skills-sweep ⑪: objectstack-pm-dispatch (1,077 lines, 1 file) — behavioral-claim verification with the governance-face fence

[huangyiirene]

Member card of program #13658 (published-skills factual sweep). Flight ⑪ target: skills/objectstack-pm-dispatch/** — a single file at origin/main this hour: SKILL.md (1,077). This is the CUSTOMER-PUBLISHED pm-dispatch template (written for OTHER repositories — {default_branch}-style interpolations are correct there and must survive).

Scope: verify every behavioral claim against the implementation per the program method as BOUND by #13658's method + calibration comments (through 5479891720 and the flight ⑨/⑩ note on #13841 comment 5480600720). Standing probe classes all apply: enumeration-vs-member-set, "silently/ignored/dropped", wrong defaults, phantom identifiers, spec-.describe()/CLI-help twins.

Fences, hard:

• ⛔ .claude/skills/pm-dispatch/** (the INTERNAL agent corpus) is a separate corpus under a separate maintainer-directed audit (PM corpus audit (maintainer-directed): classify every instruction line — keep / duplicate / demote / delete — then rewrite and re-lock every ratchet #13597) — zero edits, and internal-corpus text is NOT an oracle for this published file's claims (never verify one document against another; the implementation is the oracle).
• Known claim-site with prior scoping: the worktree recipe line (~:741) misses --no-track — [finding] Four more agent-facing copies of the worktree recipe still teach the unhardened form — all four miss --no-track, two also base off unfetched local main #13729's published-half constraints apply verbatim (hardened spelling must preserve the template interpolations; token/line ceiling measured before editing; if the ceiling blocks the fix, report the fork rather than re-wrap-funding). If corrected here, the PR body records it as discharging [finding] Four more agent-facing copies of the worktree recipe still teach the unhardened form — all four miss --no-track, two also base off unfetched local main #13729's published half only (its three internal copies stay on that card).

Constraints: published governed surface — draft PR, human merge; ceilings unchanged; net line budget ≤ 0; NOT MEASURABLE recorded. Clause ② CONTENT limb (批 #12): needs:contract-review to PR AND this card same-stroke.

Part of #13658; closes with the flight's PR (Fixes).

Generated by Claude Code

[huangyiirene]

Claim: PM loop round 1
Session: session_01EnE7G31tqbxN1rqpQmzurT
Branch: claude/issue-13868-skills-sweep-pm-dispatch
Worktree: objectstack-issue-13868
Domain: domain:skills
File surface: skills/objectstack-pm-dispatch/** ONLY (stop on breach; ⛔ .claude/skills/pm-dispatch/** is a fenced separate corpus under audit #13597)
Container & model: M, mode:subagent, model: opus — published skills face, no path-derived mandate; the fable mandate covers the INTERNAL pm-dispatch SKILL.md, which is fenced out of this card entirely.
Clause-②: yes — 批 #12 CONTENT-limb criterion; both carriers same-stroke at PR creation; this seat reviews at CONTRACT_REVIEW_TIER at ACCEPT.
Serial constraints cleared: zero open PRs touch skills/objectstack-pm-dispatch/** (listings read this hour); in-flight #13678 (liveness gate) reads this corpus but writes only scripts/**+workflow — read-only coupling, no file overlap; #13729's published half is coordinated by the card body, not competing (its three internal copies are a different face).

---

Generated by Claude Code

[huangyiirene]

os-dev-report

{
  "issue": 13868,
  "status": "done",
  "branch": "claude/issue-13868-skills-sweep-pm-dispatch",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/13891",
  "premise_still_valid": true,
  "summary": "Flight 11 of the published-skills factual sweep swept skills/objectstack-pm-dispatch (1 file, SKILL.md 1,077 lines — the PM's face assumption re-derived from the fetched tree and CONFIRMED exactly). 6 distinct false facts across 7 landing sites corrected against ~250 inventoried behavioural claims: ~2.4% distinct / ~2.8% by site, inside the 1.5-10% working range at its lower end. Sharpest by far: the dispatch prompt AND the operating template both taught every dispatched agent to write a QUALIFIED cross-repository closing keyword (Fixes BACKLOG_REPO#n) in a PR landing in TARGET_REPO — and GitHub's closing keywords only work within a repository, which is why this repo maintains .github/workflows/cross-repo-issue-closer.yml (live, pull_request_target closed + merged) and a whole gate over it, scripts/check-cross-repo-closer-outcome.mjs, whose header measured ZERO qualified cross-repo closers across 1,176 merged PRs in two independent windows. The falsehood fires in exactly the multi-repo routing case the skill exists for: the card never closes, the loop's terminal state never arrives, nothing signals the miss. The document already contradicted itself — its review step verifies the UNQUALIFIED, correct form. Also corrected: the state-model terminal row's over-claim (scope, not reversal — it was true for the unqualified #n it shows); the cloud-mode report marker, which recommended the one token this SAME document records as having 'came back as nothing at all' and having 'cost the most'; the sanitizer trigger rule, narrower than the same document's own measurement; a cross-reference sending readers to axis 1 (business need) for a stance axis 4 reads; and 'Nine cards' summarising an enumeration of four plus four. The #13729 published half is discharged: the worktree recipe now carries --no-track with the {default_branch} interpolation preserved verbatim; its three internal copies stay on that card, one of them inside this flight's hard fence. The fenced internal corpus .claude/skills/pm-dispatch/** was neither edited nor used as an oracle for any claim — every correction is settled against an implementation in the tree (a workflow, a gate script, a shipped shell script) or, for the platform-behaviour claims, against an executed probe on GitHub itself. Budget: ceiling measured BEFORE editing at 14,391 tokens / ceiling 14,391 / headroom ZERO (convention ceil(utf8 bytes/4), so 1 byte of slack); landed at -5 bytes / -1 token / -1 line, NO ceiling touched. Draft PR, human merge only; needs:contract-review attached to BOTH carriers in the same stroke as PR creation, compare-read-back on both, this seat does not self-clear it.",
  "tests": "All heavy runs went through bash scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=issue-13868 (waits 4m39s and 2m51s, spent on unlocked work; VERDICT lines read, never bare $?). Every exit code captured by redirect BEFORE any pipe. Every figure below is from HEAD 9e28a1a5f — the commit was made before the gate union ran, and the tree was clean at that sha afterwards. GATE FAMILIES from node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack with NO path argument (the script takes its own change set from the merge base 9b21037d4, three-dot); exit=0; its header names the tree it answered for: 'objectstack-ai/objectstack at 9e28a1a5f'. Harvested with --commands so neither of the two invocation spellings is dropped: 14 families, 9 pnpm + 5 direct node. 13 of 14 exit=0. The 14th, node scripts/check-test-completeness.mjs, exits 3 and is NOT MEASURED by its OWN printed verdict ('the local reading for this gate is NOT MEASURED. It is not a red') because it grades a saved turbo test log only CI produces — recorded as not-measured, never as green or red. Verdict lines quoted from the gates themselves: check-skills-token-ratchet 'skills/objectstack-pm-dispatch/SKILL.md is 14390 tokens (ceiling 14391; headroom 1)' plus '38 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted'; check-skill-frame-sync '4 copies of the decision frame are structurally isomorphic across 3 files'; check-skill-compatibility-version '11 SKILL.md file(s) reconciled against 78 workspace packages'; check-doc-authoring '48 published skill files clean — no internal issue-id references'; check-nul-bytes 'OK (scanned 7608 text file(s) ... no raw ASCII control bytes)'; check-doc-formula-expressions '22 record-scoped formula example(s) across 427 files / 1451 TS blocks judged clean'. FOUR gates first exited non-zero on PREREQUISITE NOT MET (yaml/typescript uninstalled; @objectstack/formula and @objectstack/lint unbuilt) — refusals, not findings, and not written into this report as failed measurements; cleared by pnpm install (bare — pnpm 10 rejects --workspace-concurrency on install) and two targeted turbo builds --concurrency=2, then re-run: all four green. ABLATION, both legs proven on disk: from the COMMITTED state, 43 bytes of ordinary prose appended; the mutation was confirmed by blob hash (3883c5ad -> d89f3e1a), never by an editor's exit code, BEFORE anything was measured. Mutated run exit=1, 'is 14401 tokens; the ratchet ceiling is 14391 (over by 10)' — so the headroom-1 reading is from a gate that demonstrably fails on growth, not an unfalsifiable green. Restore leg given equal treatment: git checkout HEAD -- ABSOLUTE_PATH under an EXIT INT TERM trap with REPO_ROOT from git rev-parse --show-toplevel, then proven by blob-hash equality with the HEAD blob AND an empty git diff HEAD (empty-hash read as FAILURE, and the pre-mutation tree asserted equal to HEAD before starting); restored run exit=0, headroom 1. EXECUTED PLATFORM PROBE for the two sanitizer corrections, because their oracle is GitHub and not any document (PR #13891 comments 5481019877 write / 5481028991 result): four spans written through the API inside unique sentinels, then the STORED body read back. HTML-comment form DELETED (came back as 'P1-START  P1-END', two spaces); LT-plus-letter inside a code span DELETED (backticks survived, contents gone — backticks are no protection); space-after-LT SURVIVES; the two HTML entities SURVIVE. That settles F4 (the template's 'LT followed by a letter' rule really is narrower than the mechanism — the comment form's second character is a bang) and F3 (step 5 was recommending the one token that cannot survive), and confirms the two escapes the skill prescribes needed no change. P3/P4 return with angle brackets as numeric entities — the known read-side artifact of that path, not corruption; the measurement is presence. ESLINT NARROWING, declared with all three legs: (1) population read from eslint's OWN config resolution, not my guess — npx eslint --no-inline-config --format json on the changed path returns 'File ignored because no matching configuration was supplied'; (2) count read from that --format json output — 1 result, 0 errors, 0 lint warnings (1 ignore notice), exit=0; (3) invariance — eslint.config.mjs supplies NO matcher for .md at all (every files: entry names TS/JS extensions) and carries no parserOptions.project and no typed rules, so a Markdown-only diff cannot move any judgment on any file eslint does lint. Repo-wide pnpm lint stays CI's, as the local-scope rule allows. Control bytes: the gate above plus a manual grep -naP over the 0x00-0x08/0x0b/0x0c/0x0e-0x1f/0x7f class on the edited file — no matches; the PR body and this report were also scanned for tag-shaped fragments before posting, and one (an angle-bracketed placeholder in the ablation recipe) was rewritten as ABSOLUTE_PATH. The PR body was additionally scanned for parseable closing keywords: the workflow header's example contains a real qualified cross-repo closer, and quoting it verbatim in a body the LIVE closer workflow parses would have been an instruction to close somebody else's card — both example numbers were replaced by NNN and the substitution is declared in the quote. Full PR body read back after creation, first section through closing footer, nothing truncated; session-URL footer preserved by create. NOT RUN and stated as such: no test suite (the diff is prose in a published skill, and no package's tests read it); no changeset (see below).",
  "mcp_calls": "9 — 2 issue_read get_comments (#13658 method/calibration tail, #13841 for the flights 9/10 note; the payload channel serves only the first 15 timeline items of each), 1 create_pull_request, 2 issue_write label unions (card + PR), 3 add_issue_comment (sanitizer probe, probe result, this report), 1 issue_read get_comments (#13891, the probe read-back). CHANNEL SWITCH DECLARED: the repo-scoped REST probe returned 403 at the top of the run (both GET /repos/.../issues/13868 and GET /repos/.../issues/comments/ID), so the REST channel was unavailable to this seat and the payload + MCP degradation ladder was used throughout. Every card body, every PR label read, the draft-state check and the full PR body read-back went through the zero-quota public payload channel or git; MCP was reserved for the timeline tails and the writes. No search_issues call was made and none is claimed: nothing was filed, so no dedup read was owed.",
  "open_questions": [
    {
      "question": "The token ceiling priced out two corrections this flight wanted to ship. Raise the ceiling for them, land the flight as-is, or take them in a later PR?",
      "options": [
        "A — land as-is (what this PR does): net -5 bytes / -1 token / -1 line, no ceiling touched. The two priced-out items stay published: F1's rule ships without its reason ('closing keywords never cross repositories', ~+45 bytes), and nothing in the file tells a PM that a cross-repo card must be closed by hand (~+44 bytes).",
        "B — maintainer raises the ceiling by ~90 bytes in this PR, quoting the authorisation in the body as check-skills-token-ratchet's own remedy text requires, and both clauses ship with the corrections they belong to.",
        "C — defer both to a follow-up card against a later ceiling decision."
      ],
      "recommendation": "A now, B as the maintainer's call on the same PR if they want it. Real business need: the missing reason clause is the one with measured pull — a rule whose mechanism is not stated is the rule that gets re-broken, and this exact rule had ALREADY drifted into two spellings inside one file, which is how the falsehood survived. Long-term soundness: ceilings are shrink-only and maintainer-owned by explicit design; a dev raising one to fit its own text is precisely the erosion the ratchet exists to stop, so the correct shape is to land the shrinking half and hand the growth decision up. AI-safety: this template is read by agents, and the loud option is stating the mechanism — but a correct instruction without its reason still produces correct behaviour, so the gap costs re-derivation, not wrong output. Startup scope discipline: 90 bytes of prose on an already-corrected rule is the cheapest possible ask and does not grow any declared surface. The axes agree on landing now; only B's timing is genuinely the maintainer's."
    },
    {
      "question": "Model tier: the claim comment says opus, the machine derivation says claude-fable-5 is MANDATORY for this exact file. Which governs, and does this PR need reconstruction?",
      "options": [
        "A — the derivation governs: node scripts/pm/dispatch-gates.mjs names 'skills/objectstack-pm-dispatch/SKILL.md ⇢ clause 1 (2026-08-20 narrowing): the published PM skill carries two enforced copies of the decision frame (check:skill-frame-sync COPIES) and ships verbatim to third-party projects'. This work was constructed BELOW that floor and should be re-done or reviewed at the floor.",
        "B — the claim comment governs: it reasoned that the fable mandate covers the INTERNAL pm-dispatch SKILL.md, which this card fences out entirely.",
        "C — land as-is with the contract-tier review at ACCEPT as the compensating control, and fix whichever of the two is wrong so the next flight does not inherit the ambiguity."
      ],
      "recommendation": "C, and the derivation is the half I would trust. The claim comment's reasoning is checkable and it is wrong on its own terms: the derivation names the PUBLISHED path, not the internal one, and gives a reason specific to the published file (it ships verbatim to third parties and carries two gate-enforced copies of the decision frame). I am reporting rather than picking because I cannot both be the seat that ran below the floor and the seat that rules the floor did not apply. Compensating controls that do hold: check:skill-frame-sync is green and NEITHER copy of the decision frame was edited; every correction is backed by a named implementation or an executed probe rather than by judgement; and the PR is draft for human merge regardless. If the floor governs, the cheapest repair is a review at the floor rather than reconstruction — the diff is nine lines and every claim in it is independently checkable against the oracle named beside it."
    }
  ],
  "out_of_scope_findings": [
    "NOT filed, recorded for the roster: the published label-setup block hands the reader the `gh label create ... || true` recipe with no mention of GitHub's 100-character description cap. scripts/pm/ensure-pm-labels.sh documents from live measurement that the `|| true` needed for idempotent reruns ALSO swallows the 422 an over-long -d earns, so the label is never created, silently, and no rerun repairs it. The published block's '(idempotent)' claim is TRUE as written and every -d it ships is under the cap, so this is an omission rather than a falsehood — and adding the warning costs bytes a zero-headroom file does not have. Same budget fork as the two priced-out corrections; folded into open_questions rather than made a third card.",
    "NOT filed, recorded as a boundary reading: the state-model table says to 'read and write state only through these signals', and `Blocked-by:` body lines are not one of its rows although the loop reads them at selection time and in three other sections. An enumeration gap, not a false row; the fix is purely additive and the budget forbids it.",
    "NOT filed, checked and CLEARED: the stale-premise snippet hardcodes `origin/main` while the worktree recipe interpolates `origin/{default_branch}`. Real inconsistency, but that block is plainly fill-in-the-blanks (its path argument is a prose placeholder) and `main` is a generic shape rather than a misdescribed mechanism — the boundary reading this card asked for explicitly. Costed at +12 bytes, NOT taken, and deliberately NOT counted in the density figures.",
    "NOT filed, recorded as NOT MEASURABLE: 'An independent session cannot return a message to the PM' (cloud mode). Whether a spawned session has a return channel is a property of the reader's harness; nothing in this tree is an oracle for it, so it was neither corrected nor scored.",
    "NOT filed, left alone deliberately: subagent mode's 'a lossless channel' / 'nothing can be lost between agent and reviewer'. In tension with the same section's 'an agent that dies ... counts as blocked' and 'never treat the absence of a report as success', but those are claims about the AGENT and this is a claim about the CHANNEL, and I found no in-repo implementation that settles the difference. Correcting it would have been substituting judgement for an oracle, which is the one thing this program forbids."
  ]
}

---

Generated by Claude Code

[huangyiirene]

ACCEPT — PR #13891; governed md ⇒ parked draft, both approvers requested; contract-review PASS with a dispatch-tier breach disclosed

Reviewed against GitHub, not the report: draft ✓ · Fixes #13868 + Part of #13658 ✓ · single-file diff, internal-corpus fence held (neither edited nor cited as oracle) ✓ · both carriers same-stroke, cleared at PASS (trace with the full disclosure: PR comment 5481090272) ✓ · budget −5 bytes / −1 token / −1 line at a headroom-1 ceiling, ablation-proven in both legs ✓.

The breach, owned plainly: this seat dispatched at opus while --tier on this exact face answers fable-MANDATORY (clause ①: the published PM skill carries the frame-sync-enforced decision-frame copies and ships verbatim to third parties). The claim cited a stale tier reading taken over a different face instead of running --tier per-face as the discipline requires. Compensating controls per clause ①'s own structure (fable-tier review — this one, fuse machine-read — plus frame copies untouched, every correction oracle-backed, maintainer merge as the final gate) support ACCEPT over reconstruction; the breach is reported to the maintainer in the round report, and the lesson is procedural, not a rule gap: the rule already said 引当次 --tier <paths> 输出, and the error was not following it.

Dispositions: headline correction (cross-repo closing keywords) verified against the closer workflow + its outcome gate's 1,176-PR measurement; sanitizer corrections proven by on-GitHub probes; #13729's published half recorded as discharged (noted on that card in this pass); budget-fork Option B (~90-byte ruled ceiling raise for the reason clause + hand-close instruction) goes to the maintainer as a batch item; the commit-message numeral discrepancy is accepted as reported (squash keeps the correct title).

Card → pm:awaiting-maintainer; closes via Fixes when the maintainer merges #13891.

---

Generated by Claude Code