Repository navigation
[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
Description
Filed by the domain:spec execution seat 4(座位贴 #18917,session_01AmH9bKvGoLjiY86Q4Z3og2)—— 出自 #17518 / R4 那一轮 dev 的 out_of_scope_findings,dev 明确把承接者点名为「the PM seat reading this report」= 本席。⛔ 填 bare:finding only。
⭐ 路由建议(⛔ 非本席设的标签):domain:skills —— 受管散文/流程的缺口。分诊按需改判。
Dedupe words: prior rulings corpus excludes card comments · retriage re-presents a ruled card · two contradictory maintainer rulings one card · check-prior-rulings ADR only · decision box re-entry without ruling citation
缺什么 —— 机械查重只盖 ADR,不盖卡自己线程上的裁决
SKILL.md:339 要求:「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。」
scripts/pm/check-prior-rulings.mjs 的语料按它自己的文件头是:docs/adr/**、AGENTS.md、packages/spec/src/** —— 它自陈用途是「given a decision card, list the ADR decisions that already rule on its terms (#17009)」。
⇒ ⭐ 「已裁卡」这一半没有机械面:一张卡自己线程上的维护者裁决评论不在任何语料里。而那恰恰是最便宜、最该先看的地方。
实测后果 —— 一张卡上两条互相矛盾的维护者裁决
⏱️ 本席 2026-09-18T10:10Z 自核 #17518:
| 评论 | 作者 | 时刻 | 处方 |
|---|---|---|---|
5651572469 |
os-tesla | 2026-09-13T06:10:22Z | 另声明 inert-JSON artifact 阶段;AssembledPackageBodySchema 与 composeStacks 保持不动 |
5716259259 |
hotlong | 2026-09-17T14:41:18Z | 两个 entry schema 做 artifact-stage 变体;装配体本身收窄 |
两条都带维护者原话(「other agreed」/「agreed」)。中间发生了什么:
| 评论 | 作者 | 时刻 | 引 5651572469? |
说 Prior rulings read? |
提批 #127? |
|---|---|---|---|---|---|
5706769218 |
os-bill | 2026-09-17T00:55:23Z | 否 | 否 | 否 |
5706819210 |
os-sam | 2026-09-17T01:01:39Z | 否 | 否 | 否 |
⇒ 一次 pm:retriage 把已裁过的卡移回 needs-user-decision,没有引用四天前那条裁决;卡于是被当作未裁重新呈报,维护者又裁了一次,且分诊回复里还写着维护者仍欠「one direction」。
代价是实测的,不是假设的
本席 R4 据较新那条裁决派发了 #17518。dev 实测出较新裁决按字面不可执行:它会拒掉已发布函数 composeStacks(stacks, manifest preserve) 今天的输出(仓内夹具 packages/cli/test/fixtures/option-b-collection-zoo.ts 上证明)、与 stack.zod.ts:3760-3763 逐字声明的不变量冲突、且够不到自己的验收标准。⇒ 一整轮派发以零 diff 告终,卡回决策箱等你在两条裁决间选一条。
⛔ 这一次机械面救不了,所以不是「有人忘了跑工具」
check-prior-rulings.mjs 即使跑了也查不出来 —— 5651572469 是卡上的评论,不在它的三条语料里。⇒ 责任不在那两个席位没跑工具;缺的是这条规则的机械半边根本不存在。
⭐ 而该工具的文件头自己记着它诞生的理由:那条搜索「failed twice in four days at the two points where it applies — once at FILING and once at PRESENTATION」(#17009)。⇒ 同一失效方向的第三次,只是换了语料。
可能的形状(⛔ 提请,非处方;归 skills 席裁)
- 要么把
check-prior-rulings的语料扩到被查卡自己的评论线程(找带维护者原话的裁决评论); - 要么在 retriage / 回决策箱的纪律上加一条:把一张卡移回
needs-user-decision之前,先读它自己的裁决史并在评论里点名(有则引,无则明说无)。
SKILL.md 正文(行数棘轮只降不升,进正文得有人付删减的账)。
⛔ 没量的部分
- ⛔ 没量还有多少张卡带着两条以上互相矛盾的裁决。
⚠️ 承接者要估面,candidate population 是「线程里出现过 ≥2 次pm:retriage或多次进出决策箱」的卡。 - ⛔ 没量扩语料在工程上是否便宜(要认「哪条评论是裁决」,而裁决评论没有机器可读的标记 —— 这本身可能是更根上的缺口)。
domain:spec 执行席 4 · 座位贴 #18917 · 两张表的读数为本席 2026-09-18T10:10Z 经 REST 逐条自取
Generated by Claude Code
Activity
os-elon-musk commentedon Sep 18, 2026
Claim: PM loop round 1 (wave 7)
Session:
session_01BTeBejoPUvRHN8WdAJC6oFBranch:
claude/issue-18993-prior-rulings-card-threadWorktree:
objectstack-issue-18993Domain:
domain:skillsSeat:
domain:skills#1File surface:
scripts/pm/check-prior-rulings.mjs— the card's own thread (maintainer ruling comments) joins the corpus thePrior rulings read:line answers for, read over the board route the file already uses for--card, the read's cost measured, a self-test fixture per corpus half, theunresolved/nonesplit kept — and.claude/skills/pm-dispatch/SKILL.md:339 rewritten in place, net 0 (812 / 812, ≤ 120 bytes), so the charter line names exactly what the tool now covers; one PR, one batchContainer & model:
M,mode:subagent,model: claude-fable-5-1(CONTRACT_REVIEW_TIER, path-mandated: SKILL.md isskills/**, #17285); the seat's contract-tier review at delivery; governed ⇒ draft at the four-piece terminalClause-②: no
Thread-read: 5729067026
Serial constraints cleared: SKILL.md 812 / 812 with three unlanded drafts on it — PR #18903 (old-line hunks :509–:525, :633–:675), PR #18921 (:162–:183), PR #18981 (:363–:371) — all disjoint from :339 ⇒ parallel draft, net 0; whichever lands later merges
origin/mainfirst.scripts/pm/check-prior-rulings.mjs: no open PR among the lane's thirteen (file lists read at 2026-09-18T11:05Z); last landedea1da645e(2026-09-13T08:06Z); stem scan oforiginheads (prior-rulings,18993): none; open cards naming the file — #18329 (a handshake spelling, another instrument's defect) and #18466 (repo assumption,pm:blocked) — neither this defect. Lock free.Ruling-ref: none (triage 5729067026: p1 — two closed incident chains on #17518; ⛔ no fix prescribed; the prose and the tool change in ONE batch; the cost of reading the thread is measured, not assumed)
Generated by Claude Code
os-elon-musk commentedon Sep 18, 2026
os-dev-report
{ "issue": 18993, "status": "done", "branch": "claude/issue-18993-prior-rulings-card-thread", "pr": "https://github.com/objectstack-ai/objectstack/pull/19021", "session": "session_01BTeBejoPUvRHN8WdAJC6oF", "premise_still_valid": true, "summary": "check-prior-rulings.mjs now reads the card's own comment thread as a fourth corpus on every --card run (one paged GET by page number to a short page, bounded at 50 pages, reconciled against the card's comments count, on the file's existing proxy re-exec), recognises ruling comments by the SHAPE of their first content line (charter `Ruling: batch #n item k …`, pre-charter `## Ruling recorded — …`, and `Ruling addendum …`), lists two measured near-misses uncounted, and the `Prior rulings read:` line gains a `thread:` clause spelling `2 ruling(s) (ids)` / `none` (read, empty) / `not read (no --card)` apart, with an unreadable thread being the WHOLE line `unresolved` at exit 3. SKILL.md :339 is rewritten in place (net 0, 120 bytes) to 「方向在 ADR/本卡裁决/不可重裁规则」 and decision-analysis.md :45 (the human-facing twin of the line shape, tool docblock names it) gains the clause in place at the 120-byte cap; both files stay at their ceilings. Measured on #17518 the line now names 5651572469 and 5716259259 — the two rulings the retriage never cited. Draft PR #19021, governed (.claude/**), labelled skip-changeset; assignee untouched (was os-elon-musk, the PM's dispatch step 1). Worktree removed after push.", "reader_test": { "before_card_17518": "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 (exit 0, 2,001 ms wall incl. re-exec)", "after_card_17518": "Prior rulings read: … (same ADR half) …; thread: 2 ruling(s) (5651572469, 5716259259) (exit 0, 2,353 ms wall; thread: 17 comments, 1 page, 141,623 bytes, 347–369 ms; 5651572469 os-tesla [ruling-recorded], 5716259259 hotlong [ruling-line]; near-misses 0)", "terms_only": "--terms single,posture,tenant: exit 0, line ends `; thread: not read (no --card)`", "negative_controls": "--card 12708 (director seat post): 372 comments, 4 pages, 2,212,674 bytes, 1,984 ms, thread: none, 0 near-misses; --card 6015 (longest thread on the board): 962 comments, 10 pages, 7,365,556 bytes, 4,431 ms (whole run 6,254 ms), thread: none, 0 near-misses", "shape_census": "2026-09-11T06Z–09-13T21Z (600 comments): `## Ruling recorded` 22 (os-tesla 19, claude[bot] 3), `Ruling:` 4 (os-tesla 2, claude[bot] 2), near-misses 7 (`**Ruling C landing step**` ×6 claude[bot], `## Ruling A's patch is falsified` ×1); 2026-09-18T04Z–11Z (500): `Ruling:` 26 (hotlong 26), `Ruling addendum` 2 (hotlong), `Ruling recorded` 0; quoted 「同意」/「agreed」 63 of 500 (not a face); GOVERNED_APPROVERS-author predicate would have missed every 2026-09-13 ruling incl. 5651572469 — rejected by measurement", "byte_note": "curl reads 156,874 bytes on #17518 vs the tool's 141,623: GitHub pretty-prints for a curl/* User-Agent and minifies otherwise (probed both UAs on the same URL); the tool reports what its own route pays" }, "tests": "node scripts/pm/check-prior-rulings.mjs --self-test: 99 → 144 cases pass, exit 0, live-tree D8 pin `pinned` (new: three shapes, near-misses OUT, first-content-line anchoring, fixture thread 2 rulings oldest-first, none/not-read/unresolved split, offline paged board: 2 pages / 1 page / 403 / no token / page bound / ECONNRESET / malformed page, no-Link-walk structural pin). Gates derived with `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` at cba931995: 40 families; every one run, exit captured redirect-then-$?; `--ran` reconciliation: 40 derived, 40 run, 0 NOT-MEASURED, 0 UNRUN (exit 0). First pass had ONE exit 3: `pnpm --filter @objectstack/lint run check:doc-formula-expressions` PREREQUISITE NOT MET (@objectstack/formula not built) — built formula+lint under scripts/pm/os-verify-lock.sh (VERDICT command-exit 0, held 169 s, waited 0 s), re-run exit 0. `pnpm check:pm-dispatch-gates` exceeds the foreground cap: run detached with recorded PID, waited with tail --pid, verdict `dispatch-gates self-test: 1849 cases pass`, 779.3 s, exit 0. `pnpm check:pm-half-states` (reads this file's source for the H62 writer pin): 4963 cases pass, exit 0. `check:pm-skill-ratchet` ✓ SKILL.md 812/812 headroom 0, decision-analysis.md 50/50, widest rows unchanged (157 self-test cases). `check:pm-skill-id-lint` ✓ 27 files clean. `check:nul-bytes` ✓ 8913 files; unpiped control-char grep over the three files: exit 1, 0 lines (an earlier piped reading was discarded). `node scripts/pm/check-governed-merges.mjs --test` over the three paths: exit 3 GOVERNED, 2 of 3 on the register (.claude/** ×2) — expected. eslint (root flat config, `--no-inline-config`, `--format json`) on scripts/pm/check-prior-rulings.mjs: 1 file, 0 errors, 0 warnings, 0.8 s; population = the config's `**/*.{…,mjs,…}` block which lints scripts/pm; no type-aware rules, so the diff moves no untouched file's verdict — a narrowed measurement, not a full `pnpm lint` (CI-owned). STALE TREE note: origin/main moved to 64de4c576 (5 commits, 4 os-regen artifacts in range) after derivation; the same derivation from a throwaway worktree at origin/main with the three paths yields the identical 40 families (none new, none gone), so no family is owed beyond the 40. CI on PR #19021: in_progress (not waited, per contract). No ablation owed: the change is additive and the self-test's offline board fixtures are the failing-when-wrong evidence (refusal, bound, network, malformed each classify; a shape regex drift reds the verbatim-first-line cases).", "gates": { "derived_at": "cba931995 (branch head) and 64de4c576 (origin/main, throwaway worktree) — identical 40 families", "run": 40, "exit_0": 40, "not_measured": 0, "unrun": 0, "over_cap_detached": "pnpm check:pm-dispatch-gates 779.3 s, 1849 cases pass, exit 0", "prerequisite_cleared": "check:doc-formula-expressions exit 3 → build formula+lint under the verify lock → exit 0", "ci": "in_progress on PR #19021 — not waited" }, "line_budget": { "SKILL.md": "812 / 812 before and after (ratchet headroom 0); :339 120 bytes → 120 bytes, one line rewritten in place, no other line touched", "references/decision-analysis.md": "50 / 50 before and after; :45 113 bytes → 120 bytes (cap 120), one line rewritten in place", "scripts/pm/check-prior-rulings.mjs": "1268 → 1652 lines (+430 / −46)" }, "files_changed": [ "scripts/pm/check-prior-rulings.mjs (+430/−46)", ".claude/skills/pm-dispatch/SKILL.md (1/1, :339)", ".claude/skills/pm-dispatch/references/decision-analysis.md (1/1, :45)" ], "deviations": [ "Edited references/decision-analysis.md :45 in addition to SKILL.md :339 — not in the dispatch's named surface; it states the line's shape for a human (the tool's formatPasteLine docblock names it as the twin), so leaving it would have created the next requirement/tool divergence; no open PR touches it (14 skills-lane PR file lists read).", "A third ruling shape, `Ruling addendum …`, is recognised beyond the two the dispatch hypothesised — 2 measured in the recent window, both hotlong director addenda.", "The dispatch's fallback predicate (every comment by a GOVERNED_APPROVERS login) is not used: measurement shows the director seat wrote as os-tesla, claude[bot] and hotlong within five days, so an author predicate misses 5651572469 itself.", "`--card N --terms …` now reads the card and thread (a board read where there was none before): the thread half is about the card, not the terms; `--terms` alone stays offline and its line says `thread: not read (no --card)`.", "selfTest() is now async (the offline board read awaits a fake fetch); the entrypoint awaits it and the verdict handshake is unchanged.", "origin/main not merged into the branch (5 commits behind at report time): the incoming range carries four merge=os-regen artifacts whose driver debt would hold the push on a PR touching no spec file; the gate derivation was instead repeated from a throwaway worktree at origin/main (identical 40 families). The seat's update-branch is the instrument if a refresh is wanted.", "Full `pnpm lint` not run locally (CI-owned); targeted eslint on the edited script with the three-piece narrowing evidence recorded under tests." ], "mcp_calls": "0 — no MCP GitHub tool called (reads and writes all through the REST proxy with curl / node fetch on --use-env-proxy)", "api_writes": "3 REST writes + 2 git pushes — git push (empty-branch probe, exit 0); git push (2 commits 621fb24d6, cba931995, exit 0); POST /repos/objectstack-ai/objectstack/pulls (draft, 201 → #19021, body read back byte-identical minus the trailing newline, session-URL footer intact); POST /repos/objectstack-ai/objectstack/issues/19021/labels via scripts/pm/label-write.mjs (200; read-back size/m + skip-changeset, MATCHES; independent GET agrees); POST /repos/objectstack-ai/objectstack/issues/18993/comments (this report). Reads only elsewhere: card #18993 + comments, #17518 comments (fixture source, nothing written on #17518), repo-level comments census (11 pages), #6015 / #12708 / #18883 / #18917 / #6024 / #6023 / #5280 / #6026 metadata, #6015 comments (10 pages) ×2, #12708 comments (4 pages), open PR list (2 pages) and 14 PR file lists, director.md commit history; two /search/issues attempts answered 403 (path not allowed for this session) and were not retried.", "open_questions": [], "out_of_scope_findings": [ "noted, not filed: check-prior-rulings.mjs's self-test keeps a flat case list with the verdict handshake and no battery-name floor (AGENTS.md's floor rule is stated for writing a new self-test); carrier: whoever works the 2026-09 self-test shape census — not this card.", "noted, not filed: readCard uses the global fetch while readThread takes an injectable fetchImpl — symmetry only, no behaviour; 承接者:无.", "noted, not filed: GitHub minifies JSON bodies for non-curl User-Agents (~10% fewer bytes) — an observation relevant to any byte-cost reading, not a defect; 承接者:无.", "boundary kept: #18963 / PR #18981 (grading-time dedupe by file) is the same class with a different mechanism and is not addressed here." ] }Generated by Claude Code
os-elon-musk commentedon Sep 18, 2026
ACCEPT — PR #19021 @
cba9319955(two commits: the tool's corpus half, then SKILL.md :339 + decision-analysis.md :45; +432 / −48) · skills seatsession_01BTeBejoPUvRHN8WdAJC6oF· 2026-09-18T12:03ZRead on GitHub: draft, base
main, body line 1Fixes #18993,Clause-②: noat line start; closing keywords touch this card only (closingKeywordTargets={18993 → Fixes});get_files= the three files;skip-changesetpresent (size/m,documentationare the labellers'); report 5729664156 names the head, literal first lineos-dev-report, no HTML comment, 17 keys. Gates: 40 derived / 40 run / 0 NOT-MEASURED atcba931995. Seat spot-check on a scratch worktree at the head:--self-test144 cases (base 99); ratchet 812 / 812 + 50 / 50; id-lint clean; :339 120 bytes; the live reader test on #17518 reads three rulings oldest-first (the third ruled at 2026-09-18T11:42Z, after the dev measured two);--pair 19021exit 0;check-governed-merges.mjs --test: GOVERNED (.claude/**×2). One same-file landing since the claim — PR #18921 on SKILL.md :162–:183, disjoint from :339;originheads on the stem: only this branch. CI at 2026-09-18T12:01Z: 17 success · 11 skipped · 5 in_progress, nothing red.Contract review of record (
CONTRACT_REVIEW_TIER, in seat): PR #19021 comment 5729706403 — VERDICT PASS. The card's own thread is the fourth corpus (shape-recognised, near-misses out, cost metered: 1 page / 0.35 s on the incident card, 10 pages / 4.4 s on the board's longest thread), thePrior rulings read:line carries athread:clause with its three spellings held apart andunresolvedcovering an unreadable thread, and the charter line :339 changes in the same batch — the triage's two boundaries met. Seven declared deviations accepted on the record, one of them a widening the seat agrees with (decision-analysis.md :45, the line shape's human twin).Path face: governed (
.claude/**) ⇒ this PR stays a draft at the four-piece terminal:needs-user-decisionon the PR, the final 维护者速读 posted there, review requested from os-zhuang and hotlong. On an authorized APPROVED the seat lands under ruling C (channel permitting; today the approver has been landing by hand). This card closes on the merge (Fixes).Generated by Claude Code
os-elon-musk commentedon Sep 18, 2026
Landing — PR #19021 @
cba931995entered the merge queue on this seat's record (skills seat, 2026-09-18T23:52Z)Tier S (PR #19144 / #19133 on
mainat1047fe101): every governed path of this PR lies under.claude/**, so the owning seat lands it on its standing contract-tier record — record 5729706403 (PASS at this head, unchanged since),check-clause2-carriers.mjs --pair 19021exit 0 re-run at 2026-09-18T23:49Z, every check on the head green (0 not green). Acts:needs-user-decisioncleared, ready via the CCR route at 2026-09-18T23:50Z, auto-merge SQUASH armed at 2026-09-18T23:51Z,added_to_merge_queueread on the timeline (the positive enqueue reading; the arm call's echo printsmergefor two of the four — the repo allows squash only, recorded as harmless before). The queue'smerge_groupleg judges the record; the seat clears residue on the merge.Generated by Claude Code
os-elon-musk commentedon Sep 19, 2026
Merged — PR #19021 @
cba931995→main8cfa07ab6through the merge queue on this seat's record (skills seat, 2026-09-19T00:16Z)Tier S landing (PR #19144's regime): no approval, the standing contract-tier record 5729706403 read by the queue's
merge_groupguard; merged at 2026-09-19T00:15Z. Residue:pm:dispatched→pm:done, unassigned; serials lifted: #18995 / #19001 re-read their targets onmain. This card closed on the merge.Generated by Claude Code
feat(pm): the Prior rulings read: line covers the card’s own thread —…