The reference-admission discard listener matched on followup()'s returned id,
but an agent/inbox/enqueue listener that synchronously cancels emits
agent/inbox/discard before followup() returns to assign that id. The match
then missed, leaking both the submit and discard listeners plus the attached
context per referenced prompt. Match on the content reference instead — the
same value send() carries onto the message, known before followup() runs, and
symmetric with the submit wrapper's content check.
The prior "ordinary allowed path" test passed only because the fake agent
returned a fixed 'stub' id that collided with the id the test constructed;
it now releases each wrapper through its own allowed admission, and a new
regression drives the synchronous-discard timing directly.
Fixes the node 24 / static CI failure: regenerate cordis/config catalogs and
doc graphs after this branch's line and consumer changes, and bring the
ToolRunContext concludeTurn JSDoc in the bilingual tools doc pair back in sync
with packages/core/tools/src/index.ts (re-recording the i18n pairing hash).
goal-session rides retry turns and survives admission failures. A
recovery policy closes a goal round's failed turn and reopens its
history under a retry trigger; the attempt now adopts that turn and
drops the failed turn's provisional reason, so the round settles from
the retry's own outcome instead of blocking an armed goal with
turn-error after a successful response. A downstream admission hook
that throws (rather than blocks) used to strand the queued reservation
forever; the listener now clears a still-turnless matching attempt on
the rejection path and reschedules the round.
agent-loop contains a persistently rejecting step close in the catch
path the same way the finally contains the turn close, so the
post-finally tail always publishes the terminal status — previously a
double veto escaped run(), leaving status at running while whenIdle()
resolved. The whenIdle catch arm is annotated as the backstop it now
is: every driver rejection path is contained today.
workspace-context folds an already-appended baseline from the session
log when the plugin is hot-remounted over a live session, instead of
injecting a duplicate from its fresh mount-local guard.
The TUI's reference-admission discard listener installs before
followup(): admission runs synchronously inside it on the common path,
so a listener installed afterwards missed its own cleanup and leaked
one callback per referenced prompt.
Validate replay sidecars and cross-copy failure facts, make browser console tripwires and macOS temp paths deterministic, and wait for asynchronous TUI resume details. Keep the owning docs, translations, and generated catalog aligned.
The TUI's referenced-prompt snapshot now rides the prompt's own
admission transaction instead of a pre-admission inject: while idle, a
one-shot prepended agent/prompt-submit wrapper appends the snapshot to
the allow decision's additionalContexts, so a blocking hook discards
the prompt and its attached context together instead of stranding the
snapshot in history for the next unrelated prompt. A prompt discarded
before admission releases the wrapper; steering keeps the inject path
since it bypasses admission and drains at the same boundary. The
session-reference snapshot adapter pinned the old context-before-prompt
order; the branch-wide order (prompt first, its contexts after) is now
asserted and the fixture re-recorded.
tool-tasks drops the last consumer of the removed thrown-disposed
contract: completion notices now inject unconditionally, which is
well-defined during owner teardown — the loop treats disposal like any
cancel, so the notice appends as durable idle context (persisted for
resume while the session is attached, dropped with the detached log
after). README pair and the owner-disposal tests state the new
delivery contract.
The staged-promotion machinery (concludingParents keyed by parent token
plus a pendingParentConclusions staging map) spread one fact — this call
concluded the turn — across three registry-side collections with manual
cleanup. Align it with how additionalContexts already crosses the same
boundary: concludeTurn() marks only its own execution, the marker rides
that execution's successful result (ToolExecutionFailure types
concludesTurn as never, so a policy-converted failure sheds it with the
type), and the composite that owns the nested dispatch forwards it —
Code Mode's binding does so beside its existing context forwarding.
The registry loses both parent-keyed collections and the promotion
block; the propagation decision moves to the owning boundary; the
structured-output consumer's own two-phase commit is untouched.
master's bilingual README pairing (new since this branch forked) covers
packages/core/tools and packages/client/runtime, whose EN sides this PR
edits. Translate the scheduling-contract sentences (bridge pool, SDK
overlap line, loop cross-reference, codeDispatches lifecycle) into the zh
sides — including the verbatim shared model-facing block — and re-record
both pairing records.
Master made bilingual pairing mandatory repo-wide; this PR's seam-split
edits to the tasks docs get their zh counterparts: a new pair for the
dsh-tasks-local README and minimal updates to the tasks core-data doc,
agent-spine-demo README, and the tasks family READMEs, with pairing
records re-recorded.
agent-loop lifecycle: dispose drains machine.done to true quiescence.
cancel()'s own running-to-idle transition can legitimately re-enter
through an automation listener (goal-session's idle drive runs
synchronously to its first await) and replace done with a fresh
admission after the single capture; teardown now re-cancels and
re-awaits until the slot stabilizes, so the scope never unwinds under a
live run.
tools: a nested concludeTurn() stages on its own execution and promotes
to the enclosing composite only on the call's authoritative successful
verdict. A post-execute policy that converts the nested success into an
error no longer lets a recovering composite stop the turn on a failed
terminal operation (the Code Mode structured-output shape).
goal-session: the driver owns its round durability barrier again. The
loop's persistence is eager write-behind with no turn-end flush, so the
old post-turn agent/error signal for flush failures never fires; a
settled round now sets needsCheckpoint and re-enters drive, flushing
before the next reservation and disarming on failure instead of queueing
an autonomous round on state that was never persisted.
The public classifier existed to defend an exported reader against
arbitrary signals, but its only production caller is the loop reading
its own machine-private turn signal, where cancel() is the sole aborter
and always writes one frozen canonical cause. Delete the export and its
15-line structural validation: settle() states the slot invariant with
one cast, the boolean call sites ask signal.aborted directly, and the
retry veto drops entirely because a requested window already implies a
live signal (cancel() retires the window before aborting).
The abort(reason) channel and first-wins semantics are unchanged; only
the reader's publicness is gone, and with it the paranoia it required.
Two new keyless scenarios for the functionality master gained since this
lane's base (#644 websettings, #643 workspace browser rework), both zero
model calls:
- settings-chrome: the modal shell (sidebar-foot trigger aria states,
role=dialog, aria-current section switch to the deliberately empty
Models, Escape + close-button paths, dialog aria golden); the Appearance
row as the REAL theme gesture — retiring lifecycle-chrome's
TODO(web-theme-gesture): clicking 深色 runs aria-pressed -> persisted
dsh.theme -> body[data-ds-dark-theme] -> alias-token flip, survives
reload, and 'system' follows the emulated OS scheme both ways; the
Language row switches the settings-scoped copy to English (dsh.locale
persisted, survives reload) and restores zh. Intentional reloads tear
the SSE stream, so the spec drains exactly its own reconnect warnings —
the tripwire still fails on unexpected connection loss.
- workspace-management: create-by-name twice through the region-header
dialog (host-durable via ctx.workspace.list()); rename end to end —
hover-revealed row menu (the button is display:none until the row
hovers), duplicate-name pre-check (inline role=alert + disabled primary
before any wire call), then workspace.rename through the real RPC,
row update, host durability, reload survival; the flat 'In one list'
view (section label flips, group headers drop, dsh.workspace.view
persists across reload, grouped restored); the session hover card
(dwell to open, closes on pointer leave). The one session row reuses
seeded-history's committed seed — no new recording. Deliberately not
driven: the inert menu rows and drag reorder (deferred in the note
with re-entry triggers).
Agent Note gains scenarios 8-9 and the drag-reorder deferred item in both
languages; llm-replay README's zh side catches up with the { patches }
paragraph; pairings re-recorded.
The drive-pass guard for an unsettled attempt was wrongly annotated
unreachable after admission joined the running interval. It is reachable:
a contained turn-close failure (pre-commit turn/end rejection) reaches
idle with the attempt's turn open and no terminal reason, and the idle
drive pass must yield to that attempt instead of misreading the absent
reason as settled. Replace the ignore with a test driving exactly that
path; the guard's comment now names the real producer.
The admission-ordering fix changed the event taxonomy's observable order,
so the generated event/producer catalogs and the cordis API catalog are
regenerated. goal-session's mid-drive yield (an unsettled attempt seen by
an idle-triggered drive pass) became unreachable now that claimed prompt
admission joins the loop's running interval — a drive pass requires idle,
so it can no longer observe an unsettled reservation; the guard stays as
an annotated backstop.
Three ordering fixes on the admission and settlement boundaries.
kick() installs the abort owner, marks the interval busy (running is
emitted before any listener can observe the claim), and installs the
pending done BEFORE publishing agent/inbox/dequeue: a dequeue listener
that cancels or disposes now finds live cancellation and quiescence
ownership instead of the previous activity's settled state, and claimed
prompt admission — including asynchronous prompt-submit hooks — sits
inside the running interval where cancel routing can reach it. The
admission-rejected path yields one microtask before continueOrIdle so
the idle transition cannot fire inside send()'s synchronous extent.
agent/idle now names only committed turns: a run that aborts or fails
before its turn/start commits exits without the notification, since
there is no durable turn/end for settlement consumers to act against.
The event's JSDoc states the narrowed contract.
The coverage pass left agent.ts with eight v8 ignore annotations — well
above this repo's density elsewhere. Six guarded conditions the
surrounding invariants already exclude; expressing the invariant directly
deletes the dead arm instead of excusing it: admission/run teardown
releases the abort slot unconditionally (no writer can replace the owner
mid-flight), the request-failed branch closes its always-open step
unconditionally, the recovery finally retires the window it installed,
and the driver finally drops its dead last-resort step close (every path
already closes the step). kick()'s dequeue states the non-empty-queue
invariant as a non-null assertion. The two remaining ignores are the
assertNever exhaustiveness guard and the retry()/run() slot-race guard,
both repo-wide conventions.
