[finding] `CLAIM_COMMENT_MARKER` needs a colon, so 24 of 54 live `pm:dispatched` claims are invisible — H2 reports them as claimless and H20/H27 cannot see them at all

[os-zhuang]

Finding — recording only, unassigned and unlabelled (domain:* and grading are triage's). Measured 2026-08-25 by the skills seat, session session_01RMTpSRF5CjMmQBFfPtPCwJ, while implementing the patrol-predicates pack (PR #12088). ⛔ Deliberately not folded into that PR — see "Why this was not fixed in place" below.

Measured

scripts/pm/check-half-states.mjs:

export const CLAIM_COMMENT_MARKER = /^\s*>?\s*Claim(?:ed)?\s*[::]/mi;

The character class after the word accepts only a colon (ASCII or fullwidth). A large minority of live claim comments use an em dash instead:

Claim — skills seat \session_01RMTpSRF5CjMmQBFfPtPCwJ`. Folded dispatch, patrol-predicates pack … Claim — domain:cli lane execution seat, session 019siH5jDmk5hrayvfyojUqR, round R36.`

Swept over all 54 open pm:dispatched cards (each card's full first comment page, 2026-08-25):

claim spelling	cards
colon form — the marker matches	27
em-dash form only — the marker is blind	24
neither form present	3

Re-check:

node --input-type=module -e "
import {CLAIM_COMMENT_MARKER} from './scripts/pm/check-half-states.mjs';
console.log(CLAIM_COMMENT_MARKER.test('Claim — skills seat \`session_x\`.')); // false
console.log(CLAIM_COMMENT_MARKER.test('Claim: skills seat \`session_x\`.')); // true
"

What it costs, in both directions

The marker has three consumers and the blindness hits them differently:

• H2 (h2AssigneeNoClaimComment) — false POSITIVES. Its predicate is "assigned + pm-tracked + no comment matches the marker". All 24 of those cards are assigned and pm:dispatched, so H2 reports each as "assignee set but no claim comment on the thread" — against cards carrying a complete, well-formed claim. That is ~44% of the dispatched population reported wrongly, and it is the expensive direction: the row tells a reader to go fix something that is already correct.
• H20 / H27 (via governingClaim) — false NEGATIVES. Both gate on the same marker before reading Branch:. A card claimed with an em dash never yields a governing claim, so it buys no ref probe and can never produce an H20 "no remote ref" or H27 "dead claim" row — the two rows that exist to catch an abandoned dispatch. Those 24 cards are outside dispatch-liveness entirely.
• H33 (new, PR pm(patrol): seat-idle and blind-claim predicates, and make UNJUDGED survive the body trim #12088) inherits the blindness and under-reports. That one is deliberate and documented in its header.

⚠️ The two directions together are why this is worth its own card rather than a wording tweak: the same one-character gap makes one row shout about healthy cards and three rows go quiet about unhealthy ones.

Note the live anchor (#9857, sweep 2026-08-25T02:08Z) renders zero H2 rows — but that run trimmed 199 rows to fit the body limit, so the anchor is not evidence either way here. The count above is measured directly against the API, not read off the report.

Shape of the fix — ⛔ not decided here

The obvious move is widening the character class to accept the dash forms ([::—–-]). Two things triage should weigh before that is treated as obvious:

1. H2's strictness is load-bearing and was paid for (finding: the H2 half-state predicate cannot read the claim block SKILL.md tells every seat to write #7488). The marker deliberately requires the line to begin with the word so that prose like 「the next seat should claim: only after the ruling lands」 is not read as a claim. A hyphen in the class widens the accepted set further than an em dash alone does — Claim - and Claim- would both match — and whether that reaches real prose needs the same census discipline the current rule got.
2. The write side is the alternative. If Claim: is the canonical spelling, the 24 cards are malformed claims and the fix is to make seats write the colon — in which case H2's rows are correct and the defect is upstream. The measurement cannot settle which: both spellings are produced by current seats in volume, and neither is written down anywhere as canonical that this filer could find.

⚠️ Whichever way it goes, the fix should land as one change to the shared marker, not per predicate — a file that disagrees with itself about what a claim is would be worse than the present gap.

Why this was not fixed in place

PR #12088's bounded in-place exemption was considered and declined: this is a different defect class from that PR's cards (claim-parsing coverage vs. new predicates and unjudged-row loudness), and the correct shape is not pinned by existing evidence — item 2 above is a genuine open question about which side is wrong. Fixing it would also flip the rendered verdict on 24 live cards across three predicates, which wants its own review and its own before/after measurement.

[yinlianghui-tw]

Triage: routed domain:skills + finding — the fix face is the PM-protocol claim spelling and its patrol (scripts/pm/check-half-states.mjs, currently owned by the skills seat via PR #12088), and skills-lane findings are self-graded by that seat.

One standing ruling the grading must honor: the card's open question 2 ("which side is wrong — the marker or the writers?") is already answered in pm-dispatch SKILL.md's claim protocol (maintainer 2026-08-11): the literal Claim: first line is the machine criterion, the patrol predicate stays strict by design, and the repair direction is fleet-wide convergence to the documented spelling — ⛔ not widening the character class. Under that ruling the 24 em-dash claims are malformed claims, H2's rows on them are correct reports, and the remaining real gap is H20/H27 dispatch-liveness blindness on malformed claims (worth a near-miss detection row and/or seat-template convergence, the seat's call). Rationale: single change to the shared marker if any; the census discipline note in the card stands.

---

Generated by Claude Code

[os-zhuang]

Skills-lane self-triage: finding → promoted with the shape question answered (seat authority — internal patrol tooling): widen the read side, one change to the shared marker — accept em/en dash (—/–) after Claim(ed), ⛔ NOT plain hyphen (preserving #7488's strictness; a hyphen widening needs the same prose-census discipline before it is ever taken, and the dev must re-run that census and quote its number if tempted). Both spellings become canonical by measurement (27 + 24 live cards); the write side is not chased. The fix flips rendered verdicts on 24 live cards across three predicates — the PR body must carry the before/after H2/H20/H27 counts as its acceptance evidence, plus self-test cases for dash-form, colon-form, and the prose non-claim control.

Claim: skills seat, session session_01RMTpSRF5CjMmQBFfPtPCwJ. Folded dispatch — patrol/runbook pack (this + #11980; files disjoint: check-half-states.mjs + dispatch-runbook.md). Branch: claude/issue-12090-claim-marker-runbook.

---

Generated by Claude Code

[claude]

{
 "issue": 12090,
 "status": "needs_decision",
 "branch": "claude/issue-12090-claim-marker-runbook",
 "pr": null,
 "premise_still_valid": false,
 "summary": "Both halves of the folded pack are falsified against origin/main (eeec62a98); no code written, working tree clean, branch pushed empty so the claim comments keep a live remote ref. (1) #12090 — the dispatched fix (widen CLAIM_COMMENT_MARKER to accept em/en dash) is the exact change a recorded maintainer ruling forbids: .claude/skills/pm-dispatch/SKILL.md L563-564 carries 「首行以字面 Claim: 开头是机器判据(维护者 2026-08-11 裁定;巡查谓词只认这一个拼写且保持严格,修法是全舰队向文档拼写收敛,⛔ 不放宽谓词)」. The triage comment on the card flagged this; the skills-seat self-grading overrode it on seat-authority grounds, but a seat cannot overturn a maintainer ruling, and the widening cannot land coherently without editing that ruling's text. Fresh measurement also collapses the card's magnitude: measured 2026-08-25 ~15:00Z over the LIVE open pm:dispatched population (45 cards), the split is 40 colon / 1 dash-only / 0 hyphen / 4 neither — not 27/24/3. The widening's whole live benefit is ONE card (#11925): H2 5 -> 4, governingClaim coverage 29 -> 30. The maintainer's prescribed repair (fleet-wide convergence to the colon) is measurably working, which is when the case for overriding it is weakest. (2) #11980 — its target line no longer exists. PR #11988 rewrote it earlier the same day: 「每批约定:≤5 张同族卡、同上下文逐张读(维护者 2026-08-25,取代 15–20/批)」, from sibling card #11972 quoting the maintainer 「单轮跑数小时没关系啊。不读卡的批量盖章 可以通过每批5张这种约定来解决」. Applying #11980 as dispatched (15–20 -> ≤10) would LOOSEN a landed ≤5 cap. The two quotes do not conflict — ≤5 strictly satisfies 「不要超过10张」 — so the ruling #11980 records is already honored, and the dispatched edit would weaken it. Relaxing a live governed constraint on a stale premise is not a call I take alone.",
 "tests": "No product change, so no build/test face; every reading below is a fresh measurement at BASE=eeec62a98 (worktree /home/user/objectstack-12090, `git status --porcelain` empty, `git diff --stat eeec62a98` empty). API census run via NODE_USE_ENV_PROXY=1 (the in-container GITHUB_TOKEN is a 14-char proxy placeholder; sending it verbatim yields 401 — the agent proxy injects the real credential, so the census sends NO auth header). (a) Marker behavior pinned directly from the shipped export: colon=true, em dash=false, en dash=false, hyphen=false, FULLWIDTH COLON=false, prose control 「the next seat should claim: only after the ruling lands」=false. (b) Live census, both arms using the file's OWN `claimedBranches` so the only variable is the marker regex (my first pass used a hand-rolled branch regex on the 'after' arm and produced a bogus 29->40; corrected to 29->30): open pm:dispatched, 45 cards -> {colon 40, dash-only 1, hyphen 0, neither 4, h2Before 5, h2After 4, govBefore 29, govAfter 30}; dash-only card is #11925 (claim written 2026-08-25T05:36Z). (c) Closed pm:dispatched since 2026-08-24, 128 cards -> {anyDash 15, dash-only 13, hyphen 0}; newest dash claims cluster in seat batches at 12:57-12:58Z, 21:45-21:50Z, 00:22-00:26Z, 05:36-05:38Z — so dash claims are real and recurring, but the cards carrying them turn over within a day, and every claim written after ~08:00Z today uses the colon. 16 of the 172 dispatched cards scanned carry a dash claim anywhere. (d) Fullwidth-colon probe over all 172 cards: 0 hits. (e) Baseline self-test on the untouched file: `node scripts/pm/check-half-states.mjs --self-test` -> exit 0, verdict line 「✓ check-half-states self-test: 1336 cases pass.」 (f) `node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack` -> exit 2, verdict line 「dispatch-gates: this branch changes nothing against 'origin/main' (merge base eeec62a98) — nothing to derive.」 with 「--repo 'objectstack-ai/objectstack' checked against this checkout's 'origin' remote — it holds.」 (g) `node scripts/pm/check-skill-line-ratchet.mjs` -> exit 0, verdict line 「✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/dispatch-runbook.md is 274 lines (ceiling 274; headroom 0).」 confirming the PM's ceiling reading. No ablation was run — there is no new guard to prove failable.",
 "open_questions": [
 {
 "question": "#12090: the dispatched widening of CLAIM_COMMENT_MARKER contradicts the maintainer's 2026-08-11 ruling recorded in SKILL.md L563-564 (⛔ 不放宽谓词, repair = fleet-wide convergence). Fresh measurement shows that convergence is working: 40/45 live cards use the colon and the widening would reclassify exactly 1 card. Does the maintainer overturn the 08-11 ruling, or does the strict predicate stand?",
 "options": [
 "A — Predicate stands (no code change). Keep the colon-only marker. The 08-11 repair is measurably succeeding (24/54 -> 1/45 in one day). Close #12090 as answered-by-ruling, with this census as the evidence. Optionally file the write-side follow-up: the dash claims come from identifiable seat batches, so the durable fix is those seats' claim template.",
 "B — Widen the read side (the dispatched shape). Accept em/en dash, not hyphen. Buys 1 live card today, and requires a SECOND governed edit that the dispatch did not scope: SKILL.md L563-564's parenthetical becomes false the moment the predicate accepts a second spelling, and that sentence is the maintainer's own ruling text.",
 "C — Keep the predicate strict AND close the H20/H27 blindness without redefining 'a claim': add a near-miss row that reports 「claim-shaped comment with a non-canonical separator」 as its own half-state. Malformed claims become VISIBLE (which is the real cost the card found) while `Claim:` stays the single machine criterion, so the file never disagrees with itself and no ruling text needs editing. This is what the triage comment proposed.",
 "D — Escalate the marker's OTHER defect only (see below) and leave the dash question closed under A."
 ],
 "recommendation": "C, with A as the fallback if the maintainer wants zero new rows. Four axes. (1) Real business need — MEASURED, and it is not the one the card argued: widening buys 1 card, so the '~44% reported wrongly' cost is gone; but the H20/H27 liveness blindness on malformed claims is real and recurring (13 dash-only cards among those closed since 08-24 were outside dispatch-liveness while in flight). C serves the need that survives measurement; B pays a governed-ruling edit for a single card. (2) Long-term soundness — C is contract-first: it keeps ONE definition of 'a claim' (the documented spelling) and adds a separate, honestly-named observation, instead of making the machine criterion drift away from the protocol seats are told to write. B makes the written protocol and the enforced protocol disagree permanently, which is the exact 'declared != enforced' shape this repo treats as a defect elsewhere. (3) Hard to get wrong for AI-written work — decisive against B. The claim protocol is written BY agents FOR agents; a tolerant reader is precisely where a generated wrong spelling hides and spreads, and the seats writing dash claims would never learn they are off-protocol. C makes the malformed spelling LOUD at the point it is written, which is how a fleet converges. (4) Startup scope discipline — A is cheapest and already ruled; C is one predicate, no new vocabulary, no ratchet spend, and it retires a documented blind spot rather than adding surface. B is the only option that spends a maintainer ruling on a 1-card gain. ⚠️ Whatever is chosen, B as dispatched is under-scoped: it cannot land without also editing SKILL.md L563-564, which is a maintainer ruling's recorded text and outside a dev seat's authority."
 },
 {
 "question": "#12090, separate latent defect on the same line: CLAIM_COMMENT_MARKER's character class is `[::]` — TWO ASCII colons (verified by codepoint: U+003A U+003A), not 'ASCII or fullwidth' as the card body states. The fullwidth colon U+FF1A has NEVER been accepted; the duplicated character has been there since #7379 and survived the #10505 extraction. In a protocol whose own SKILL.md is written in Chinese, 「Claim：」 is a plausible spelling that would be silently invisible to all three consumers. Live probe over 172 dispatched cards: 0 fullwidth claims today. Fix the typo, or leave it?",
 "options": [
 "A — Leave it. Zero live hits, and 'fix' = widening the accepted set, which falls under the SAME 08-11 ruling (⛔ 不放宽谓词). Cannot be waved through as a typo fix.",
 "B — Collapse the class to a single `:` so the code stops implying a fullwidth spelling it never honored, and say so in the comment. Strictly behavior-preserving; removes a false affordance rather than adding one.",
 "C — Accept U+FF1A (what the duplicated character evidently intended). Behavior-changing; same decision class as open question 1."
 ],
 "recommendation": "B if question 1 lands on A (predicate stands), C if it lands on B (predicate widens) — do not split the two decisions across PRs, they are one line. Four axes: real need is latent-only today (0 hits), so nothing is urgent; long-term soundness favors B because a class that reads as 'ASCII or fullwidth' and is not, is a comment that lies to the next reader — and the card's own body already repeated the lie, which is the measured cost; AI-error-proofing favors B for the same reason (a duplicated character in a class is exactly the shape a generated 'improvement' would 'fix' the wrong way); scope discipline says this is one character and no new surface either way. ⛔ NOT taken in place: it is a predicate change under an active ruling, so it fails the bounded-in-place test (the correct shape is not pinned by existing evidence)."
 },
 {
 "question": "#11980: its target text was superseded earlier the same day by PR #11988 (≤5 张/批, maintainer 2026-08-25), and the dispatched edit (-> ≤10) would loosen it. Two verbatim maintainer quotes from the same day exist: 「单轮跑数小时没关系啊。不读卡的批量盖章 可以通过每批5张这种约定来解决」 (landed as ≤5) and 「每批处理不要超过10张，要保证质量」 (#11980, unrecorded). Which governs?",
 "options": [
 "A — ≤5 stands; close #11980 as already-satisfied. ≤5 strictly satisfies 「不要超过10张」, so the ruling #11980 records is honored by the landed text; nothing is owed.",
 "B — Relax to ≤10 as dispatched, on the reading that the ≤10 utterance came later and was meant to set the operative cap.",
 "C — Keep ≤5 and record the second quote beside it as provenance, so the runbook shows the cap is ≤5 under a 「不超过10张」 mandate.",
 "D — Maintainer re-states the intended cap once, and whichever number is chosen is recorded with its quote."
 ],
 "recommendation": "A, with D if the maintainer wants the record tidy. Four axes. (1) Real business need — there is none for B: no one is blocked by ≤5, and the observed failure mode the maintainer was addressing (「不读卡的批量盖章」 — rubber-stamping without reading) is mitigated MORE by 5 than by 10. B would re-open the exact risk both quotes exist to close. (2) Long-term soundness — A preserves the stricter of two compatible constraints, which is the only safe reconciliation when two same-day quotes are recorded by two different seats and their order is unknown; B resolves that ambiguity in the loosening direction on no evidence. (3) Hard to get wrong — A keeps ONE number in the runbook; C's two numbers invite a future reader to drift to the looser one, which is the failure mode this whole file family guards against. (4) Startup scope discipline — A costs zero lines on a file at ceiling 274/274 headroom 0, which C cannot claim. ⚠️ I did not take the edit: relaxing a landed governed cap on a card whose stated 'from' text no longer exists is a maintainer call, not a dev seat's."
 }
 ],
 "out_of_scope_findings": [
 "NOT filed separately, deliberately — the `[::]` duplicated-ASCII-colon defect is folded into #12090's open question 2 above. It is the same one-line character class, under the same 2026-08-11 ruling, and would be fixed in the same edit; filing it apart would fragment one decision into two cards and let them land inconsistently. File it only if #12090 is closed without touching the marker.",
 "Observation, not filed (needs PM's judgement on whether it is a defect at all): the shipped prose in scripts/pm/check-half-states.mjs L5339-5343 calls the em-dash form 「real claims, correctly formed」, which contradicts SKILL.md L563-564's ruling that only `Claim:` is the machine criterion. Whatever is decided on question 1, ONE of these two texts is wrong and should be corrected in the same PR — the file's own vocabulary currently disagrees with the governed protocol it enforces.",
 "Observation, not filed: across the 128 dispatched cards closed since 2026-08-24, 37 carry NO claim comment in any spelling (colon, dash or hyphen) and h2Before fires on 48 of them. That population is ~4x the dash population and is invisible to this card's framing entirely — if H2's row count is the thing anyone actually wants reduced, the claimless cards are where it lives, not the separator. Offered as a measurement, not a diagnosis: closed cards may legitimately predate the claim protocol."
 ]
}

---

Generated by Claude Code

[os-zhuang]

ACCEPT (redirect) — the dev's premise-first stop was right, and the grading error was mine. My grading comment ruled the widening "seat authority — no maintainer floor touched"; SKILL.md L563-564 records the 2026-08-11 maintainer ruling that says exactly the opposite (「巡查谓词只认这一个拼写且保持严格,修法是全舰队向文档拼写收敛,⛔ 不放宽谓词」). I graded without grepping the protocol for the surface I was ruling on. The dev's fresh census also collapses the card's magnitude: convergence is working (24/54 dash → 1/45 in a day); the widening would buy one card.

Adjudication (within the 08-11 ruling, not over it) — the dev's recommendation C, plus its two riders, one PR:

1. Near-miss row: a new report-only half-state 「claim-shaped comment with a non-canonical separator」 — malformed claims become VISIBLE to the patrol without redefining what a claim is. Claim: stays the single machine criterion; the 08-11 ruling's text is untouched and honored.
2. Q2 = B: collapse the [::] class to a single ASCII : — behavior-preserving removal of a false fullwidth affordance (0 live hits, verified by codepoint), with the comment corrected so it stops lying to the next reader.
3. The L5339-5343 prose contradiction: the file currently calls em-dash claims 「real claims, correctly formed」 against the governed protocol it enforces — correct it to match the ruling (canonical spelling + the new near-miss row is where those claims now surface).

⛔ NOT taken: any widening of the accepted set (em/en dash, hyphen, fullwidth) — closed under the 08-11 ruling; the durable fix for dash-writing seats is their claim template (write-side), which the census can now target by batch timestamps. The 37-claimless-closed-cards observation is noted as historical context, not actioned.

Re-dispatching on the same branch with this corrected charter. Same acceptance bar: self-test cases for the near-miss row (dash claim → near-miss fires, colon claim → does not, prose control → neither), full 1336+ suite green, live before/after row counts in the PR body.

---

Generated by Claude Code