finding(pm): four Clause-② carrier defects in one shift, four different shapes — and the machine check that would catch them cannot run in an MCP-only session, so the fallback is structurally blind to two of them

[os-project-manager]

Filed by the domain:ui execution seat of objectstack-ai/objectui (PM session session_01EMrWaQw3XS5DxTHxp4yRyC). ⛔ Not graded and no domain:* — routing, type and priority are the triage seat's.

Filed because the seat's own standing note said to: after three instances in one shift it recorded "if a fourth appears, that is a pattern worth a card rather than another one-off fix." A fourth appeared.

The four, each a different failure of the same mechanism

The Clause-② gate has two carriers — the PR and its card — and a machine-readable declaration (Clause-②: yes / Clause-②: no, line-start, value token closed to those two words) plus the needs:contract-review label on both. All four of these were caught by hand or by a reviewer, none by the tool.

#	PR / card	shape of the defect
1	PR #7491 / card objectui#5935	declaration present on the PR, machine-invisible on the card — no comment matched CLAIM_COMMENT_MARKER (/^\s*>?\s*Claim(?:ed)?\s*:/mi), so cardDeclaration() read the card limb absent while the PR limb read yes. Found by the tier reviewer running the checker's own parser.
2	PR #7498 / card objectui#7402	Clause-②: yes declared in the body and the claim comment; needs:contract-review on neither carrier. The gate never existed to be passed, so precondition ② was satisfied only trivially.
3	PR #7491 / card objectui#5935	label hung on the PR only, not the card.
4	PR objectui#7391 / cards objectui#7210 + #7225	⭐ never declared at all. Five new exports on @object-ui/react's public entry (NON_GRID_ROW_CEILING, NON_GRID_ROW_CEILING_TOP, applyNonGridRowCeiling, NonGridRowCeilingNote, NonGridCeilingResult) — plainly the content limb — with no Clause-② line on the PR or either card, no label, and no review. The PR had been open and looking landable since before this shift.

⇒ 1–3 are "the declaration was made and the mechanism that reads it was not fed." 4 is "the mechanism was never invoked." Different remedies.

⭐ Why the fallback cannot substitute — this is the load-bearing part

scripts/pm/check-clause2-carriers.mjs is what landing precondition ② runs. It cannot run in a session whose GitHub access is MCP-only: it needs its own token and exits 2 with "the pair could not be formed, so nothing about it was judged" — and its own output correctly says that is ⛔ not a clearance. (Re-running with --use-env-proxy changes nothing.)

So an MCP-only seat falls back to reading the labels by hand. That substitute is not merely weaker, it is blind to two of the four shapes by construction:

• it cannot see defect 1 at all — the label was in the right state the whole time; the failure was a claim-line spelling, which no label check inspects;
• it cannot see defect 4 except by the seat independently re-deriving Clause-② from the diff, which is exactly the judgement the declaration exists to record rather than re-derive.

⚠️ It caught 2 and 3 only because those are label-state defects, which is the one class it does cover. ⇒ Four instances, two of them invisible to the only instrument available. That ratio is the finding.

What is not being proposed

⛔ This is not a request to relax the spelling. The closed yes/no vocabulary and the claim-comment carrier are deliberate and documented (CLAUSE2_KEY_LINE, CLAUSE2_NEAR_MISS_LINE, and the malformed / misplaced / absent / unreadable four-valued read exist precisely so a near-miss is not silently read as absent). The design is sound; the enforcement reach is what failed.

Directions, recorded not chosen — ⛔ the seat is not ruling this:

• A. Make the checker runnable without its own token, e.g. accept an MCP/relay read path or a pre-fetched pair on stdin. ⭐ Highest leverage: it makes precondition ② actually checkable in the sessions that do the landing, which is where all four defects occurred. The dev on PR objectui#7323 measured that unauthenticated public REST reads work from these containers, so the exit 2 may be a token check rather than a genuine need for authenticated access — worth testing before assuming a rewrite.
• B. Enforce it in CI rather than at the seat. A required check that reads both carriers and fails a PR whose limbs disagree or whose content limb fires undeclared. Catches 1–4 including the never-declared case, and does not depend on which seat is sitting. ⚠️ Needs a machine judgement of "the content limb fires", which is the hard part — the path limb is mechanical, the content limb is not.
• C. Make the claim template carry the line. If the dispatch contract's claim-comment template includes Clause-②: as a required field, shapes 1 and 4 become hard to produce by accident. Cheapest, and it closes the two shapes the fallback is blind to. ⛔ Does nothing about 2 and 3.

Honest limits of this filing

• ⛔ Four instances in one shift in one lane is not a rate. Other lanes and earlier shifts are unmeasured; this may be a local cluster, and the seat that found them is the same seat that produced two of them. Whoever triages should sample another lane before sizing the work.
• Every instance here was caught before landing — nothing shipped with a false or missing declaration. The cost was a held PR and seat time, not a bad merge. That argues for fixing the reach, not for urgency.
• ⚠️ The seat's own carrier readings this shift were manual and labelled as such on every landing comment. That labelling is the only reason this pattern was visible at all; it is worth keeping regardless of what happens to this card.

Related

objectstack/scripts/pm/check-clause2-carriers.mjs · objectstack/scripts/pm/check-half-states.mjs (CLAIM_COMMENT_MARKER) · the Clause-② enqueue gate in .claude/skills/pm-dispatch/SKILL.md and references/contract-review.md · instances: objectui#7491 / #5935, objectui#7498 / #7402, objectui#7391 / #7210 + #7225

[zhuangjianguo]

Evidence, a fifth shape from the domain:spec seat (session_0174WZTU6XcFcS7g2kykC53i, seat post #6017), 2026-09-03 — recorded here rather than as a new finding (this card is the open home of the family; #13914 closed the earlier 2-of-3 measurement).

Shape 5 — the declaration is on the PM's claim, not the dev's. PR #15077 / card #14092: the PM claim comment 5530925467 carries Clause-②: yes in the fixed spelling; the dev's claim comment 5531072392 (the os-dev's own Claim: comment, written from the os-dev.md template) carried no Clause-② line at all. node scripts/pm/check-clause2-carriers.mjs --pair 15077 (run 2026-09-03T20:2xZ from a checkout at origin/main be416187, network available) exited 4: "the Clause-② declaration is MISPLACED: the fixed spelling appears on the thread (PM claim) but NOT in the card's claim comment (the dev's claim 5531072392), which is the carrier the enqueue gate's content limb reads". Repair: the dev edited its claim in place (REST PATCH, 20:31:48Z) appending the line; re-run at 20:32Z: exit 0. The pair was reviewed and cleared normally after that.

Where the shape comes from (mechanizable, the (a)-type fix): .claude/agents/os-dev.md on origin/main be416187 contains no Clause-② field in the claim-comment template (grep for Clause-②, Clause ②, 条款②: zero lines), so a dev's claim carries the declaration only when the PM's dispatch text says to add it by hand. The two dispatches this seat made after the reading (#14103 at 20:04Z and #14104 at 20:46Z) carry that instruction explicitly; that is a (c)-type check to remember, not a fix. The fix is one template line in os-dev.md — Clause-②: yes | no copied from the dispatch order — so the dev's claim is a legible carrier by construction. Skills lane's grading.

Positive control for the reading (run 2026-09-03T20:50Z): --pair 15055 (open, queued, Clause ② no) exits 0 — "readable in the fixed spelling and both carriers agree"; --pair 15029, 15053, 15054 (already merged) exit 2 — "not open … the pair could not be formed", so their enqueue-time readings are not re-derivable now and are not claimed here.

---

Generated by Claude Code