- tool-jobs: a completion notice carries the job's unread output when it fits inlineOutputBytes (default 2048); job_output withdraws a notice still pending, and the label is its first line bounded to 120 characters
- tool-bash, tool-pwsh: the promotion result keeps only the still-running marker
- subagent: a settlement notice omits the closing message when the child's last action was a message to its parent
Delete @deepseek-ai/dsh-invariants, every package's ./invariant companion
and its tests, the invariant gates and Vitest setup, the scoped-event
resolver generator, and the INVARIANT-coded listener rethrow paths.
Update docs, READMEs, Agent Notes, and add an upgrade guide.
Add a path-check reminder to the bash and pwsh tool descriptions. pwsh
also warns against assigning automatic variables such as $HOME; bash
also asks agents to guard path variables with ${VAR:?}.
Delete constraints the model learns from the call result, shorten
implementation detail, and move per-parameter rules into the owning
parameter descriptions. Bash, pwsh, and fs share one
sandbox_permissions description in dsh-sandbox.
The standard preset's first-turn prompt drops from 7985 to 6551 tokens.
Compaction threshold snapshots lower their context windows by the
1548-token tool-schema estimate reduction so each trigger keeps its
position; their fixtures move to v4 generations.
* feat(approval): localize prompts and allow consent for one turn
* test(approval): update cast baseline and Web PTC prompt pin
* fix(approval): cover PTC language and clarify turn consent
* fix(approval): group turn consent under allow menu
* fix(approval): keep turn consent in memory only
* test(snapshot): refresh owners for merged justification guidance
* fix(approval): drop duplicate allow-once from the allow menu
* refactor(approval): derive turn consent expiry from one check
Replace the turn approval bookkeeping (revocation state, incremental log
cursor, and turn-signal listeners) with one side-effect-free expiry check
over the log and lifetime flags. Events now only withdraw open
presentations early; the expiry check decides whether an answer or turn
consent still applies, including for Sessions without event publication.
* refactor(approval): drop one-turn consent and keep localization only
Remove the allowed-turn answer, the live turn-consent state, the allow
menu, and their tests, notes, and recorded Web scenario. The approval
service returns to the master request path; the request keeps the
localized displayReason for presentation, and the card resolves it in
the active UI language.
* test(snapshots): record the localized justification schema text in new scenarios
---------
Co-authored-by: Turtle <[email protected]>
Co-authored-by: Kaige-Gao <[email protected]>
With a job registry composed, tool-bash and tool-pwsh now register every
command with ctx.jobs as it starts and wait on the job with the registry's
own bounded wait: a command that finishes in time returns the ordinary
foreground result and its record is removed, one that outlives the timeout
returns the same job id it already had, and a kill from outside the call is
reported as `[stopped: <reason>]`. Nothing changes hands at the deadline, so
the shell seam loses `onExpiry: 'offer'`, `ShellExecution.promotion`, and
`ShellPromotionOffer`, and both local executors drop their offer arm.
The registry stays optional: the tools register a foreground-only definition
while no registry is composed and swap to the job-backed one inside a
`ctx.inject(['jobs'])` fork, so a composition without jobs keeps a plain bash
with no job tools in the model's schema. A refused registration at the start
(admission, no controller) runs under the executor's deadline kill, and a wait
that expires before the process spawned ends as the preparation timeout.
JobRegistry gains `remove(id, caller?)` and the `settled` event reports
`awaited`, so dsh-tool-jobs no longer keeps its own wait ledger; it withholds
notices for awaited settlements and for kills the model requested. The tool
catalog harvest composes the registry so the shipped schema stays documented,
and the two compaction fixtures move their context window from 10000 to 10300
tokens because the branch's larger tool schemas pushed their condensed history
over the old trigger, which requested a fourth scripted model call.
Run the dsh-archive-agent-notes audit over every active Agent Note on
current master, judging each record by whether its rationale still guides
work rather than by size or age.
- Archive 453 implemented bilingual triplets (417,882 English words):
completed UI chrome, narrow adapters, closed bug fixes, implementation
walkthroughs whose package READMEs, docs pages, generators, or successor
notes now carry the useful behavior, and 51 records fully superseded by
a later active note. Keep 201 implemented notes whose ownership rules,
negative guarantees, durable or wire semantics, security rules,
reintroduction conditions, or still-tempting rejected alternatives
remain useful.
- Reject 7 proposals whose premise is gone or whose work shipped in
amended form under other records; delete 2 rejected notes that no
longer prevent a plausible mistake.
- Retarget every remaining inbound link to the archived path, and repair
active prose that named an archived record as the owner of a live fact:
parenthetical citations drop, ownership sentences redirect to the
README, docs page, or active note that states the fact, and history
citations say so. Chinese files link the English archived path because
the pairing gate treats the frozen tree as outside the bilingual corpus.
- Seal 1,359 new frozen artifacts; existing seals are unchanged and
outbound links from archived notes are neither inspected nor repaired.
- Regenerate docs/config-catalog.md after the hook-bridge comment edits
shifted two source line numbers.
The tool-bash and tool-pwsh READMEs still said a promoted run is mirrored into ctx.activities, a seam the jobs consolidation deleted, and the shell README still called the observed readers optional and named the activity plane. The promoted run now feeds the registry's ring from the observed streams' current offsets and the Web list streams it through job.rows and job.observe; observed readers are required and carry the spawn-failure note.
tool-bash and tool-pwsh no longer render their own background reads.
processSources() wraps the process's observed offset readers as registry
pull sources, pumped at dsh-jobs-local's pumpPollMs, and processOutcome()
folds the settled sandbox facts into the terminal detail. observe.ts and
the recordPollMs config field are gone; READMEs and pairings follow.
Review corrections on the observation record:
- `pumpJobOutput` subscribes to `done` once and keeps one pending timer;
racing the same pending promise every poll round retained a reaction
per round for the life of the job.
- The `ctx.jobOutput` effect disposer is async and awaits every open
carrier's `dispose()`, so the fiber reports unloaded only once the old
iterators are quiescent.
- `SessionJob.record: true` replaces `outputTotal`: `append` never commits
a roster change, so the mirrored count was stale while output flowed;
the live offsets ride the observation stream's `opened` anchor.
- The record is documented as a best-effort live preview whose
cross-stream reordering is bounded by the poll interval.
- The web e2e scenarios hold their background job on a test-owned barrier
file instead of a wall-clock sleep, and each is one test rather than a
chain sharing a job id.
Carries the master sync forward (invariant-companion cleanup, alpha.3 bump);
regenerates the client api catalog so ISession lists both killJob and the new
loadThrough, and refreshes the new ptc-python-turn scenario's system-prompt
pin for this branch's bash timeout-promotion schema text.
Absorbs the repo-wide no-op invariant companion removal (ui-jobs loses its
placeholder companion, export, and devDep), the 0.1.2-alpha.3 release bump,
and the tool-subagent-report package retirement; regenerates the client slot
catalog for the shifted contract line numbers.
bash and pwsh background runs declare record: true and pump their observed
streams straight into the job through observeProcessRecord — the open/end
pairing, outcome mapping, and registry-absent degradation branches are gone,
and the pump's final drain folds into hooks.done so the record holds its
last bytes before settlement closes it. activityPollMs becomes recordPollMs.
tool-workflow drops its foreground activity mirror deliberately: a
foreground run's live phase/log narration had the mirror as its only
carrier, and the transcript keeps the recorded run/member lifecycle events.
Live narration returns with backgrounded runs when workflow gains
run_in_background.
The executor seam collapses run()/start() (and the planned begin()) into one
resolve() + execute(): the handle IS the live process, and the foreground
result, the background cursor, and the deadline's promotion offer are
projections over the same spawn. ShellExecSpec.onExpiry ('kill' default /
'offer' / 'none') replaces the implicit method split; one stdout budget applies
everywhere, the 'start ignores timeoutMs' wart dies, and sync spawn throws are
contained into the same settled-killed handle as async rejections (result()
rejects with the original error, identity preserved for the sandbox
classification).
The promotion offer settles exactly once and never rejects — offer at an
'offer' deadline while running, undefined when the process settles first — so
'await execution.promotion' alone distinguishes the outcomes. An offer must be
answered synchronously; an unanswered one auto-declines back to the
kill-on-timeout behavior. accept() ends the deadline and detaches the caller's
abort signal; decline() kills and classifies timedOut.
tool-bash and tool-pwsh promote on timeout by default (promoteOnTimeout: true,
inert without run_in_background or a job registry): at the offer the running
handle registers as a job with the same hooks as run_in_background, mirrors
into ctx.activities (the web task list streams it, stop control included), and
the call returns { kind: 'promoted', jobId, timeoutMs, output } carrying the
output so far — the job cursor continues exactly after it. Any promotion
failure declines and falls back to the plain timeout kill. Descriptions and
the timeoutMs schema text advertise the semantics only when they hold;
recorded-session and web snapshot fixtures re-recorded with them.
Also repairs the four CI failures from the human-kill push: ui-activity's
missing dsh-session peer declaration (static + windows observational), the
observe owner-arm/outer-catch and killJob-stub coverage gaps, tool-workflow's
owned-run arm, and the timing-dependent projection-cache interval flush test
(now poll-based for coarse Windows timers).
Peer evidence and rejected alternatives (Codex's session model, a third seam
method, a second threshold knob) in the Agent Note
2026-08-26-shell-execute-timeout-promotion.
Add the activity capability seam: ctx.activities is a pure observation
plane where long-running work (background bash/pwsh, workflow runs) opens a
handle, appends transient output into a byte-offset ring buffer, and settles
exactly once. Nothing model-visible changes: reads are non-consuming and
lossy-by-design (absolute UTF-8 offsets survive eviction and report gaps),
correlation carries {callId, jobId}, and visibility is owner-fenced like jobs.
The seam is strictly optional. Producers tap it through ctx.get('activities')
and contain every observation failure, so an unregistered or broken registry
never affects the underlying job (proven by broken-registry tests).
Ship the full path in one change:
- @deepseek-ai/dsh-activity: Service Definition, pumpActivityOutput pull
substrate, invariant companion; dsh-activity-local: ring-buffer provider
with retention config and owner cleanup.
- shell: ShellProcess.observed exposes non-consuming stdout/stderr readers;
tool-bash/tool-pwsh mirror background jobs into activities.
- tool-workflow: phase/log/agent lines mirrored as a workflow activity.
- api-activity-controller: Typert Remote control (roster baseline+buckets)
and observe (opened/output/status frames, coalesced and byte-capped)
streams plus the client feed service with reference-counted observation.
- ui-activity: session-header action rendering live output through
TerminalBlock, which now shows output while a command is still running.
Docs: activity subsystem page, architecture row, regenerated catalogs, and
the Agent Note (2026-08-24-activity-observation-seam), all bilingual. A
keyless web e2e replays a real run_in_background stream to ARIA goldens.
turtle1999