mirror of
https://github.com/turnstonelabs/turnstone.git
synced 2026-08-27 06:14:48 -06:00
perf/webui-transcript-windowing
20 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Patrick Buckley
|
110d44b07e |
refactor(tools): remove man, math, and plan_agent built-in tools
`man` and `math` duplicated capabilities already reachable through `bash`; `plan_agent` is better expressed as a `task_agent` running a planning skill, and carried a large amount of special-case machinery (plan-review gate, refinement loop, per-kind model routing). Removing all three shrinks the tool surface and cuts per-call token cost. Also removed, as dead-once-the-tools-are-gone: - the `math` sandbox executor (`turnstone.core.sandbox`) and its `[sandbox]` extra; the eval analyst now runs bash-only - the read-only `AGENT_TOOLS` sub-agent tool set and the `agent` tool-metadata key (`task_agent`/`TASK_AGENT_TOOLS` retained) - the plan-review protocol end to end: the `on_plan_review` UI hook, `resolve_plan`, `POST /v1/api/plan` + `POST /v1/api/route/plan`, the `plan_review`/`plan_resolved` SSE events, and their Python SDK / TypeScript SDK / OpenAPI / frontend / Discord+Slack bindings - the `model.plan_alias` / `model.plan_effort` settings and the registry `plan_model` / `plan_effort` routing fields TOOLS 31->28, TASK_AGENT_TOOLS 13->11; COORDINATOR_TOOLS unchanged. BREAKING CHANGE: removes the `man`, `math`, `plan_agent` tools, the plan-review SSE/HTTP/SDK surface, and the plan_* model-routing settings from the experimental 1.6 line. |
||
|
Patrick Buckley
|
b02e4312ac |
feat(tools): make notify dual-kind, expose to coordinator sessions (#559)
notify was interactive-only — a coord with a natural "fan-out complete" or "batch failed" beat could only post by spawning a child for the single message, which is a lot of ceremony. Routing is session-kind- agnostic in _prepare_notify / _exec_notify; this is a metadata flip that adds the coord flag (plus the explicit interactive flag the loader needs once coordinator is set) and updates the dual-kind whitelists, coord tool-set assertions, and skill-author docs accordingly. Adds two coord-session tests pinning the prepare dispatch contract (needs_approval=False matches notify.json auto_approve) and the exec → channel-gateway path. |
||
|
Patrick Buckley
|
471d48abd9 |
feat(skills): unify skill + list_skills into dual-kind action-multiplexed tool
Replaces the legacy `skill` (load + search) and `list_skills` tools with a single `skills(action=...)` tool serving both interactive and coordinator sessions. Stacks on the model.skills.write permission introduced in PR 1. Tool surface - `find`: filter by category/tag/risk_level/enabled_only/limit with optional BM25 query ranking; auto-approved on both kinds; kind-scoped at the storage filter (interactive sees interactive+any, coord sees coordinator+any). - `get`: fetch a single skill including content; cross-kind misses collapse to "not found" so a model can't enumerate the other surface by name-probing. - `load`: activate a skill in the current session (interactive-only; coord sessions get an explicit hint pointing at spawn_workstream). - `create`/`update`/`enable`/`disable`: require approval AND model.skills.write; permission re-checked at exec time to catch a revocation between approval and write. - No `delete` — hard-delete stays admin-UI exclusive; tool description documents the soft-delete-via-disable pattern. Defenses on the write surface - Approval cards surface projected risk_level (scanner re-run against the proposed final state) and warn explicitly when allowed_tools + auto_approve combine (auto-fire-on-load consequence is spelled out, not just shown as raw field values). - Toggle preview surfaces existing risk_level + allowed_tools count so re-enabling a critical-tier skill is never a one-click bypass. - Update path now re-fetches the row at exec to catch a readonly flip between approval and write, filters updates back to the runtime-only set if so, refuses if no fields survive. - Update path rejects empty content (hollow-out via emptying bypassed the soft-delete-via-disable invariant), non-list tags, and empty category — failures are loud rather than silent. - Permission denials audit `skill.write_denied` with actor_source=model so probing the permission state leaves a trail. Audit failures log at error (not warning) — a successful write without a row is the exact gap the trail exists to surface. - `_skill_hint` routes both message and system_reminder through escape_wrapper_tags so caller-controlled values can't close the <system-reminder> envelope and let the model fabricate directives in its own future context. Shared validation - `parse_skill_session_config` lifted from console/server.py to turnstone/core/skill_field_validation.py; both the HTTP admin path and the model-tool path consume it. Single source of truth so field rules can't drift between layers. - `SKILL_RUNTIME_CONFIG_FIELDS` lifted similarly (was duplicated as _SKILL_RUNTIME_CONFIG_FIELDS in server.py and _SKILLS_READONLY_FIELDS on ChatSession). - `notify_on_complete` validator now accepts list input from the JSON schema's `array` type — previously rejected because str() of a list yields Python repr that json.loads then refuses. Performance - Update prepare skips the projected-risk scan when neither content nor allowed_tools is changing (storage re-scans on write authoritatively). Metadata-only updates no longer pay the ~25 regex-pass scan cost. Cleanup - CoordinatorClient.list_skills deleted (-91 lines); model-tool path talks to storage directly via list_skills_filtered. - Roles admin UI gains a Model section exposing model.skills.write. - tests/test_load_skill.py renamed to tests/test_skills_tool.py and rewritten for the new tool — 48 tests covering registration, prepare dispatch, permission gating (including TOCTOU-revoked exec deny), audit actor_source on create + disable + permission-denied probe, BM25 ranking, invalid-kind branches, audit-failure swallow, and <system-reminder> envelope injection resistance. |
||
|
Patrick Buckley
|
9a30530d41 |
feat(coord): surface child errors, isolate tool exceptions, add memory tool (#443)
* feat(coord): surface child errors, isolate tool exceptions, add memory tool
Closes four coordinator gaps identified during operator triage:
1. Child workstream errors now surface in inspect/wait. Worker-thread
exception text is sanitized (URL userinfo masked, sk-/Bearer/ghp_/
github_pat_/AKIA tokens redacted, capped at 1024 chars) and persisted
to workstream_config.last_error before _emit_state("error") fires, so
coord polling never sees state=error with a missing cause. The row
is cleared on recovery transitions (idle/running) so a once-leaked
exception body doesn't outlive the failure. inspect_workstream and
wait_for_workstream return last_error for state=error rows; the
wait surface prefers it over the assistant-tail walk.
2. Tool exceptions now return as tool_results with sibling-aware
guidance. ChatSession._safe_prepare_tool wraps every per-call
_prepare_tool invocation; a buggy preparer becomes an error item
for that call only — sibling parallel tool_calls keep going,
never orphaning the assistant message's tool_calls block.
run_one's runtime exception path includes the exception class
and a short note that other tool calls in the batch completed
independently so the model can recover.
3. Memory tool exposed to coordinator with a coord-only scope.
memory.json gains coordinator: true + interactive: true + per-kind
kind_variants. Coord sessions see scope enum ["coordinator"] and
an orchestration-flavored description; IC sessions see ["global",
"workstream", "user"] and the existing flavor. Coord-scope rows
are private to the coordinator session (children cannot read or
write them), closing the cross-session prompt-injection lane that
an adversarially-steered child would otherwise have. Coord
visibility is also restricted to coord-scope only — coords no
longer see global / workstream / user memories that belong to the
user's interactive sessions.
4. Per-call exception isolation in tool batches. _safe_prepare_tool
was previously the implicit shield; now it's an explicit method
with documented invariants. KeyboardInterrupt / GenerationCancelled
re-raise so the cooperative cancel path still works.
Other notable changes:
- LAST_ERROR_CONFIG_KEY + persist_last_error / clear_last_error /
load_last_error / sanitize_error_text moved to turnstone.core.memory
(the storage facade hub) — readers in coordinator_client.py import
the constant.
- Memory scope tuples extracted to module constants
_VALID_MEMORY_SCOPES and _IMPLICIT_SCOPE_WALK; seven inline
duplicates collapsed.
- tools.py grows _apply_kind_variant for the per-kind tool surface;
tools without kind_variants pass through unchanged (no spurious
deep-copies).
- Session adds _coordinator_scope_id, _default_memory_scope,
_implicit_scope_walk, and _record_fatal_error chokepoints so the
worker-thread fatal path is one site rather than three.
- Removed duplicate on_error / on_state_change emits from
session_routes.py and coordinator_adapter.py — session.send()'s
_record_fatal_error owns the sequence now.
Tests: 4742 pass (no live), +30 net since the baseline. Ruff + mypy
clean on every modified production file.
* fix(coord): redact secrets in tool error paths via output_guard
Copilot review flagged two paths where ``str(exc)`` flowed back into
the model-facing tool_result without going through the credential-
redaction the new fatal-error path applies:
- ``ChatSession._safe_prepare_tool``: a preparer-side exception
becomes an error item whose ``error`` field embedded the raw
exception text.
- ``ChatSession._execute_tools.run_one``: a runtime tool exception
became an ``Error executing X: <e>`` tool_result, again with
the raw exception text.
Both now route through ``sanitize_error_text`` (sanitised log line +
sanitised tool_result), and ``sanitize_error_text`` itself was
refactored to delegate to ``output_guard.redact_credentials`` instead
of carrying its own parallel regex catalog — the audit log + post-tool
guard already use that pattern set, so the credential definition
stays in one place.
Also extended ``_RE_CONNECTION_STRING`` in ``output_guard`` to cover
``http(s)://user:pass@host`` so a misconfigured ``OPENAI_BASE_URL``
that lands in an httpx ``ConnectError.__str__`` is redacted by every
caller of ``redact_credentials`` (audit details, close-reason
persistence, last_error, the two tool error paths). The
host (useful for triage) survives; only the password is replaced
with the standard ``[REDACTED:password]`` marker.
Tests: full suite (4745 pass), ruff + mypy clean. Two new tests pin
the redaction behaviour in both tool error paths so a future refactor
can't drift back to leaking ``str(exc)`` verbatim.
|
||
|
Patrick Buckley
|
dea2729292 |
refactor(coordinator): rename task_list → tasks, doc/prompt sweep (#437)
Four themes from a coordinator-feature shakedown:
1. Correctness fixes (return shapes / examples / behavior)
- tools_coordinator.md: drop fake skill names from spawn examples;
fix wrong kwarg ``node_id=`` → ``target_node=``.
- wait_for_workstream.json: document ``message`` + ``truncated``
per-ws fields (always enriched in the client; the JSON shape
lagged the docstring).
- cancel_workstream.json: document the conditional ``dropped``
payload — ``was_running`` always present when ``dropped`` is,
``pending_approval`` and ``queued_messages`` conditional sub-shapes.
- spawn_workstream.json: document full return shape including
``routing_strategy ∈ {rendezvous, target_node, resume}`` and
``status``.
- close_all_children.json: clarify ``skipped`` covers BOTH
hard-deleted children AND already-closed-and-evicted children
(wire shape doesn't distinguish); drop incorrect "echoed back
in response" claim — server returns ``{status, closed, failed,
skipped}``, never echoes ``reason``.
- console/server.py: comment in ``_fanout_on_children`` clarifying
that the 400 "No session" branch fires for cancel-cascade
callers and is unreachable from close_all_children (close
handler 404s instead).
- coordinator_client._utc_now_iso(): switch to bare ISO format
matching the rest of the storage row format used in the codebase.
2. Tightened the 11 longest tool descriptions (~23% cut on the
coord set). Removed ALL-CAPS emphasis, normalised em-dashes,
dropped informal phrasing. No new claims.
3. Removed static approval annotations from descriptions.
Approval is governed at runtime by the unified ``approve_tools``
body and admin-defined ``tool_policies`` (#436); static
"Auto-approved" / "Approval required" / per-action approval
tags become a stale signal. Field names (``pending_approval``)
and operational verb behaviour ("cancel unblocks pending
approvals") stay.
4. Renamed ``task_list`` coord tool → ``tasks``. The previous name
compounded the bare word ``task`` (which collides with chat-template
channels on local models — same reason ``task_agent`` carries
the suffix); the plural form sidesteps the collision and reads
more accurately, since the tool acts on the whole list rather
than a single task. Sweep covers tool JSON, Python methods (5
client methods + 2 session methods + 1 helper + 1 constant),
audit event name (``task_list.update`` → ``tasks.update``), log
tag (``task_list.corrupt_envelope`` → ``tasks.corrupt_envelope``),
frontend SSE event matcher, prompts, docs, and tests. CHANGELOG
entry added.
Plus: dropped the ENV block (Output Environment / Available
rendering / Formatting principles) from coordinator system
prompts. Coordinators orchestrate rather than render rich output
to the user, so the rendering capability matrix is not actionable
for them. Coord prompt drops ~29% (6309 → 4493 chars).
SDK regeneration via ``generate-types.py`` updates both
``openapi-console.json`` (the rename's downstream change) and
``openapi-server.json`` (PR #436 drift — its merge added
``pending_approval_detail`` + ``recent_auto_approvals`` fields to
the Python schemas but didn't regenerate the JSON artifact).
## Behavior changes (operator-visible)
- Audit event name: ``task_list.update`` → ``tasks.update``.
Audit dashboards / SIEM filters / log greps that pinned the old
prefix should update.
- SSE ``tool_result`` events now ship ``name="tasks"`` for the
scratchpad tool. The bundled coord-tree UI is updated atomically;
external consumers reading SSE events by tool name need to update.
- Existing task envelopes in production storage have ``+00:00``
timestamps from the old ``_utc_now_iso``. New writes are bare;
old rows are not backfilled. Within an envelope you may briefly
see mixed formats until each row is re-touched. No code path
string-compares timestamps within an envelope, so this is
cosmetic.
## Validation
- ``ruff check`` + ``ruff format --check`` clean
- ``mypy turnstone/`` clean (175 source files)
- ``pytest -m "not live"`` — 4679 passed, 3 deselected
|
||
|
Patrick Buckley
|
7d61f9a37c |
feat(coordinator): phase 8 PR A — spawn_batch + close_all_children batch tools (#386)
* feat(coordinator): phase 8 PR A — spawn_batch + close_all_children batch tools
Adds two model-facing batch tools so a coordinator can fan out without burning one approval per child:
- `spawn_batch` — create up to 10 child workstreams in a single approval. Serialised
spawns so sibling ordering (by created_at) stays deterministic. Returns
`{results: {idx: {ws_id, name, node_id, status}}, denied: [{idx, reason}]}`.
Per-item validation / spawn failures surface in `denied[]`; the batch hard-errors
on >10 rather than silent truncation.
- `close_all_children` — soft-close every direct child in one approval. Server-side
Sem(16) fan-out via `coord_client.close_workstream`; `reason` propagates to every
closed child's audit + workstream_config. Response mirrors `stop_cascade`'s cascade
idiom: `{closed, failed, skipped}` where `skipped` is upstream-404 / already-gone.
Shipped infra:
- New console endpoint `POST /v1/api/coordinator/{ws_id}/close_all_children`
(gated `admin.coordinator`, `allow_service_bypass=False`, 512-char reason cap,
`coordinator.closed_all_children` audit).
- Shared `_fanout_on_children` helper — both `stop_cascade` and `close_all_children`
now delegate to it (one place to own the snapshot → semaphore-gather → bucket-split
skeleton).
- `CoordinatorClient.close_all_children(reason)` plus a `_post_url` seam that
`_post` now reuses (no more duplicated transport-error handling).
- `_emit_batch_event` — best-effort SSE emitter modelled on `_emit_wait_event`.
Emits `batch_started` / `batch_ended` pairs keyed by call_id. Throttled
`batch_progress` deferred to a follow-up.
- OpenAPI request + response schemas, endpoint spec entry, TS SDK regenerated.
- Persona doc (`tools_coordinator.md`) covers the two new patterns.
Bulk-endpoint shape policy (codified in PR C later): split by semantic category —
`{results, denied, truncated}` for bulk-read / bulk-create-with-payload (cluster/ws/live,
spawn_batch), `{<bucket>, failed, skipped}` for cascade-mutation (stop_cascade,
close_all_children). No retrofit needed on stop_cascade.
Tests: new `test_coordinator_close_all_children.py` (8 endpoint tests), expanded
`test_coordinator_tools.py` (session-side prepare/exec, coord_client=None guards,
batch SSE events), expanded `test_coordinator_client.py` (route map, client method,
transport errors), tool-count assertions updated.
Deferred (not this PR): per-item selective-deny approval UI, throttled batch_progress
SSE, coordinator-skills doc + bulk-endpoints doc (PR C), spawn budget / rate limit (PR B).
* fix(coordinator): address PR #386 copilot review
- coordinator_client.close_all_children: pass the unformatted path template
as log_path so telemetry aggregates don't fragment per session (ws_id
still lives in the real URL).
- session.py: drop dead spawned_ids accumulator in _exec_spawn_batch —
leftover from an eager-register path that got removed earlier.
- close_all_children tool JSON: document the 512-char server-side cap on
reason and that reason is echoed back in the response payload. Added
maxLength:512 on the schema property so the LLM sees the constraint.
- CoordinatorCloseAllChildrenRequest: add Field(max_length=512) so the
OpenAPI schema reflects the runtime 400-on-overflow constraint.
|
||
|
Patrick Buckley
|
553d73109b |
feat(coordinator): phase 5 — harness-test polish + wait_for_workstrea… (#378)
* feat(coordinator): phase 5 — harness-test polish + wait_for_workstream + judge fix
Closes the bug list surfaced by the 2026-04-17 coordinator harness test
plus the post-phase-4 wait_for_workstream ask, and folds in three
adjacent cleanups that landed in the same window. Tightens defense-in-
depth on the model-invoked mutating ops, fixes the LLM judge silent
no-op, kills the inspect-poll token burn, and rounds out a handful of
observability / docstring / spec gaps.
The session-factory pre-resolve at console/session_factory.py and
server.py was rewriting `judge.model` from an alias (e.g. `judge-mini`)
to the resolved underlying id (e.g. `gpt-5-mini`). IntentJudge then
checked `model_registry.has_alias(config.model)`, found nothing, and
fell back to the SESSION's provider/client with that bare model id —
silent `llm_fallback / "did not return a verdict"` whenever the
coordinator and judge alias resolved to different providers.
Pass the alias through unchanged; IntentJudge's existing alias-
resolution path picks up the matching client + provider. Validate
the alias exists so an obvious typo still surfaces, but don't replace
the model field.
Regression: `test_alias_uses_registry_provider_not_session_provider`
constructs an alias whose provider differs from the session's and
asserts the judge picks up the alias's provider/client/model;
`test_coordinator_tool_call_returns_llm_verdict_not_fallback` asserts
the verdict tier is `llm` (not `llm_fallback`) on the happy path.
New `cancel_workstream` tool (approval required, primary_key=ws_id) —
cancels in-flight generation, unblocks any pending approval / plan,
moves the child to idle, leaves the row in storage so a fresh
send_to_workstream lands cleanly. Re-uses the existing
`/v1/api/route/cancel` route + `route.cancel` audit namespace; no
new server endpoint.
`CoordinatorClient.cancel/close_workstream/delete/send` now enforce
a tenant guard inline (`_is_own_subtree`) — only the coordinator
itself or one of its own children is targetable. Foreign ids return
the same 404-shape inspect/wait_for_workstream use, so the model
can't distinguish foreign from missing (no existence oracle).
Defense-in-depth — the upstream node enforcement is the perimeter,
this is the second line.
`list_workstreams` advertised `state="deleted"` and an
`include_closed=true` that surfaced deleted rows. Hard-deletes
cascade the workstream + conversation rows out of storage, so
deleted is unreachable in normal operation. Doc-only fix; the
synthetic-test path that registers `state="deleted"` rows still
works (terminal-state filter still excludes them via
`_terminal_states = {"closed", "deleted"}` in list_children).
Documented that the 120s service-registry heartbeat window means a
node returned by list_nodes can drop out before a follow-up
`spawn_workstream(target_node=…)` lands — the spawn fails with "No
available node for routing" rather than falling back. Two-line
clarification on each tool. No code change (a code fallback is a
bigger discussion deferred to 1.6).
`close_workstream` accepts `reason`; the upstream server handler now
persists it to `workstream_config.close_reason` (capped at 512 BYTES,
sliced on UTF-8 not code points so a CJK / emoji-heavy payload can't
4× the documented budget). `CoordinatorClient.inspect()` reads it
and surfaces as `close_reason` in the result dict — only for
terminal-state children (closed/error/deleted) so the live-child hot
path doesn't pay a per-inspect DB round-trip.
Tests: server-side persistence covers success / no-reason /
length-cap / non-string / storage-failure / multi-byte-utf8 paths;
client-side surface covers terminal vs. live workstreams.
For idle children whose node-dashboard live counter is 0 (the live
block only surfaces in-flight token counters), fall back to
`SUM(prompt_tokens + completion_tokens)` from `usage_events` so the
inspect output reflects cumulative spend.
New `storage.sum_workstream_tokens(ws_id) -> int` on the protocol +
both backends. The fallback is folded INTO `_fetch_cluster_live` so
the merged live block (with persisted total applied) is what gets
cached — back-to-back inspects of an idle child amortize through
the existing 2s LRU cache instead of each firing a fresh aggregation.
`CoordinatorClient.list_skills()` now projects `allowed_tools` per
skill — capped at 20 with a `+N more` sentinel so a skill that
whitelists a wide MCP surface doesn't bloat the per-row payload.
Reads the existing `prompt_templates.allowed_tools` column; no
storage change. Coordinators no longer have to guess what tools a
skill brings.
`route_create` now sets `routing_strategy: "hash_ring" | "target_node"
| "resume"` on the spawn response so the coordinator's spawn
response (and the `spawn_workstream` tool output) carries why a
given node was chosen. 3 lines + 3 covering tests in
test_console_routing_proxy.py.
New coordinator tool `wait_for_workstream(ws_ids, timeout=60,
mode='any'|'all')` that absorbs the wait into a single tool call —
the model sees one call + one result regardless of how long the
children take. Kills the busy-poll inspect loop that burned 20+
turns on a 3-child fan-out.
Storage-poll loop with batched primitives —
`get_workstreams_batch` + `sum_workstream_tokens_batch` issue exactly
two storage calls per tick regardless of N. At the cap (32 ws_ids /
600s / 0.5s tick) that's ~2400 round-trips for a full wait, down
from ~38k under the naive per-id shape.
Validation single-source-of-truth: the client owns mode whitelist,
ws_ids dedup + cap, timeout coerce + clamp. The session preparer
is a thin pass-through that builds the header + dispatches; bad
input surfaces at exec time as a tool error via `result.get("error")`.
Tenant-isolation collapse: missing-row and cross-tenant cases both
return `state="denied"` so wait can't be used as an existence oracle
(matches the 404-mask contract `inspect` uses).
Prompt-side: tools_coordinator.md adds a `wait_for_workstream`
pattern + an explicit "PREFER wait_for_workstream OVER a loop of
inspect_workstream" line in the workflow-shape section.
Replaces the quote-bracketed substring LIKE/ILIKE pattern with proper
JSON-array containment. The previous shape effectively did
`LOWER(tags) LIKE '%"<lower-tag>"%'`, which broke for tag values
containing `"` (the JSON encoder escapes it to `\"` and the literal-
substring search misses), `\` (encoded as `\\`), or non-ASCII
characters that the encoder rendered as `\uXXXX`. Also exposed a
small spoofing surface — `tags=["foo\","bar"]` would have matched a
query for `bar`. Real-world tag values are alphanumeric+dash today
so it hadn't fired in production, but the fix is small.
- SQLite: `EXISTS (SELECT 1 FROM json_each(prompt_templates.tags)
WHERE lower(value) = lower(:tag))` (JSON1 extension; SQLite 3.38+).
- PostgreSQL: `EXISTS (SELECT 1 FROM jsonb_array_elements_text(
prompt_templates.tags::jsonb) AS jat(elem) WHERE lower(jat.elem) =
lower(:tag))`.
Three new tests prove the substring pattern was broken for
quoted / backslash / unicode tag values; the existing case-fold +
wildcard tests continue to pin the contract.
Phase 1 added the coordinator workstream API; phase 2 added only
`/open` to the OpenAPI catalog and missed every other coordinator
endpoint plus phase 3's `/children`, `/tasks`, and the
`/cluster/ws/{ws_id}/detail` aggregator. SDK consumers + operators
browsing `/docs` couldn't discover the surface. Doc-only addition:
12 endpoints + 9 new Pydantic models, all under the `Coordinator`
OpenAPI tag so /docs groups them together.
Sidebar re-fetches `GET /tasks` on every `task_list` `tool_result`
SSE event. A model that runs `add → list` (or any back-to-back
mutation pair) double-fetches the same envelope. Coalesced into
one fetch per 150ms window via a new `loadTasksDebounced` wrapper;
direct UI actions (refresh button, page load) keep calling
`loadTasks` directly so user clicks aren't delayed.
- `ruff check turnstone tests` — clean
- `mypy turnstone` — clean (157 source files)
- `pytest -m "not live"` — 4284 passed, 3 deselected (was 4226 on
main; +58 new tests across coordinator client, tools, judge,
storage, console routing proxy, server close-handler,
storage_skills_filtered, OpenAPI catalog, server close-reason
persistence)
- New tools added: 2 (cancel_workstream, wait_for_workstream) —
TOOLS count 28 → 30; coordinator subset 9 → 11; auto_approve adds
wait_for_workstream; primary_key adds cancel_workstream
- New OpenAPI endpoints: 12 (every phase-1/2/3 coordinator route +
the cluster-inspect aggregator)
- New storage protocol methods: 3 (sum_workstream_tokens,
sum_workstream_tokens_batch, get_workstreams_batch)
All phase 1 / 2 / 3 / 4 invariants preserved: COORDINATOR_TOOLS /
INTERACTIVE_TOOLS disjoint; coordinator sessions have no MCP surface;
list-style tools return {items, truncated}; route-proxy emits
route.<action> audit on 2xx; 404-mask on ownership failures; tenant
filters pushed into SQL; per-coordinator JWT carries scope context.
* fix(coordinator): address Copilot review on PR #378
Three valid Copilot findings on the wait_for_workstream surface:
1. ``wait_for_workstream.json`` description claimed the tool returns a
top-level mapping ``ws_id -> {state, tokens, updated}`` plus
elapsed/complete/mode at the same level, but the actual shape is
``{results: {ws_id: {...}}, elapsed, complete, mode}``. Description
now matches the implementation. Also adds ``deleted`` to the
advertised terminal-state list (it's in ``_WAIT_REAL_TERMINAL_STATES``;
the doc and runtime now agree).
2. ``CoordinatorClient.wait_for_workstream`` docstring listed
``idle / error / closed`` as the real terminal set but the constant
includes ``deleted``. Same fix — list ``deleted`` with a parenthetical
noting it's unreachable in normal operation (hard-delete cascades the
row).
3. Storage protocol docstring math: ``sum_workstream_tokens_batch``
claimed "from ~38k to ~1200" round-trips per wait at the cap, but
``wait_for_workstream`` issues TWO storage calls per tick
(``get_workstreams_batch`` + this one), so 1200 ticks × 2 = ~2400.
Updated to "~2400" with the math spelled out.
Also a clean rebase onto today's main (PR #377 — the rebalancer node_id
snapshot doc — landed since phase 5's last push). Single conflict in
``inspect_workstream.json`` resolved by keeping both notes (rebalancer
node_id binding semantics + the new ``close_reason`` surface from phase
5); ``spawn_workstream.json`` auto-merged.
The github-code-quality bot also flagged three items on
``_protocol.py`` asking to replace ``...`` with ``pass`` in Protocol
method bodies. Refuted: ``...`` is the canonical PEP 544 idiom for
Protocol method bodies and the rest of the file uses it consistently.
The bot's lint rule misfires for ``Protocol`` classes.
Verification:
- ``ruff check turnstone tests`` clean
- ``mypy turnstone`` clean (158 source files)
- ``pytest -m "not live"`` — 4308 passed, 3 deselected (no test count
change; pure doc/comment edits)
|
||
|
Patrick Buckley
|
8650370790 |
Feat/coordinator phase2 audit (#369)
* feat(coordinator): audit middleware on routing proxy — phase 2
Adds per-tool-call audit attribution to the multi-node routing proxy
handlers so coordinator → server hops land observable rows in
``audit_events``. Phase 1 preserved the ``src="coordinator"`` claim
through ``_proxy_auth_headers``'s upstream re-mint; this commit
closes the recording side. Was the last real security gap from
phase 1 — an enterprise deployment with ``admin.coordinator``
granted got only the three console-side
``coordinator.{create,close,cancel}`` rows; per-tool-call
attribution was missing.
## Action-naming scheme
route.workstream.create POST /v1/api/route/workstreams/new
route.workstream.send POST /v1/api/route/send
route.workstream.close POST /v1/api/route/workstreams/close
route.workstream.delete POST /v1/api/route/workstreams/delete
route.approve POST /v1/api/route/approve
route.cancel POST /v1/api/route/cancel
route.command POST /v1/api/route/command
route.plan POST /v1/api/route/plan
Action-name conventions documented in ``turnstone/core/audit.py``
module docstring alongside the existing namespaces — the docstring
is now ``<resource>.<verb>`` shaped (non-exhaustive) rather than
trying to enumerate every prefix.
## Recording rules
- ``record_audit()`` fires only on a 2xx upstream response. 4xx/5xx
are observable via ``_record_route``'s metrics path; doubling the
audit-events table size for failure rows would dilute signal
without giving operators much extra value.
- ``detail`` JSON carries ``{src, node_id, coord_ws_id?}`` — ``src``
lands verbatim from ``auth.token_source`` so non-coordinator
origins (``"jwt"``, ``"console-proxy"``) also get attribution;
``coord_ws_id`` only appears when the inbound JWT carried it.
- Wrapped in ``try/except`` + ``log.debug("route.audit_failed", ...)``
defence-in-depth. ``record_audit`` itself is fire-and-forget;
the outer try guards against a programmer error in the call site.
## Routing-proxy specifics
- ``route_create``: emits at the post-multipart/JSON convergence
``if resp.status_code == 200`` block. Both branches set
``audit_ws_id`` correctly — multipart from the query-string ws_id,
JSON from ``body["ws_id"]`` (post-503-retry) or ``body["resume_ws"]``.
- ``route_proxy``: emits the URL-method-mapped action. ``ref`` is
reassigned to ``new_ref`` after a successful 404→cache-refresh
retry so audit attribution uses the retried node, not the failed
first node.
- ``route_workstream_delete``: emits on 2xx using the ws_id from
the request body.
- ``route_attachment_proxy``: out of scope (upstream attachment
endpoints emit their own ``workstream.attachment.*`` rows;
auditing here would double-count).
## Tests
16 new tests in ``tests/test_route_proxy_audit.py`` covering:
- Coordinator-origin emission with full detail payload.
- 502 / 400 / 503-retry-final-node-id paths.
- Parametrised method→action mapping for the 6 ``route_proxy`` URLs.
- Plain-JWT origin (no ``coord_ws_id`` in detail).
- Delete handler 2xx + 502.
- Audit-storage exception swallowed (proxied response unchanged).
- ``auth_storage`` absent → no-op (existing route-handler tests
unaffected).
Verification: ``ruff check`` clean, ``mypy turnstone`` clean
(156 source files), ``pytest -m "not live"`` 4087 passed,
3 deselected (live-backend), 0 failed.
* feat(coordinator): discovery tools and /open parity — list_nodes, list_skills, POST /coordinator/{ws_id}/open
Adds the read-side surface coordinators need to make informed
orchestration decisions plus an explicit rehydration endpoint
matching the server's ``POST /v1/api/workstreams/{ws_id}/open``.
## list_nodes (auto-approved)
``list_nodes(filters={key: value, ...})`` reads ``node_metadata`` via
``storage.filter_nodes_by_metadata`` + ``get_all_node_metadata`` —
one query each, no N+1. Each row carries its full metadata dict so
the coordinator has both auto keys (``arch`` / ``cpu_count`` /
``fqdn`` / ``hostname`` / ``os`` / ``os_release`` / ``python``;
always present) and operator-supplied user keys (``capability`` /
``region`` / ``tenant`` / ``role``) without a second round-trip.
Tool description enumerates the auto keys explicitly so the model
knows what's always available vs deployment-specific.
Storage stores metadata values as JSON-encoded strings (the write
path in ``server.py`` / ``admin.py`` / ``console/server.py`` all go
through ``json.dumps``). The client re-encodes filter values
before the stored-text comparison and decodes stored values before
returning them to the model — so ``{"capability": "gpu"}`` is the
natural form the model uses, not ``{"capability": "\"gpu\""}``.
Ints round-trip as ints.
Returns ``{nodes, truncated}``; ``truncated=True`` when the page
was full.
## list_skills (auto-approved)
``list_skills(category?, tag?, scan_status?, enabled_only?, limit?)``
surfaces the skill registry so coordinators can discover worker
profiles. New storage protocol method ``list_skills_filtered(...)``
on both SQLite and PostgreSQL backends pushes filters into SQL.
``tag`` filter matches against the JSON-array ``tags`` column with
quote-bracketed substring (``%"tag"%``) — quote-safe against
``foo`` vs ``foobar`` collisions on both backends.
Returns ``{skills, truncated}`` with ``name`` / ``category`` /
``tags`` (decoded to list) / ``version`` / ``description`` /
``model`` / ``enabled`` / ``scan_status`` / ``activation`` — the
discovery projection, not the full row.
## POST /v1/api/coordinator/{ws_id}/open
Explicit rehydration endpoint. Lazy ``GET`` rehydration works for
the UI; this gives SDK callers and operators a way to warm a
coordinator without browsing to it. Same ownership / 404-on-
mismatch / correlation-id-masked error semantics as
``coordinator_detail``. Returns ``{ws_id, name, already_loaded?}``.
Registered in ``turnstone/api/console_spec.py`` with a dedicated
``CoordinatorOpenResponse`` Pydantic model so the OpenAPI schema
matches the wire shape.
## Tests
- ``tests/test_storage_skills_filtered.py`` — 8 cases validated on
BOTH SQLite and PostgreSQL backends (``pytest --storage-backend
postgresql``). Covers no-filter ordering, category exact-match,
tag quote-safety (``"foo"`` matches ``["foo","bar"]`` but not
``["foobar"]``), scan_status, enabled_only, limit, AND semantics,
empty result.
- ``tests/test_coordinator_client.py`` — 11 new cases covering
node/skill shape decoding, JSON-encoded filter round-trip (the
``"gpu"`` vs ``'"gpu"'`` case), int filter encoding, truncation,
no-match empty, no N+1 (``get_prompt_template`` /
``get_node_metadata`` call counts asserted zero).
- ``tests/test_coordinator_tools.py`` — 11 new cases for
``_prepare``/``_exec`` dispatch, filter type-drop, limit clamping
(``limit=0`` falls back to 100, negatives clamp to 1),
truncation-signal summary.
- ``tests/test_coordinator_endpoints.py`` — 8 new cases for
``/open``: ``already_loaded`` on in-memory hit, 404 on ownership
mismatch, lazy rehydrate on miss, admin bypass, unknown ws_id,
503 on ``coord_mgr`` unavailable, 500 with correlation-id mask on
factory failure, 503 passthrough on ``ValueError``.
- ``tests/test_workstream_kind.py`` / ``test_tools_schema.py``
updated to include ``list_nodes`` and ``list_skills`` in the
disjoint-namespace regression guard and the tool-count check.
Verification: ``ruff check`` clean, ``mypy turnstone`` clean
(156 source files), ``pytest -m "not live"`` 4122 passed, 3
deselected (live-backend), 0 failed. Postgres backend storage
tests green (``pytest --storage-backend postgresql
tests/test_storage_skills_filtered.py`` 8 passed).
* feat(coordinator): task_list tool — persistent planning state
Adds a coordinator-only ``task_list`` tool persisted on the
coordinator's own ``workstream_config`` row. Gives coordinators a
scratch surface for work decomposition that survives restarts so the
UI can render planned-vs-done state once the tree view lands.
## Tool surface
``task_list(action, ...)`` with five actions:
- ``list`` auto-approved read. Returns ``{tasks, truncated}``;
truncated=True when the list exceeded the 200-row
page cap.
- ``add`` needs approval. ``title`` required; optional
``status`` and ``child_ws_id``. Title clamped at
200 chars. Capacity cap at 500 tasks — hitting the
cap is an explicit signal to prune done/blocked rows.
- ``update`` needs approval. Mutate by ``task_id``; fields
``title`` / ``status`` / ``child_ws_id`` optional.
- ``remove`` needs approval. Drop by ``task_id``.
- ``reorder`` needs approval. Pass ``task_ids``; validated as an
exact permutation of the current set (rejects
partial, extra, or substituted ids — prevents silent
task loss).
Status enum: ``pending`` / ``in_progress`` / ``done`` / ``blocked``.
``child_ws_id`` links a task to the child workstream spawned for it
(no enforcement; the coordinator owns the relationship).
## Persistence
Stored as a single JSON-envelope value on ``workstream_config`` —
``{"version": 1, "tasks": [...]}``. No new table; the kanban v2
work will supersede this row via a format migration keyed on
``version``. ``_save_task_list`` writes only the ``tasks`` key so
concurrent writers to other ``workstream_config`` keys (e.g. the
admin Settings UI updating ``reasoning_effort``) aren't clobbered
by a read-modify-write on the full row.
## Corrupt-envelope safety
A hand-edited or legacy config row that doesn't parse as the
expected shape logs a warning and returns an empty envelope from
``task_list_get``. Mutators refuse to overwrite corrupt data —
they detect the sentinel and return a clear error so the operator
can inspect or clear the row rather than losing work silently.
## Concurrency
Per-(ws) ``threading.Lock`` cached on the client. The worker
thread is single-threaded for tool execs so this is mostly
defence-in-depth against future maintenance-script call sites.
Cache never grows beyond one entry per coordinator session because
the scope guard short-circuits foreign ``ws_id`` before the lock
is acquired.
## Malformed-JSON recovery
``_prepare_tool`` fallback-1 regex-extract allowlist extended with
``action`` / ``status`` / ``task_id`` / ``title`` (alphabetized) so
slightly-malformed ``task_list`` calls get the same
self-correction behaviour as the other coordinator tools.
## Tests
- ``tests/test_coordinator_client.py`` — 15 new cases covering:
fresh-envelope shape, add/get roundtrip, empty-title + invalid-
status rejection, 200-char title clamp, update by id + missing
id, remove semantics, reorder permutation validation (partial +
extra + wrong id + valid), cross-ws scope violation, corrupt-
JSON read recovery, corrupt-envelope write refusal (all four
mutators), 500-task capacity cap, workstream_config key
preservation across ``_save_task_list``.
- ``tests/test_coordinator_tools.py`` — 12 new cases covering the
dispatch layer: list auto-approved, each mutating action needs
approval, unknown-action / missing-required-arg errors, list
returns tasks, page-cap at 200 with truncated signal, add
dispatches to client, reorder surfaces permutation error,
remove-not-found.
- ``tests/test_tools_schema.py`` / ``tests/test_workstream_kind.py``
extend the tool-count + disjoint-namespace + primary-key
regression guards with ``task_list``.
Verification: ``ruff check`` clean, ``mypy turnstone`` clean
(156 source files), ``pytest -m "not live"`` 4148 passed,
3 deselected (live-backend), 0 failed.
|
||
|
Patrick Buckley
|
e42add1b77 |
feat(coordinator): coordinator workstream kind — phase 1 (#368)
* feat(coordinator): coordinator workstream kind — phase 1
Adds a new ``kind="coordinator"`` workstream that runs inside the
``turnstone-console`` process (first ChatSession hosted on the console)
with a dedicated tool set for spawning and driving child workstreams.
Supersedes the external ``turnstone-coordinator`` MCP side-car for new
installs; the extension is marked deprecated in
``examples/mcp-cluster-ops/README.md`` but still works on 1.4-and-earlier
clusters.
Phase 1 ships: the workstream class, 6 lifecycle tools, console hosting,
9 HTTP endpoints, per-user audit attribution, and a one-pane web UI at
``/coordinator/{ws_id}``. Node/skill discovery tools, task-list tool,
tree-view UI, and routing-proxy audit middleware follow in a later PR.
## Schema
Migration 039 adds ``kind`` / ``parent_ws_id`` columns + indexes to
``workstreams``. Both SQLite and PostgreSQL backends take the new
kwargs on ``register_workstream``; empty-string ``parent_ws_id``
normalises to ``NULL`` at the storage edge. PostgreSQL uses
``INSERT ... ON CONFLICT DO NOTHING`` to match SQLite's ``OR IGNORE``
and close a pre-existing SELECT-then-INSERT TOCTOU window.
``list_workstreams`` gains optional ``parent_ws_id`` / ``kind`` filters;
new ``get_workstream(ws_id)`` returns the full row (the existing
``get_workstream_metadata`` stays untouched for back-compat).
## Core session + kind routing
- ``ChatSession.__init__`` accepts ``kind`` / ``parent_ws_id`` /
``coord_client``. On ``kind="coordinator"`` it swaps
``_tools = COORDINATOR_TOOLS`` and zeros sub-agent tool lists.
- ``Workstream`` dataclass extended with ``user_id`` / ``kind`` /
``parent_ws_id``. Both ``WorkstreamManager`` and the new
``CoordinatorManager`` use the same type — no parallel hierarchy.
- ``_SessionFactory`` Protocol + server / cli factory closures thread
the new kwargs. ``POST /v1/api/workstreams/new`` rejects
``kind != "interactive"`` with 400; ``POST
/v1/api/workstreams/{ws_id}/open`` refuses coordinator rows so a
server node can't accidentally rehydrate one.
## Coordinator tool set
Six tools (``spawn``, ``inspect``, ``send``, ``close``, ``delete``,
``list_workstreams``) with a ``coordinator: true`` metadata flag,
scoped to coordinator-kind sessions only. ``inspect`` and ``list`` are
auto-approved reads; the four mutators need approval. ``list`` returns
``{"children": [...], "truncated": bool}`` so the model can detect
post-filter under-fill and paginate.
## CoordinatorClient (in-process, sync)
Mutating ops HTTP-POST to the console's own ``/v1/api/route/*`` on the
local bind URL so every existing middleware (auth, rate-limit) runs.
Read ops hit ``storage.list_workstreams`` / ``get_workstream`` /
``load_messages`` directly — the routing proxy doesn't expose
list/inspect paths. URL paths are a validated constant table (avoids
an httpx ``base_url``-merge trap). A new
``/v1/api/route/workstreams/delete`` proxy handler joins the existing
route-proxy endpoints.
## Per-session coordinator JWT
``CoordinatorTokenManager`` mints short-lived JWTs with ``sub=<real
user>`` (attribution preserved), ``src="coordinator"``,
``aud="turnstone-console"``, ``coord_ws_id=<ws>`` custom claim.
``_proxy_auth_headers`` preserves ``src`` + ``coord_ws_id`` across the
upstream re-mint so server-side middleware sees coordinator-origin,
not ``console-proxy``. ``AuthResult.extra_claims`` carries
non-reserved claims through validate→remint; ``create_jwt``'s
reserved-claim set (now including ``nbf`` / ``jti``) is symmetric with
``validate_jwt``.
## Console hosts the ChatSession
- New ConfigStore settings: ``coordinator.model_alias`` (required),
``reasoning_effort``, ``max_active`` (default 5),
``session_jwt_ttl_seconds``.
- Console lifespan builds a ``ModelRegistry`` +
``CoordinatorManager``. Missing / unresolvable alias returns **503**
with remediation text — never 500.
- ``CoordinatorManager``: placeholder-slot reservation under lock,
rollback on factory failure, per-ws_id rehydration lock to serialise
concurrent lazy-opens, ``max_active`` enforced via ``close_idle``
eviction semantics.
- ``ConsoleCoordinatorUI`` is a thin ``SessionUI`` implementation — no
global broadcast, no per-node metrics, shared
``_APPROVAL_WAIT_TIMEOUT`` constant across approval + plan paths.
- No eager startup rehydration: persisted coordinator rows load lazily
on first ``GET /v1/api/coordinator/{ws_id}``.
## Console coordinator API
Nine endpoints under ``/v1/api/coordinator/*`` gated by ``approve``
scope + new **``admin.coordinator``** permission (added to
``_VALID_PERMISSIONS``; not in any builtin role — operators opt in
explicitly). Ownership failures return **404, not 403** and use
strict equality so empty-owner rows don't leak across tenants.
Correlation-id masking on every factory-raising path
(``coordinator_create`` + ``coordinator_detail`` lazy rehydrate) — no
stack traces to the client.
## Audit attribution
Three console-side events (``coordinator.create`` / ``.close`` /
``.cancel``) with the real creator's ``user_id`` plus
``detail={coord_ws_id, src="coordinator"}``. No schema migration
required. Per-tool-call audit across the routing proxy is deferred
(needs either a ``source`` column on ``audit_events`` or
``record_audit`` calls wired into the route-proxy handlers).
## Web UI (``/coordinator/{ws_id}``)
One-pane chat served by the console. Reuses ``shared_static``
(``base.css``, ``auth.js``, ``theme.js``, ``toast.js``, ``utils.js``,
``kb.js``) and the server UI's ``renderer.js`` pipeline (KaTeX, Mermaid,
highlight.js already bundled).
- SSE to ``/v1/api/coordinator/{ws_id}/events`` with exponential-
backoff reconnect; status line carries a leading glyph
(● / ○ / ⚠) so state isn't conveyed by colour alone.
- Renders content, reasoning (dimmed italic
``.role-reasoning``), tool_result, approve_request, intent_verdict,
output_warning.
- Child ws_id references auto-wrap to
``/node/{node_id}/?ws_id={child}`` links — both ids regex-validated
before interpolation, everything else HTML-escaped.
- Non-modal approval bar (``role="region"``) with a batch header
("Approve N tool calls"), initial focus on the approve button,
buttons disabled during the in-flight POST, red-bordered deny.
``aria-live`` flips to ``off`` during streaming.
- "New coordinator" button on the dashboard header — permission-gated
on the UI side, matching the backend 403.
- Mobile composer capped under ``@media (max-width: 700px)``.
## Tests
~120 new tests across 8 files: workstream-kind storage + dataclass
semantics, CoordinatorClient URL map + token minting + storage reads +
truncation signalling, tool prepare/exec dispatch and approval gating,
CoordinatorManager create / rollback / eviction / lazy rehydration +
concurrency, HTTP endpoint auth + 404-on-ownership + 503-on-misconfig,
proxy-auth ``src`` preservation, full lifecycle end-to-end, coordinator
page HTML-injection guard. ``test_tools_schema.py`` widened to 25
tools (19 existing + 6 coordinator).
Verification: ``ruff check`` clean, ``mypy turnstone`` clean
(156 files), ``pytest`` 4054 passed (5 pre-existing failures unrelated
to this change — confirmed against ``main``).
* polish(coordinator): address PR review + CI + tool-namespace isolation
CI:
- `ruff format`: two files reformatted, matches the in-repo pre-commit config.
- `wheel-completeness`: add `turnstone/console/static/coordinator/*.html` +
`*.js` to the hatch wheel-include list. Without this the coordinator UI
was missing from published wheels.
- `test (3.11/3.12/3.13)` + `test-postgres`: three `TestExecReadImage`
tests were masking a real bug — my 6 new tool JSONs pushed tool count
19→25, crossing the default `tool_search.auto` threshold (20), which
made `ChatSession.__init__` construct a `ToolSearchManager` and cache
`_cached_capabilities` during init. Tests that later patched
`session._provider.get_capabilities` saw the cached value instead.
Root-cause fix: the tool-search threshold code path now reads
capabilities through `_resolve_capabilities(...)` directly — no cache
populate — so the patch takes.
Tool-namespace isolation (bigger fix than CI symptoms suggested):
- `TOOLS` was the union of all loaded tool JSONs including the 6 new
coordinator tools. Interactive sessions were getting coordinator
tools in their function-calling surface (which is nonsense — they
require a console-hosted `coord_client`), and coordinator sessions
counted against the interactive tool-search threshold. Fix:
- New `INTERACTIVE_TOOLS` / `INTERACTIVE_TOOL_NAMES` in
`turnstone/core/tools.py` exclude anything with `coordinator: true`
metadata. `TOOLS` stays as the union for schema introspection +
eval catalog.
- `ChatSession.__init__` selects tool set by kind: coordinator gets
fixed `COORDINATOR_TOOLS` (no MCP merge, no listeners registered);
interactive gets `INTERACTIVE_TOOLS` (+ MCP if configured).
Coordinators are meta-orchestrators that spawn child workstreams;
MCP tools / resources / prompts live on the children, not on the
coordinator's own surface.
- `_on_mcp_tools_changed` no-ops for coordinator sessions
(defence-in-depth in case listeners were registered).
- `always_on_names` on `ToolSearchManager` is now the set of builtin
tools actually present in the session (kind-aware) rather than the
full `BUILTIN_TOOL_NAMES` frozenset.
- `turnstone/eval.py` uses `INTERACTIVE_TOOLS` (coordinator tools
aren't in scope for the eval harness which tests interactive agent
behaviour).
- Regression tests in `tests/test_workstream_kind.py`:
- `INTERACTIVE_TOOLS ∩ COORDINATOR_TOOLS == ∅` and their union is
`TOOLS`.
- Interactive `ChatSession._tools` does not include any
coordinator tool name.
- Coordinator `ChatSession._tools` contains `spawn_workstream` but
not `bash` / `edit_file` / `memory`; sub-agent lists are empty.
- Coordinator `ChatSession` with an MCP client attached does NOT
merge MCP tools and does NOT register any MCP listeners.
PR review findings:
- **#10 / #11** (Copilot): coordinator UI claimed to reuse the server
renderer pipeline but loaded none of its JS. Mirrored
`turnstone/ui/static/renderer.js` into
`turnstone/console/static/coordinator/renderer.js` (flagged in-file
as a cleanup candidate to promote into `shared_static/`), added
`katex.min.js` / `highlight.min.js` / `renderer.js` script tags to
`coordinator/index.html`. `coordinator.js` now buffers raw markdown
via `textContent` during streaming, then swaps to `renderMarkdown` +
`postRenderMarkdown` on `stream_end`.
- **#7** (Copilot): N+1 query pattern in
`CoordinatorClient.list_children()` — per-row `storage.get_workstream`
just to read `skill_id`. Pushed `skill_id` + `skill_version` into
the `list_workstreams` SELECT projection on both backends; the
client reads them from `row._mapping` directly. New
`test_list_children_skill_filter_avoids_n_plus_one` pins the
behaviour (asserts `storage.get_workstream` call count is 0).
- **#8 / #9** (Copilot): `spawn_workstream` tool JSON said "if empty,
the workstream is created idle" but the prepare method rejected
empty and the field was marked required. Resolved by allowing
empty end-to-end: removed from `required`, prepare builds a
"spawn idle workstream" header + empty preview when empty,
updated `test_spawn_prepare_allows_empty_initial_message`.
- **#1–#5** (github-code-quality): five asserts with side-effecting
method calls in `test_coordinator_manager.py` (`mgr.close`,
`mgr.open`, `mgr.create` in a dead `_c = ...`). Extracted each
call to a local variable so `python -O` can't strip the side
effect.
Verification:
- `ruff check turnstone tests` clean.
- `mypy turnstone` clean (156 source files).
- `pytest -m "not live"` — 4063 passed, 3 deselected (live-backend
tests), 0 failed. The 3 image tests that were failing on this
branch now pass; wheel + lint both green locally.
* polish(coordinator): address Copilot re-review findings
Two findings from the re-review of #368 after the first polish commit.
**user_id wired into `mgr.create()` at the server handlers.** Phase 1
added ``user_id`` to the ``Workstream`` dataclass and
``WorkstreamManager.create()`` signature, but the two call sites in
``turnstone/server.py`` forgot to pass the authenticated caller
through. Result: interactive workstreams created via
``POST /v1/api/workstreams/new`` (including coordinator-spawned
children, which route through this handler) were landing with blank
``user_id``, defeating ownership-based access control on subsequent
sends / approvals / closes (``_require_ws_access`` treats blank
owners as legacy/allowed). Two changes:
- ``server.py:create_workstream`` forwards ``user_id=uid`` — the same
``uid`` already resolved from the auth result (with trusted-service
forwarding preserved).
- ``server.py:open_workstream`` prefers the persisted owner on the
workstream row over the rehydrating caller so reloading someone
else's workstream doesn't silently re-parent it. Falls back to
the authenticated caller when the stored row has no owner
recorded (pre-phase-1 rows).
Regression test in ``tests/test_workstream.py`` pins
``WorkstreamManager.create(user_id=X)`` → ``ws.user_id == X`` so the
manager seam can't regress silently on a future refactor.
**Malformed-JSON recovery allowlist expanded for coordinator args.**
``_prepare_tool()`` has a two-stage salvage path for models that
emit malformed JSON: a regex-extract (fallback 1) and a bare-string
→ primary_key wrap (fallback 2). The fallback-1 key list didn't
include coordinator argument names, so a slightly malformed
``spawn_workstream`` / ``send_to_workstream`` / etc. call would
hard-fail instead of salvaging into a minimal-args dict for retry.
Added ``ws_id`` / ``message`` / ``initial_message`` / ``parent_ws_id``
to the allowlist (kept alphabetised) so the coordinator tools get
the same model-self-correction behaviour as the interactive tools.
Fallback 2 already covers the ``ws_id``-primary-key tools via
``PRIMARY_KEY_MAP``; the regex path matters when the model emits
``{"ws_id": "abc", "message": "..."}`` with a trailing syntax error.
Verification: ``ruff check`` clean, ``mypy turnstone`` clean
(156 source files), ``pytest -m "not live"`` → 4065 passed, 3
deselected (live-backend), 0 failed.
* fix(coordinator): address ultrareview findings on coordinator workstream kind
Security
- Cross-tenant leak: CoordinatorClient.inspect/list_children now constrain
to the coordinator's own ws_id + direct children; an LLM coerced via
prompt injection can no longer exfiltrate other tenants' workstreams.
- Empty-owner short-circuit bypass: strict equality at coordinator.py
ownership gate and at the storage-fallback branch in coordinator_history;
orphan/system-owned coordinator rows can no longer be rehydrated by
arbitrary holders of admin.coordinator (DoS + history disclosure vector).
- Closed coordinators no longer silently resurrect on subsequent GET —
the Close button is now actually durable across URL revisits and tab
refreshes; rows with state in {closed, deleted} refuse rehydration.
Correctness
- ChatSession.close() now releases the CoordinatorClient httpx.Client
pool; previously every closed/evicted coordinator dropped a connection
pool on the floor until non-deterministic GC.
- open_workstream rehydration now forwards parent_ws_id + kind, so
coordinator-spawned children survive node restart / idle eviction
with their parent link intact instead of becoming silent orphans.
- list_children truncated flag now signals whenever the SQL fetch hit
the page cap (previously permanently False in the no-filter case,
causing confident-but-incomplete summaries from the coordinator).
- ConsoleCoordinatorUI.approve_tools: per-tool auto-approve now checks
auto_approve_tools independently of the blanket auto_approve flag,
so 'Always approve this tool' actually works on the next invocation.
Concurrency
- _spawn_worker no longer falls through to start a second concurrent
worker thread on the same ChatSession when queue.Full fires; instead
send() returns False and the endpoint surfaces HTTP 429.
- _open_locks entries are now refcounted under self._lock and only
popped when the last waiter releases — eliminates the race where a
rehydration-failure path lets two threads serialize on different lock
instances for the same ws_id and trip the "already tracked" guard.
Tests: +6 regression cases covering closed-coordinator refusal,
empty-owner non-admin refusal, queue.Full no-duplicate-worker,
inspect/list_children cross-tenant rejection, and truncated semantics.
|
||
|
Patrick Buckley
|
cfef3616e6 |
feat: add diff_file tool for comparing files and content (#234)
* feat: add diff_file tool for comparing files and content
New read-only tool that shows unified diffs between two files or
between a file and provided content. Useful for verifying edit_file
changes and comparing file versions. Auto-approved (no side effects).
Configurable context lines (default 3). Available to task agents.
* refactor: extract _read_text_lines helper, share across read_file and diff_file
Copilot review: diff_file duplicated file-loading and lacked binary
detection. Extract _read_text_lines() that handles realpath
resolution, null-byte binary detection, and error handling. Used by
both _exec_read_file and _exec_diff for consistent behavior.
* fix: address code review — agent flag, resolved shadowing, read_files
- Add agent: true to diff_file schema so plan agents can use it
- Fix resolved variable shadowing in _exec_read_file (use _ for
unused return from _read_text_lines)
- Register diffed files in _read_files so edit_file read guard
is satisfied after diff_file
- Move difflib import to module level (stdlib, no lazy-load needed)
- Fix description wording ("provided string" not "previous version")
* fix: stream diff with early cutoff, expand paths before header
- Stream difflib output and stop collecting after tool_truncation
chars to avoid large intermediate allocations on big diffs
- Expand paths with expanduser before building the approval header
so display matches actual execution paths
|
||
|
Patrick Buckley
|
361876b17a |
feat: eval pipeline improvements + tool description optimization (#174)
Eval pipeline: - --optimize-tools mode freezes system prompt, optimizes tool descriptions only - Analyst sees available tool list (prevents hallucinated "tool not in schema") - Analyst sees current tool descriptions in --optimize-tools mode - Three-layer timeout defense: httpx timeout + _cancelled event + client.close() - Filter MCP-only tools (read_resource, use_prompt) from headless eval - Pattern-over-rules framing in optimizer, analyst, and observer prompts - Tool description diffs logged after each iteration - Tool optimizer failure retries instead of stopping the loop Tool renames (avoid chat template channel collision on local models): - create_plan -> plan_agent - task -> task_agent Tool descriptions (from eval-driven optimization, 79% -> 98%): - bash: environment question examples, disambiguation from write_file/man - edit_file: multi-file workflow, prerequisite clarification, docstring example - man: "questions about flags are tool-use tasks" prefix - math: simple example up front - plan_agent: "delegate to sub-agent" framing, negative boundary for direct edits - read_file: multi-file workflow hint - search: trigger phrases, prerequisite clarification, disambiguation - write_file: immediate action framing, placeholder example System prompt: enriched tool patterns from eval results, added identity opener Test suite: plan-before-refactor -> plan-when-asked (simplified) Docs: comprehensive eval.md rewrite covering all current features |
||
|
Patrick Buckley
|
dc464ac313 |
feat: Agent Skills standard compliance + frontend spec fields
Brings skills implementation into full compliance with agentskills.io: Parser: - Read `allowed-tools` (hyphenated, standard) only; stored as `allowed_tools` internally — no underscore fallback - Reject consecutive hyphens in skill names - Extract author/version from standard `metadata:` map with top-level fallback; null-safe (no "None" string for bare YAML keys) - Truncate description at 1024 chars, compatibility at 500 chars (spec caps) with log warnings - Lenient parsing mode (lenient=True) for cross-client import: sanitizes names, returns None on skip, malformed-YAML colon-value retry - Type overloads: strict mode returns ParsedSkill, lenient returns ParsedSkill | None Session: - `<available-skills>` XML catalog in system messages for activation="search" skills (disabled ones filtered out, capped at 30) Tool rename: - `load_skill` tool → `skill` (JSON, session preparers/executors, approval labels, tests, docs) Storage (migration 023): - Add `license` and `compatibility` columns to prompt_templates - skill_license / compatibility params on create_prompt_template across protocol, SQLite, PostgreSQL backends - Add to SKILL_MUTABLE for update_prompt_template API + server: - SkillInfo, CreateSkillRequest, UpdateSkillRequest: license + compatibility fields - Create/update/install endpoints extract and persist both fields - Install endpoint maps parsed.license + parsed.compatibility from imported SKILL.md (previously discarded) - _skill_to_response() includes both fields Admin UI: - Create + edit modals: version, license, compatibility fields - Readonly (imported) skills: "edit" → "view" button, modal title "View Skill", all fields disabled, Save hidden, Cancel → "Close", collapsibles auto-expand, focus on Close button - :disabled CSS for dark-theme modal inputs (bg-highlight, cursor not-allowed, dimmed text) - Fix addEventListener stacking on auto-approve checkboxes → .onchange SDK: license + compatibility on SkillInfo, CreateSkillRequest, UpdateSkillRequest TypeScript interfaces Docs: governance.md, judge.md, tools.md, README, diagram updated |
||
|
Patrick Buckley
|
1010f163f0 |
feat: load_skill built-in tool — model-driven skill discovery and act… (#112)
* feat: load_skill built-in tool — model-driven skill discovery and activation Two-action tool: 'search' finds skills by multi-word query with substring matching on name/description/tags/category (auto-approved, read-only); 'load' activates a skill by name via set_skill() (requires approval). Guards: filters disabled skills from search + load; short-circuits when skill is already active; approval_label includes skill name for granular tool policies (load_skill__<name>); main session only (excluded from sub-agents). Logs storage errors in search path. 25 tests covering registration, preparer validation, executor logic, disabled/already-active edge cases, multi-word queries, approval labels. * refactor: use BM25 relevance ranking for load_skill search Replace substring matching with BM25Index from turnstone/core/bm25.py, matching the pattern used by memory relevance and tool search. Handles multi-word queries, term frequency, and document length normalization. * fix: address copilot review — BM25 tags parsing, primary_key, test cleanup - Parse JSON tags into space-separated text before BM25 indexing so individual tag terms match queries (was passing raw '["foo","bar"]') - Add primary_key: "name" to load_skill.json for PRIMARY_KEY_MAP - Remove dead resolve_workstream patch from test helper - Update diagram: "substring match" → "BM25 ranking" |
||
|
Patrick Buckley
|
723cad24bb |
feat: structured memory system — typed/scoped memories with BM25 rele… (#53)
* feat: structured memory system — typed/scoped memories with BM25 relevance and metacognitive prompting Replace flat key-value memories table with structured_memories (migration 014). Four memory types (user/project/feedback/reference), three scopes (global/workstream/user). Consolidate remember/recall/forget into two tools: memory (action-based: save/search/delete/list) and recall (conversation history only). BM25 relevance scoring (extracted to turnstone/core/bm25.py) selects top-5 memories for system message injection based on conversation context. Metacognitive prompting injects ephemeral nudges after corrections, tool denials, workstream resume, and completion signals. Scope isolation enforced: system message injection and nudge counts filtered to visible memories only (global + current workstream + authenticated user). User scope requires authentication. Content capped at 32KB. ILIKE/LIKE metacharacters escaped in both backends. 113 new tests (2053 total). * fix: CI failure + copilot review feedback - Fix time.monotonic() cooldown: use None sentinel instead of 0.0 default (monotonic clock starts at boot, not epoch — fresh CI runners have uptime < 300s so cooldown check always triggered) - Catch sa.exc.IntegrityError specifically in upsert instead of broad Exception (copilot review) - Preserve existing description/type on upsert when caller doesn't explicitly set them (copilot review) - Add last_accessed + access_count columns to schema/migration for future LRU/LFU eviction support |
||
|
Patrick Buckley
|
be165c1971 |
feat: MCP resource and prompt discovery with read_resource tool (#44)
* feat: MCP resource and prompt discovery with read_resource tool Extends MCPClientManager with resource and prompt discovery alongside existing tool support. Resources and prompts are discovered on connect, cached per-server with copy-on-write rebuilds, and refreshed via push notifications, periodic polling, or manual /mcp refresh. New read_resource built-in tool reads MCP resources by URI. Requires user approval (same as MCP tool calls) since resources are served by external MCP servers. Resource catalog injected into system message with XML delimiters. Error messages sanitized to prevent leaking server internals to the model. Prompt discovery stores prefixed names (mcp__server__prompt) and exposes get_prompt_sync() for future use_prompt tool (Chunk D). /mcp command now shows tools, resources, and prompts. Docs and diagrams updated. * feat: MCP prompt governance sync with origin tracking and readonly guards Migration 009 adds origin, mcp_server, and readonly columns to prompt_templates. MCP prompts discovered by MCPClientManager are automatically synced into the governance table as read-only templates with origin="mcp". Sync engine handles: create on connect, update on prompt refresh, delete when prompts are removed from server. Manual templates take precedence on name collision (MCP prompt skipped with warning). Admin API returns 403 on update/delete of readonly templates. Console UI shows MCP origin badge and disables edit/delete buttons. Storage backends gain get_prompt_template_by_name, list_prompt_templates_by_origin, and delete_prompt_templates_by_server methods. Also addresses PR #44 review feedback: concurrent.futures.TimeoutError handling in sync dispatch, XML-escape resource catalog descriptions, resource template entries excluded from _resource_map, URI collision warnings, needs_periodic capability-aware computation, malformed JSON primary key fallback for read_resource. * feat: use_prompt tool, prompt catalog, and PR review hardening New use_prompt built-in tool invokes MCP prompt templates by name, expanding them into messages. Requires user approval (external MCP servers). Prompt catalog injected into system message with XML delimiters (up to 30 prompts, HTML-escaped). Prompt listener registered in session for catalog rebuild on changes. Addresses PR #44 review feedback: - _init_system_messages() now uses copy-on-write (build locally, assign atomically) so background thread callbacks never see partial system messages - sync_prompts_to_storage() serialized behind _sync_lock to prevent races between set_storage() (main thread) and MCP background thread - shutdown() clears listener lists to release callback references Docs and diagrams updated for 18 built-in tools. * feat: granular tool policies for MCP resources, prompts, and tools Policy evaluation now uses approval_label (falling back to func_name) for fnmatch pattern matching, enabling fine-grained per-URI and per-server policies: - read_resource: mcp_resource__{normalized_uri} - use_prompt: mcp__{server}__{prompt} (prefixed name) - MCP tools: mcp__{server}__{tool} (was static "mcp_tool") URI normalization resolves .. path segments to prevent traversal bypasses in policy matching. Resource templates filtered from system message catalog (not directly readable). use_prompt arguments validated as dict with string coercion. TypeScript SDK PromptTemplateInfo gains origin, mcp_server, readonly fields. Governance docs updated with MCP policy patterns. * feat: MCP visibility in server and console UIs Server health endpoint includes mcp.servers, mcp.resources, mcp.prompts counts. Server UI status bar shows magenta MCP indicator with tooltip. Console cluster status bar shows MCP metrics with magenta LED dot. Console node detail view shows per-node MCP summary. Console collector aggregates MCP counts across nodes in overview. Uses var(--magenta) design token with new --magenta-glow for theme adaptation. ARIA roles on MCP status elements. Tooltips on console MCP metric labels. Node MCP summary hidden on mobile (< 700px). New diagram: 20-mcp-architecture.puml covering full MCP lifecycle (connection, discovery, refresh, governance sync, policy, UI). * fix: McpStatus in health schema, count properties, catalog name fidelity Adds McpStatus model to HealthResponse (Python + TypeScript SDKs) so typed clients see the mcp field from /health. Addresses Copilot review feedback: - resource_count/prompt_count properties avoid list allocation on /health and /metrics polls - get_tools/resources/prompts return shallow-copied dicts to prevent callers from mutating internal cache - Prompt names and arg names in system message catalog are NOT HTML-escaped (model must use exact strings in use_prompt calls); only descriptions are escaped * fix: OpenAPI spec McpStatus + diagram approval column accuracy Adds McpStatus schema and optional mcp field to HealthResponse in openapi-server.json, matching the Python schema and TypeScript types. Fixes tool pipeline diagram: math, web_fetch, web_search correctly shown as auto-approve (not "Yes" for approval). |
||
|
Patrick Buckley
|
de64535221 |
Feat/eval improvements (#37)
* ci: add GitHub Release creation on tag push
* refactor: rename plan tool to create_plan
Rename plan → create_plan to resolve cross-provider tool selection
failures. Models consistently treated "plan" as a reasoning concept
rather than a callable tool. The new name is an unambiguous verb+noun
action. Also rename the parameter from prompt → goal for clarity,
add web_search to the default system prompt tool patterns
* feat: eval harness improvements inspired by autoresearch patterns
Major enhancements to turnstone-eval:
- Per-test timeout (--test-timeout, default 300s) and suite timeout
(--suite-timeout) prevent stuck runs from blocking the suite
- Fast-fail skips remaining runs after ceil(n/2) consecutive zeros
- Summary table with colored PASS/WEAK/FAIL and append-only TSV output
- Progress reporting with running pass rate, token count, and ETA
- Parallel test execution via ProcessPoolExecutor (--parallel N)
- Per-role model assignment: test/optimizer/observer can use different
models and providers (--optimizer-model, --observer-model, etc.)
with auto-detection from base URL
- Improved optimizer and observer system prompts with structured
failure-mode diagnosis, keep/discard rules, and trend analysis
- Fixed token counting (prompt tokens use last-turn value, not sum)
- Added math-calculation and web-search-query test cases
- Fixed multi-file-edit test (both files now contain the target string)
* fix: address Copilot review feedback
- Revert prompt token counting to sum (reflects billed usage)
- Add tool_args to fast-fail skipped run dicts for schema consistency
- Align approval_label with func_name ("create_plan")
- Add timeout to future.result() in parallel path (test_timeout + 30s)
- Document thread-leak trade-off on serial timeout path
|
||
|
Patrick Buckley
|
187d004033 |
feat: watch tool — periodic command polling within workstreams (#36)
* feat: watch tool — periodic command polling within workstreams
Add a new `watch` tool that lets the model (or user) set up periodic
polling of a shell command. Results inject as synthetic user messages
that trigger LLM turns, enabling reactive workflows like PR monitoring,
CI/CD status tracking, and deployment health checks.
Key design:
- Single tool with create/list/cancel actions
- Python expression DSL for stop conditions (restricted eval)
- Server-owned WatchRunner daemon (DB-persisted, survives eviction + restart)
- Three dispatch paths: idle, busy, and evicted workstream restore
- REST API for console visibility (GET /v1/api/watches, POST cancel)
- Migration 007, 8 storage CRUD methods, 75 new tests (1383 total)
* fix: address Copilot review — condition errors, restore deadlock, docs
- Condition eval errors now deactivate the watch immediately instead
of silently looping until max_polls
- Restored (evicted) workstreams set auto_approve=True to prevent
approval deadlocks with no connected user
- Tool description clarifies first-poll baseline behavior for change
detection mode
- Diagram updated: DELETE → POST /v1/api/watches/{id}/cancel
|
||
|
Patrick Buckley
|
e7fe8fca9d |
Add channel notification tool with security hardening (#27)
* Add channel notification tool with security hardening Implements the `notify` tool allowing the LLM to send notifications to Discord channels/users via the channel gateway. Includes fixes for 11 review findings: JWT auth on the gateway endpoint, first-healthy gateway delivery with retry+backoff, rate limiting only on success, SSRF URL scheme validation, Discord mention sanitization, SQLite ON CONFLICT upsert preserving created timestamps, advertise URL resolution for 0.0.0.0 bind, randomized service IDs, generic error messages to prevent internal state leakage, and partial direct-target validation. Service registry with heartbeat-based health filtering (migration 005). Channel gateway registers on startup, heartbeats every 30s, deregisters on shutdown. 70 new tests covering tool prepare/execute, HTTP endpoint auth (static + JWT), storage CRUD, and retry behavior. * Add notify documentation, diagrams, and review fixes Documentation: - New sequence diagram 17-notify-flow.puml showing end-to-end delivery - Updated 16-channel-architecture.puml with services table, notify HTTP path, and Notification Flow note - channels.md: Notifications section (targeting, delivery flow, service registry, security) and new config table entries - tools.md: notify tool reference, updated counts/tables (14→15 tools) - security.md: channel gateway row in service-to-service auth table - architecture.md: notification subsystem paragraph Review fixes (copilot): - _http.py: fail closed when auth unconfigured (401 instead of pass- through), strip whitespace on message/title, generic error messages for user-not-found vs no-linked-channels - session.py: parse gateway response JSON and require at least one result with status=="sent" before counting as success - _postgresql.py: use index_elements instead of constraint for upsert |
||
|
Patrick Buckley
|
9be155b97a |
Quality overhaul: code tooling, CI/CD, architecture diagrams, UI rede… (#1)
* Quality overhaul: code tooling, CI/CD, architecture diagrams, UI redesign, and legacy cleanup - Add ruff (lint+format) and mypy (strict) with zero errors across 37 source files - Add GitHub Actions CI (lint, typecheck, test matrix 3.11/3.12/3.13) and PyPI publish workflow - Create 12 PlantUML architecture diagrams with PNG renders covering all subsystems - Refresh README and docs with badges, diagram links, and current descriptions - Refactor test_server_live.py with mock streaming helpers for deterministic CI testing - Update dependencies to current versions (openai>=2.24, httpx>=0.28, redis>=7.2) Console dashboard: - Move state indicators from top cards to fixed bottom status bar with cluster metrics - Replace flat 50-node list with hostname-prefix grouped nodes (expand/collapse, up to 1000) - Apply "Instrument Panel" visual redesign: IBM Plex Mono + Outfit fonts, warm amber accent, LED glow state indicators, deep charcoal surfaces, WCAG AA contrast compliance - Add render cache, stale indicator, active filter highlight, loading states Server web UI: - Apply matching Instrument Panel aesthetic for visual consistency with console - Fix branding (pcode → turnstone), extract inline styles to CSS classes - Rename pcode localStorage keys and history state to turnstone Legacy cleanup: - Remove persona-model-specific --persona flag and /persona slash command - Remove model_identity from chat_template_kwargs (vLLM-specific mechanism) - Refactor plan agent to use standard developer message instead of model_identity - Remove dead code (unused date/has_tools variables, noqa suppressions) * Fix CI typecheck: add mypy overrides for optional sympy/numpy imports The math sandbox optionally imports sympy and numpy at runtime (try/except ImportError). In CI these packages are not installed, so mypy raises import-not-found rather than import-untyped. Add mypy overrides to ignore missing imports for these optional dependencies. * Fix Copilot review findings: ARIA role, status bar cache, and pulse opacity - Change #node-table from role="tree" to role="list" and group elements from role="treeitem" to role="listitem" (proper ARIA semantics) - Include currentView and currentFilter.state in renderStatusBar cache key so active pill highlight updates when switching views - Align pulse animation to 0.35 opacity (already applied in CSS) |
||
|
Patrick Buckley
|
0d6252dd7d | Initial commit — turnstone multi-node AI orchestration platform. |