[PM corpus] rules only, plain language: strip provenance narratives and incident post-mortems from the pm-dispatch skill corpus (maintainer 2026-09-04) — member 1: SKILL.md

[claude]

Filed by the domain:skills seat (session session_019RfFHiRCSs3JXLK4cwcfox, os-steve) on a maintainer ruling given in the live PM chat, 2026-09-04 13:1xZ. Execution family under the corpus-audit anchor #13597 (Phase 3 continues under this ruling, which supersedes the narrower X0 quote policy where the two differ).

The ruling (verbatim, untranslated)

The maintainer asked whether the PM skills should be re-reviewed for readability — 「要不要把所有的项目经理 skills 再复核一遍，语言改的通顺一点，便于人类审核，目前我看起来好累」 — and, on the seat's core-vs-secondary reading, ruled:

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

⇒ Provenance narratives and incident post-mortems do not belong in the skill corpus at all. The corpus carries rules only, in plain language a human can review.

What this changes

• Every rule keeps its operative content (what to do, what never to do, the thresholds and names it depends on). What goes: who said it when, which incident produced it, which day it was measured, the "this is why" story — unless the story IS the rule's condition (then one clause, no date, no quote).
• A maintainer quotation stays only where the quoted words are themselves the criterion the reader tests against (the four-axes frame block is one such: it is watched byte-for-byte by check:skill-frame-sync and is ⛔ untouched). Everything else quoted becomes the rule in plain words.
• Language: each file keeps its current language (the pm-dispatch corpus and os-dev.md are Chinese with English identifiers; AGENTS.md is English). One rule per sentence; ⛔ only for prohibitions; no ⚠️/⭐ emphasis; no issue numbers in operational text (check:pm-skill-id-lint).
• Net shrink only — every ratchet ceiling stays where it is; the rewrite never buys lines. The seat re-locks ceilings once after the family lands (programme rule).

Members (one governed PR each, in this order; the seat dispatches serially)

#	file(s)	size today (a3fbd200)	flight
1	.claude/skills/pm-dispatch/SKILL.md	971 lines · 142 ⛔ · 124 quotes · 97 dates	this card
2	references/dispatch-runbook.md · review-checklist.md · contract-review.md · landing-operations.md · seat-post-protocol.md · state-machine.md · decision-analysis.md	≈720 lines	member card
3	references/platform-readings.md · rest-channel.md · lanes/*.md · release-aftercare.md · true-green.md · compile-surfaces.md	≈900 lines	member card (readings ledger: facts stay, narratives go)
4	.claude/agents/os-dev.md	450 lines · 41 ⛔ · 53 quotes	member card
5	AGENTS.md	1161 lines (English)	member card, last

Human review, by design

Each PR body carries a rule inventory — one row per rule: section · rule as it now reads · what was removed (narrative / quote / date) · "operative content unchanged" — so the maintainer reviews the table, not the diff. The seat's contract-tier review checks the inventory against the diff before requesting the two approvers.

Gates every member must keep green

check:pm-skill-ratchet (line count and 120-byte lines; shrink only) · check:skill-frame-sync (the frame block byte-identical in its two copies) · check:pm-governed-prose (the instruction surfaces still name the five governed surfaces) · check:pm-skill-id-lint · check:nul-bytes · check:pm-governed-merges · the whole-repo lint · whatever dispatch-gates --commands derives for the paths.

Refs: #13597 (anchor; Phase 1–2 tables and rulings) · #14685 items 2–5 (the earlier X0 / excerpt / frame / move rulings) · #12911 (the thesis).

---

Generated by Claude Code

[claude]

Step 1 re-ordered on the maintainer's word — the digest first. In the same chat, on the seat's three options (A: a plain-language 《关键条款摘要》 the maintainer reviews instead of the corpus, then style-only rewrites; B: restructure core vs detail; C: delete constructs), the maintainer ruled, verbatim: 「同意先做 A 的摘要」.

So member 1 of this family is now the digest, not the SKILL.md rewrite: one new reference file, .claude/skills/pm-dispatch/references/core-rules.md — every binding rule of the PM corpus (SKILL.md + references + os-dev.md's PM-facing clauses) as one plain Chinese sentence each, grouped by the corpus's own sections, ≤150 lines, no dates, no quotations, no incident stories, ⛔ only for prohibitions; the PR body carries the rule inventory (digest line → source section) so the maintainer reviews the digest and the seat verifies completeness against the corpus. The SKILL.md rewrite (rules only, matching the approved digest) becomes member 1b and waits for the digest's approval.

Claim: PM loop round 5
Session: session_019RfFHiRCSs3JXLK4cwcfox
Branch: claude/issue-15379-pm-core-rules-digest
Worktree: objectstack-issue-15379
Domain: domain:skills
File surface: .claude/skills/pm-dispatch/references/core-rules.md (new) + the ceiling row for it in scripts/pm/check-skill-line-ratchet.mjs pinned at the landed count (a split adds a row, never edits another). ⛔ No edit to SKILL.md, the other references, os-dev.md or AGENTS.md; ⛔ the four-axes frame's fingerprint phrases never appear in the digest (it refers to the frame by section, check:skill-frame-sync stays at two copies).
Container & model: L (a 4,300-line read, a 150-line write), mode:subagent, model: opus
Clause-②: no (a human-facing digest; the protocol's semantics are unchanged)
Serial constraints cleared: no open PR touches references/; PR #15340 (AGENTS.md) and #15355 (os-dev.md) are landing and touch neither the digest nor its sources' rules. Governed (.claude/**) ⇒ draft PR, in-seat completeness review, both approvers requested by the seat, human merge.

Decision re-read (13:33Z): the ruling above plus the 13:1xZ ruling (rules only, no narratives) are the whole mandate; no further ruling owed before the digest exists.

---

Generated by Claude Code

[claude]

os-dev-report

{
  "issue": 15379,
  "status": "done",
  "branch": "claude/issue-15379-pm-core-rules-digest",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/15402",
  "premise_still_valid": true,
  "summary": "Added one new reference file, .claude/skills/pm-dispatch/references/core-rules.md: 150 lines carrying 134 binding rules, one rule per line, grouped under the corpus's own 13 sections, every line within the 120-byte cap (widest 120, mean 97.1), with no date, no quotation, no issue number and no marker other than the prohibition sign. Sources read in full: SKILL.md (971 lines), all 21 files under references/ including lanes (1,659 lines), and the PM-facing clauses of .claude/agents/os-dev.md (450 lines) — about 3,080 lines. Roughly 2,930 lines are not represented, in nine declared classes listed in the PR body's second table: provenance and ruling metadata; incident narrative and rationale tails; measured platform facts (platform-readings, the rest-channel per-operation table, the true-green run recipes); lookup tables and rosters; templates and forms; per-lane job descriptions; intra-corpus duplicates; retired tombstones; and the dev agent's own working rules, which the card puts out of scope. Two rule-level details were folded out for space and are named in that table rather than left silent. The escalation decision frame is referred to by section with its axes named in short words, and check:skill-frame-sync still reports exactly two copies. The ratchet map gains one CEILINGS row at the landed count of 150 plus one max-table-row pin of 0; no other row moves. On the premise: the card body still describes member 1 as the SKILL.md rewrite, while the claim comment re-orders member 1 to this digest on the second ruling — the premise holds as claimed, not as the card body reads.",
  "tests": "All readings taken on the final commit 3345687b; the whole union was re-run after that commit and every exit code captured by redirect before any pipe. Reconciliation: node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --ran RANFILE printed 'dispatch-gates --ran: 28 derived famil(ies) accounted for — 28 run, 0 NOT-MEASURED.' (exit 0), plus nine families run beyond the derivation. Ratchet, the new file: 'check-skill-line-ratchet: .claude/skills/pm-dispatch/references/core-rules.md is 150 lines (ceiling 150; headroom 0).' and 'widest table row is 0 bytes (pin 0; headroom 0).'; map self-test 'check-skill-line-ratchet self-test: 155 cases pass.' (exit 0). Frame sync: pnpm check:skill-frame-sync exit 0 — two copies, untouched by this diff, and the digest carries no fingerprint phrase. Id lint: pnpm check:pm-skill-id-lint exit 0. NUL bytes: 'check-nul-bytes: OK (scanned 7460 text file(s) ... no raw ASCII control bytes).' Governed predicate: node scripts/pm/check-governed-merges.mjs --test on both paths exits 3 with 'governed-surface predicate: 1 of 2 path(s) hit the register' and 'GOVERNED — a human merge is the review record for this PR'. Whole-repo lint through the shared verification lock (OS_VERIFY_LOCK_SLOT=issue-15379-heavy): 'os-verify-lock: VERDICT command-exit 0 · held the lock 77s (1m17s) · waited 0s · SHARED-BOX SECONDS'. check:doc-formula-expressions first returned exit 3, PREREQUISITE NOT MET, so nothing was measured; after two workspace builds through the same lock (@objectstack/formula then @objectstack/lint, build verdict 'command-exit 0 · held the lock 155s') it passed: 'check:doc-formula-expressions: 22 record-scoped formula example(s) across 428 files / 1371 TS blocks judged clean' (exit 0). No ablation applies — the change is prose plus one ratchet row, and that row is exercised by the map self-test above. CI was not waited on: the report lands at draft-PR time, and PR gate status is in_progress.",
  "mcp_calls": "0 — every GitHub read and write went through repo-scoped REST with the container token after a probe; zero MCP GitHub calls",
  "open_questions": [],
  "out_of_scope_findings": [
    "not filed as new cards, per the card's own stop condition: two corpus passages state the same rule with different names, and both are reported in the PR body's third table for the seat to raise — (1) whether a single-lane repo seat may produce the issue type and the first grading of a finding (SKILL.md's multi-repo preamble and rule 4 grant it; references/lanes/hotcrm.md withholds it at three sites; the digest takes the stricter, later SKILL.md reading), and (2) who releases a clause-two contract PR after review (SKILL.md's ACCEPT fork and references/contract-review.md put it in the dispatching seat's in-seat review, while references/lanes/director.md still frames the director duty around it). Both are already recorded as correction-class rows in the corpus audit's Phase-1 tables, so filing them again would be a duplicate."
  ]
}

---

Generated by Claude Code

[claude]

Patch round landed: head fb5fecb

Three seat-review corrections on PR #15402, all re-verifiable from the tree.

1. State-set count. The mutual-exclusion rule (digest line 31) counted seven states. The corpus has five pm:* state labels — queue, dispatched, blocked, on-hold, awaiting-maintainer — plus needs-user-decision, six in total. It now reads: 五个 pm 状态标签加 needs-user-decision 共六态互斥,转换恒一笔 replace,队列卡逾三天欠转换。 One line, 120 bytes.
2. Narrative voice. The handover rule (digest line 71) said 归档本席派出的会话; it now says 归档自己派出的会话. grep -c '本席\|上一席' on the file is 0.
3. Model identifier in the PR body. Table 1 row 93's last column named a model tier by name; it now reads "date; narrative (the top tier audits and does not build)". The whole stored PR body greps 0 for every banned model-name spelling; the body was re-read after the PATCH and the sent text is a strict prefix of what GitHub stored.

The file stays at exactly 150 lines, headroom 0, every line inside the 120-byte cap, and the inventory table in the PR body was regenerated from the tree so rows 25, 61 and 93 match the edited lines.

Verdict lines on fb5fecb9, each exit code captured by redirect before any pipe (all exit 0):

check-skill-line-ratchet: .claude/skills/pm-dispatch/references/core-rules.md is 150 lines (ceiling 150; headroom 0).
check-skill-line-ratchet: .claude/skills/pm-dispatch/references/core-rules.md: widest table row is 0 bytes (pin 0; headroom 0).
check-skill-frame-sync: 2 copies of the decision frame are structurally isomorphic across 2 files
check-skill-id-lint: 24 file(s) clean (pattern /#[0-9]{3,}/g).
check-nul-bytes: OK (scanned 7460 text file(s) -- 7460 tracked, 0 untracked-not-ignored; skipped 7 binary; no raw ASCII control bytes).

Hygiene greps on the digest, all 0: dates, corner-bracket quotations, straight double quotes, three-or-more-digit issue references, warning and star markers, the frame's fingerprint phrases, and model names.

The PR is still a draft, no reviewers requested by this run, no auto-merge. Same branch, no force-push. Worktree kept for the next round.

Generated by Claude Code

---

Generated by Claude Code

[claude]

ACCEPT — governed in-seat review of PR #15402 (family member 1, the digest), head fb5fecb9 after one patch round (skills seat, 2026-09-04T14:46Z).

Implemented-by: claude/issue-15379-pm-core-rules-digest (os-dev subagent)
Reviewed-by: session_019RfFHiRCSs3JXLK4cwcfox (skills seat)

Verified in a detached compare worktree at the PR head, not from the report:

• Surface. git diff 27875fc7 --stat = the new .claude/skills/pm-dispatch/references/core-rules.md (150 lines) + 7 lines in scripts/pm/check-skill-line-ratchet.mjs (one CEILINGS row at 150, one MAX_TABLE_ROW_BYTES row at 0; no other row moves). Nothing else touched; SKILL.md, the other references, os-dev.md and AGENTS.md are byte-identical to the base.
• The card's style gates, measured on the file: 150 lines, widest line 120 bytes; dates 0, 「」 quotations 0, " 0, issue/PR numbers 0, ⚠️ 0, ⭐ 0, frame fingerprint phrases 0, narrative voice (本席/上一席) 0; ⛔ appears 54 times, on prohibitions only. The digest refers to the decision frame by section and names the four axes in its own words.
• Gates at the head, exit codes captured by redirect: check-skill-line-ratchet exit 0 (core-rules.md is 150 lines (ceiling 150; headroom 0), widest table row 0); check-skill-frame-sync exit 0 (binding sentence present in all 2; 73 markdown files scanned for undeclared copies); check-skill-id-lint exit 0 (24 files clean); check-governed-merges --test on both paths exit 3 (GOVERNED — a human merge is the review record, 1 of 2 paths on the register). The dev's report adds dispatch-gates --ran 28/28 accounted, the whole-repo lint through the verification lock, check:doc-formula-expressions after building the lint closure, check:nul-bytes clean.
• Completeness, sampled from the corpus into Table 1 of the PR body (the rule inventory, 136 rows): the cheapest-inventory-first opening act; the six mutually exclusive states and the one-write replace; the four-step label write with read-back; the two-check hold release; no angle-bracket placeholders; the Claim: first-line convention; the three-zone dispatch brief; the 45-minute probe threshold; all-green as the enqueue qualification; the governed fork (draft, two approvers, never approve from the seat); the Chinese round report; the ≥50% weight of the long-term axis; packages/spec routed to the spec seat; the report contract's authoritative shape in os-dev.md. Every one is present with its source section named. Table 2 (nine classes left out, with the two folded details named) and Table 3 (two same-rule-different-name conflicts, reported and not resolved) are what the card asked for.
• Patch round (one, 14:36Z → 14:36Z push): the state-count line now reads five pm:* labels + needs-user-decision = six states; the one narrative-voice word is gone; the one model name in the PR body is gone (body grep for model names: zero).
• Report hygiene: dev report posted first on this card with the literal os-dev-report first line, footer intact; PR body starts Fixes #15379, no model identifier; zero GitHub MCP calls.

Two conflicts from Table 3 are the seat's to raise, not the digest's: (1) whether a single-lane repo:* seat produces the issue type and first grading (SKILL.md grants it; references/lanes/hotcrm.md withholds it at three sites — the digest takes SKILL.md); (2) the director lane file still frames its contract-review duty around a release that SKILL.md's ACCEPT fork gives to the dispatching seat. Both are already correction-class rows in the #13597 Phase-1 tables; they go into the lane-file member of this family rather than a new card.

Landing regime: governed (.claude/**) ⇒ the PR stays a draft; reviewers os-zhuang + hotlong requested in this same act; human merge. The maintainer reviews the digest itself (150 lines) and Table 1, not the corpus. On the maintainer's word on the digest, member 1b (the SKILL.md rules-only rewrite, matched to the approved digest) is dispatched; ⛔ not before. This family card stays pm:dispatched across its members.

---

Generated by Claude Code

[claude]

Claim: PM loop round 5 (canonical first-line form of the claim already posted as 5541175284 — same flight, same branch; posted so the patrol predicate that reads only a comment's first line finds it)
Session: session_019RfFHiRCSs3JXLK4cwcfox
Branch: claude/issue-15379-pm-core-rules-digest
Worktree: objectstack-issue-15379
Domain: domain:skills
File surface: .claude/skills/pm-dispatch/references/core-rules.md (new) + one ceiling row and one table-row pin in scripts/pm/check-skill-line-ratchet.mjs; later members of this family re-claim per member.
Container & model: L, mode:subagent, model: opus
Clause-②: no
Serial constraints cleared: as read on 5541175284; member 1 delivered as PR #15402 and accepted (5542044383).

---

Generated by Claude Code