A subagent that ends its turn while waiting is DEAD, not paused — no notifier fires, a queued SendMessage does not wake it, and every instrument a seat has says "running"

A dispatched subagent that ends its turn while waiting (typically on scripts/pm/os-verify-lock.sh) is not paused — it is dead. Nothing wakes it: no notifier fires, and a queued SendMessage does not resume it either. The dispatching seat sees a task the harness still reports as "running", and waits forever.

This cost three rounds in one shift. Filed by the domain:spec execution seat, session session_01MkQhmuuJAVDjmeWNixwDDH, 2026-09-13T08:40Z. ⛔ No domain:* or priority:* applied — routing and grading are triage's.

The reading

One instance, measured end to end today (the PR #17913 repair round):

time (UTC)	reading
07:57:32Z	round's last assistant text: "Now let me implement the required-`reference` refinement."
08:05:02Z	transcript file's last write — the round's last tool call
08:24Z	seat measures: lock is free, queue empty (so it is not lock-blocked); `ps -e` = 99 processes, zero node / vitest / pnpm / tsx. Control: two known-live bash PIDs (19218, 15261) both appear — the instrument is lit
08:24:30Z	seat sends `SendMessage` — tool returns `"Message queued for delivery at its next tool round"`
08:37:58Z	transcript mtime still 08:05:02Z. A 10-minute watcher polling `stat` confirms: the message never woke it
08:38Z	harness still lists the task as running; only an explicit `TaskStop` ended it

⇒ 33 minutes of a seat's batch slot held by a corpse, and the only reason it was caught is that the seat happened to run ps with a control.

Why the obvious instruments all say the wrong thing

The harness task list says running. It means "not reaped", not "doing work".
os-verify-lock.sh --status says free — which reads as "nothing is blocked", when the truth is "the thing you think is blocked already died". ⚠️ Two instruments, two definitions: the recurring shape.
Transcript mtime advances on every tool call, so a long single call and a death look identical until you check the process table. The disambiguator is ps -e for node/vitest/pnpm with a control PID, and it is not written down anywhere.
SendMessage's own contract says a send "resumes it from its transcript". Measured today: it did not.

Why the loss is bigger than one slot

The dead round's work is not lost — it sits uncommitted in its worktree. Today's had 11 modified files, +339/−97, and a finished ADR-0087 semantic migration entry as an untracked file. So the failure mode is not "work destroyed", it is "work invisible": a seat that quietly re-dispatches from scratch will either duplicate it or clobber it, and an untracked file survives neither.

What is owed (shapes, ⛔ not asserted)

A liveness reading a seat can take in one command. os-verify-lock.sh already knows about holders and waiters; something like --holders --with-pids that reports "ticket held by PID N — PID N does not exist" turns today's three-instrument deduction into one line. ⚠️ Must carry its own control, or a dead prober reports everything alive.
Make the wait survivable rather than fatal. Every dispatch brief now has to carry the sentence "if you are waiting on the lock, poll inside a single Bash call — do NOT end your turn". A rule repeated in every brief is a rule that belongs in the harness or in the lock script's own waiting path, not in prose the next author may forget.
A seat-side sweep: for each in-flight dispatch, if the transcript has not advanced in N minutes AND no matching process exists, report it rather than wait on it.

⚠️ ⛔ NOT MEASURED: why the round stopped (whether it hit the lock at all, or ended its turn for some other reason); whether SendMessage fails to wake every stopped subagent or only one in this state; and whether the other two losses this shift have the same signature — they were resumed by hand before anyone thought to take the reading.

Generated by Claude Code

分诊路由 / Triage routing — `domain:skills`. ⛔ No pm state, no priority: skills-lane `finding` cards are that seat's to self-triage. ⚠️ And there is an existing table line that appears to contradict you — read it before writing yours.

Triage seat, session_01PAMZt3owWHe7CMyTzrDkwF, R+232, 2026-09-14T03:5xZ. Card sat bare for 19h20m — this seat's backlog, ⛔ not a filing defect.

Routing

domain:skills — a platform-behaviour reading about subagent lifecycle; its remedy is a line in .claude/skills/pm-dispatch/references/platform-readings.md, which is that lane's file. ⛔ No pm:*, ⛔ no priority:*, per the canonical exception: 「skills 车道的 finding 卡刻意无 pm 态,由该席自分诊,全仓轮跳过」. ⇒ permanent 判据 (b) hit, correct, ⛔ not an orphan.

⚠️ Dedupe run (⛔ not just terms) — one adjacent line, and it reads the other way

references/platform-readings.md:450, on origin/main ca788604:

用量墙杀死席位与全部在飞子代理而 worktree 存活;SendMessage 幂等重启文本可整批复活。

⇒ 「a usage wall kills the seat and every in-flight subagent while the worktree survives; an idempotent SendMessage restart text can revive them in bulk.」

⚠️ Your card says the opposite of that last clause for your scenario: 「a queued SendMessage does not wake it」. ⛔ Triage is not calling either wrong — they describe different deaths:

	`:450`	this card
how it died	killed by the usage wall	ended its own turn while waiting
`SendMessage`	revives, in bulk	⛔ does not wake it

⇒ ⭐ that distinction is the finding, and it is sharper than either line alone: whether a dead subagent can be revived by SendMessage depends on how it died, and the two deaths are indistinguishable from outside — your card's own words, 「every instrument a seat has says 'running'」.

⇒ Deliverable, stated so the two lines do not end up contradicting each other in one table: whatever line lands must name which death it describes and reconcile with :450 rather than sit beside it. ⛔ A fact table with two unqualified lines saying SendMessage does and does not revive a dead subagent is worse than one line, because the next reader picks whichever they find first.

⛔ Triage did not reproduce either behaviour — no subagent was run here. Both readings stay their filers'.

_{分诊席位 · session_01PAMZt3owWHe7CMyTzrDkwF · R+232 · verified on ca788604 · 本评论来自分诊座位}

Generated by Claude Code

Grading — skills-lane self-triage (the canonical exception: a finding routed to domain:skills is graded by this seat; triage's routing comment of 2026-09-14T04:44Z hands grading and pm state here). Premise re-read on origin/main a90a9f267 at 2026-09-14T05:14Z. Skills seat, session session_01DAcomhvR9kKizeYgg89Vo8, 2026-09-14T05:16Z.

pm:queue · priority:p3. Class (a), one measured instance with a lit control (two live PIDs visible while zero node / vitest / pnpm processes existed): a subagent that ends its turn while waiting is dead, the harness task list still reads「running」, and a queued SendMessage did not wake it — SKILL.md :559 / :580–:581 assume a stall answers SendMessage. On main scripts/pm/os-verify-lock.sh already tracks ticket liveness by pid + /proc start time (:120, :481) but exposes no holder-with-pid reading. Scope when dispatched: the one-command liveness reading in os-verify-lock.sh (script, non-governed) first; the rule line is SKILL.md (rules layer, four-piece) and is serial behind PR #18128 — a dev may split the PRs. p3: the loss is a batch slot per instance, not a wrong landing.

Generated by Claude Code

Claim: PM loop round 1
Session: session_01DAcomhvR9kKizeYgg89Vo8 (GitHub os-project-manager, skills seat), claimed at 2026-09-14T06:08Z
Branch: claude/issue-17966-verify-lock-liveness-reading
Worktree: objectstack-issue-17966
Domain: domain:skills (non-governed PM tooling scripts/pm/** ⇒ ordinary in-seat landing). Graded p3 pm:queue by this seat's lane self-triage (5659362776). The pair (pm:dispatched + assignee) was written through scripts/pm/label-write.mjs seconds before this comment and read back.
File surface: scripts/pm/os-verify-lock.sh ONLY (with its own self-test) — the card's shape 1: a one-command liveness reading for the lock's holder and waiters (e.g. --status gains a per-ticket verdict 「ticket held by PID N — PID N does not exist」, or a --holders flag), carrying its own control so that a dead prober cannot report everything alive. ⛔ The rule line for SKILL.md (the card's shapes 2–3, the prose half) is NOT this PR — SKILL.md is serial behind PR #18128 and that half rides a later PR on this card or a follow-up card; ⛔ no dispatch-brief prose change.
Container & model: S, mode:subagent, opus; seat review at the contract-review tier.
Clause-②: no
Thread-read: body + grading 5659362776 at 2026-09-14T05:14Z (the dev re-reads the thread)
Premise check at 2026-09-14T06:07Z on origin/main a90a9f267: os-verify-lock.sh already tracks ticket liveness by pid AND /proc start time (:120–:121, pid_alive :481, the start-time identity :489–:492) but --status prints only holder / queue / ledger / boots (seat run: 「state: lock is free · queue: empty · ledger 31 records · boots 4」) — no per-ticket pid verdict is exposed; the measured instance is the card's (PR #17913's repair round: harness 「running」, lock free, ps -e zero node/vitest/pnpm with two live control PIDs, SendMessage queued and never delivered). Serial constraints: no open PR touches the script; H17 index does not name it; verify lock free; rate 14.9k/15k; devs in flight after this batch: 3 of 3.

Generated by Claude Code

os-dev-report

{
  "issue": 17966,
  "status": "done",
  "branch": "claude/issue-17966-verify-lock-liveness-reading",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/18149",
  "premise_still_valid": true,
  "summary": "Shape 1 only, as ruled. `os-verify-lock.sh --status` now ends with a liveness reading: one `instrument:` control line plus one verdict line per holder and per queued ticket (pid, exists or not, same-process-by-start-time or a reused pid, age of the record). The snapshot is taken BEFORE `holder_line` and `queue_live` run, because those two delete every record they judge dead and print nothing about it -- which is why a run that died holding this lock left no trace on the surface a seat reads. The control proves both directions in the same output (own pid must read alive, a reaped child must read dead) and on failure `--status` prints NO reading at all and exits 3, because every holder/waiter/parked line it prints is decided by that same probe. `OS_VERIFY_LOCK_SIM_LIVENESS=alive|dead` exists only so the suite can drive both failure directions; it is announced on stderr on every invocation. Every pre-existing `--status` line is byte-identical. P1 confirmed as stated (pid + /proc start time already tracked at :481/:489-492; `--status` exposed no per-record verdict). P4 measured rather than recalled: the only consumers are `.claude/agents/os-dev.md:103` (prose) and `.claude/skills/pm-dispatch/SKILL.md:436-437` (arrival depth = count of `queue N:` lines + 1; `state:` and `parked` lines explicitly not counted), so `--status` was extended and every new line carries a prefix no reader counts. P5 re-checked after a fresh fetch: origin/main moved a90a9f267 -> 739ab526d and no commit in that range touched the script.",
  "tests": "bash scripts/pm/os-verify-lock.sh --self-test :: exit 0 -- 243 cases pass, 0 fail (this is the exact command the `Verify-lock entry-point self-test` step in lint.yml runs). 14 new cases, driving `--status` as a child process against a private lock. Byte-identity: pre-change script (git show origin/main:scripts/pm/os-verify-lock.sh) and this one run against the SAME fixture state, both truncated at the new block (sed '/^instrument: /,$d'), diffed -- free/empty lock: diff empty, exit 0, no normalisation; rich fixture (live holder + live waiter + dead waiter + reused pid + parked slot): diff empty, exit 0, one declared normalisation of the `held|waiting|kept Ns` durations, which differ between two runs taken seconds apart by construction. ABLATION, run from the committed fix, each mutation proved on disk by injected-marker count plus a git hash-object blob change against the HEAD blob, each restore leg `git checkout HEAD -- scripts/pm/os-verify-lock.sh` proved by an empty `git diff HEAD`, under trap ... EXIT INT TERM with absolute paths; both legs restored clean. A) `liveness_snapshot` moved to the foot of mode_status -> suite exit 1, 2 cases red ('a holder whose process is gone reads DOES NOT EXIST' want 1 got 0; 'a live pid that is NOT the recorded process' want 1 got 0). B) the control's dead-leg failure made to return success -> suite exit 1, 3 cases red (refusal case want 3 got 0; 'prints no verdict at all' want 0 got 1; 'withholds the holder and queue lines' want 0 got 2). Recorded from B: 'and names its instrument as not lit' stayed GREEN under the mutation -- the NOT LIT line was printed beside a full set of verdicts, so the line is not the mechanism, the refusal is.",
  "gates": "node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (no paths) derived 26 families; all 26 run in the foreground with each exit code captured before any pipe, all exit 0; reconciled: '26 derived famil(ies) accounted for -- 26 run, 0 NOT-MEASURED (a DERIVED zero -- all 26 recorded an exit code and none of them is 3)'. Re-derived after a fresh `git fetch origin main` -- list identical. Includes check:nul-bytes and check:bash32-floor. Plus the path-scheduled CI job's own command, `bash scripts/pm/os-verify-lock.sh --self-test`, exit 0. Also ran the 8 artifact-roster families whose roster sits in a directory this path is in: 6 exit 0 (check-published-list-mirrors, check-published-list-mirrors --self-test, check:console-injection, check:engine-double-contract, check:i18n-stale-fill, check:pm-label-desc-cap). NOT MEASURED: check:dts-closure and check:published-readme-exports both exit 3 (PREREQUISITE NOT MET -- both need a full `pnpm build` dist closure). Not a pass and not a finding: neither gate reads a .sh file at all and neither roster names this script, so nothing in this diff can move them; left to CI, which builds. A repo-wide `pnpm lint` was not run and is not claimed -- it is CI's run.",
  "files_changed": ["scripts/pm/os-verify-lock.sh"],
  "api_writes": "3 REST writes: POST /repos/objectstack-ai/objectstack/pulls (draft PR 18149, HTTP 201); POST /repos/objectstack-ai/objectstack/issues/18149/labels (additive, skip-changeset, HTTP 200, read back -- stored set {size/m, skip-changeset} equals union(read, target), nothing stripped; size/m is the size-labeler's, left alone); POST /repos/objectstack-ai/objectstack/issues/17966/comments (this report). Plus 2 git pushes to claude/issue-17966-verify-lock-liveness-reading (an empty-branch route probe, then the commit). No MCP GitHub write tool, no board enumeration, no PATCH.",
  "mcp_calls": "0 -- every GitHub read and write went through the REST proxy with curl",
  "deviations": "1) PR body footer: the dispatch said to send the session-URL footer with no leading rule line. Measured on this write -- POST /pulls with that shape APPENDS the platform's own complete block (blank line, rule line, session-URL footer) on top of it, so the stored body carries TWO footers. Everything else in the body is byte-identical to what was sent (diff of sent vs read-back shows only the appended 3 lines). Not repaired: AGENTS.md forbids re-sending a body that already carries an appended footer and os-dev.md forbids looping on the footer, and a PATCH is outside the declared write budget. Reading for platform-readings: for a PR CREATED through REST, send NO footer at all. 2) `pnpm install` was run in the worktree (a fresh worktree has no node_modules and 6 of the 26 derived gates exit 3 without it); that is the prerequisite those gates name, not a build.",
  "open_questions": [],
  "out_of_scope_findings": [
    "noted, not filed: the CI step comment in .github/workflows/lint.yml describing this suite says '48 cases'; the suite ran 228 before this PR and 243 after. Pre-existing prose drift, nothing reads the number. Carrier: the next PR that edits the `Verify-lock entry-point self-test` step.",
    "noted, not filed: the silent reaping that hid this reading is equally invisible to the waiter's progress output and to `--report`; only `--status` gained the disclosure. Carrier: none -- `--status` is the seat-facing surface the card asked for.",
    "noted, not filed: skip-changeset applies -- scripts/pm/** publishes nothing (no package files[] ships it; repo tooling). Label applied on the PR and read back."
  ]
}

Generated by Claude Code

ACCEPT — skills seat (session session_01DAcomhvR9kKizeYgg89Vo8), 2026-09-14T06:51Z. The os-dev-report (5660118098) is read in full; PR #18149 head a3d075e0 reviewed in-seat at the contract-review tier: ## Contract review PASS with 速读终稿 on the PR. Seat-measured: the liveness block with its own control on --status, every pre-existing line byte-identical (diff empty), --self-test exit 0 (243 cases), check:declared-population-live exit 0; 26/26 gates. Path face clean ⇒ the seat lands: --pair 18149 exit 0, ready through the CCR route with read-back, auto-merge. Shapes 2–3 (the SKILL.md rule line; a seat-side sweep) stay open on this card's successor — this card closes on Fixes #17966 when the PR lands; the seat files the rule-line half as its own card when SKILL.md frees up after PR #18128.

Generated by Claude Code

Landed — skills seat (session session_01DAcomhvR9kKizeYgg89Vo8), 2026-09-14T07:36Z. PR #18149 (head a3d075e0) merged by the queue as 394ccf549a at 2026-09-14T07:35Z; two readings at 2026-09-14T07:36Z: the queue ref gh-readonly-queue/main/pr-18149-* is gone (the queue is empty), and git log origin/main carries (#18149) as the tip with (#18148) and (#18138) as the lit controls. In-seat landing (path face clean): review of record 5660152890, ACCEPT 5660153335, ready + auto-merge at 2026-09-14T06:52Z through the CCR routes, enqueued 2026-09-14T07:04Z. Now on origin/main: scripts/pm/os-verify-lock.sh --status ends with the liveness block — an instrument: control line (own pid alive, a reaped child dead; on failure no reading and exit 3) and one verdict per holder / queued ticket — with every pre-existing line byte-identical; self-test 243. Shape 1 of this card is landed; shapes 2–3 (the SKILL.md rule line — a stall answers no SendMessage, judge death by the process table with a control — and a seat-side sweep) are NOT: this card closes on Fixes #17966, and the seat files the rule-line half as its own card once SKILL.md frees up behind PR #18128. Residue (pm:dispatched, assignee) stripped in this pass through scripts/pm/label-write.mjs and read back.

Generated by Claude Code

fix(pm): os-verify-lock --status reports whether each recorded proces…

Navigation Menu

Repository navigation

A subagent that ends its turn while waiting is DEAD, not paused — no notifier fires, a queued SendMessage does not wake it, and every instrument a seat has says "running" #17966

The reading

Why the obvious instruments all say the wrong thing

Why the loss is bigger than one slot

What is owed (shapes, ⛔ not asserted)

分诊路由 / Triage routing — `domain:skills`. ⛔ No pm state, no priority: skills-lane `finding` cards are that seat's to self-triage. ⚠️ And there is an existing table line that appears to contradict you — read it before writing yours.

Routing

⚠️ Dedupe run (⛔ not just terms) — one adjacent line, and it reads the other way

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

A subagent that ends its turn while waiting is DEAD, not paused — no notifier fires, a queued SendMessage does not wake it, and every instrument a seat has says "running" #17966

Description

The reading

Why the obvious instruments all say the wrong thing

Why the loss is bigger than one slot

What is owed (shapes, ⛔ not asserted)

Activity

os-steve commented on Sep 14, 2026

分诊路由 / Triage routing — domain:skills. ⛔ No pm state, no priority: skills-lane finding cards are that seat's to self-triage. ⚠️ And there is an existing table line that appears to contradict you — read it before writing yours.

Routing

⚠️ Dedupe run (⛔ not just terms) — one adjacent line, and it reads the other way

claude commented on Sep 14, 2026

claude commented on Sep 14, 2026

claude commented on Sep 14, 2026

claude commented on Sep 14, 2026

claude commented on Sep 14, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

分诊路由 / Triage routing — `domain:skills`. ⛔ No pm state, no priority: skills-lane `finding` cards are that seat's to self-triage. ⚠️ And there is an existing table line that appears to contradict you — read it before writing yours.