docs(agents): AGENTS.md + CLAUDE.md rules only — narratives, dates and quotations out; issue ownership written by role

[claude]

Fixes #15670, fixes #15125, fixes #14913, fixes #14509, fixes #15241, fixes #15357
Part of #15379
#15287 remains open until its rule is readable on main: the ownership paragraph in this PR is now its only carrier (PR #15290 was closed by the maintainer as superseded); its three criteria are rows 30a–30c of the inventory below.

Family member 5 of the rules-only rewrite (maintainer ruling recorded on #15379: 「各条规则的出处叙事、事故复盘 根本不重要啊，不需要写入skills啊」). AGENTS.md and CLAUDE.md keep every rule as one executable sentence in its own section and its own register; every incident retelling, provenance story, ruling date, verbatim quotation and "why we learned this" leaves. The ownership paragraph is rewritten by role under the 2026-09-05 ruling on issue ownership (「待你定 1（情形 C）: agent 无法区分是你还是 PM 可以接受」 / 「待你定 2（情形 H 与配额）: 暂时不区分」). Six cards fold in as one rule each. Ceilings move down only.

Governed surfaces (AGENTS.md, CLAUDE.md, scripts/pm/**) — draft PR, human merge; skip-changeset (nothing publishes). No model identifier appears in the diff, this body, the commit or the comments.

Census (before → after, measured on head 9a3a44190d)

reading	AGENTS.md	CLAUDE.md
lines	1161 → 1058 (ceiling 1162 → 1058)	36 → 41 (ceiling 86 → 41)
bytes	84947 → 75258	2095 → 2618
widest table row (bytes)	1081 (translations row) → 768 (same row, narrative out; pin 1081 → 768)	0 → 0
widest non-table line (bytes, LC_ALL=C awk)	365 — line 8, the untouched v5.0 rename blockquote (exempt shape); every prose line ≤ 120 bytes by the ratchet's own classifier: 0 offenders	97
lines carrying a date stamp	21 → 0	0 → 0
「」 ruling quotations	5 → 0	0 → 0
⚠️ / ⭐ markers	10 / 0 → 0 / 0	0
⛔ rules	39 → 47 (34 ⛔-bearing lines → 40; every original ⛔ re-stated, see row 20 for the one retired with its object)	4 → 6
#NNN issue citations	0 → 0 (check:pm-skill-id-lint green)	0 → 0
headings	20 → 20, byte-identical (## Communication · ## Build & Test · ### A test that reads outside its own package … · ### Running the dev server · ### Frontend (Studio UI) … · ## Prime Directives · ## Multi-agent working discipline · ## Monorepo Layout · ## Protocol Domains … · ## Kernel · ## Documentation Guardrails · ### Touched packages/spec? … · ## Context Routing … · ## Skills (skills/) · ## Patterns · ## Route & surface ownership · ## Degradation log levels … · ## Startup registry reads … · ## Post-Task Checklist · ## Edit Sizing)	4 ⛔ excerpts kept, same subjects

Byte-identical regions: the RECOGNISED_PATH_SPELLINGS mirror block (check-published-list-mirrors green), the Monorepo Layout / Protocol Domains / Kernel / Context Routing tables, the dev-server and spec-artifact tables, every code fence, the attribution-footer block, and the four-axes frame block appears in neither file (check:skill-frame-sync green, 2 declared copies).

Where each folded rule landed

card	rule, as landed	file:line
#15670 + #15287 (R1)	the ownership paragraph by role — assignee = presence bit, Claim: = identity, both written by the owning session, a dispatched executor inherits and verifies, never writes the field, never yields; release is an explicit act; findings filed unassigned; the maintainer's own work is an accepted blind spot; no per-seat identities	AGENTS.md:393-410; one-paragraph form CLAUDE.md:7-17
#15125	the model-free agent commit trailer pair (Claude-Session: session URL + the Co-authored-by: Claude noreply line); no model identifier in any content artefact; the harness-written co-author trailer is the one exemption	AGENTS.md:440-444
#14913	the provenance sentence in the rules-only form (no dates, no quotations, no issue numbers; provenance lives in the landing PR)	AGENTS.md:10-15; the second site, the check-skill-id-lint.mjs header sentence, at scripts/pm/check-skill-id-lint.mjs:18-19 (pattern, roots and verdicts untouched — self-test 14 cases pass, live 24 files clean). The card's third limb, references/rest-channel.md, is member 3's file (PR #15641) and is not touched here
#14509	an ADR lives in the repository whose code it governs; cloud records cited as cloud ADR-NNNN, ⛔ never bare; a cloud decision whose mechanism half governs open code gets a local ADR with a ## Provenance section; files never move, numbers never reassigned	AGENTS.md:249-254, inside Prime Directive #13 (no new heading; the provenance line lives here, not in the file)
#15241	the PR-body footer rule as measured in this repository: the session-URL footer survives an edit and the platform appends a bare block — send an edited body with no footer, read it back, ⛔ never re-send a body that already carries the appended footer; the "normalises DOWN to bare" sentence and its date are gone, and the claim about which footer survives is stated once, with the repository named	AGENTS.md:428-439
#15357	a reading you cite is a count plus the repository and commit it was taken against, or it is not a reading; "measured over the corpus" names examples/app-crm at its sha, never an application it is not; a moving ref is not an anchor, for a corpus exactly as for a git diff (the same line as #15647 section 2's second principle, written once)	AGENTS.md:354-357, beside the BASE/moving-ref practices

Rule inventory — AGENTS.md (one row per rule paragraph; "operative content unchanged" unless the row says otherwise)

#	section · lines now	rule as it now reads (lead)	what left
1	header · 1–8	file role, AGENTS.md wins, v5.0 rename	nothing (byte-identical)
2	header · 10–15	rules only; no narrative, date, quotation or issue number; gate; provenance in the landing PR; script headers are the authority	the 2026-08-12 ruling quote and "rulings keep their date and verbatim quote" — content changed on purpose: #14913
3	Communication · 19–31	Chinese in chat with the maintainer; English on GitHub; quoted Chinese rulings untranslated; repo artefacts keep their language; one rule per channel	the 2026-08-08 quotation block and the "a merged rule was measured to fail" story
4	Build & Test · 35–42	the command block	nothing
5	Build & Test · 44–49	pnpm lint is the only style authority; ⛔ never prettier --write; --check is no second opinion	the 2026-09-01 measurement (--file-info, --find-config-path, prettier 3.8.1)
6	Build & Test · 51–52	⛔ push a WIP commit before every step that takes minutes	"lost 2,081 lines"
7	Build & Test · 54–60	type-check coverage/debt ratchets and their four invariants	nothing
8	Build & Test · 62–80	the three tsconfig/@ts-expect-error/TS7006 principles	"worth knowing before you fight them", one parenthetical aside
9	cross-package test inputs · 82–90	what the gate keeps equal, why a source scan sees only known spellings, seed forms	the "on main, with every PR reporting green" and "a detector with no dependencies" rationale
10	cross-package test inputs · 92–110	the published spellings block	nothing (byte-identical; mirror gate green)
11	cross-package test inputs · 112–120	mirror gate holds the block; extend detector + self-test + block in one edit, ⛔ never route around; the two unflagged shapes; judged on the shallowest point	"an unseen read is the defect above…", the sibling-descent example
12	cross-package test inputs · 122–123	clocked windows measure behaviour, never loading	nothing
13	dev server · 125–145	table, --fresh, the port/kill/script rules, the -- separator rule	"(measured; …)", "(other agents/the user may be using it)"
14	Frontend · 147–168	backend-only, refresh workflow, ⛔ never hand-edit dist/cache, pin's second half, fast iteration	the ⚠️ marker (now ⛔)
15	PD 1–9 · 172–210	unchanged	nothing
16	PD 10 · 211–216	file findings; never advertise an undelivered capability; trim / implement / keep the claim as narrow as enforcement; a case label is not enforcement, check every call site incl. bulk	the write-path-validator incident; gh issue create spelling (not every seat has gh)
17	PD 11 · 217–229	worktree-first, both hooks, override, precision-first Bash guard, stash/cache not isolated	"(it still gets switched under you)"
18	PD 12 · 230–241	contract-first: fix at the producer, reject at publish, ⛔ never a lenient alias/?? in a consumer, internal contract, change the spec deliberately, ADR-0087 entry, stored rows replay the full chain	the cfg.filter ?? cfg.filters history and the worked example; "do not" → ⛔
19	PD 13 · 242–248	an accepted ADR binds until superseded; grep the ADRs; leave the id in code; one anchor JSON per path, no index	the "three ADRs reversed by a patch changeset" incident
19b	PD 13 · 249–254	new — #14509 (see the fold table)	—
20	PD 14 · 255–288	⛔ governed surfaces confirmed and merged by the maintainer; the five surfaces named; register + prose gate + print command; authoring open, landing reserved (four ⛔); file list not description; approval path; byte-exact regen clearance; ⛔ no agent approval; draft-back procedure; the barrier is the directive, ⛔ not a relaxation; prevention/detection mechanisms; violation handling; no exception	the three verbatim rulings, the 2026-08-27/09-04 approver-set history and quotation, the 2026-09-01 date, the "drafted ADR merged twice" incident, the audit-header pointer. Retired with its object: "⛔ never edit the quotation above to satisfy a gate" — no quotation remains; the surfaces are now editable prose and the gate's own failure text still carries the prohibition
21	PD 15 · 290–310	⛔ releases are the maintainer's; release-adjacent work open; every reserved act (seven ⛔); Version Packages PR facts; ⛔ no counter-publish; environment: release mechanism; merging that PR queues a deployment in the maintainer's name	the 2026-08-07 quotation and 「这个要写入规范」, the "69 packages twice in one week" incident, the workflow_dispatch-barrier history, "since ADR-0125 (2026-08-20)"
22	Multi-agent · 313–316	intro	nothing
23	Multi-agent · 318–334	⛔ never bare git stash push/pop; shared stack; hook behaviour; replacements block	"it has really happened, mid reverse-verification"
24	Multi-agent · 336–352	⛔ the worktree isolates four ref namespaces only; shared refs/remotes/*; FETCH_HEAD per checkout, fails by absence; the four practices; ⛔ no hook backs the moving-ref half	"(measured next door: three moves in ten minutes)", "measured on git 2.43", "exactly backwards"
24b	Multi-agent · 354–357	new — #15357 (see the fold table)	—
25	Multi-agent · 359–362	⛔ nor the build cache	nothing
26	Multi-agent · 364–371	reverse verification: commit first; prove identity by hash	"has cost changes repeatedly"
27	Multi-agent · 373–385	⛔ git checkout REF -- PATH stages; restore forms; ⛔ never trust git diff HEAD alone	"(one real PR: 157 such deletions…)"
28	Multi-agent · 387–391	windowed history goes through git-history.mjs	nothing
29	Multi-agent · 393–410	the ownership paragraph, by role — (a) shared identity, assignee = presence bit, Claim: = identity record with its first-line spelling; (b) the owning session writes both: a self-picking seat takes only an unassigned card, assigns and claims first; a PM dispatch sets both on the dev's behalf; (c) a dispatched executor inherits both records — the assignee (step 1) and the dispatch's Claim: naming its branch (step 2) — verifies the newest Claim: names its branch (mismatch ⇒ stop and report), posts NO second claim (the dispatch's Claim: is its identity; its own record is the report comment); ⛔ never writes the assignee, ⛔ never yields; (d) re-read the comments — the comments decide: a Claim: from another session or branch means taken, ⛔ never reassign; a bare assignee with no Claim: is a dispatch's step 1, not a foreign claim; (e) release is an explicit act — clear the assignee, post a Release: line; (f) findings filed unassigned, assign when you start; (g) the maintainer's own work carries no claim and is indistinguishable from a dispatch — accepted; no per-seat identities	"two agents burn the same hours twice", "implemented twice in one morning", the gh issue edit spelling. Content changed on purpose: the by-role ruling. Agrees with pm-dispatch SKILL.md 认领 (assignee 字段归 PM; a Claim: from a session other than the PM's reads as a foreign claim) and os-dev.md rule 2 (assignee 归 PM; the PM's claim satisfies claim-first for the dev); quotes neither
30a	↳ #15287 criterion 1	"A dispatched executor inherits that assignee — the dispatch's step 1, already done"	—
30b	↳ #15287 criterion 2	satisfied at the owning-session level: "a PM dispatch sets the assignee (step 1) and posts the Claim: naming the dev's branch (step 2), both on the dev's behalf" — the executor "posts no second claim — the dispatch's Claim: is its identity and its own record is the report comment"	—
30c	↳ #15287 criterion 3	"the comments decide, not the field: a Claim: from another session or branch (other than the dispatch that sent you) means taken whatever the field says … a bare assignee with no Claim: under it is a dispatch's step 1, not a foreign claim — not taken for the executor it was dispatched to, and not free for a seat picking for itself"	—
31	Multi-agent · 412–418	state you did not set belongs to another actor; a rewritten body is evidence about the body only, ⛔ never extend it to the draft flag; read the timeline actor or ask	the worked draft-flip failure
32	Multi-agent · 420–426	the attribution footer block	nothing (byte-identical)
33	Multi-agent · 428–439	PR body and issue comment footer rules — #15241 form (see the fold table); comment rules unchanged	the 2026-08-26 / 08-20 / 09-03 measurements and probe counts; "how a body comes back in a shape nobody typed"
33b	Multi-agent · 440–444	new — #15125 (see the fold table)	—
34	Multi-agent · 446–452	GitHub mutates body bytes: spell poison tokens in words; read back; ⛔ don't "repair" a short API reading	nothing
35	Multi-agent 1–6 · 454–474	unchanged	nothing
36	Multi-agent 7 · 475–516	merge queue: arm only green + accepted; ⛔ two classes never enter; green = success; the six required contexts by name; re-arm awareness; stale red; gh pr merge unmeasured; fallback	"supersedes the older blanket auto-merge ban", "by maintainer decision (2026-08-07, extended 2026-08-09)", "much longer than docs/adr/** and it grows fast"
37	Multi-agent 8 · 517–521	own temporary server, never stop someone else's	nothing
38	Multi-agent 9 · 522–543	refresh build state after pulling main; the stale-artefact table; the prescription; check:dev-prereqs	"which is why it is worth recognising in one step rather than re-diagnosing per gate"
39	Multi-agent 10 · 544–560	clean merge ≠ working merge; full suite before the PR; scoped re-checks	"~15 minutes per lap while main lands a PR every few", "the guards exist largely because…"
40	Multi-agent 11 · 561–592	generated artifacts: merge driver, deferral, ⛔ never --no-verify, sharding, per-clone registration, excluded baselines	"(maintainer ruling 2026-08-12)", "the old spelling"
41	Layout / Domains / Kernel · 595–655	tables	nothing (byte-identical)
42	Documentation Guardrails · 657–664	the four rows; content/docs/releases/ RELEASE-OWNED (audit-scope gate green); translations regeneration rules	releases row: "hottest conflict magnet (three PRs raced…)"; translations row: "a stale bundle entry always beat an edited source string…" (1081 → 768 bytes)
43	Touched packages/spec · 666–738	eight artifacts, one job, sequential; the gate table; the three commands; the five wrapper principles; declaration-parity reads two declarations, exits 1 without a manifest, manifest from the pin bump, ⛔ no skip, no workflow wiring; two ungated generators	"one red build per artifact", "which is precisely what CI cannot do", "(dead blocks have shipped…)", "settled (maintainer ruling 2026-08-07)", ⚠️ markers
44	Context Routing · 741–764	table	nothing (byte-identical — PR #15427's pending row would apply cleanly)
45	Skills · 767–779	filesystem is the catalog; internal entries; ⛔ both roots governed	"(two such lists drifted stale as skills landed)"
46	Patterns · 781–806	Zod and Plugin patterns; no onInstall/onEnable/onDisable hooks	"an older example … was retired for it"
47	Route & surface ownership · 808–878	the five rules; closed query-parameter set adopted incrementally, ⛔ never as one big batch; the three measurement rules (read points incl. helpers, both pins, open routes excluded by name, per-lane conversion); recognition before arity; verify on the provisioned composition	"each paid for by a real bug", the retired-duplicate re-learning story, the 2026-08-12 quotation, "all of them measured rather than assumed", the localised-export 400 narrative (helper rule kept with extractLocale), "the failure mode the ruling rejected", "answered wrongly three times"
48	Degradation log levels · 880–957	the one question; functional → warn, durability → error (prefer failing to falling back, say what was lost); an error owes consequence + fix, once; do not over-apply; caller-propagation declared, not logged; both sets of teeth with their vocabularies and ledgers; narrow scan surface, widening is its own issue	the founding suspended-run-store incident, "which is what made the founding incident's warn unreadable", the keystroke example, the 2026-08-06 「裁 3 —— 收窄先行」 quotation
49	Startup registry reads · 959–1004	the question; three parts; read-only probe legal; three cures; teeth	the three-instance showcase cold-start incident
50	Post-Task Checklist · 1006–1054	seven steps incl. ⛔ governed exception, changeset rules, ADR-0087 marker, sibling pin, spec artifacts	"Why it is asked of you at all:" lead
51	Edit Sizing · 1056–1058	~20KB payloads	nothing

CLAUDE.md

lines	excerpt	change
1–5, 19–41	intro, worktree-first, never stash, never edit releases, outro	untouched — four ⛔ subjects kept, heading pointers unchanged
7–17	⛔ Claim the issue before you write any code	the one-paragraph form of the by-role rule (assignee = presence bit, Claim: = identity; self-picking seat assigns + claims first; PM dispatch sets the assignee and posts the Claim: for its dev, who inherits both, checks the newest Claim: names its branch, posts no second claim, ⛔ never writes the assignee or yields; comments decide, ⛔ never reassign; bare assignee = a dispatch's first step); still points at AGENTS.md → Multi-agent working discipline, paragraph Claim the issue BEFORE you write any code.

Scripts

• scripts/pm/check-skill-line-ratchet.mjs: the family's re-pin convention — own rows only: AGENTS.md 1162 → 1058 and 1081 → 768, CLAUDE.md 86 → 41 (its 0 table pin untouched), one comment each; no other row moved, no raise. Self-test: ✓ check-skill-line-ratchet self-test: 155 cases pass.
• scripts/pm/check-skill-id-lint.mjs: one header sentence ([finding] After ruling B (#14859), AGENTS.md:12-13 and the scripts/pm/check-skill-id-lint.mjs header still describe the pm-dispatch provenance form as "date + verbatim quote" — the two off-surface sentences PR #14911 could not touch #14913's second site) — the bounded in-place fix the card's disposition assigns to this member: same defect class, mechanical, the form pinned by the card; the gate's pattern, roots and verdicts untouched (--self-test 14 cases pass; live 24 file(s) clean).

Verification (head 13fc873b29 = my commit a52aceadb1 + origin/main 5a21d73af5 merged; exit codes captured by redirect, verdict lines quoted from the gates)

• node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (change set from the merge base, 4 paths) → 36 commands, every one run, exit 0; --ran reconcile: ✓ dispatch-gates --ran: 36 derived famil(ies) accounted for — 36 run, 0 NOT-MEASURED.
• node scripts/pm/check-skill-line-ratchet.mjs --self-test → 155 cases pass; live → ✓ AGENTS.md is 1058 lines (ceiling 1058; headroom 0) · ✓ AGENTS.md: widest table row is 768 bytes (pin 768; headroom 0) · ✓ CLAUDE.md is 41 lines (ceiling 41; headroom 0) · ✓ CLAUDE.md: widest table row is 0 bytes (pin 0; headroom 0).
• pnpm check:pm-governed-prose → ✓ 2 instruction surface(s) name all 5 registered governed surfaces (docs/adr/** · .claude/** · skills/** · AGENTS.md · CLAUDE.md) and claim no others.
• pnpm check:required-contexts → ✓ 6 required context name(s) pinned across 2 workflow(s); 6 instruction surface(s) scanned (all six names kept verbatim in Multi-agent §7).
• pnpm check:skill-frame-sync → ✓ 2 copies of the decision frame are structurally isomorphic across 2 files … 73 markdown files scanned for undeclared copies.
• pnpm check:pm-skill-id-lint → ✓ 24 file(s) clean; node scripts/check-published-list-mirrors.mjs (+ --self-test, 37 cases) → exit 0; pnpm check:docs-audit-scope → ✓ release-owned pages are in scope and read-only; pnpm check:agent-test-spelling → 0 violations; pnpm check:nul-bytes → exit 0; pnpm check:pm-governed-merges → exit 0; pnpm check:watch-hint-literal, check:cli-command-ids, check:cross-package-test-inputs, check:entry-guard, check:parse-guard, check:pnpm-filter-targets, check:ratchet-remedy-authority, check:refd-timer-probe, check:bash32-floor, check:driver-memory-census, check:pm-dispatch-gates and the direct-node families → all exit 0.
• All five .claude/hooks/*.selftest.sh → 0 failed (50 / 121 / 116 / 51 / 36 passed): no pinned spelling touched.
• Whole-repo pnpm lint through scripts/pm/os-verify-lock.sh (slot issue-15379-m5) → VERDICT command-exit 0 · held the lock 69s.
• Roster gates flagged beside the scripts/ paths, run for completeness: pm-label-desc-cap, console-injection, i18n-stale-fill, engine-double-contract exit 0; dts-closure and published-readme-exports print PREREQUISITE NOT MET (no dist/ in this worktree) — NOT MEASURED, not red; they read nothing this PR changes.
• git merge-tree --write-tree --name-only origin/main HEAD before the push: clean. PR docs(agents): state the app-vs-platform boundary once, so it stops being re-derived #15427 (one Context Routing row + one CEILINGS row) has not landed; its hunks touch lines this PR leaves byte-identical, so it re-resolves trivially on top of this.

---

Generated by Claude Code

[os-steve]

ACCEPT — contract-tier in-seat review of PR #15856, head 9a3a44190 (skills seat, 2026-09-05T12:0xZ). Family member 5 of #15379.

Implemented-by: claude/issue-15379-agents-claude-rules-only (os-dev subagent)
Reviewed-by: session_019RfFHiRCSs3JXLK4cwcfox (skills seat)

Verified in a detached compare worktree at the PR head, three-dot against origin/main, not from the report: four files — AGENTS.md 1161 → 1058 lines with the widest table row 1081 → 768 bytes, CLAUDE.md 36 → 41, scripts/pm/check-skill-line-ratchet.mjs (only the two files' rows re-pinned, down-only, at the landed counts), scripts/pm/check-skill-id-lint.mjs (#14913's header sentence). Rules only, as the family's ruling requires: date-stamped sentences 21 → 0, ruling quotations 5 → 0, warning markers 10 → 0, ⛔ rules 39 → 47 with every original re-stated, the twenty headings byte-identical. The ownership paragraph (AGENTS.md 393–410, CLAUDE.md 9–17) is written by role under the maintainer's 2026-09-05 ruling: the assignee field a presence bit and the Claim: comment the identity; the owning session writes both — a self-picking seat assigns itself and claims before any other action, a PM dispatch sets the assignee and posts the claim for its dev; the dispatched executor inherits both, verifies the newest Claim: names its branch, posts no second claim (the dispatch's claim is its identity, its record is the report comment — the seat's correction on 9a3a44190, which is what os-dev.md line 35 and the SKILL.md line 81 reading already say), never writes the assignee and never yields; the comments decide, a bare assignee with no Claim: under it is a dispatch's step 1 and not a foreign claim; release is an explicit act; findings are filed unassigned; the maintainer's own work is an accepted blind spot and per-seat identities are not introduced. hotlong's three #15287 criteria are present as rules. Six folds, one rule each: #15125 (440–444), #14913 (10–15 plus the id-lint header), #14509 (249–254), #15241 (428–439, in the measured form), #15357 (354–357), #15670 (the paragraph). PR #15427's Context Routing row and its ratchet row are byte-identical here, and merge-tree against main is clean, so that PR re-resolves trivially whichever lands second. No model names, no dates; governed on both surfaces (check-governed-merges --test exit 3); the seat chose the ceiling model for this paragraph at its discretion.

Seat gate runs on this head in the compare worktree: skill-line-ratchet exit 0 at the re-pinned ceilings (1058 / 768, 41 / 0), governed-prose exit 0 (two instruction surfaces name all five governed surfaces), governed exit 3; frame-sync and required-contexts are NOT MEASURED locally (the compare worktree has no installed tree) — the evidence is the dev's runs (two frame copies, 73 files scanned; six required contexts pinned) and CI. The dev's evidence beyond that: 36 derived families all exit 0 and reconciled 36 of 36 on both heads, the five hook self-tests 0 failed (no pinned spelling touched), whole-repo lint through the lock, and CI on 9a3a44190: 33 check runs completed, 25 success, 8 skipped, none failing (seat's read at 12:02Z).

Open question ruled B: the seat rewrote the body's first line to Fixes #15670, fixes #15125, fixes #14913, fixes #14509, fixes #15241, fixes #15357 (read back; the rest of the body byte-identical up to the platform footer) so all six close on merge; Part of #15379 stays on line 2; #15287 is named as carried and stays open until the paragraph is readable on main, when the devx lane closes it. The dev's follow-up #15857 (os-dev.md still states the pre-#15241 footer reading) is graded by the seat and waits behind this PR.

Landing regime: governed surface ⇒ this PR stays a draft; reviewers os-zhuang and hotlong requested by the seat; the human merge is the review record; no seat flips it ready, enqueues it, or arms auto-merge.

---

Generated by Claude Code

[os-steve]

维护者速读(skills 席,2026-09-05T13:2xZ)

改了什么:把 AGENTS.md(所有 agent 的总规则)从 1161 行压到 1058 行、CLAUDE.md 只留规则——事故复盘、裁定日期、原话引用全部删掉,每条规则一句话,20 个标题字节不变。同时按你今天的裁定重写「issue 归谁」这一段:指派字段只表示「有人占着」,认领评论才是身份;PM 派发时替 dev 写好这两样;dev 不再自己发第二条认领;离手必须显式释放;你本人的卡与 PM 派的分不出来,已接受为盲区。顺手修掉六条散落的规则错误(提交 trailer、过期出处、ADR 登记规则、PR 页脚规则、「测过语料」要带仓 + commit、指派矛盾),并了结 hotlong 的 #15287(其 PR #15290 已按你指示关闭)。

为什么改:你 9-4 裁定「出处叙事、事故复盘根本不重要,不需要写入 skills」;今天三个开发席撞上「已指派 = 别人的」与「dev 永不写指派」的矛盾,浪费飞行。

风险与代价:这是 agent 的宪法,改动面大(+288 / −377),但 ⛔ 规则 39 → 47 无一丢失、五个 hook 自测无一红、所有门禁与 CI 绿;席位预读抓出并改掉了一处与 os-dev.md 冲突的句子。落地后 hotlong 的 #15427 要在其上重新解一次冲突(两处字节相同,已通知,属小事)。可回滚:一次 revert。

席位意见:建议合并。

你要做的:批准 + 入队(或让 os-zhuang 代做)。

---

Generated by Claude Code