* fix(agents): wake top-level requester when its last parallel child settles [AI]
A top-level session (normal chat/dashboard, not itself a subagent) that
spawns parallel subagents and waits via sessions_yield never receives the
"all descendants settled -> synthesize" wake: wakeSubagentRunAfterDescendants
only targets orchestrators with a subagent-registry run record. The parent
only sees passive per-child announces, commonly mis-tracks the outstanding
set on the final completion turn (or never hears results whose announce gave
up), and then parks until a human sends a message.
Fix: when a child reaches a terminal settle (announce delivered, give-up, or
delivery suspended) and its requester has no more descendants awaiting
settle, deliver a one-shot "all spawned subagents settled - synthesize and
deliver now" wake to the parked top-level requester through the existing
announce delivery pipeline (active requesters get it steered into the live
turn; parked ones get a new origin-routed turn).
- hasDescendantRunAwaitingSettle: early-exit drain check where a suspended final
delivery counts as settled (suspension is terminal for automatic retries).
- maybeWakeRequesterAfterAllChildrenSettled (subagent-announce.ts): scopes
the batch to the settling child's parallel wave (the connected component of overlapping run
lifetimes), skips nested/cron requesters (owned by the descendant-settle
wake), skips single delivered completions and fire-and-forget children,
and dedupes concurrent last-sibling settles via a batch-stable
announce idempotency key (requester-settle:<requester>:<sorted runIds>).
- Trigger fires from completeCleanupBookkeeping (all cleanup-terminal paths)
and suspendPendingFinalDelivery, and is skipped by the suspended-delivery
discard sweep, so the wake is outcome-independent: a child whose announce
gave up with NO_REPLY still counts toward - and can trigger - the drain.
Companion to the nested-orchestrator wake; covers the top-level case it
left open.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rukuut98qvDwzSwMp3u4kN
* fix(agents): retry the requester settle wake on transient turn failures [AI]
The settle wake is the only event that ever fires after a fan-out drains, so
a wake turn lost to a transient infra failure re-parked the requester
permanently. Observed live: the wake turn survived dispatch, reconciled all
investigators, and then died to a provider stream stall (LLM idle timeout)
mid-synthesis — with nothing left to retry it.
Bounded recovery: up to 3 attempts with 30s/120s backoff. Each retry uses a
fresh `:retry-N` idempotency suffix because the gateway dedupe caches
terminal run outcomes per key, so re-dispatching the same key would no-op. A
legitimately silent wake reply already classifies as delivered and never
retries; terminal failures and an abandoned requester stop immediately.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rukuut98qvDwzSwMp3u4kN
* fix(agents): ledger retired delete-cleanup rows for the requester settle wake [AI]
cleanup="delete" retires a child's registry row in the same funnel pass
that schedules the settle wake, so a pure delete-mode fan-out could never
rebuild its drained wave from live rows: empty batch, no wake, requester
parked — the exact incident class this wake exists to close.
Retiring cleanup paths (delete branch, reconciled-killed tombstone) now
pass settledRowRetired, and the wake ledgers the in-hand record before
its first await: concurrent last-sibling settles all see the same
ledgered rows by the time any batch computation runs, keeping batch
membership and the idempotency key stable — the guarantee keep-mode gets
from rows persisting in the registry. Ledgered rows merge into the batch
candidates (a live registry row wins by runId); entries clear on
delivered/terminal wakes and the nested/requester-gone exits, with a 24h
TTL prune as the memory backstop. In-memory only: a restart degrades a
mid-wave delete fan-out to waking with the surviving rows.
* fix(agents): snapshot delete-mode child results before cleanup clears them [AI]
The delivered-finalize path clears completion.resultText (delete mode) and
the frozen delivery payload before cleanup bookkeeping runs, so the
requester-settle wake ledgered a row whose findings had already been
emptied: a cleanup="delete" fan-out woke its requester with "(no output)"
for successfully delivered children.
Capture an immutable snapshot of the entry (completion + delivery payload)
at the top of finalizeSubagentCleanup and in the resumed-announce give-up
path — before any clearing — and pass it through completeCleanupBookkeeping
to the settle wake for the retired-row branches. Keep-mode is unchanged:
its registry rows keep their result text and win over ledgered copies.
Regression: a lifecycle-driven test drives the real delivered delete-mode
cleanup (completeSubagentRun -> announce -> finalize -> funnel) and asserts
the wake's settledEntry still carries the child's result text after the
live entry was cleared and the row retired.
* fix(agents): release drained no-wake batches from the settle ledger [AI]
The zero-required and single-delivered guard exits decline to wake a wave
that has already fully drained. Ledgered delete-mode rows in such a batch
can never join a later wake (a child spawned after the drain cannot
lifetime-overlap rows that already ended, and any still-running overlap
would have failed the drain gate before these exits), so holding their
child-result snapshots for the 24h TTL was pure retention: high-volume
delete-mode fan-outs could pin large result texts in process memory.
Release the batch at both exits. No behavior change: waves with pending
overlapping children return earlier at the drain gate, which still keeps
their rows.
* test(agents): route e2e session fixture through the sqlite accessor [AI]
Upstream flipped session reads to the sqlite-backed session accessor
(#98236), which bypasses the loadSessionStore mock this e2e used for its
in-memory session fixture — requester entry lookups came back empty and
the settle wake exited at the usable-session guard. Serve loadSessionEntry
from the same fixture; everything else in the accessor stays real.
* refactor(agents): move the requester settle wake into its own module [AI]
Two pieces of fallout from rebasing across upstream tooling changes:
- The max-lines lint budget (#107315) has no grandfathered suppression for
subagent-announce.ts, and the settle wake + retired-row ledger pushed it
past the cap. The wake is a coherent unit, so it moves to
subagent-announce.requester-settle-wake.ts (matching its test file) with
its own registry-runtime seam; the registry's lazy dep and the lifecycle
type alias point at the new module. announce.ts drops back well under
the budget with no suppression needed.
- The announce read-path refactor rewired the runtime barrels onto
subagent-registry-announce-read.js, orphaning the
hasDescendantRunAwaitingSettle re-export on subagent-registry.ts, which
the deadcode-exports gate now flags. Removed; the retry-grace e2e reads
the announce-read implementation directly.
* fix(agents): admit the requester settle wake as tracked gateway root work [AI]
The settle wake was launched as a detached promise from cleanup
bookkeeping, so registry cleanup or shutdown could reach quiescence
before the wake admitted its gateway turn and the last-child completion
could still be lost during restart or teardown. Route the wake through
runWithGatewayIndependentRootWorkContinuation: a live cleanup parent
reserves the root synchronously, and restart drain now waits for the
in-flight wake. Adds a deterministic quiescence-race regression.
* fix(agents): persist requester settle recovery
Store requester-settle wake obligations on subagent run rows, replay them after restart, and retire cleanup rows only after a durable outcome. Persist admitted attempts and retry deadlines, coalesce live restores, drain disconnected waves, and let the sweeper recover stranded processors.
Co-authored-by: smthfoxy <263563487+smthfoxy@users.noreply.github.com>
* test(agents): allow durable settle persistence
Keep the bulk-cancellation regression focused on recovery after the injected persistence failure instead of assuming an exact write count; requester-settle bookkeeping now adds a legitimate durable write.
Co-authored-by: smthfoxy <263563487+smthfoxy@users.noreply.github.com>
* style(agents): declare settle wake scheduler
Use a hoisted function declaration for the lifecycle scheduler so lint accepts the intentional callback cycle without a mutable binding.
Co-authored-by: smthfoxy <263563487+smthfoxy@users.noreply.github.com>
* fix(agents): harden requester settle recovery
Co-authored-by: smthfoxy <263563487+smthfoxy@users.noreply.github.com>
* fix(agents): release settle wake during retry backoff
Co-authored-by: smthfoxy <263563487+smthfoxy@users.noreply.github.com>
* test(agents): type settle wake fixtures
Co-authored-by: smthfoxy <263563487+smthfoxy@users.noreply.github.com>
* style(agents): bind settle wake transition callbacks
Co-authored-by: smthfoxy <263563487+smthfoxy@users.noreply.github.com>
* test(agents): restore past-due settle wake
Co-authored-by: smthfoxy <263563487+smthfoxy@users.noreply.github.com>
---------
Co-authored-by: smthfoxy <263563487+smthfoxy@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* feat(mattermost): add thread.requireExplicitMention to opt out of thread auto-follow
Mattermost treats any reply in a thread the bot has participated in as an
implicit mention, so requireMention only gates the first message and the bot
then answers follow-ups addressed to other people for the participation TTL.
Slack exposes channels.slack.thread.requireExplicitMention for exactly this;
Mattermost had no equivalent and its strict schema rejected the key.
Add channels.mattermost.thread.requireExplicitMention (channel + per-account),
mirroring Slack. When set, thread participation no longer counts as a mention.
Default (unset/false) keeps today's auto-follow behavior unchanged.
Related: #108269
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* feat(channels): add implicit mention policy foundation
* feat(channels): unify implicit mention policy
Co-authored-by: leon <dodoma0919@gmail.com>
* refactor(channels): keep implicit policy inside evaluator
* fix(channels): use exported implicit mention type
* chore(channels): satisfy extension lint
* fix(config): break implicit mention type cycle
* fix(plugin-sdk): account for implicit mention config export
* refactor(config): isolate implicit mention schema
* chore(plugin-sdk): align implicit mention surface budget
* fix(config): remove unused schema re-export
---------
Co-authored-by: leon <dodoma0919@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(process): clarify command lane timeout cause
* fix(process): complete timeout cause formatting
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(cli): filter Commander value placeholders from Fish completion flags
The Fish completion generator was treating Commander.js value placeholders
(like <path|->, <file>, [optional]) as command-line tokens, causing the
angle-bracket values to be interpreted as shell redirects when the
completion script was sourced.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(cli): use parsed option flags for completions
Co-authored-by: 丁宇婷0668001435 <ding.yuting@xydigit.com>
* fix(cli): support long option aliases in Fish completion
Co-authored-by: 丁宇婷0668001435 <ding.yuting@xydigit.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
Co-authored-by: Peter Steinberger <peter@steipete.me>
* test(slack): drop rotating loading-message fixture expectations
* refactor(slack): drive progress drafts through the shared compositor
Slack's status_final progress path now uses the shared channel compositor
(preamble headline, reasoning, tools, plan checklist, commentary
arbitration) instead of its hand-rolled gate/merge/render pipeline, and
native task cards consume compositor snapshots with stable row identity
(line ids first, fixed reasoning id, occurrence-suffixed content hashes).
Finals in progress mode post as fresh messages that follow the draft's
thread; the working draft then collapses into a compact activity receipt,
while failed or error finals leave the draft as the turn record. The
rotating native loading messages are gone in favor of the plain typing
status. Queued and assistant-boundary turns re-arm with clean delivery
state, and native completion status survives transient append failures.
* fix(proxy-capture): guard body-less arrayBuffer reads against oversized responses
* fix(proxy-capture): exercise body-less fallback in bounded read tests
New tests use mock clones with body: null plus arrayBuffer spies to
prove the content-length precheck guards the !body path. A real
Response clone exposes body.getReader in Node 24, so the prior test
only exercised the streaming branch and would stay green even if
the precheck were deleted.
* chore: retrigger CI
* fix(proxy-capture): reject non-safe content-length before arrayBuffer (#101268)
ClawSweeper P2: the body-less fallback used Number(content-length), so a
huge digit-only Content-Length value could overflow to Infinity, bypass the
Number.isFinite guard, and still call arrayBuffer() — leaving an OOM path in
the hardening PR.
Add declaredContentLengthExceedsCap, which accepts only plain digit strings,
treats any value longer than Number.MAX_SAFE_INTEGER as oversized, and
compares safe-integer parsed values against the cap. Non-numeric or malformed
values fall through to the post-read length check.
Adds a regression test for a 100-digit Content-Length that would previously
have bypassed the guard.
* fix(proxy-capture): normalize zero-padded Content-Length before digit-count guard
* fix(proxy-capture): fail closed without response streams
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(sandbox): use Buffer.byteLength for env var value size limit
validateEnvVarValue checked value.length (UTF-16 code units) against
the 32768-byte limit, so multi-byte CJK values like "值".repeat(11000)
passed the check despite exceeding 33 KB in UTF-8. Switch to
Buffer.byteLength(value, "utf8") so the limit matches the actual byte
count the OS and child processes see.
* test(sandbox): simplify env byte-limit coverage
Co-authored-by: 唐梓夷0668001293 <tang.ziyi@xydigit.com>
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
Peter Steinberger