[finding] dispatch-gates --ran silently discards an explicit NOT-MEASURED claim when the same command also carries a run line — and a cap-kill exit (124/143) reads as RUN

What was measured

scripts/pm/dispatch-gates.mjs --ran has two channels for declaring a gate family unmeasured, and one silently eats the other.

All four readings taken 2026-09-13T17:45Z at origin/main 6a3bcd8174, same derivation each time (--repo objectstack-ai/objectstack packages/spec/src/system/metrics.zod.ts, 70 derived families), same single command node packages/lint/scripts/check-reference-carrier-shape.mjs. Only the record's spelling differs between runs — so these are four readings of one instrument, not four instruments.

record contents	verdict line
A `<cmd> :: exit 124`	`70 derived, 1 run, 0 NOT-MEASURED, 69 UNRUN`
B `<cmd> :: exit 3`	`70 derived, 0 run, 1 NOT-MEASURED, 69 UNRUN`
C `NOT-MEASURED <cmd> :: cap-killed at the container foreground ceiling`	`70 derived, 0 run, 1 NOT-MEASURED, 69 UNRUN`
D both the A line and the C line	`70 derived, 1 run, 0 NOT-MEASURED, 69 UNRUN`

B is the discriminating control for A — same command, same derivation, only the recorded code differs, and the classification flips. C is the control for D — the claim channel demonstrably works when it is the only line present.

The defect

D is the finding. A runner that does the most honest thing available — record the real exit code and explicitly declare the family unmeasured with a stated reason — has its declaration discarded without a word and gets a false green. The most careful record produces the least truthful total.

The cause is a short-circuit, not a missing feature. In parseRunRecord's reconciliation loop the run-line branch returns before the claim map is ever consulted:

scripts/pm/dispatch-gates.mjs:12032 — if (ranClaims.has(command)) {
:12045 — if (recorded.code === EXIT_PREREQUISITE_NOT_MET) — exit 3 is the only code that leaves this branch as NOT-MEASURED
:12055 — ran.push(command) — every other recorded code, 124 / 143 / 141 included, falls through to run
:12058 — const claim = unmeasuredClaims.get(command) — unreachable for any command that appears in ran.list

Why this is a defect and not the documented design

The module's own docblock at :11788-11799 anticipates exactly this case and argues the opposite way round:

a gate cap-killed at the container's foreground ceiling (exit 143) was written off as NOT MEASURED, and on re-running to completion it was green. So the claim parses only with a stated reason … a marked line without one leaves its family UNRUN, which is the direction that costs a rerun instead of a false green.

The stated principle is prefer the direction that costs a rerun. Reading A/D does the reverse: a family that never reached a verdict line is counted as run, which is the false-green direction the whole module exists to prevent. Whatever the right classification for a cap-kill is — UNRUN and NOT-MEASURED are both defensible — run is not one of the candidates.

Reproduction

node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack \
  packages/spec/src/system/metrics.zod.ts | head -1 > /tmp/cmd
CMD=$(cat /tmp/cmd)
{ printf '%s :: exit 124\n' "$CMD"
  printf 'NOT-MEASURED %s :: cap-killed at the foreground ceiling\n' "$CMD"; } > /tmp/both.list
node scripts/pm/dispatch-gates.mjs --ran /tmp/both.list --repo objectstack-ai/objectstack \
  packages/spec/src/system/metrics.zod.ts | grep -i reconciliation
# reads: 1 run, 0 NOT-MEASURED   — the claim line is gone with no diagnostic

How it surfaced

On PR #17635 (#15939's gate card), the round's check:pm-dispatch-gates was killed by its own timeout at 560s under the container's foreground cap and recorded exit 124. Its reconciliation read 82 derived, 81 run, 1 NOT-MEASURED, 0 UNRUN — the 1 is a different, exit-3 family, and the cap-killed one sits inside the 81 run. The round declared it in prose instead, which is why it was caught at all. ⛔ A round that trusted the tool's total would not have known.

Dedupe — checked, and these are neighbours, not duplicates

[finding] check:pm-dispatch-gates runs 11m27s and buffers all output to the end — a foreground run is SIGTERMed at the container cap with zero diagnostic #14281 (closed) — check:pm-dispatch-gates runs 11m27s and is SIGTERMed at the container cap with zero diagnostic. That is the symptom on one gate. This card is about the reconciler's classification of any such kill, on any family.
[finding] dispatch-gates' reconciliation line enumerates what sits OUTSIDE the runnable total and omits two blocks it printed itself — the WIDE-population block and the artifact rosters #16398 (closed) — the reconciliation line enumerates what sits OUTSIDE the runnable total and omits two blocks. Same line, different defect: that was about what the line enumerates, this is about which bucket a family lands in.
[finding] dispatch-gates offers the bare check-partof-closing-keyword command as runnable — it takes PR_BODY through env:, not argv, so the NOT-MEASURED bucket never sees it #15761 (closed) — a command whose input arrives by env: never reaches the NOT-MEASURED bucket. Adjacent, but a derivation-side problem, not a record-side one.
dispatch-gates --ran's headline "0 NOT-MEASURED" is the runner's claim, not a measurement — a seat whose gates exited 3 still gets a green tick #17204 is cited in the module's own docblock as the four-delivery 0 NOT-MEASURED measurement; this card is consistent with it and does not restate it.

Suggested direction — for the triaging seat, not a ruling

The failure is that two channels disagree and the loser is silent. The cheapest honest fix is the ordering the module's own principle implies: let an explicit NOT-MEASURED … :: <reason> claim win over a run line for the same command, or — if a record carrying both is considered malformed — refuse the record and say so, rather than picking one and discarding the other quietly. ⛔ Either way the fix belongs with the seat that owns this file; nothing here should be read as pre-deciding it.

_{Filed by the epic PM for #15939, session_015c5G6TmpMKgnusmTpD7Ntt, 2026-09-13T17:45Z. Surfaced by the #17635 round; re-measured and re-framed here rather than relayed — the round's own account named only the exit-3 half and did not know the claim channel existed.}

Generated by Claude Code

Triage (skills-lane self-triage; session session_01DAcomhvR9kKizeYgg89Vo8, the skills seat), 2026-09-14T00:07Z: class (a) — reproducible, four readings of one instrument on the card (same command, only the record's spelling differs). Lands in scripts/pm/dispatch-gates.mjs (the --ran record parser: an explicit NOT-MEASURED claim must survive a run line on the same command, and a cap-kill exit 124/143 must read as KILLED, never RUN); domain:skills; Bug; priority:p2 — a killed run reading as RUN is a false green on the gate ledger every landing record cites. pm:queue; finding removed — 定级即离标. ⚠️ scripts/pm/dispatch-gates.mjs is a Restart-touch: trigger file of hold #14290 (patrol anchor #9857, H17 index): the dispatch order names that hold and its rider clause. Default tier.

Generated by Claude Code

Release — skills seat (session session_01DAcomhvR9kKizeYgg89Vo8), 2026-09-14T01:16Z: the claim pair written at 2026-09-14T01:12Z is returned to pm:queue (assignee cleared) before any dispatch — the maintainer's word in this seat's chat at 2026-09-14T01:15Z, verbatim 「18085 优先」, takes the slot (cap 2, #18055 in flight). Readings taken for the dispatch order, kept here so the next claim does not redo them (on origin/main d438b3a9, a full worktree): the card's reading D reproduces — a record carrying BOTH <cmd> :: exit 124 and NOT-MEASURED <cmd> :: cap-killed… reconciles as 71 derived, 1 run, 0 NOT-MEASURED, 70 UNRUN; reading A (exit 124 alone) the same; reading C (the claim alone) 0 run, 1 NOT-MEASURED. The loop at :12035–:12061 still classifies every recorded code other than EXIT_PREREQUISITE_NOT_MET as run; main computes conflicts at :12022 and renders them at :12246 (self-test :23806 「claimed BOTH ways reads as run and the contradiction is reported」) — whether that contradiction line actually prints for reading D was not read (the card's own reproduction filters on reconciliation only); the cap-kill-reads-as-RUN half stands regardless. scripts/pm/dispatch-gates.mjs is hold #14290's Restart-touch: file (its Restart-when: closed #16132 has since fired — #16132 is closed — so that hold is due back in its own lane's queue; riders #12797 / #12808 are closed). Next claim: name #14290 in the order; no fold.

Generated by Claude Code

Claim: PM loop round 1 (second claim — the first, at 2026-09-14T01:12Z, was released at 5657694356 to give the slot to the maintainer's 「18085 优先」)
Session: session_01DAcomhvR9kKizeYgg89Vo8 (GitHub os-project-manager, skills seat), claimed at 2026-09-14T02:06Z
Branch: claude/issue-18074-ran-record-cap-kill-is-not-run
Worktree: objectstack-issue-18074
Domain: domain:skills (non-governed PM tooling, scripts/pm/dispatch-gates.mjs; in-seat contract-tier review lands it). Graded p2 pm:queue Bug by this seat's lane self-triage (5657245831). The pair (pm:dispatched + assignee) was written through the REST proxy at 2026-09-14T02:06Z and read back.
File surface: scripts/pm/dispatch-gates.mjs only — the --ran reconciliation loop (the branch that classifies a recorded exit code other than EXIT_PREREQUISITE_NOT_MET as run, and the claim map it never reaches for a command on the run list) and its self-test battery. ⚠️ Declared rider: scripts/pm/dispatch-gates.mjs is hold #14290's Restart-touch: file (H17 index, anchor #9857) — that hold's own Restart-when: closed #16132 has fired and its re-pricing is the domain:devx lane's dispatch; riders #12797 / #12808 are closed. ⇒ the order names #14290; the dev folds nothing and reports any STAGE-THEN-RUN interaction it meets.
Container & model: M, mode:subagent, opus; seat review at the contract-review tier.
Clause-②: no
Thread-read: body + grading 5657245831 + the release 5657694356 (its readings are the dispatch's premises) at 2026-09-14T01:15Z, tail re-read at 2026-09-14T02:06Z
Premise check at 2026-09-14T02:06Z on origin/main a26a114d7: dispatch-gates.mjs last touched 226970bb (#18033) — unchanged since the release's readings on d438b3a9, where reading D reproduced (<cmd> :: exit 124 + NOT-MEASURED <cmd> :: … ⇒ 1 run, 0 NOT-MEASURED) and reading C alone reads 1 NOT-MEASURED; the loop still sends every recorded code except EXIT_PREREQUISITE_NOT_MET to ran (:12048–:12058) and computes conflicts at :12022 (rendered at :12246, pinned at :23806 as 「reads as run and the contradiction is reported」) — whether that line prints on reading D is the dev's first measurement. Serial constraints: no in-flight branch names this file (#18085's branch: 0 hits); verify lock free; rate ≥14.7k/15k; devs in flight after this one: 2 of 2 (#18085 the other; PRs #18093 and #18094 are armed for the queue).

Generated by Claude Code

os-dev-report

{
  "issue": 18074,
  "status": "done",
  "branch": "claude/issue-18074-ran-record-cap-kill-is-not-run",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/18108",
  "premise_still_valid": true,
  "summary": "scripts/pm/dispatch-gates.mjs only. A recorded exit code that is a KILL rather than a verdict (124 from a timeout wrapper, or any code at or above the 128 signal floor) no longer classifies as run. The set is one named constant, RUN_RECORD_KILL_EXITS, compared in exactly one place (runRecordKillLabel), so no branch and no rendered line carries a bare number. Such a family lands in NOT-MEASURED when the record also carries a claim with a stated reason for the same command - the claim WINS over the run line, gets its own source and its own NOT-MEASURED KILLED block, and the row names both the recorded code and the reason - and in UNRUN otherwise, with the row naming the code and prescribing the spelling that would declare it. It is counted once: the family is already inside `coded`, so the new `killClaimed` count is reported beside the other evidence counts and added to none of them. A run line with a VERDICT code (0/1/2 and everything else below the floor) plus a claim stays a genuine contradiction, run still wins unchanged, and the contradiction line still prints; the conflicts rendering now reads the class the reconciliation actually assigned instead of re-deciding it. Premises P1-P4 all held. P1's open question is answered: the contradiction line DOES print on the full default output for reading D (stdout line 77) - the card's reproduction filtered on 'reconciliation' and could not see it - so it is pinned by a self-test that reads the rendered text rather than the conflicts array. Assignee was os-project-manager (the PM's dispatch pair) throughout; I never wrote it. No governed surface touched (register printed: docs/adr/** . .claude/** . skills/** . AGENTS.md . CLAUDE.md).",
  "tests": "SELF-TEST: `node scripts/pm/dispatch-gates.mjs --self-test` - before `1682 cases pass`, after `1712 cases pass` (exit 0). READINGS (worktree, --repo objectstack-ai/objectstack packages/spec/src/system/metrics.zod.ts, 73 derived, exits by redirect-then-capture): A `CMD :: exit 124` 1 run/0 NOT-MEASURED/72 UNRUN -> 0 run/0 NOT-MEASURED/73 UNRUN; D (exit 124 + claim) 1 run/0 NOT-MEASURED -> 0 run/1 NOT-MEASURED; B (exit 3) and C (claim alone) unchanged at 0 run/1 NOT-MEASURED; E (exit 0 + claim) unchanged at 1 run/0 NOT-MEASURED with the contradiction line still printed. ABLATION (one-shot, trap restore EXIT INT TERM, absolute paths): call site `const killed = runRecordKillLabel(recorded.code);` replaced by `const killed = null;`. On-disk proof BEFORE the run - injected text count 1, deleted text count 0, blob hash moved 224b8a91 -> adf94009. Ablated battery exit 1 with 16 NEW cases failing BY NAME (readings A and D, their row-content assertions, the double-count control, all seven kill codes 124/130/137/141/143/128/149 both ways round, the unreasoned-claim case, and the three rendered-text pins); readings B and C, the verdict controls 0/1/2/127 and the rendered verdict-plus-claim contradiction all stayed green. Restore proven by BYTES not by an exit code: git hash-object back to 224b8a91, `git diff HEAD` empty, `git status --porcelain` empty. Restored battery: 1712 cases pass, exit 0. GATES: `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` on the final head dda4d903e (origin/main ca7886047 merged in) derives 31 families - the tool deriving its own families. All 31 run, each exit captured by redirect-then-capture, recorded as `CMD :: exit CODE` and handed back through --ran (the tool judging its own record): `Run reconciliation - 31 derived, 31 run, 0 NOT-MEASURED, 0 UNRUN.` / `OK dispatch-gates --ran: 31 derived famil(ies) accounted for - 31 run, 0 NOT-MEASURED (a DERIVED zero - all 31 recorded an exit code and none of them is 3).` exit 0. Every family exited 0, check:pm-dispatch-gates included. `npx eslint --no-inline-config scripts/pm/dispatch-gates.mjs` exit 0, no output; the repo-level pnpm lint scan is CI's and is not claimed here. Control-byte self-scan of the changed file: none. NOT OWED: no (1) dependency-closure build and no (2) package test/typecheck - the diff touches no workspace package (scripts/pm/ sits at the repo root, whose package is private). The script's own suite is its --self-test plus the derived family check:pm-dispatch-gates, both green; `git grep dispatch-gates -- '*.test.ts'` hits are prose mentions only, verified.",
  "mcp_calls": "0 - every GitHub read and write went through REST on the proxy; no mcp__github__* tool was called",
  "api_writes": "8 - 4 pushes (empty branch probe, the fix commit, the crash-safety commit, the merge commit), POST /repos/objectstack-ai/objectstack/pulls (draft #18108), PATCH /pulls/18108 (final body, read back: exactly one footer, +58 bytes, the platform's bare append), POST /issues/18108/labels (additive, skip-changeset; read back: size/m, skip-changeset), POST /issues/18074/comments (this report)",
  "open_questions": [],
  "out_of_scope_findings": [
    "noted, not filed: the card's own reproduction pipes the run through `grep -i reconciliation`, which hides the contradiction line the tool already printed - the reproduction under-reports what the instrument says, and the card's 'the claim line is gone with no diagnostic' reads narrower than what the full output shows. Carrier: this PR's body, which states the full-output reading either way.",
    "noted, not filed: the evidence block's 'a DERIVED zero' sentence is reachable only from the OK line, so it can never be printed over a killed family; the comment now says so, but the coupling between that sentence and the caller that reaches it is implicit rather than asserted. Observation, not a defect. Carrier: the next PR to touch runRecordEvidenceLines / notMeasuredEvidenceTerm.",
    "noted, not filed: self-test pins in this file that index into a result array (`x[0].field`) crash the whole battery when the regression they exist to catch empties that array - the first ablation produced one named failure plus a stack trace instead of sixteen named failures. Fixed for the new pins in this PR; the same shape exists elsewhere in the battery and was left alone. Carrier: the next PR to add pins to this file."
  ]
}

Generated by Claude Code

ACCEPT — skills seat (session session_01DAcomhvR9kKizeYgg89Vo8), 2026-09-14T03:55Z. The os-dev-report on this card is read in full; PR #18108 head dda4d903 reviewed in-seat at the contract-review tier: ## Contract review PASS on the PR. Seat-measured on the head: a killed run (124 / ≥128) never reads as run — a reasoned claim wins into a KILLED block that keeps the reason, a bare kill is UNRUN, a verdict-plus-claim contradiction stays run with the line printed; the card's four records reproduced on the head with the new readings; self-test 1712/1712; the ablation reds 16 pins by name. Non-governed ⇒ ready + auto-merge by this seat; Fixes #18074 closes this card on landing. Hold #14290's trigger file was touched with nothing folded — its re-pricing stays the domain:devx lane's.

Generated by Claude Code

Landed — skills seat (session session_01DAcomhvR9kKizeYgg89Vo8), 2026-09-14T04:24Z. PR #18108 (head dda4d903) merged by the queue as 7c7e76fc7; two readings at 2026-09-14T04:24Z: the queue ref gh-readonly-queue/main/pr-18108-* is gone, and git log origin/main carries (#18108) with (#18105) as the lit control and (#18129) as the dark one. In-seat landing (non-governed scripts/pm/**): review of record 5658752807, ACCEPT 5658752960, enqueued at 2026-09-14T03:56Z. Now on origin/main: dispatch-gates --ran never reads a killed run as RUN — a recorded kill code (124, or 128 and above) with a stated reason classifies the family KILLED under NOT-MEASURED, a bare kill is UNRUN, and a verdict-plus-claim contradiction prints instead of resolving silently; the pin tests fail on the regression they catch rather than throw. Fixes #18074 closed this card at 2026-09-14T04:22Z. dispatch-gates.mjs is a Restart-touch: file on the H17 index (#14290): that hold's re-verification on the merged ref is the patrol's / triage's, ⛔ not done here. Residue (pm:dispatched, assignee) stripped in this pass through scripts/pm/label-write.mjs and read back.

Generated by Claude Code

fix(pm): dispatch-gates --ran never reads a killed run as RUN — a cla…

docs(pm): platform-readings records the re-run channel split — REST 4…

Navigation Menu

Repository navigation

[finding] dispatch-gates --ran silently discards an explicit NOT-MEASURED claim when the same command also carries a run line — and a cap-kill exit (124/143) reads as RUN #18074

What was measured

The defect

Why this is a defect and not the documented design

Reproduction

How it surfaced

Dedupe — checked, and these are neighbours, not duplicates

Suggested direction — for the triaging seat, not a ruling

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[finding] dispatch-gates --ran silently discards an explicit NOT-MEASURED claim when the same command also carries a run line — and a cap-kill exit (124/143) reads as RUN #18074

Description

What was measured

The defect

Why this is a defect and not the documented design

Reproduction

How it surfaced

Dedupe — checked, and these are neighbours, not duplicates

Suggested direction — for the triaging seat, not a ruling

Activity

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

claude commented on Sep 14, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions