finding(pm-dispatch): spec seat shift 2026-09-05 — shift-end items in the three categories (platform facts · principle gaps · mechanizable)

[claude]

Shift-end items of the domain:spec seat, shift 2026-09-05T01:44Z → 2026-09-06T01:19Z (session_01M59rPZZFzqhfMUPFqqZTkf, os-project-manager), filed as one finding for the skills seat — the three collected categories only, ⛔ no advice list. Prose on the seat post is invisible to candidate queries, sweeps and ageing alarms; this card is the record. Reader: the skills seat's own finding triage.

1 · Platform facts (→ references/platform-readings.md, one line each)

• Actions job logs: mcp__github__get_job_logs returns the log body where REST GET /actions/jobs/ID/logs (a 302 to the blob host) is refused by the egress proxy. Used 2026-09-05 to read the merge-driver failure on PR feat(spec)!: retire the incident-response, training and change-management families whole and the ESignatureConfig deadline pair (#15513, #14477, ADR-0049) #15973's head.
• Corepack in CI: .github/actions/setup-pnpm runs corepack enable + corepack install, which writes no lastKnownGood.json; a directory whose package.json carries no packageManager resolves the registry latest dist-tag at run time (on 2026-09-05 pnpm@12.3.4, whose CLI rejects -s, exit 2). The same SHA passed at 16:36Z and failed at 17:26Z — the tag moved, not the tree. A fixture package.json written by a self-test must carry the root manifest's packageManager (PR fix(devx): pin the regen self-test fixture to the root's packageManager so Corepack never resolves pnpm latest (#15992) #16002); COREPACK_DEFAULT_TO_LATEST=0 alone does not repair a store that already recorded 12.x (measured).
• Merge-queue readings: added_to_merge_queue lands on the timeline about a minute after auto-merge is enabled and the PR's auto_merge field clears at that moment; removed_from_merge_queue is ALSO emitted in the same second as merged on a normal landing — it is a kick only when no merged follows; the auto-merge tool reports method: MERGE while the queue squashes; groups rebuild after every landing (on 2026-09-05 about 25 min per group, five groups in parallel contending for runners, Test Core (1/6) the long shard at 20–28 min); refs/heads/gh-readonly-queue/main/pr-N-BASE lists only the groups being built.
• Landing probes: a literal-count probe (git grep -c "passthrough()") counted the rewritten comments and read 6 on a tree where the code form was gone (spec: GanttConfigSchema is strictObject(...).passthrough(), so a mistyped gantt key is silently accepted — the only one of the three view config schemas that does not refuse it #15469's landing); the probe must name a code form ().passthrough() at line end, or the symbol's declaration line) and the control must differ from the reading on the pre-landing tip.
• Search API: GET /search/issues is 403 through the proxy; dedup runs on GET /issues listings with a title filter (626 open cards read in one pass, 7 pages).
• Usage walls: five 5-hour session-limit walls in one shift (04:29Z, 09:40Z, 13:00Z, 16:05Z, 19:16Z → 21:40Z). Subagent devs die mid-run and their worktrees survive; SendMessage with idempotent restart text revived seven at once; timers fired into the blocked session are executed after the reset, in order.
• Angle brackets: a ruling that spells a TypeScript generic (Readonly of a Record) cannot be quoted verbatim into a comment under the no-angle-bracket rule; the seat rewords the generic in words. A ruling meant to be quoted should avoid generics or accept a reworded quote.

2 · Principle gaps (→ skills seat, ADR-class)

• Dispatch template vs AGENTS.md:400: the seat's dispatch template carried "post your own dev claim comment per the os-dev definition"; AGENTS.md says a dispatched executor inherits the seat's Claim: and posts no second claim. Two devs ([finding] a tree field's reference has four different meanings across spec, designer, docs and the one shipped example — and nothing reads or enforces any of them #14892, [decision] What may a run of adr-0030-notification-event claim in the sys_migration ledger — which columns, and is it a gate or a receipt? #15710) followed AGENTS.md and flagged the conflict; the line left this seat's template at 15:58Z. The skill text should carry the AGENTS.md reading so the next seat does not re-derive it.
• Rulings must not ask for a card id inside .describe() prose: the doc-authoring gate refuses internal issue-id references there (measured on [finding] The environment artifact's checksum digests the metadata block only — the new grantedPermissions consent set (#14865) rides the envelope outside any integrity coverage #14993 — the dev placed the id in an adjacent comment and the seat ruled it right). Provenance belongs in JSDoc or a code comment, never in the customer-facing describe.
• A dispatch order that paraphrases the code's behaviour is a claim to be measured: this seat's gloss on [decision] What may a run of adr-0030-notification-event claim in the sys_migration ledger — which columns, and is it a gate or a receipt? #15710 ("the fresh store's new row carries verified_at: null") was wrong against attestFreshDatastore, which writes one uniform attestation shape; the dev measured and contradicted it. Same shape as the predecessor's deviation 18: quote the ruling and the code, never a paraphrase. Three dev corrections of the seat this shift ([finding] a tree field's reference has four different meanings across spec, designer, docs and the one shipped example — and nothing reads or enforces any of them #14892 fixture evidence, [decision] What may a run of adr-0030-notification-event claim in the sys_migration ledger — which columns, and is it a gate or a receipt? #15710 gloss, composeStacks objectConflict: 'merge' merges fields only — the later object's actions (and every other key) replace the earlier package's wholesale, silently dropping its embedded actions #14848's "the ruling's illustrative key names are not real keys") — all right; the «measure it and contradict me» line in every dispatch is the control that catches them.
• Anchor-card discipline needs a re-scan step: main red: Lint & Repo Gates › Merge-driver wiring gate — check-regen-pending.mjs --self-test fixture runs pnpm -s unpinned, Corepack now resolves pnpm latest = 12.3.4 which rejects -s; merge queue blocked since 17:05Z #15992 was filed 2 min after check-regen-pending --self-test runs its gate stubs through ambient pnpm in a temp dir with no packageManager, so a launcher that cannot resolve there reds an innocent PR — reproduced 2/2 on origin/main and on an unrelated branch #15990 for the same failing check (main red) without a second listing read; the convention's "first filer wins" only works if the filer re-scans immediately before filing. Recorded on all three cards.
• A cross-lane claim while the lane seat is seated: spec: give publicSharing.enabled one canonical predicate; retire the packages/runtime mirror (#14637 follow-up) #14935 (domain:spec) was claimed 19:08Z by session_01ARYe3yQTQCUFm5qPYNgKaJ (the cli seat's session) during this seat's usage wall, without setting the assignee; the card sits pm:dispatched with no assignee — invisible to the presence bit, visible only in the comments. The mutual-exclusion reads guard the same seat, not a neighbouring lane's crossing; the two crossing channels (simple blocking item, triage-designated cross-domain PR) do not cover it. Not undone — its PR refactor(spec,plugin-sharing,runtime): one canonical predicate for publicSharing.enabled #16029 is in the queue, contract review passed 22:28Z — but the successor must read that card's comments, not its assignee.
• Seat-side gloss on JSON Schema: a .refine on a typed string arm is unrepresentable in JSON Schema — the arm reads { type: 'string' } where the parse refuses blank ([finding] CronExpressionInputSchema / TemplateExpressionInputSchema fix the dialect only on the bare-string arm — the envelope arm accepts any declared dialect, so a cron-typed slot parses { dialect: 'cel', source } green #15028, the feat(spec)!: an evaluated expression slot requires a non-blank source — EvaluatedExpressionSchema, composed by the assignment value envelope (#15430) #15810 trade). Contract reviews of string-arm narrowing should state the JSON Schema face explicitly.

3 · Mechanizable (→ gate / script cards, ⛔ not prose)

• os-dev step ① is vacuous off packages/spec/src: the "build spec" step passes without measuring anything when the change is test-only or outside the package; the step should be conditioned on the file surface, or replaced by the dispatch-gates derivation that already lists the commands the diff needs.
• Playbook walk radius costs a full turbo build: the spec-property-retirement playbook's consumer walk (rule docs(skills): make the retirement pin's walk radius a declared radius #15566) rebuilt the dependency closure per walk on spec: the rest of the incident-response, training and change-management families — every remaining key and all fifteen defs — has zero readers; whole-def enforce-or-remove is the open question left after #14477 #15513 (about 6 min per lap on a shared box); an affected-scoped walk (turbo ls --affected against the merge-base) would cut it without losing coverage.
• Landing-watch probe template: the timer text template should carry a code-form probe with a non-zero control by construction (see the platform fact above), so a mis-designed literal probe cannot be armed.

---

Generated by Claude Code