Files
turnstone/docs/diagrams/09-workstream-states.puml
Patrick Buckley 480a1426b3 Fail-closed history-commit handoff (#1005)
* fix(session): fail-closed history-commit handoff (#981)

The deleted-workstream discovery is now a terminal, ws_id-keyed latch:
keyed conversation commits refuse admission once the durable parent is
gone (convergence finalizers and force-abandon are exempt), history
handoff refuses to mint a proof token so /history fails closed with a
503 instead of silently wiping the pane, and the SSE stream carries a
workstream_gone resync reason. Discarded commits leave a forensic log
of commit keys and roles, never content.

Conversation rows gain a commit_key (migration 071): keyed saves are
idempotent under retry, validated against the full commit identity, and
refused when they would cross a workstream deletion. The prune orphan
category now requires a NULL alias plus a two-hour updated grace, with
cutoffs computed at discovery time and carried into both dialects'
rechecks.

The mid-turn interjection queue is owner-partitioned with no per-site
mode flags: pops take the acting principal's and unowned rows, other
participants' rows are structurally retained, and enforcement lives at
queue admission plus the shared before_spawn gates. The retraction
ledger is bounded by open pop windows: pops open a window atomically
with the queue delete, restores close their ids atomically with the
ledger consume, every other exit closes through one helper, and misses
for unheld ids record nothing. The workstream-gone latch refuses
unattended wakes at all three gates (watcher spawn, claim, delivery
pre-pop), and the retry dispatcher regained its pre-envelope
cancel/error convergence net.

Persistence-state reporting derives through the session bound to each
UI instead of a registry lookup by id that failed open to healthy
during tombstone retention. The dashboard roster no longer re-inserts
ghost entries from trailing activity events, the history tool-outcome
scan tolerates interleaved non-turn rows, and the shared
handoff-deadline handle owns its own retirement.

Single-sourced across call sites: keyed-commit row values, attachment
save wrappers, tail-truncation and conflict-resolution bodies for both
storage dialects; worker-slot lifecycle field sets; the direct-commit
admission frame; queued-row layout accessors; the string-aware comment
stripper shared by every JS harness suite.

Refs #981 #964

* fix(session): sweep handoff fixes to their sibling surfaces

The interactive replay loop treated a system row as a tool-batch
boundary, so every tool result after an interleaved row vanished from
that pane while the coordinator rendered the same history correctly.
Only a conversational turn ends the batch window now, matching the
shared outcome index.

Accepted user turns clear the composer's attachment chips on the same
viewer policy that settles optimistic bubbles rather than on having
matched a local bubble, so a workstream created with an upload no
longer keeps a chip for an attachment the create dispatch already
consumed. The coordinator's raced-Stop arm emits the stream-end hook it
inherits alongside the idle state, leaving no unfinalized bubble or
unflushed tool output. Ending a session surfaces a failure toast when
the request never lands or answers with a non-JSON body.

The per-second persistence reconcile now probes each session without
blocking: a workstream whose generation and handoff locks are held is
skipped until the next pass instead of contending the locks every
commit needs. The one-shot repair that gates workstream creation at
capacity keeps a definite probe — it has no next pass, and the sessions
likeliest to be contended are the ones whose unresolved journals
emptied its candidate list.

Single-sourced: the attachment lane builds its conversation row through
the shared commit-identity builder; the ordinary worker exit releases
its slot through the lifecycle owner; both operator surfaces snapshot
their counters through one non-consuming helper; the replay preamble
loses its per-kind wrappers and its config hook; the browser harness
suites share one brace walker; and each in-flight history attempt is
one record carrying both its abort controller and its deadline.

Refs #981 #964
2026-08-11 04:18:36 -07:00

142 lines
5.2 KiB
Plaintext

@startuml
!theme plain
title Turnstone — Workstream State Machine
skinparam state {
BackgroundColor<<lifecycle>> #ECEFF1
BackgroundColor<<idle>> #E8F5E9
BackgroundColor<<thinking>> #E3F2FD
BackgroundColor<<running>> #FFF3E0
BackgroundColor<<attention>> #FCE4EC
BackgroundColor<<error>> #FFCDD2
}
state "CREATING (persisted only)" as creating <<lifecycle>> : Hidden durable reservation.\nNot returned by ordinary list/open/history.
state "IDLE" as idle <<idle>> : Waiting for user input.\nNo active LLM call or tool execution.
state "THINKING" as thinking <<thinking>> : LLM streaming response.\nTokens flowing (reasoning + content).
state "RUNNING" as running <<running>> : Tools executing.\nThreadPoolExecutor active.
state "ATTENTION" as attention <<attention>> : Blocked on user action.\nTool approval needed.
state "ERROR" as error <<error>> : Exception occurred.\nRecoverable on next send().
state "CLOSED (persisted only)" as closed <<lifecycle>> : Unloaded, explicitly reopenable row.\nNot a live WorkstreamState member.
[*] --> creating : register exact incarnation\nstate="creating"
creating --> idle : finalize + publish create\nemit ws_created
creating --> [*] : immediate exact-token rollback\n(no lifecycle birth emitted)
creating --> [*] : stale >2h recovery\natomic hard delete; no close event
idle --> thinking : send() called\n_emit_state("thinking")
thinking --> running : Tool calls detected\nin LLM response\n_emit_state("running")
thinking --> idle : No tool calls\n(final answer)\n_emit_state("idle")
thinking --> error : Exception during\nstreaming
running --> attention : approve_tools() called\n_emit_state("attention")
attention --> running : User approves\n(tools execute)\n_emit_state("running")
attention --> running : User denies\n(denial recorded)\n_emit_state("running")
running --> thinking : Tool results appended,\nnext LLM call\n_emit_state("thinking")
running --> error : Exception during\ntool execution
error --> thinking : New send() call\n_emit_state("thinking")
idle --> closed : close / eviction\n[journal reconciled]
error --> closed : close\n[journal reconciled]
thinking --> closed : close\n[journal reconciled]
running --> closed : close\n[journal reconciled]
attention --> closed : close\n[journal reconciled]
closed --> [*] : hard delete
closed --> idle : open / rehydrate
note right of closed
Before every soft-close / eviction transition,
the total accepted conversation-row journal must
be durably reconciled. An unresolved row makes an
explicit close return HTTP 409 (eviction refuses),
and the workstream remains loaded in its live state.
end note
thinking --> idle : cancel() called\nstream aborted\n_emit_state("idle")
running --> idle : cancel() called\n_emit_state("idle")
attention --> idle : cancel() unblocks\napproval wait\n_emit_state("idle")
note left of idle
**Generation-scoped Stop:**
• Sets the active generation event.
• Closes its SDK stream; aborts child model
scopes and judges; kills subprocess groups.
• Sweeps every approval cycle owned by the
cancelled workstream operation.
• Every later send/model live or durable commit
re-checks generation ownership.
**force=true:** also abandons the stuck worker
slot and emits stream_end + IDLE immediately.
An orphaned send/model generation may unwind
but cannot publish into a successor generation.
Quick slash-command workers are a best-effort
escape hatch: without generation checkpoints,
one may finish an in-place mutation concurrently.
**Capacity eviction:** an IDLE candidate is only
a hint. Per-ID + object lifecycle lanes and the
workstream lock revalidate it as worker- and
send-barrier-free,
then install a terminal claim before slot swap.
end note
note right of thinking
**Emitted via:**
session._emit_state(state)
→ ui.on_state_change(state)
→ SessionManager state tail
**Propagation:**
• WebUI → global SSE queue (ws_state)
• Console → cluster event / HTTP state
• CLI → SessionManager.set_state()
Non-terminal persistence may use StateWriter;
a per-id tail orders storage + subscribers and
prevents a late state from overwriting CLOSED.
end note
note left of attention
**Blocking mechanisms:**
• TerminalUI: input() prompt
• WebUI: one Event per ApprovalCycle
• ChannelBot: SSE event + Discord button
• NullUI: auto-approve (never reaches)
end note
note right of creating
CREATING and CLOSED are storage lifecycle
values, not members of WorkstreamState. The
live enum remains IDLE / THINKING / RUNNING /
ATTENTION / ERROR.
**Crash-abandoned CREATING recovery:**
• Boot pass, then every 5 min even when idle
eviction is disabled.
• Only rows >2h old; manager loaded/pending
IDs and live remote owners are protected.
• The current stable node ID is not a live-owner
exemption, allowing restart recovery.
• Unknown liveness/storage fails closed. Deletion
is atomic across dependents and attachment refs.
• Tokenless legacy/corrupt rows are locked,
reaped, and logged with a warning.
A loaded hard delete closes publication, drains
admitted session durability + state tails, then
conditionally removes the exact durable token.
end note
@enduml