docs(pm-dispatch): six protocol-line repairs, line-neutral on both governed text surfaces by os-steve · Pull Request #16339 · objectstack-ai/objectstack

os-steve · 2026-09-06T14:38:54Z

Fixes #16004
Fixes #16064
Fixes #15897
Fixes #15961
Refs #15975
Refs #14768

Six protocol-line repairs on the two governed pm-dispatch text surfaces. No code, no
behaviour: these files are read in full by every seat session, and each repaired line was
measured to be producing wrong runs.

Both files are LINE-NEUTRAL — every added line is paid by a fold inside the same
section, never by a ceiling raise:

file	lines	widest table row
`.claude/skills/pm-dispatch/SKILL.md`	811 / 811 (headroom 0)	342 / 342 (headroom 0)
`.claude/skills/pm-dispatch/references/core-rules.md`	150 / 150 (headroom 0)	0 / 0

Every added line is ≤ 119 bytes against the 120-byte rule. The four-axis decision-frame
block in SKILL.md is untouched, and check:skill-frame-sync confirms the two copies are
still structurally isomorphic.

What changed, per card

#16064 — SKILL.md:538 ordered every dispatch to demand a second claim comment

AGENTS.md:400 binds a dispatched executor to post no second claim; SKILL.md:538
ordered the PM to demand one. The skill's own tiebreak makes AGENTS.md the winner, so
:538 was the defective line. Five devs across two repos hit it; four refused, one complied
— identical orders, opposite outcomes.

Rewritten in place, 1:1, to the AGENTS.md reading: the dispatch states the dev posts no
second claim, the PM's Claim: is its identity, and the dev verifies the newest one names
its branch and stops with a report on a mismatch.

⚠️ Worth a reviewer's eye: the triage comment on #16064 also named SKILL.md:479
(家族派发的折叠认领) as a second site ordering a claim per member card. Reading that line in
its block (467–485), its subject is the PM's atomic pair, not the dev — it describes the
PM's own folded claims across member cards, which is what this flight's own claims are. It
therefore does not carry the same defect and is left untouched. Flagging the reading rather
than acting on it, since it differs from the triage note.

#15897 — the dispatch asked for the human-readable gate list and got it scraped

The old line named a bare dispatch-gates.mjs run, whose prose output has to be
scraped; a dev's regex silently dropped the whole convention-triggered section at exit 0,
five gates with it, and CI went red on check:i18n.

Rewritten in place, 1:1, to name the machine form: take the list from
dispatch-gates.mjs --commands, run each command, land the exit code before any pipe, and
reconcile with --ran; never scrape the human-readable output. The reconciliation is what
now mechanically guarantees no family is missed, so it carries that guarantee instead of
prose (rules-only register).

#16004 — the seat mutex's reading 3 was blind to a queued card's branch

Reading 3 filtered on pm:dispatched only, so a session pushing to a pm:queue card's
claude/issue-N-* branch produced no reading at all; one collision was found by hand, off
the branch tip, not by the mutex.

Widened to two lines: the card set becomes pm:dispatched ∪ pm:queue, and a branch push
since the last shift-end brief counts as a claim in its own right (one git ls-remote per
card) — a push is a claim, and owes the comment reading 3 counts.

Paid by folding the cross-seat-handwriting line and the readings-only-add-a-yield line into
one, inside the same block.

#15975 (Refs) — a single-carrier gate hang makes a clear indistinguishable from a strip

The clause-② gate is a dual carrier; hung on one side only, the eventual legitimate clear
can only ever produce the one-removal strip signature. The defect is committed at hang
time and only surfaces at landing.

One line added beside the existing carrier rule: a Clause-②: yes claim hangs the gate on
the card in the same write as the claim, and when the delivering PR opens the seat reads
check-clause2-carriers --pair N at exit 0 before requesting any review — so a split is
repaired at hang time, never at landing.

Paid by folding the --tier hint line and the enqueue-refusal line into one, inside the
same 复核 block. The other half of this card (the os-dev.md line) rides its own flight, so
this PR only refs it.

#14768 (Refs) — the unexecutable "report, then act" fence

SKILL.md:543 carried 「报备后」 — report for the record, then fix — which presupposes a
mid-run channel a one-shot dev does not have, and so hands it the whole three-way choice the
author thought they had fenced. This is the census's hit 3, in the allow-with-trace class.

Rewritten in place, 1:1, to the executable form: fix on the producer side and report the
landing point; never patch on the consumer side. Only the SKILL.md hit is in scope here —
the census's other three hits ride flight B.

#15961 — the core-rules digest owed the two decision-frame rules

One digest line added to the 升级与决策 block, in the digest's plain-language register: every
decision card and director presentation carries Governing text:, and the protocol is the
baseline — a spec-vs-code mismatch defaults to changing the code, and a protocol change is
its own card rather than an option on a defect card.

Paid by folding the two recommendation-order lines into one. ⚠️ The fold drops two clauses
that are derivable rather than independent — 「其余投不翻它」 (a consequence of the ≥50% weight
stated in the same line) and 「防错轴打破余下平局」. The second is a genuine rule and it is
not lost from the corpus: SKILL.md keeps it. It leaves the digest only, which is what a
fold in a digest costs.

Verification

Gate union run at d88c29c76a, this PR's head, after the final commit.

node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 16
commands from the tree (never from a hand-written path list). Every one was run with its
exit code captured before any pipe, then reconciled:

✓ dispatch-gates --ran: 16 derived famil(ies) accounted for — 16 run, 0 NOT-MEASURED.

16 of 16 green. One needed a build first and is worth naming, because its first exit was not
a finding:

pnpm --filter @objectstack/lint run check:doc-formula-expressions first returned exit
3 — the gate's own PREREQUISITE-NOT-MET code, in its words "Nothing was measured … It is
NOT a finding". After turbo run build --filter=@objectstack/formula --filter=@objectstack/lint
it re-ran at exit 0.

The dispatch-named gates, quoted from their own verdict lines rather than from a bare $?:

✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md is 811 lines (ceiling 811; headroom 0).
✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md: widest table row is 342 bytes (pin 342; headroom 0).
✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/core-rules.md is 150 lines (ceiling 150; headroom 0).
✓ check-skill-id-lint self-test: 14 cases pass.
✓ check-skill-id-lint: 25 file(s) clean (pattern /#[0-9]{3,}/g).
✓ check-skill-frame-sync self-test: 13 cases pass …
✓ check-skill-frame-sync: 2 copies of the decision frame are structurally isomorphic across 2 files
✓ check-governed-prose self-test: 28 cases pass.
✓ check-governed-prose: 2 instruction surface(s) name all 5 registered governed surfaces …
✓ check-governed-merges --self-test: 274 assertions …
✓ check-half-states self-test: 2535 cases pass.

check-governed-merges --test on this PR's real paths returns exit 3 — GOVERNED, as
expected for .claude/**:

governed-surface predicate: 2 of 2 path(s) hit the register (5 surfaces, repo-agnostic).
  ⛔  GOVERNED — a human merge is the review record for this PR (#9495 regime).

Whole-repo pnpm lint (eslint . --no-inline-config) ran in full through
scripts/pm/os-verify-lock.sh (slot issue-16004) and returned:

os-verify-lock: VERDICT command-exit 0 · held the lock 116s (1m56s) · waited 0s

No narrowing was taken, so no narrowing needs declaring.

H-row check. Each of the 10 sentences this diff removes was grepped over scripts/ and
.claude/hooks: 0 hits, with a positive control proving the grep reaches that corpus
(scripts/pm/check-half-states.mjs does quote SKILL.md-register phrases, just not these).
No patrol row pins a sentence this PR rewrites.

维护者速读(草稿)

一、这次动的是什么。 只改两份 PM 协议文本,不动任何代码、不动任何行为。改的是六条把
派发跑错的规则行:每条要么在实测中直接产生了错误的运行,要么让一次合法操作变得无法自证。
两份文件都严格行数不变——新增的每一行都由同一段落里的一次合并偿还,没有抬任何天花板。

二、最贵的一条是"第二条认领"。 协议要求每份派发令命令 dev 再发一条认领评论,而仓库根
的 AGENTS.md 明令 dev 不得再发。两处相互矛盾,而 AGENTS.md 按规则获胜。代价是实测出来的:
跨两个仓、两条车道,五个 dev 撞上同一条矛盾指令,四个拒绝执行、一个照做——同样的命令,相反
的结果。真正的成本不是那四次拒绝,而是那些默默挑一边、没人看见的运行。

三、门禁清单被当成散文抓取。 派发令原来让 dev 跑工具的"给人看"的那种输出,再用正则去
抓。一个 dev 的正则静默丢掉了整整一段(五个门禁),退出码仍是 0,读起来像跑完了,直到 CI 变
红。现在派发令点名机器可读形式,并要求逐条跑、退出码先落盘、再用工具自己的对账模式核。

四、两处让"证据"失效的机制。 其一,开轮互斥的第三个读数只看已派发的卡,于是有人在排队
卡的分支上干活时它一个读数都产生不了——实测撞过一次车,是靠人读分支尖端发现的,不是靠互斥。
现在分支推送本身即认领。其二,契约复审的门禁标签本该同时挂在卡和 PR 两处,只挂一处时,日后
那次合法的清除只会留下"被剥离"的签名——缺陷在挂标那一刻就已铸成,却要到落地才暴露。现在挂
标即同笔挂卡,并在 PR 开出时先对账。

五、剩下两条是收口。 一条把"先报备、再动手"这种一次性 dev 根本没有通道执行的围栏,改成
本轮内能闭合的形式;一条把决策框架的两条新规则补进核心条款摘要,让摘要和主协议不再各说各话。
需要维护者留意的是第六条的合并代价:摘要里少了一句"防错轴打破余下平局"——该规则在主协议
SKILL.md 中完整保留,只是不再在摘要中重复。

席位意见:

🤖 Generated with Claude Code

Generated by Claude Code

- reading 3 counts `pm:dispatched` u `pm:queue` and a branch push (a push is a claim and owes the comment), paid by folding the cross-seat/readings pair - the dispatch names `dispatch-gates.mjs --commands`, exits captured before any pipe, `--ran` reconciling; the human-readable form is never scraped - the dispatch tells the dev it posts no second claim, per AGENTS.md - the producer-side fence made executable: fix and report, no notify-then-act - a `Clause-<2>: yes` claim hangs the gate on the card in the same write and reads `--pair N` before requesting review, paid by folding the tier pair - core-rules digest gains the `Governing text:` + protocol-baseline line, paid by folding the two recommendation-order lines SKILL.md 811/811, table pin 342 untouched; core-rules.md 150/150. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019RfFHiRCSs3JXLK4cwcfox

…ill-protocol-lines

claude · 2026-09-06T15:07:50Z

ACCEPT — contract-tier review of PR #16339, head d88c29c76a (skills seat, 2026-09-06T15:1xZ).

Implemented-by: claude/issue-16004-skill-protocol-lines (os-dev subagent, opus under the measured quota exemption recorded on #16004 and #14768 at 14:05Z)
Reviewed-by: session_019RfFHiRCSs3JXLK4cwcfox (skills seat)

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

Two governed files, +10/−10, both line-neutral, every rule as ruled. SKILL.md (811/811, table pin 342, the four-axis frame block byte-identical to main): reading 3 of the seat mutex now reads pm:dispatched ∪ pm:queue and a branch push since the shift-end brief is a claim that owes a Claim: (finding(skills): the seat mutex's reading 3 counts Claim: only on pm:dispatched cards — a session re-engaged on a pm:queue card's branch is invisible to it (collision measured 2026-09-05) #16004; paid by folding the cross-seat-handwriting and readings-only-yield lines into one, no sentence lost); the dispatch-gates line names --commands, exits landed before any pipe, --ran reconciling, never the scraped human-readable form (the os-dev dispatch template asks for dispatch-gates in the human-readable form and gets it scraped — a whole gate section can be dropped silently at exit 0 #15897); the second-claim line is now the AGENTS.md reading — the dev posts no second claim and verifies the newest Claim: names its branch (finding(skills): SKILL.md:538 orders every dispatch to demand a second claim comment; AGENTS.md:400 / CLAUDE.md:13 bind the dev to post none — two devs refused the order in one round, a third complied #16064); the 文件面 fence is allow-with-trace — fix on the producer side and report the landing, never a consumer-side patch (Census of dispatch templates and skills text for the unexecutable "report before acting" fence form, now that the executability criterion is landed #14768's hit); one line beside the needs:contract-review rule hangs the gate on the card in the same write as a Clause-②: yes claim and reads check-clause2-carriers --pair N at 0 before any review request ([finding] a single-carrier needs:contract-review hang makes the eventual clear indistinguishable from a strip — the defect is committed at hang time and only surfaces at landing #15975's SKILL.md half), paid by folding the --tier-hint and enqueue-refusal lines into one with both clauses intact. core-rules.md (150/150): one digest line for Governing text: + protocol-as-baseline (core-rules.md digest owes one line for the two decision-frame rules PR #15946 lands (Governing text, protocol as baseline) #15961), paid by folding the two recommendation-order lines. The fold drops three digest fragments — 「逐轴分析不是可选项」, 「其余投不翻它」, 「防错轴打破余下平局」 — and the seat checked each survives in SKILL.md's frame lines (754 and 756): the digest is a subset, nothing leaves the corpus. Every added line measures 97–119 bytes (LC_ALL=C awk).
Gates at the head: check-skill-line-ratchet exit 0 (SKILL.md 811/811, pin 342; core-rules.md 150/150); check-skill-frame-sync self-test 13 cases + live exit 0; check-skill-id-lint 25 files clean; check-governed-prose exit 0; check-half-states --self-test 2535 (the ten removed sentences are pinned by no H-row — the dev's grep with a positive control); check-governed-merges --test on the two paths exit 3 (GOVERNED). The dev's report adds the derived union 16/16 with --ran at 0 NOT-MEASURED and the whole-repo pnpm lint through the verify lock (exit 0, 116 s, unnarrowed). PR checks on d88c29c76a: 17 success, 12 skipped, 0 failed (29 runs, converged).
Report hygiene: os-dev-report on all six cards (full on finding(skills): the seat mutex's reading 3 counts Claim: only on pm:dispatched cards — a session re-engaged on a pm:queue card's branch is invisible to it (collision measured 2026-09-05) #16004), JSON closes; PR body opens with the four Fixes lines and the two Refs lines exactly as ruled ([finding] a single-carrier needs:contract-review hang makes the eventual clear indistinguishable from a strip — the defect is committed at hang time and only surfaces at landing #15975 and Census of dispatch templates and skills text for the unexecutable "report before acting" fence form, now that the executability criterion is landed #14768 keep their os-dev.md halves in flight B); ## 维护者速读(草稿) present; no model identifier (the only claude- strings are the two footer URLs); skip-changeset read back; one MCP call declared. The branch merged origin/main once after a stale-tree warning from dispatch-gates — the diff against main is still the ten lines.

Ruling on the open question — A. SKILL.md's 「家族派发的折叠认领:共享分支按链首卡命名;每张成员卡各留认领评论并点名该分支」 sits in the PM's atomic-pair block (its neighbours are 「派发前按序执行原子对」 and 「② 认领评论(英文)」) and orders the PM's folded claim per member card — exactly what this flight's six claims are. It states nothing AGENTS.md contradicts and is not copied into dispatch orders, so it is not the defect #16064 measured; no edit, and no follow-up card for a subject marker — the block's subject is explicit two lines above.

Collision, recorded for the approvers: the devx seat's draft PR #16271 (the #14944 depth-hold rules) edits the same two files; git merge-tree between the two heads conflicts in SKILL.md's 候选与批次 block. Neither waits for the other; whichever merges second owes a patch round that merges main, re-runs the line ratchet live and re-folds (seat note 5560027553 on that PR).

Landing regime: governed (.claude/**) ⇒ the PR stays a draft; reviewers os-zhuang + hotlong requested in this same act, needs-user-decision added, the final 维护者速读 posted as the next comment; human merge. On the merge: #16004 · #16064 · #15897 · #15961 close by Fixes; #15975 and #14768 stay pm:dispatched for flight B.

Generated by Claude Code

claude · 2026-09-06T15:07:53Z

维护者速读

事情:只改两份 PM 协议文本(SKILL.md 与核心条款摘要),不动代码、不动行为。改的是六条实测把派发跑错的规则行,两份文件严格行数不变——新增的每一行都由同段落一次合并偿还,没有抬任何天花板。

最贵的一条是「第二条认领」:协议原来要求每份派发令命令 dev 再发一条认领评论,而仓库根 AGENTS.md 明令 dev 不得再发。跨两仓两车道五个 dev 撞上这条矛盾指令,四个拒绝、一个照做。现在协议改成 AGENTS.md 的读法:PM 那条认领就是 dev 的身份,dev 只核对最新一条点名自己的分支。

门禁清单不再靠人读输出:派发令原来让 dev 抓工具「给人看」的输出,一个 dev 的正则静默丢了五个门禁、退出码仍是 0。现在点名机器可读形式,逐条跑、退出码先落盘、再用 --ran 对账。

两处让证据失效的机制补上:开轮互斥的第三个读数原来只看已派发的卡,有人在排队卡的分支上干活时它读不到——现在分支推送本身即认领;契约复审标签原来可以只挂 PR 不挂卡,日后合法清除就留下「被剥离」的假签名——现在挂标同笔挂卡,PR 开出即对账。另两条是收口:「先报备再动手」改成一次性 dev 能闭合的形式;决策框架的两条新规则补进摘要。

要请您留意的:摘要里少了「防错轴打破余下平局」这半句,规则本体在 SKILL.md 四轴段完整保留(席位逐句核过),只是摘要不再重复。另外 devx 席的 PR #16271 改同一段落,两份草案互相冲突,后合的一方补一轮合并,不影响本 PR。

席位意见:建议合并。四张卡由本 PR 关闭(#16004 · #16064 · #15897 · #15961),#15975 与 #14768 的另一半在 flight B 的 os-dev.md 里。

Generated by Claude Code

The merge of origin/main brought SKILL.md to 813 lines against a ceiling of 811: both sides were line-neutral against the merge base, but the folds this branch used to pay for its four depth-hold lines landed in regions main had re-folded independently, so the resolution kept both. Pay the +2 with two more same-topic folds inside the same file — no rule removed, no ceiling raised, no cross-file move: - the two seat invariants "循环必须能从全新会话恢复" and "PM 不写文件、不写代码" join on one line (112 bytes); the separator before 唯一例外 drops from ; to , so the exception still reads as the no-writing rule's, not the recovery one's. - the two conditional-clause items 拒收用例 and 过滤/谓词语义 join on one line (119 bytes) inside the enumeration they already belong to. Ratchet reads SKILL.md 811/811 headroom 0, widest table row 342 (pin 342), core-rules.md 150/150. Every prose line stays within the 120-byte budget. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TezFG8ZMrNH6n5VTNpPpdH

claude added 2 commits September 6, 2026 14:23

Merge remote-tracking branch 'origin/main' into claude/issue-16004-sk…

d88c29c

…ill-protocol-lines

github-actions Bot added size/s documentation Improvements or additions to documentation labels Sep 6, 2026

claude Bot added the skip-changeset PR has no user-facing published change; bypasses the changeset gate label Sep 6, 2026

claude Bot requested review from hotlong and os-zhuang September 6, 2026 15:07

claude Bot added the needs-user-decision label Sep 6, 2026

os-zhuang approved these changes Sep 7, 2026

View reviewed changes

os-zhuang marked this pull request as ready for review September 7, 2026 00:49

os-zhuang enabled auto-merge September 7, 2026 00:49

os-zhuang added this pull request to the merge queue Sep 7, 2026

Merged via the queue into main with commit 8da8d50 Sep 7, 2026
35 checks passed

os-zhuang deleted the claude/issue-16004-skill-protocol-lines branch September 7, 2026 01:18

os-zhuang mentioned this pull request Sep 7, 2026

[PM seat] director(项目总监) — 🟢 os-zhuang（第 15 场，契约复审 + 落地）· ⏹️ hotlong（第 17 场已收班）· 第 16 场已收班 #12708

Open

os-zhuang removed the needs-user-decision label Sep 7, 2026 — with Claude

os-musk mentioned this pull request Sep 7, 2026

[finding] os-dev subagents end their turn to "wait for the Monitor notification" after a background gate sweep — four times in one round, each needing a PM stall reset #16332

Closed

Navigation Menu

Uh oh!

docs(pm-dispatch): six protocol-line repairs, line-neutral on both governed text surfaces - #16339

docs(pm-dispatch): six protocol-line repairs, line-neutral on both governed text surfaces#16339
os-zhuang merged 2 commits into
mainfrom
claude/issue-16004-skill-protocol-lines

os-steve commented Sep 6, 2026

Uh oh!

claude Bot commented Sep 6, 2026

Uh oh!

claude Bot commented Sep 6, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

os-steve commented Sep 6, 2026

What changed, per card

#16064 — SKILL.md:538 ordered every dispatch to demand a second claim comment

#15897 — the dispatch asked for the human-readable gate list and got it scraped

#16004 — the seat mutex's reading 3 was blind to a queued card's branch

#15975 (Refs) — a single-carrier gate hang makes a clear indistinguishable from a strip

#14768 (Refs) — the unexecutable "report, then act" fence

#15961 — the core-rules digest owed the two decision-frame rules

Verification

维护者速读(草稿)

Uh oh!

claude Bot commented Sep 6, 2026

Uh oh!

claude Bot commented Sep 6, 2026

维护者速读

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants