feat(pm): the Prior rulings read: line covers the card’s own thread — tool corpus half and charter :339 in one batch

[os-elon-musk]

Fixes #18993
Clause-②: no

scripts/pm/check-prior-rulings.mjs answered the Prior rulings read: line out of three tree corpora (docs/adr/**, AGENTS.md, packages/spec/src/**) while SKILL.md:339 required it to cover 「已裁卡」 — a card's own thread, where the cheapest ruling lives, was in no corpus. Measured consequence on #17518: a director-seat ruling (5651572469, 2026-09-13) was followed by a retriage that re-presented the card as unruled, a second maintainer ruling with a different prescription (5716259259, 2026-09-17), and a dispatch round that ended with zero diff. This PR makes the card's own thread a corpus half of the tool and rewrites the charter line in the same batch, so the requirement and the tool never diverge again.

What changed

scripts/pm/check-prior-rulings.mjs (report-only, no write path, unchanged exit register)

• Every --card run now reads the card's comment thread over the board route the file already uses (its own proxy re-exec guard, taken once, never on --self-test), one paged GET …/issues/{n}/comments?per_page=100&page=P walked by page number to a short page, bounded at MAX_THREAD_PAGES (50), reconciled against the card's comments field.
• A ruling comment is recognised by the shape of its first content line, three measured shapes: the charter's Ruling: batch #n item k … (ruling-line), the pre-charter ## Ruling recorded — … heading (ruling-recorded, the shape 5651572469 is written in), and Ruling addendum … (ruling-addendum). Every hit prints its id, author, timestamp and shape.
• Two measured near-misses are pinned OUT and listed uncounted (**Ruling C landing step** — …, ## Ruling A's patch is falsified …).
• The paste line gains a thread: clause with three spellings kept apart: thread: 2 ruling(s) (id, id); thread: none (read, no ruling); thread: not read (no --card) (a --terms-only run). A thread that could not be read is the WHOLE line unresolved, exit 3, exactly as an unreadable corpus — never none.
• The report prints the read's cost (pages, bytes, ms) and, when a ruling stands, the 〈升级与决策〉③ warning; with two or more rulings it says to read every one and to return a disagreement to the director seat.
• Self-test: 99 → 144 cases (the three shapes, the near-misses, first-content-line anchoring, the fixture thread with two rulings oldest-first, none / not read / unresolved, and an offline paged board: two pages, one page, a 403, no token, the page bound, a network failure, a malformed page; plus a structural pin that no Link header is followed). check-half-states.mjs's source-text pin on the line's writer (H62) still holds: 4963 cases pass.

.claude/skills/pm-dispatch/SKILL.md :339 — rewritten in place, net 0, 120 bytes → 120 bytes, 812 / 812 lines before and after:

before: - 方向落在 ADR/已裁卡/不可重裁规则:贴 check-prior-rulings.mjs的Prior rulings read: 行,或不点。
after: - 方向在 ADR/本卡裁决/不可重裁规则:贴 check-prior-rulings.mjs的Prior rulings read: 行,或不点。

The three places a direction may already be ruled now read exactly as the tool's line covers them: an ADR, a ruling on this card, a non-re-rulable rule.

.claude/skills/pm-dispatch/references/decision-analysis.md :45 — the human-facing statement of the same line shape (the tool's formatPasteLine docblock names it as the twin), rewritten in place, 113 → 120 bytes, 50 / 50 lines:

before: - 四棱后一行:`Prior rulings read: TERMS → N hits; ADR-xxxx Dn|none`;未搜成整行写 unresolved。
after:  - 四棱后:`Prior rulings read: TERMS → N hits; ADR Dn|none; thread: K|none`;未搜成整行写 unresolved。

(The file spells the placeholders in angle brackets; they are written as capitals here because the platform mutates angle-bracket fragments in a body.)

This file was not in the dispatch's named surface; it is included because it states the line's shape and would otherwise be the next 「a requirement and a tool that cannot answer it」 pair. No open PR touches it (file lists of the 14 skills-lane PRs read 2026-09-18T11:3xZ).

Hypotheses measured

1. Ruling shape is stable; author is not. Census over 1,100 comments (REST, reads only): 2026-09-11T06Z–09-13T21Z (600): ## Ruling recorded — … 22 (os-tesla 19, claude[bot] 3), Ruling: … 4 (os-tesla 2, claude[bot] 2); 2026-09-18T04Z–11Z (500): Ruling: … 26 (hotlong 26), Ruling addendum 2 (hotlong). The charter line (director.md: 「裁决评论首行固定一条:Ruling: batch #N item K · letter X · maintainer 「VERBATIM」 UTC」 — placeholders capitalised here, the file spells them in angle brackets) landed 32a3214 2026-09-13T11:27Z; the heading shape is the same seat's earlier spelling. The fallback the dispatch named — 「every comment by a GOVERNED_APPROVERS login」 — is falsified: it would have missed every 2026-09-13 ruling, 5651572469 included. A quoted 「同意」/「agreed」 predicate hits 63 of 500 recent comments (claims, ledgers, dev reports) and names nothing.
2. Cost, the tool's own meter (node fetch on the re-exec'd route, pages serial, 2026-09-18T11:3xZ):

thread	comments	pages	bytes	thread read	whole run	rulings / near-misses
#17518 (the incident)	17	1	141,623	347–369 ms	2,353 ms (before: 2,001 ms)	2 / 0
#12708 (director seat post)	372	4	2,212,674	1,984 ms	3,669 ms	0 / 0
#6015 (longest thread on the board)	962	10	7,365,556	4,431 ms	6,254 ms	0 / 0

curl on the same pages reads ~10% more bytes (156,874 on #17518): GitHub pretty-prints for a curl/* User-Agent and minifies for any other, measured directly; the tool reports what its own route pays. The fixed cost per dedupe on the incident card is one page and about a third of a second.

3. The line, before / after, on --card 17518 through the proxy (exit 0 both):

before: Prior rulings read: spec,assembledpackagebodyschema,declares,callable,custom,branches,inert-json,artifact,hold,schema,embedding,loses (+1 more) → 117 hits; ADR-0087 D4, ADR-0130 D4, ADR-0029 D6, ADR-0039 D7, ADR-0056 D5, ADR-0076 D3, ADR-0076 D9, ADR-0087 D1, ADR-0087 D3, ADR-0087 D7
after:  Prior rulings read: spec,assembledpackagebodyschema,declares,callable,custom,branches,inert-json,artifact,hold,schema,embedding,loses (+1 more) → 117 hits; ADR-0087 D4, ADR-0130 D4, ADR-0029 D6, ADR-0039 D7, ADR-0056 D5, ADR-0076 D3, ADR-0076 D9, ADR-0087 D1, ADR-0087 D3, ADR-0087 D7; thread: 2 ruling(s) (5651572469, 5716259259)

and the report's thread block:

the card's own comment thread (ruling comments, by first-line shape): 2 ruling(s) of 17 comment(s) — read in 1 page(s), 141,623 bytes, 347 ms
    5651572469  os-tesla  2026-09-13T06:10:22Z  [ruling-recorded]  Ruling recorded — B: `packages/spec` declares the inert-JSON artifact stage; the read A…
    5716259259  hotlong  2026-09-17T14:41:18Z  [ruling-line]  Ruling: batch #149 item 1 · letter B (artifact-stage JSON-only variants of `FlowFunctio…

4. --terms single,posture,tenant (no card): exit 0, line ends ; thread: not read (no --card).

Gates

Derived with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack from the worktree (three-dot change set: the three files above): 40 families, every one run, exit codes captured redirect-then-$?, reconciled with --ran (40 derived, 40 run after the one PREREQUISITE NOT MET was cleared by its declared build: check:doc-formula-expressions exited 3 for an unbuilt @objectstack/formula; built under the verify lock, re-run, exit 0). check:pm-dispatch-gates exceeds the foreground cap and ran detached (tail --pid): 779 s, 1849 cases pass, exit 0. Also: node scripts/pm/check-prior-rulings.mjs --self-test 144 pass; check:pm-skill-ratchet ✓ (812 / 812, 50 / 50, widest rows unchanged); check:pm-skill-id-lint ✓ 27 files clean; check:nul-bytes ✓; check-governed-merges.mjs --test over the three paths: GOVERNED (2 of 3 on the register — expected, .claude/**); eslint on the edited script: 1 file, 0 errors, 0 warnings (--no-inline-config, the root flat config's **/*.mjs block; no type-aware rules, so the diff moves no untouched file's verdict). Readings taken at cba931995, the branch head. The same derivation from a throwaway worktree at origin/main 64de4c5 (five commits past this branch's base, four of them touching family-defining files) yields the identical 40 families — none new, none gone. This branch is deliberately not merged with main here: the incoming range carries four merge=os-regen artifacts whose driver debt would hold the push on a PR that touches no spec file; the seat's update-branch is the right instrument if a refresh is wanted before landing.

No published package moves (scripts/pm/** and .claude/** ship in no files[]): skip-changeset.

Acceptance notes

• noted, not filed: the file's self-test keeps a flat case list with the verdict handshake and no battery-name floor (the AGENTS.md --self-test section's floor rule is stated for writing a new one); the 2026-09 self-test shape census owns that class — carrier: whoever works that census; not this card.
• noted, not filed: readCard still uses the global fetch while readThread takes an injectable one; symmetry only, no behaviour — carrier: none.
• Boundary kept: no merge with [finding] two cards on one file, one defect, graded two days apart into two lanes and both dispatched — #18314 (spec, 2026-09-15) and #18844 (skills, 2026-09-17) each produced a PR for check-single-claim-paths.mjs's proxy re-exec; #18844's triage named no sibling, and the second PR closed as superseded #18963 / PR skills(pm-dispatch): dedupe at grading by file/mechanism over the open board of both repos — same finding ⇒ duplicate_of, never a second dispatch #18981 (grading-time dedupe by file) — same class, different mechanism; [finding] two cards on one file, one defect, graded two days apart into two lanes and both dispatched — #18314 (spec, 2026-09-15) and #18844 (skills, 2026-09-17) each produced a PR for check-single-claim-paths.mjs's proxy re-exec; #18844's triage named no sibling, and the second PR closed as superseded #18963 is not addressed here.

维护者速读(草稿)

改了什么:查重工具 check-prior-rulings.mjs 多读一处 —— 被查那张卡自己评论线上的裁决评论;贴进卡面的那一行 Prior rulings read: 多出一段 thread: …,写明本卡已有几条裁决、评论 id 是哪几条;读不到线程时整行仍写 unresolved,绝不写成「没有」。章程 SKILL.md:339 与 decision-analysis.md:45 各原位改一行,行数不变,同一批落地。

为什么改:#17518 上四天内出现两条互相矛盾的维护者裁决,一整轮派发白跑 —— 原因是重新分诊的席位没看到卡上四天前的那条裁决,而工具即使跑了也查不到:卡的评论线不在它任何语料里。这次把「已裁卡」这半边做成机械面,并把章程那句话改成工具真能答的范围。

风险与代价(含回滚):每次查重固定多一次线程读取 —— 事故卡 1 页、约 0.35 秒;全板最长线程(962 条评论)10 页、约 4.4 秒,已实测。识别靠裁决评论首行的固定写法(三种实测形态),不靠作者身份;写法漂移会以「近似项」在报告里显形而不是静默漏掉。回滚 = revert 本 PR 三个文件,无发布物、无数据迁移。

席位意见:(留空)

你要做的:确认三种裁决写法的识别边界可接受;若同意,按受管面流程合并。

---

Generated by Claude Code

[os-elon-musk]

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: cba9319955

① Derived judgments

• Surface: scripts/pm/check-prior-rulings.mjs (+430 / −46), .claude/skills/pm-dispatch/SKILL.md :339 in place (1 / 1, 120 bytes measured, 812 / 812), .claude/skills/pm-dispatch/references/decision-analysis.md :45 in place (1 / 1, 120 bytes, 50 / 50). Governed: .claude/** ×2 (check-governed-merges.mjs --test, seat-run); the tool is off the register. The third file is outside the claim's named surface — accepted: it is the human-facing twin of the paste line's pinned shape (the tool's own docblock names it), and leaving it would have recreated the requirement / tool divergence this card exists to close; no open PR touches it. One landing on SKILL.md since the claim: PR skills(pm-dispatch): a passing control certifies the instrument, not the question — a zero-hit reading names the instrument's reach and one known target outside it #18921 (37ca54a11, hunks :162–:183) — disjoint from :339; PR skills(pm-dispatch): key the clause-② contract review by lane — spec and skills owe it on every round, other lanes owe none #18903 (:509–:525, :633–:675) and PR skills(pm-dispatch): dedupe at grading by file/mechanism over the open board of both repos — same finding ⇒ duplicate_of, never a second dispatch #18981 (:363–:371, queued) likewise.
• The corpus half. A ruling comment is recognised by the SHAPE of its first content line — Ruling: … (the charter line in director.md), ## Ruling recorded — … (the pre-charter director shape; the incident's own 5651572469 is written in it), Ruling addendum … — measured over 1,100 live comments in two windows; two near-miss shapes measured and pinned OUT (**Ruling C landing step**, ## Ruling A's … is falsified), listed uncounted so a spelling drift shows as a near-miss rather than an empty thread. The dispatch's fallback (an author predicate on GOVERNED_APPROVERS) was falsified by measurement — the director seat wrote as three logins in five days and that predicate would have missed 5651572469 itself; the quoted-「同意」 predicate names nothing (63 of 500 comments quote one). Correct readings, and the falsification is the seat's own hypothesis, as invited.
• The read. One paged GET …/issues/{n}/comments?per_page=100&page=P walked by page number to a short page (never Link:, per rest-channel.md), bounded at MAX_THREAD_PAGES = 50 (a filled bound is unresolved, never a clean read of its first pages), reconciled against the card's own comments count (a mismatch is flagged as 「one landed between the two reads; re-run」 — the finding(skills): a suspended account's comments are REMOVED from threads while the issue's comments count lags behind — a seat's entire written record can vanish, and the mutual-exclusion tail read then says "vacant" #18052 transient), no retry on any status, and the cost metered off the wire: finding(spec): AssembledPackageBodySchema declares callable and custom branches an inert-JSON artifact cannot hold, and every schema embedding it loses its JSON Schema #17518 1 page / 141,623 B / ~0.35 s; [PM seat] director(项目总监) — ⏹️ 第 23 场(续)已收班 2026-09-16(决裁批 #133–#139 裁毕)· 第 22 场(续)session_01QsCVSivtpwT6ZXs5Rtvqxe 块保留 · 第 16–21 场已收班 #12708 4 pages / 2.2 MB / 2.0 s; [PM seat] triage (objectstack-wide) — 🟢 session_01KYzNJPin9Ar4oMMkSh4S9h · 就座 2026-09-20 · 自绑 Routine trig_016NBra1kZGHk7UteTziFbrK · ⛔ 接手先读正文「大扫除试运行班收班交接台账」再定级 #6015 (the longest thread) 10 pages / 7.4 MB / 4.4 s — the fixed cost the triage asked to be measured, measured. Read BEFORE the tree corpora, so an exit 3 still means 「nothing was searched」.
• The line. Prior rulings read: … ; thread: k ruling(s) (ids) / thread: none (read, empty) / thread: not read (no --card) (a --terms-only run, which stays offline) — three spellings kept apart, and an unreadable thread makes the WHOLE line unresolved at exit 3, exactly as an unreadable corpus does. The report adds, when a ruling already stands, the SKILL.md 〈升级与决策〉③ sentence (the card is execution, cite the ruling on Ruling-ref:) and, for two or more, 「a conflict goes back to the director seat, not to a dev as a choice」. :339 now reads 「方向在 ADR/本卡裁决/不可重裁规则:…」; :45 pins the clause. Requirement and tool change in one batch — the triage's condition, met.
• Seat re-measure on the head (scratch worktree at cba9319955): --self-test 144 cases (main 99), the live-tree D8 pin pinned; ratchet ✓ 812 / 812 and 50 / 50; check-skill-id-lint ✓ 27 files; the live reader test PM_SWEEP_REPO=objectstack-ai/objectstack node scripts/pm/check-prior-rulings.mjs --card 17518 at 12:00Z: exit 0, thread: 3 ruling(s) (5651572469, 5716259259, 5729478920) — the third is hotlong's batch Remove redundant field name declarations in object protocol #159 item 2 (「the 2026-09-13 ruling stands」), ruled at 11:42Z after the dev's measurement and read by the tool oldest-first with the two-ruling advice printed: the reader test held on a thread that moved under it. On main the same run prints no thread clause at all. --pair 19021 exit 0; closingKeywordTargets(body) = {18993 → Fixes}; commit trailers model-free; no model string in the body.

② Semver level

• skip-changeset is correct (scripts/pm/** and .claude/**, nothing published moves). Clause-②: no on the claim (5729187152) and in the body; --pair 19021 at 2026-09-18T11:58Z: exit 0, both carriers agree, no widening tell.

③ Boundary flags

• Dev flags: no open questions. Seven declared deviations — the third file (accepted, ①); the third shape (measured, accepted); the author predicate rejected (measured, accepted); --card N --terms … now reads the thread too (the thread is about the card, not the terms — accepted); selfTest async with the handshake unchanged (the seat's run reached the verdict); origin/main not merged into the branch because the incoming range carries four merge=os-regen artefacts (the queue rebuilds on main; disjoint hunks verified above — accepted); no full pnpm lint (targeted eslint 0 / 0 on the edited script; CI's). Four noted, not filed — the flat self-test case list without a battery floor (carrier: the self-test shape census), fetch-injection asymmetry, GitHub's User-Agent-dependent JSON minification, the [finding] two cards on one file, one defect, graded two days apart into two lanes and both dispatched — #18314 (spec, 2026-09-15) and #18844 (skills, 2026-09-17) each produced a PR for check-single-claim-paths.mjs's proxy re-exec; #18844's triage named no sibling, and the second PR closed as superseded #18963 boundary kept — agreed.
• Gates: 40 derived / 40 run / 0 NOT-MEASURED at cba931995 (one exit 3 cleared by building formula + lint under the verify lock, re-run 0); check:pm-dispatch-gates 779.3 s detached, 1849 cases; check:pm-half-states 4963 cases (the H62 writer pin reads this file). Derivation repeated from a throwaway worktree at origin/main 64de4c576: the identical 40 families. 0 MCP calls; 3 REST writes + 2 pushes; label-write.mjs was NOT denied for this subagent (the [finding] duplicate-fix-guard's branch-name advisory matches issue-N as an undelimited substring, so a fix branch that names a DIFFERENT, longer card number passes silently (.github/workflows/duplicate-fix-guard.yml :85; same class PR #18918 closed in AGENTS.md) #18922 subagent's denial of the same tool was not deterministic — recorded for [finding] os-dev.md says nothing about a classifier denial of a tool the dispatch names — an os-dev subagent, denied label-write.mjs by the auto-mode classifier, re-implemented the label write by hand over the REST proxy #19004).
• CI on this head at 2026-09-18T12:01Z: 17 success · 11 skipped · 5 in_progress — nothing red; the landing act reads it again.
• Landing: governed ⇒ draft at the four-piece terminal (ACCEPT on [finding] 机械查重只盖 ADR,不盖卡自己线程上的裁决 —— 一次 retriage 把已裁卡当未裁重新呈报,#17518 因此带着两条互相矛盾的维护者裁决,并让一整轮派发以零 diff 告终 #18993, needs-user-decision on this PR, the 维护者速读 below, review requested from os-zhuang and hotlong). On an authorized APPROVED: ruling C by the seat, channel permitting — or the approver's own hand, as today.

Implemented-by: claude/issue-18993-prior-rulings-card-thread
Reviewed-by: session_01BTeBejoPUvRHN8WdAJC6oF

VERDICT: PASS

---

Generated by Claude Code

[os-elon-musk]

维护者速读(终稿)· skills 席 · 2026-09-18T12:03Z

这个 PR 做什么:PM 循环在呈报决策卡前要机械查一遍「有没有已经裁过的」。查重工具 check-prior-rulings.mjs 以前只读三份树上的语料(ADR、AGENTS.md、spec 源码),读不到卡自己线程上的维护者裁决 —— #17518 因此被当作未裁重新呈报,您裁了两次、互相矛盾,一整轮派发零 diff 收场。现在工具每次 --card N 都把卡的评论线程当第四份语料读一遍,按裁决评论的首行形状识别(总监席的三种写法都认,两种「以 Ruling 开头但不是裁决」的写法明确排除),粘贴行多一段 thread: k ruling(s) (评论 id);读不到线程整行写 unresolved,⛔ 不会假装线程是空的。章程 SKILL.md :339 与 decision-analysis.md :45 同批改成对应的字,各一行、净 0。

代价,实测:每次查重多一次分页读线程 —— 事故卡 #17518 是 1 页、0.35 秒;板上最长的线程(962 条评论)10 页、4.4 秒。

验收:复核记录 5729706403(PASS)、ACCEPT 5729706770;自测 99 → 144;派生门禁 40 / 40 绿;在 head 上实跑 #17518 读出三条裁决(含 hotlong 今天 11:42Z 新裁的一条),最早的在前。

请您做的一件事:受管文本(.claude/**)⇒ 需要您或 os-zhuang 的 APPROVE;批了之后席位落地,或您顺手 ready + squash 合并。没有要问的字。

---

Generated by Claude Code