docs(pm-dispatch): point every PR-body rewrite at the issue_patch door (#31058)

<!-- claude-projects-attribution: {"github_login":"huangyiirene"} -->
_Requested by **Irene** · [project
thread](https://claude.ai/code/project/chan_01TJXtQ1tVD26ecLDZxFS67x?thread=cmsg_01TJXtQ1tVD26ecLDZxFS67xRZfRWwVbM1oLGd4rdqg3kL)_

Fixes #31050

Clause-②: no
Metadata-protocol: none

`dispatch-gates --tier` reads a MANDATORY model tier, from clause ①
(`.claude/agents/os-dev.md`, the dev-agent definition). This card ran on
the project's default model: that mandate is waived by the maintainer's
standing order in this project (2026-11-03), and the dispatching seat
runs the contract review.

## 维护者速读

- 改了什么:席位改写 PR 正文的其余三处(`rest-channel.md`、`platform-readings.md` 的 flake
记因、`os-dev.md` 的席位代写)都指向 `seat-post-protocol.md` 的 PR 正文行;中继改 PR
正文不追加页脚这一读数补进 `platform-readings.md`;`check-handover.mjs`
头注分别点名交回规则的两处去处。
- 为什么改:#31049 只在两处教了 PR 正文门;`rest-channel.md` 仍写「直调 `dispatch.mjs`
仅落地」,与 ACCEPT 前改正文相抵,另两处改正文不点门。
- 风险与回滚:只改内部 agent 规程文字与一行脚本注释,不发包(`skip-changeset`),四个文件行数不变;回滚即
revert。
- 先看:`rest-channel.md` L16–17(新例外行);`platform-readings.md` L343(新页脚读数,排在
L337–342 的裸 REST 改侧格之后)。
- 两分钟验证:`node scripts/pm/check-skill-line-ratchet.mjs`,预期退出
0,`rest-channel.md` 82 行、`platform-readings.md` 458 行、`os-dev.md` 405
行,各等于上限。
- 需要你做的:无,席位落地。

## What changes

**Item 1, `.claude/skills/pm-dispatch/references/rest-channel.md` L16.**
The line said a direct `dispatch.mjs` call is for landing only, but a
seat rewrites a PR body before ACCEPT. It is split in two:

- L16 keeps the op table, the read side and the batch rule: `闭合 op 表住
fleet-write/ops.mjs;读侧不走中继;批量恒走逐 op 工具。`
- New L17 names the exception and points to the door, with no second
spelling of it: `直调 dispatch.mjs 仅落地,与 PR 正文单笔
issue_patch(seat-post-protocol.md 正文行,不送页脚)。`

**Item 2, `scripts/pm/check-handover.mjs` L33–34 (header comment only,
no code).** PR #31049 deleted the old home (`seat-post-protocol.md`,
收班), and its two facts now live in two places, so the header names both:

- the refusal inside (a), a `needs_decision` card handed to the decision
box and never back to `pm:queue` → `seat-post-protocol.md` 交接收尾清单 step 2
(L47, `needs_decision 进决策箱`);
- the stay-behind narrowing → `seat-lifecycle.md` 离任留守 (L44, 离任留守只写点名的已
ACCEPT 待队列或待人批 PR).

The comment now reads: `The sign-off hand-back rule
(seat-post-protocol.md, 交接收尾清单 step 2) names one refusal inside (a);
seat-lifecycle.md 离任留守 narrows (a)'s stay-behind (section below):`.

**Item 3, the two rewrites that named no door.**

- `references/platform-readings.md` L290 (new, under the `Check
Changeset` flake line): `⇒ 该正文改走 seat-post-protocol.md 的 PR 正文行,不送页脚。`
- `.claude/agents/os-dev.md` L66, rewritten in place: `PR 正文 dev 只在
pr_create 写,⛔ 不 PATCH;改法报告点名,席位经 seat-post-protocol.md 代写。` The meaning
is the old line's (the dev writes the body once, at the create; the
report names the change; the seat writes it), and it now names where the
seat's rewrite goes.

**Item 4, the footer reading (`platform-readings.md` L336–341).**
Settled from code and stored bodies only; no test write was made.

- How the relay reads back: after a `success` run, `sendFleetWrite`
reads every action that sent a `body` (`READ_BACK_LOCATORS.issue_patch`
= `GET /repos/{repo}/issues/{issue}`) and judges it with post-stamped's
`classifyReadBack` (`judgeStoredBody`,
`scripts/pm/fleet-write/dispatch.mjs` L1081). The class `identical` is
`stored === sent` (`declaredNormalisation`,
`scripts/pm/post-stamped.mjs` L2843). Nothing is stripped before that
comparison, and `storedBytes` is the raw stored length. A stored body
with the bare footer appended would print `footer-appended`, not
`identical`.
- What it read: PR #31049's body, 8149 bytes sent and 8149 stored,
`identical`. PR #29225's edit was read back `identical` too (its landing
comment on #29145).
- What the stored bodies show today (one `GET
/repos/objectstack-ai/objectstack/issues/{n}` each, read only): #31049
is 8149 bytes and #29225 is 6785 bytes, and both end with the `<!--
signed: … -->` trailer and a newline. Neither carries a `_Generated by
…_` footer.
- So the relay `issue_patch` appends no footer, while main L338–341
measured the raw-REST `PATCH` cell (+58). Per AGENTS.md, the cells
differ by channel and are not generalised. Main L341 (now L340) stays as
it is. New L343 records the relay cell after the whole raw-REST
edit-side band (L337–342), so the strip-both-spellings and attribution
lines still read with their own cell: `中继 issue_patch 不在此列:dispatch.mjs
读回 PR 正文存 = 送(identical),无追加。`
- `seat-post-protocol.md` L31 is unchanged. Adding 不送页脚 there does not
fit at 120 B (the shortest spelling measured is 125 B), and the file has
no restatement to delete to pay for a second line. For now the no-footer
word rides on the pointers: `rest-channel.md` L17 and
`platform-readings.md` L290 carry it, and `SKILL.md` L313 already did.
`os-dev.md` L66 is at exactly 120 B and carries the pointer only. Giving
the rule its one home at L31 (and dropping the pointers' copies) is left
to #31059.

### Deleted restatements (deletion only; each fact's surviving home)

| File | Deleted line (main) | Surviving home |
|---|---|---|
| `rest-channel.md` | L42 `⛔ post-stamped/label-write 永不接进管道再 &&:拒收读成
0;看尾先落文件或 set -o pipefail。` | `scripts/pm/post-stamped.mjs` header, 〈⛔
Never pipe this tool, then `&&` the write that follows〉 (L112–127:
redirect then capture, or `set -o pipefail`; "it covers every writer
beside it — `label-write.mjs` included"); `scripts/pm/label-write.mjs`
header, 〈Exit codes — capture them BEFORE any pipe〉 (L156) |
| `platform-readings.md` | L36 `⛔ 永不用 -c merge.os-regen.driver=
覆盖:驱动不是被关掉而是跑失败,路由路径全报冲突。` | AGENTS.md, Multi-agent discipline §11,
corollary ("⛔ Never with `-c merge.os-regen.driver=` … git runs the
configured program, it fails, and every routed path is reported
conflicted"); `scripts/pm/os-regen-merge.sh` header L303–304 |
| `platform-readings.md` | L313 `否则 repair-first
被正确地应用到完好的卡上,重写毁掉的是正确内容。` | AGENTS.md, 〈GitHub mutates body BYTES〉 ("⛔ A
body reading short only through the API is probably intact — check the
rendered page before "repairing" it; a rewrite destroys a correct
card"); the prescription it followed, L312 (now L311), stays |

## Tests

None: the diff is instruction text and one comment line, with no runtime
behaviour to test. The gates below are the checks.

## Acceptance notes

### Line budget (`scripts/pm/check-skill-line-ratchet.mjs`, exit 0)

| File | Lines (main → head) | Ceiling | Bytes (main → head) |
|---|---|---|---|
| `references/rest-channel.md` | 82 → 82 | 82 | 7,772 → 7,739 |
| `references/platform-readings.md` | 458 → 458 | 458 | 47,066 → 47,039
|
| `.claude/agents/os-dev.md` | 405 → 405 | 405 | 38,411 → 38,414 |
| `scripts/pm/check-handover.mjs` (not ratcheted) | 891 → 891 | — |
65,299 → 65,301 |

Every touched line: `rest-channel.md` L16 120 → 88 B, new L17 118 B;
`platform-readings.md` new L290 78 B, new L343 102 B; `os-dev.md` L66
117 → 120 B; `check-handover.mjs` L33 111 → 113 B and L34 31 → 78 B. No
new section; no fence line touched.

### Not changed here (found while reading)

- `scripts/pm/fleet-write/dispatch.mjs` header (〈A pull request's body
…〉, L454–456 on `main`) still says "an edit of a pull's body gets the
bare footer block appended after the trailer (`platform-readings.md`,
the edit side), so send no footer of your own". The relay read-backs
above show no append on that path. The advice (send no footer) still
holds: a footer after the trailer would be the body's last non-empty
line, and the door would refuse the body as unsigned. Only the stated
reason is stale. This is outside the card's file surface; it is filed as
#31059 and not changed here.

## Verification

<details><summary>Local gates on 3ac6083: 41 derived, 41 run with exit
0, 0 NOT MEASURED (`dispatch-gates --ran`)</summary>

- `node scripts/pm/dispatch-gates.mjs --commands --repo
objectstack-ai/objectstack` derived the same 41 commands for the 4
changed paths as on round 1 (33bac6a). Each ran under a 900 s timeout,
its exit code captured before any pipe; all 41 exited 0 on the first run
(the `@objectstack/formula`, `@objectstack/lint` and
`@objectstack/objectql` builds from round 1 were in place).
- They include `check:pm-skill-ratchet`, `check:pm-skill-id-lint`,
`check:pm-governed-prose`, `check:pm-governed-merges`,
`check:pm-half-states`, `check:pm-post-stamped`, `check:doc-authoring`,
`check:skill-frame-sync`, `check:commit-card-trailers`,
`check-comment-mask-corpus.mjs`, and the self-tests of
`check-governed-queue-guard.mjs` and `check-harness-current.mjs`.
- `--ran` printed: `41 derived famil(ies) accounted for — 41 run, 0
NOT-MEASURED (a DERIVED zero — all 41 recorded an exit code and none of
them is 3)`.
- Round 1 (33bac6a): 41 of 41 exit 0, the formula gate after the same
builds.

</details>

- `node scripts/pm/check-governed-merges.mjs --branch
claude/project-thread-0ili91`: GOVERNED, Tier S (席内达档复核落地): `.claude/**`
×3; `scripts/pm/check-handover.mjs` is not on the register; 14 changed
lines (+7 / −7), under 3000.
- `node scripts/check-commit-card-trailers.mjs --range
origin/main..HEAD`: exit 0 before every push.

## Human Minimum

- [x] 1. Schemas start as Zod; `packages/spec` holds no runtime logic.
(No schema touched.)
- [x] 2. Every name I added uses the casing for its kind. (No names
added.)
- [x] 3. No formatter ran. (Markdown and one comment line; no lintable
source changed.)
- [x] 4. One issue, one branch off an up-to-date `main`, one pull
request.
- [x] 5. Nothing publishes (`.claude/**` and a `scripts/pm/**` comment),
so the PR takes `skip-changeset`. Both declaration lines are filled in
above.
- [x] 6. `packages/spec` not touched; nothing to regenerate.
- [x] 7. No hand edits to `content/docs/references/`,
`packages/*/CHANGELOG.md` or `content/docs/releases/`.
- [x] 8. Title, body and comments are in English. (The 维护者速读 section is
Chinese, per the lane template.)
- [x] The *Before Submitting* list in CONTRIBUTING.md is done.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_01J8VVAkpKySQ5xGrSjjjnVw

<!-- signed: domain:skills#3 · huangyiirene ·
session_01J8VVAj7CvjkGXvgXzd6RZ7 -->

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: objectstack-fleet[bot]

.claude/agents/os-dev.md

@@ -63,7 +63,7 @@ model: opus
    - 写预算三笔:`git push`、`pr_create`(draft,含 labels 与 assignees)、`os-dev-report` 评论。
    - 卡与线程只走 payload 档(公开仓单卡网页内嵌 JSON,拼写住 platform-readings)或单卡 REST 读。
    - 三类发现进报告交席位代立,不 `POST /issues`、⛔ 不静默弃报;PM 去重读数当既有事实用。
-   - PR 正文 dev 只写一次,在开 PR 那一笔,⛔ 不 `PATCH`;事后要改的报告点名改法,席位代写。
+   - PR 正文 dev 只在 `pr_create` 写,⛔ 不 `PATCH`;改法报告点名,席位经 `seat-post-protocol.md` 代写。
    - 报告记 `api_writes`(次数 + 端点清单)与 `mcp_calls`;越界真写了的照列注明缘由,⛔ 不漏记。
    - 归挂不散落:落在已排队 issue 范围内的发现,报告点名为它的 sub-issue(自动进派发池)。
    - 只是依赖它的,报告点名独立单带 `Blocked-by:` 行;立在修复落地的仓,带回链;席位代立。

.claude/skills/pm-dispatch/references/platform-readings.md

@@ -33,7 +33,6 @@
 - 注册按克隆(`pnpm install` 的 prepare),服务端一个驱动都不跑 ⇒ 两侧答的不是同一个问题。
 - 驱动接手的路径上 exit 0 只说内容判断被推迟,⛔ 不是无冲突:它对内容什么都没说。
 - ⇒ 冲突证明探无驱动裸克隆,配方与读法见 `scripts/pm/os-regen-merge.sh`:exit 1 无树 id = 缺对象
-- ⛔ 永不用 `-c merge.os-regen.driver=` 覆盖:驱动不是被关掉而是跑失败,路由路径全报冲突。
 - 入队决策点才整对象 `get` 一次;挂了 flip 定点到点读,⛔ 不又查又等。
 - 状态核验用最小字段(search/list 加 `fields`)或等事件。
 - 聚合读数(`blocked`/`dirty`)只作阴性筛查再定位;放行按名定向读单条 job,⛔ 不拉全表。
@@ -288,6 +287,7 @@
 - 经 `rerun-failed.mjs` 重跑 PR `ci.yml` 失败 job 仅一次,先在 PR 或卡写明 flake 理由;二败即真红。
 - `run_rerun_failed` 只收 `ci.yml`;他 workflow 的 flake 走下两条,⛔ 永不空提交、永不关了重开。
 - `pr-automation.yml`(`Check Changeset` 取消无后继):席位改正文记下 flake 理由,其 `edited` 即重跑。
+- ⇒ 该正文改走 `seat-post-protocol.md` 的 PR 正文行,不送页脚。
 - `lint.yml`(检出超时):`main` 前进后中继 `pr_update_branch` 出真新 head 整套重跑;齐平即空操作。
 - 同一 head 上轻量兄弟 workflow `success` 加重量级载体 `cancelled` 是普通取代的预期签名。
 - 同名 `failure` 也会被带另一诊断的后一次 `failure` 取代 ⇒ 任何判定前先按名取最新一次。
@@ -310,7 +310,6 @@
 - doctype 开标记与 object 标签形状同触发;网页全文完好,⛔ 不修复只在 API 读短的卡。
 - 同坑第三条读路径:MCP `issue_read` 与 PR 读路径把行内反引号里的尖括号片段整个吃掉。
 - GitHub 存储字节完好、raw REST 一字不差 ⇒ ⛔ 不单凭 MCP 读判截断,先取 raw 或渲染页核对。
-- 否则 repair-first 被正确地应用到完好的卡上,重写毁掉的是正确内容。
 - 唤醒中继渲染会把 `{` `}` `<` 转义 ⇒ 判已发布产物按 `GET` 取存储体,⛔ 永不按中继正文。
 - 判据 = 数空的行内代码跨度:一个空跨度恰是短尖括号片段被吃掉的签名。
 - 判据是尖括号不是反引号:有的正文本来就把标识符不带尖括号写。
@@ -341,6 +340,7 @@
 - 该追加按面不按路由:`PATCH /issues/{n}` 打在 PR 上同样 +58 追加裸页脚,而 draft 位不动。
 - 页脚两拼写:裸版与 session-URL 版都要剥,漏剥的卡在正文中段,而 58 字节差照常。
 - ⇒ 代价是归属:裸形无 session id,按此剥净的 PR 正文不载明哪个会话写的;另置见 AGENTS.md。
+- 中继 `issue_patch` 不在此列:`dispatch.mjs` 读回 PR 正文存 = 送(`identical`),无追加。
 - 第四形:建 PR 两通道同判 —— 送出体尾部不是 `---` 加页脚块时,追加一条同形页脚。
 - 该追加是横线加一条 session-URL 形页脚,送出体是存储体的严格前缀;字节数非常量(90/91)。
 - 尾部已是该块则一字不追加,两通道各实测两向 ⇒ 建侧通道不是变量,判据是送出体尾部。

.claude/skills/pm-dispatch/references/rest-channel.md

@@ -13,7 +13,8 @@
 - ⛔ 它、赋值与 `cd … &&` 前缀都落分类器;多仓云会话首调单独裸 `cd /home/user/objectstack`。
 - 活着 = 工作流文件在 `main` 且 Actions 状态 `active`;维护者在 Actions UI 停用即整队回 `direct`。
 - dispatch 带会话令牌发往 objectstack;run 在 runner 上以逐次铸造、窄到目标仓的 App 令牌执行。
-- 闭合 op 表住 `fleet-write/ops.mjs`;读侧不走中继;批量恒走逐 op 工具,直调 `dispatch.mjs` 仅落地。
+- 闭合 op 表住 `fleet-write/ops.mjs`;读侧不走中继;批量恒走逐 op 工具。
+- 直调 `dispatch.mjs` 仅落地,与 PR 正文单笔 `issue_patch`(`seat-post-protocol.md` 正文行,不送页脚)。
 - 正文超 60,000 字节的那一笔 `post-stamped` 写走 `direct` 并印一行,身份是席位自己的用户。
 - run 失败记 5(post-stamped 4,close-cards 有步落地或读不回 4),UNCONFIRMED 6:只读,⛔ 永不重试。
 
@@ -39,7 +40,6 @@
 
 - ✓ 评论 `POST .../issues/{n}/comments`;改评论 `PATCH .../issues/comments/{id}`。
 - ✓ 标签加法 `POST .../issues/{n}/labels`,定向删 `DELETE .../issues/{n}/labels/{name}`;加法优先。
-- ⛔ `post-stamped`/`label-write` 永不接进管道再 `&&`:拒收读成 0;看尾先落文件或 `set -o pipefail`。
 - ⛔ 永不 MCP `issue_write`(锁 1 已拒);会话分类器拒改动 ⇒ 无通道,交有通道席位立卡。
 - ✓ 建卡带标签 `POST .../issues` · 改正文 `PATCH .../issues/{n}` · 认领 `POST .../issues/{n}/assignees`。
 - ✓ 该 `PATCH` 带 `state` 关卡/重开,`state_reason` 交付 `completed`、撤单 `not_planned`,走裸 REST。

scripts/pm/check-handover.mjs

@@ -30,8 +30,8 @@
  *                         newer than it, and whose state labels (`PM_EXCLUSIVE_STATE_LABELS`, the six
  *                         mutually exclusive states of `SKILL.md`) are anything but `needs-user-decision` alone.
  *
- * The sign-off hand-back rule (`seat-post-protocol.md`, 收班) names one refusal inside (a) and narrows (a)'s
- * stay-behind (section below):
+ * The sign-off hand-back rule (`seat-post-protocol.md`, 交接收尾清单 step 2) names one refusal inside (a);
+ * `seat-lifecycle.md` 离任留守 narrows (a)'s stay-behind (section below):
  *
  *   (a′) NEEDS_DECISION   a card (a) holds whose NEWEST `os-dev-report` (`latestDevReport`, the patrol's newest-of
  *        HELD             reading) is newer than the claim and declares `"status": "needs_decision"`, while the card