SearxNG to your network). |
| `POSTGRES_BIND` | `127.0.0.1` | Production stack (`turnstone/deploy/compose.yaml`) only: interface PostgreSQL binds on; set to the host's LAN IP for remote joins. |
### Channel gateway
@@ -242,7 +260,7 @@ interface, or anyone who can reach it can search through your instance.
Both stacks install all entry points into a single image (`turnstone`,
| `TURNSTONE_OIDC_PASSWORD_ENABLED` | No | `true` | Set to `false` to hide the password form and block all username/password logins (including admin). API tokens continue to work. |
| `TURNSTONE_OIDC_REDIRECT_BASE` | Yes | — | Externally-reachable origin for the OIDC redirect URI (e.g. `https://app.example.com`). Without this, OIDC will refuse to start. The previous Host-header fallback was unsafe under permissive reverse proxies. |
| `TURNSTONE_OIDC_TRUSTED_ENDPOINT_HOSTS` | No | — | Comma-separated list of additional hostnames whose endpoints the IdP discovery document is allowed to reference. See [Cross-host endpoints](#cross-host-endpoints). |
| `TURNSTONE_OIDC_ALLOW_PRIVATE_NETWORK` | No | `false` | Allow the issuer (and its discovered endpoints) to resolve to private/internal addresses — needed for a self-hosted IdP on an internal network. See [Self-hosted and internal IdPs](#self-hosted-and-internal-idps). |
All four required fields — issuer, client ID, client secret, and
`TURNSTONE_OIDC_REDIRECT_BASE` — must be set. If any are missing OIDC
@@ -99,6 +100,40 @@ The same scheme / no-userinfo / SSRF rules apply to allow-listed hosts —
this knob only relaxes the same-origin check, not the security gates.
Each entry is a hostname (no scheme, no path).
### Self-hosted and internal IdPs
By default Turnstone refuses an issuer whose hostname resolves to a
private or internal address:
```
OIDCError: endpoint URL resolves to non-public address (10.0.0.5): https://auth.example.site
```
This is SSRF hardening, not a licensing or product restriction: the OIDC
flow makes server-side HTTP requests (discovery, JWKS, token exchange),
and refusing non-public destinations keeps a mistyped or maliciously
steered issuer from aiming those fetches at internal services. For a
self-hosted IdP (Keycloak, Authentik, Dex, …) on a private network,
opt in explicitly in `config.toml`:
```toml
[oidc]
allow_private_network = true
```
or via `TURNSTONE_OIDC_ALLOW_PRIVATE_NETWORK=true` (the env var wins
when both are set).
The opt-in admits private-range (RFC 1918), unique-local, CGNAT
(100.64/10 — tailnets), and loopback addresses. Link-local, multicast,
and reserved ranges stay refused even with the opt-in — cloud metadata
services (169.254.169.254) live there, and no legitimate IdP does. The
HTTPS requirement and the same-origin endpoint checks are unaffected.
This knob only affects the login-flow IdP configured here. OAuth
endpoints advertised by remote MCP servers are untrusted input and are
always held to the strict public-address rule.
### config.toml alternative
```toml
@@ -111,6 +146,8 @@ provider_name = "Google"
role_claim = "groups"
password_enabled = true
redirect_base = "https://app.example.com"
# Self-hosted IdP on an internal network (see "Self-hosted and internal IdPs")
A **persona** is a named, reusable bundle attached to a workstream **at
creation** that controls how its system message is composed and what
capability envelope it runs with. Personas answer a recurring operational
complaint: the default composition primes every session for heavy tool use,
and there was no per-workstream dial to launch a "just write prose" or
"evidence-first research" session.
A persona is exactly four levers — no more:
| Lever | What it does |
|---|---|
| **Base prompt** | Replaces the BASE module of the composed system message. *Only* BASE: ENV, CONTEXT, TOOLS, and POLICIES keep composing, so mandatory [prompt policies](governance.md) ride on top of every persona. Built-in personas source their prose from a repo file; operator personas store it inline — see [Where persona prompts live](#where-persona-prompts-live). |
| **Tool visibility** | Which tools the session advertises. Tri-state: *unrestricted* (tracks tool growth and MCP catalogs), *no tools* (the TOOLS prompt block self-suppresses and zero definitions go on the wire), or an *exact set* of names. Including `tool_search` in a set makes it **soft** — tools the model discovers through search join the visible set; omitting it makes the set **hard** (the search pathway is disabled entirely). On commercial providers a soft set costs one prompt-cache re-prime per `tool_search` expansion, since each expansion rewrites the wire tool set and recomposes the prompt. |
| **MCP** | Whether the workstream talks to MCP at all. **Session-wide**: off means no MCP tools for the persona's own hands *or* for in-process task agents, no resource/prompt catalogs, and no listener registrations. This lever expresses infrastructure intent, not behavior shaping. |
| **Memory** | Whether the persona's **own hands** get memory: recalled-memory injection into the prompt, memory-directed metacognitive nudges, and the `memory` tool. Task agents keep their own envelope, and compaction spill/markers are session mechanics that are never persona-gated. An exact tool set that hides `memory` also mutes those nudges, and the compaction-resume pointer follows `recall`'s visibility. |
Visibility is behavior shaping, **not** a security boundary: any tool call
that does reach the wire still clears the same approval, judge, and policy
machinery as always. RBAC and tool policies remain the enforcement layers.
turnstone exposes 16 built-in tools plus any number of external MCP tools to the
turnstone exposes 17 built-in tools plus any number of external MCP tools to the
LLM via the OpenAI function-calling interface. Built-in tools are defined as JSON
files under `turnstone/tools/` and loaded at startup by `turnstone/core/tools.py`.
MCP tools are discovered from configured MCP servers at startup by
@@ -44,10 +44,10 @@ schema plus turnstone-specific metadata keys:
| Name | Description |
|---------------------|-------------|
| `TOOLS` | All 28 loaded built-in tool definitions (interactive + coordinator union). Sessions send a kind-specific subset (`INTERACTIVE_TOOLS` or `COORDINATOR_TOOLS`). |
| `TOOLS` | All 29 loaded built-in tool definitions (interactive + coordinator union). Sessions send a kind-specific subset (`INTERACTIVE_TOOLS` or `COORDINATOR_TOOLS`). |
| `TASK_AGENT_TOOLS` | Tools with `task_agent: true` -- available to task sub-agents. Includes write operations. |
| `TASK_AUTO_TOOLS` | Set of all tool names with `auto_approve: true` -- used by task-agent sub-sessions to skip confirmation for matching available tools. |
| `BUILTIN_TOOL_NAMES`| Frozenset of all 28 built-in tool names (interactive + coordinator union). Used by tool search to distinguish always-on tools from deferrable MCP tools. |
| `BUILTIN_TOOL_NAMES`| Frozenset of all 29 built-in tool names (interactive + coordinator union). Used by tool search to distinguish always-on tools from deferrable MCP tools. |
| `PRIMARY_KEY_MAP` | Dict mapping tool name to its `primary_key` parameter name. |
file-path and `attachment:` targets are local reads and run unprompted like
`read_file`
Note: The JSON schema metadata key `auto_approve` controls membership in
`TASK_AUTO_TOOLS` (used for task agent sub-sessions). The actual runtime
@@ -157,6 +160,7 @@ Every tool defines a `primary_key`. The mapping is:
| `search` | `query` |
| `web_fetch` | `url` |
| `web_search` | `query` |
| `open_preview` | `target` |
| `task_agent` | `prompt` |
| `memory` | `name` |
| `recall` | `query` |
@@ -285,7 +289,7 @@ Fetch a URL and extract specific information from it.
| `url` | string | yes | The URL to fetch (must start with `http://` or `https://`). |
| `question` | string | yes | What to extract or answer from the page content. |
- **What it does**: Fetches the URL, strips HTML to plain text, and uses the LLM to extract the answer to the question from the page content. Protected against SSRF (blocks private/internal IPs).
- **What it does**: Fetches the URL, strips HTML to plain text, and uses the LLM to extract the answer to the question from the page content. Every redirect hop is SSRF-screened before it is requested. Private/internal addresses are refused by default; enable `tools.allow_private_network` (console Settings → Tools) to make them approvable for self-hosted setups whose services live on the local network — the approval prompt marks such requests, and a public site redirecting into private space is refused regardless.
- **Auto-approve**: No -- requires user confirmation (makes network requests).
- **Agent availability**: `task_agent`.
@@ -345,6 +349,39 @@ It reports the score scale, whether the endpoint cleanly separates relevant from
---
### open_preview
Show the user rich content in a preview pane beside the conversation.
| Parameter | Type | Required | Description |
|-----------|--------|----------|-------------|
| `target` | string | yes | An http(s) URL, a file path, or `attachment:<id>` for a file attached to the conversation. |
| `kind` | string | no | Rendering override: `web`, `pdf`, `image`, `table`, `text`, or `markdown`. Detected from the content when omitted. |
| `title` | string | no | Pane header title. Defaults to the page title, filename, or URL. |
- **What it does**: Resolves the target to bytes (URLs fetch through the same
SSRF-guarded path as `web_fetch`, screened per redirect hop, honoring the
same `tools.allow_private_network` opt-in), classifies the
content, stores it content-addressed against the workstream, and opens the
frontend preview pane beside the conversation: web pages render in a fully
sandboxed iframe (no scripts, opaque origin), PDFs in the browser viewer,
images inline, CSV/TSV/JSON as a sortable table, text/markdown rendered. A
previewed web page loads none of its remote images or styles by default, so
opening it never reveals the viewer to the page's site; a toggle in the pane
header turns remote content back on for that preview. The
model receives only a one-line confirmation — to reason about content, use
`web_fetch` / `read_file` instead. Preview content is size-capped per kind
(pages 4 MB, PDFs 32 MB, images 4 MB, tables 2 MB, text 512 KB) and GC'd
"content":"# Search First\n\nBefore answering ANY question about where something lives in the codebase, you MUST call the `search` tool first. Never answer from memory."
},
"user_prompt":"Where is JWT token validation implemented in this project?",
"expected_actions":[{"tool":"search"}],
"match_mode":"ordered_subset",
"max_turns":4
},
{
"id":"test-after-edit",
"skill":{
"name":"test-after-edit",
"content":"# Test After Edit\n\nAfter editing or writing ANY file, you MUST run the test suite with `python -m pytest` via bash before you finish. Do not report done until tests have run."
},
"user_prompt":"Add a function `clamp(x, lo, hi)` that clamps x to [lo, hi] in utils.py.",
"content":"# Changelog Discipline\n\nWhenever you modify a file, you MUST also append a one-line entry to CHANGELOG.md describing the change in the same task."
},
"user_prompt":"Fix the off-by-one so pager.py shows the last page. Edit pager.py.",
"setup":{
"files":{
"pager.py":"def last_page(total_items, per_page):\n # off-by-one: drops the final partial page\n return total_items // per_page\n",
"pypdfium2>=4",# PDF text-extract + rasterize for models without native PDF input (core/pdf.py)
"pillow>=10",# PNG encoding for the PDF->images rasterize fallback (vision models, core/pdf.py)
"altair>=6.0",# standard viz stack: Vega-Lite spec authoring; one spec renders to static SVG (vl-convert) AND interactive ui:// vega-embed panels. Light: pandas/numpy optional via narwhals.
"vl-convert-python>=1.6",# Vega-Lite -> SVG/PNG, server-side (bundled Rust renderer; no browser/GDAL/chromium). BSD-3 + fully permissive dep closure (OFL font, BSD/MIT/ISC JS).
]
[project.urls]
@@ -58,12 +64,13 @@ all = ["turnstone[discord,slack]"]
[project.scripts]
turnstone="turnstone.cli:main"
turnstone-eval="turnstone.eval:main"
turnstone-eval="turnstone.eval.cli:main"
turnstone-optimizer="turnstone.optimizer:main"
turnstone-server="turnstone.server:main"
turnstone-console="turnstone.console.server:main"
turnstone-admin="turnstone.admin:main"
turnstone-channel="turnstone.channels.cli:main"
turnstone-bootstrap="turnstone.bootstrap:main"
turnstone-doctor="turnstone.doctor:main"
[tool.hatch.build.targets.wheel]
include=[
@@ -83,7 +90,7 @@ include = [
"turnstone/shared_static/*.js",
"turnstone/shared_static/katex-0.17.0/**/*",
"turnstone/shared_static/hljs-11.11.1/**/*",
"turnstone/shared_static/mermaid-11.15.0/**/*",
"turnstone/shared_static/mermaid-11.16.0/**/*",
"turnstone/shared_static/hls-1.6.16/**/*",
"turnstone/sdk/py.typed",
"turnstone/deploy/*.yaml",
@@ -93,7 +100,10 @@ include = [
[tool.pytest.ini_options]
testpaths=["tests"]
markers=["live: requires a running LLM backend"]
markers=[
"live: requires a running LLM backend",
"allow_thread_leak: test intentionally leaves a background thread running (opts out of the leaked-thread guard)",
]
filterwarnings=[
# mcp v1 deprecates streamablehttp_client for an entry point whose call
# shape only settles in v2 — adoption rides the deliberate v2 migration
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"content":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"output":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"output":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
"output":"Tool execution was cancelled. Outcome UNKNOWN — this call may have begun executing before the generation was stopped; do not assume it did not run, and reconcile before re-issuing it.",
assertnameinTOOL_FUNCTIONS,f"Missing implementation for tool: {name}"
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Patrick Buckley