The injected instruction persists as a durable user message, so an
unscoped 'Do not call any more tools.' stays in every later request's
history. Scope it: '...in this run; further work waits for the user's
next instruction.' A/B probes on deepseek-v4-pro show the scoped wording
is non-inferior in-turn (4/4 zero tool calls, closing quality unchanged)
and next-turn tool use is unaffected under both wordings; the scoped
form states the instruction's actual lifetime.
- fromRequest placeholder: the last two braces of a consecutive } run now
terminate the placeholder, so patterns may end with a brace quantifier
(bot warning; the truncated pattern could even silently mis-match since
an unclosed { is literal in JS regexes)
- document that derived JSONL entries pass through the same resolution
- widen ToolRunContext/deferContext seam docs beyond composite-only usage
(source JSDoc, README pair, core-data-structures type-equiv blocks)
- pin direct-human blocked as uninstructed, completing the
goal-round/direct-human x complete/blocked test quadrant
A goal round reporting complete or blocked used to conclude the turn at
the tool result, so the model never spoke after the call and sessions
ended on a bare update_goal card. The terminal update now defers one
plugin-sourced <goal_complete>/<goal_blocked> instruction onto its
result asking for a grounded closing message without further tool
calls; the turn then ends through the ordinary no-tool-calls stop.
Direct-human mutations stay uninstructed. Wording chosen by A/B
sampling on deepseek-v4-pro; one extra request per goal lifecycle.
New keyless ACP snapshot goal-wrapup drives the shipped app through
create -> round one -> autonomous complete and pins the injection, the
same-turn closing message, and the completed turn end.
goalChangeRef returned the full GoalSnapshot for every snapshot operation,
so foldGoal(...).lastRef and the goal/changed notification ref carried
objective, phase, and maxGoalRounds fields instead of the declared
GoalRef { id, revision }. Only the clear tombstone was bare. Emit an exact
{ id, revision } ref for snapshot changes and pin the contract with a
regression test covering create/edit/block notifications and the fold.
- projection.spec covers applyGoalProjection's non-message early return and
the foreign-change-kind arm directly (the pure-function posture the unit
requires — a throwing apply would tear down the registry drive).
- New ui-goal browser-plugin spec exercises the plugin body over a real
cordis Context with fake faces: dock registration (id/order), verbs
reading the CAS ref from the current projected value, the
no-current-goal short circuit, RPC-error mapping, HMR disposal, the
GoalDock adapter's three projection states, and the inert node half.
dsh-goal/types becomes the client-safe pure outlet (tool-todo dual-outlet
shape): GoalId/GoalRef/GoalPhase/GoalBlockReason/GoalSnapshot, the new
GoalProjection payload, and the SessionProjectionMap 'goal' key merge — zero
host imports. Host-coupled vocabulary (GoalView/activation, change metas,
message source, folds, GoalError codes, the scoped goal/changed event) moves
to src/domain.ts, re-exported from the package root. ./client re-exports the
outlet for client aggregates.
GoalService registers the projection unit under ctx.inject(['sessionProjections']):
applyGoalProjection is a projection-grade last-wins fold — plain-JSON state,
same-reference return on non-goal or malformed events (a throwing apply
would tear down the registry drive; strict validation stays with the write
side and foldGoal). Activation is process-local and deliberately absent from
the projection value.
An out-of-band zero-step turn (durable command lifecycle on an idle log)
advances the log's turn numbering behind ReactLoopAgent's cached lastTurn,
so the next real turn reused a stale number and tripped the session
invariant (turn/start expected N, got 1) — hanging the TUI after any idle
slash command. The log is the numbering authority: take max(cached, logged)
+ 1 at open. The command-goal stub's inject helper also gains the one-shot
injection turn wrap the real agent performs, restoring turn enclosure in
its log assertions.
CommandService.execute now returns a CommandExecution — the normalized
result plus the commandId minted for its command/run/command/done records —
and the wire admission value carries commandId exactly when matched, so the
issuing client can correlate its RPC acknowledgment with the flow node the
lifecycle events produce. apiproxy api/schema/handler, the connection
fixture, and the TUI/plan/goal consumers follow the new shape.
CommandService.execute appends the log-only pair around every resolved
handler — run before invocation, done at settlement, including thrown and
aborted handlers (kind:'error'); admission misses log nothing. commandId is
minted monotonically per instance; per-session appends serialize through a
tail queue over SessionStore.appendOutOfBand (zero-step wrap on an idle log,
direct join inside an open turn). The invariant companion now asserts the
pairing relation (unique run ids; a done requires a prior in-log run).
CommandSource is a minimal merge-extensible map (user variant only).
Dependent benches mount SessionStore; TUI/e2e snapshots re-recorded for the
executor's durable-append timing and the /status event counts.
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.
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 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.
