docs(pm-skill): the sixth readings increment — seven measured platform facts, ceiling 388 → 395 (ruled) by os-musk · Pull Request #16615 · objectstack-ai/objectstack

os-musk · 2026-09-07T12:37:54Z

Fixes #16566
Refs #14265
Refs #14014
Refs #13561
Refs #13326

Session (durable attribution in prose, because a PR body's trailing block does not survive an edit): https://claude.ai/code/session_018dxq7YqsLDMeZDZ5AzsgJX

Seven deduplicated readings land as one rule line each in references/platform-readings.md, one further reading REPLACES a row its own measurement falsified (line-neutral), and the ceiling row in scripts/pm/check-skill-line-ratchet.mjs moves 388 to 395 — the landed count — under the STANDING one-file exception in pm-dispatch SKILL.md, recorded as a ruledRaises record rather than a decision card, which is what that exception directs. was is untouched, no other ceiling row moves, and nothing is paid by re-wrapping.

Dedup ledger

Every candidate was checked against the ledger BEFORE it was written, by reading the neighbouring rows rather than by title.

#	candidate	verdict	why
1a	a listing returning a bare array and no total cannot announce its own truncation, so its zero is not a reading	landed	the neighbouring row covers only the full-first-page case for one endpoint; the response-shape property is the half that does not rot
1b	an existence question uses the single-point 200/404 read, or `git ls-remote --heads`, never a listing page	landed	the remedy half was absent everywhere in the file
2	`issue_write` clears every field it is not passed, assignees included	landed	the labels half is on the ledger; this is the other half, and it does not fit inside that line under the 120-byte rule
3	an MCP body edit sent with no footer of its own comes back with no platform-appended footer	landed, line-neutral	measured again on this PR (below); it falsifies the row that proposed the opposite mechanism, so it replaces that row rather than adding one
4 · #14265	a standalone marker line is dropped by the COMMENT read path with the bytes still in storage	landed	the recorded read-drop covers inline code spans only; the non-inference (absence is not evidence of a write-side strip) was not stated for that surface
4 · #14014	MCP's read limiter and write limiter are independent of each other	landed	the ledger has MCP-vs-GitHub bucket independence; read-vs-write independence inside MCP was absent, and both polarities are measured on the card
4 · #13561	assignability is a repo+account property that moves mid-session, with a 204/404 probe and a read-back	landed	nothing in the file mentions assignability as a probe-able property
4 · #13326	a control word's hit does not license a different term's zero — the defect is term-shaped	landed	the control-word rule is on the ledger as a NECESSARY condition; the measurement that it is not SUFFICIENT was not
4 · #13573	MCP `issue_read` truncates a long body silently; REST returns it whole	already-present	the two-read discriminant, the deterministic trigger and the empty-code-span detector are all on the ledger
4 · #13583	read PR labels via `pull_request_read get`, never `issue_read get_labels`	already-present	the three-leg row states it, and states more: leg 2 is itself unreliable, which supersedes the card's proposed rule
4 · #13141	the `get_labels` miss makes the mandated read-modify-write strip concurrent labels	already-present	both halves are on the ledger (the three legs, and the whole-set replacement with its stripping consequence)
4 · #13384	an issue body ending in a rule plus footer loses everything from the rule onward on create	already-present	four channel/action forms plus the create-append asymmetry are recorded; the single-PR relocation refinement is governed by the standing refusal to compress an observation into a law
4 · #13417	a contract-review PASS leaves no trace on the PR	refused	not a platform reading, and its landing file is `references/contract-review.md` — outside this flight's file surface; its own `Restart-when:` names the contract-review row, not the readings row
4 · #11742	the PR-lookup rate-limit signature, its window and the rate-not-budget knob	hold-not-met	its LAST transition comment (2026-09-04, director) moved the exit to "the discriminating `/rate_limit` reading exists" and states the hold stays until then; that reading is not on #15275, which is still open

Counts — candidates 13 · landed 7 · already-present 4 · refused 1 · hold-not-met 1. Eight lines written, seven of them new: 388 to 395. No line is deleted to pay for them, nothing is re-wrapped, and the one replaced row is replaced by its own refutation at equal line count.

⚠️ Stated for the ACCEPT rather than buried: the hold cards' Restart-when: lines name corpus audit #13597, and #13597 is still OPEN (read this run). Landing them rests on the dispatch's ruling that Phase 2's ruling on the readings row was executed by the rules-only rewrite, not on the literal text of those cards. #11742 is the one card whose NEWEST transition names a different, unmet gate, so it is held back.

The measurement candidate 3 required

Run on this PR's own body, through the same channel the fleet writes with, before the line was written.

step	what was sent	what read back
create (MCP `create_pull_request`)	body ending in the session-URL footer block	byte-identical, ONE footer, no appended block
edit (MCP `update_pull_request`, `draft: true`)	the same body with no footer of its own	no footer at all — nothing synthesised, nothing appended

⇒ Second independent reading, agreeing with the one taken on PR #16591. The ledger row that proposed the opposite mechanism ("the tool synthesises a footer only for a body carrying none") was explicitly marked 未实测; it is now measured, and false, so it is replaced by the reading.

⚠️ This disagrees with AGENTS.md, whose edit-side sentence says the session-URL footer survives and the platform appends a bare block under whatever you send. AGENTS.md is outside this flight's declared file surface, so the conflict is reported here rather than silently resolved — it is the seat's call whether AGENTS.md's edit-side sentence now needs the same correction.

The line #11742 would carry, if the seat rules its hold met

- MCP 变更 PR 的限流窗实测 ≥18 分钟,退避按 ~20 分钟;拒时主池读 idle ⇒ 按速率不按预算。

Gates

All 38 families derived by node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack from the real change set were run on the final head ca1fba4d2, and reconciled with --ran:

✓ dispatch-gates --ran: 38 derived famil(ies) accounted for — 38 run, 0 NOT-MEASURED.
✓ check-skill-line-ratchet self-test: 157 cases pass.
✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/platform-readings.md is 395 lines (ceiling 395; headroom 0).
✓ check-skill-line-ratchet: cross-file move into .claude/skills/pm-dispatch/references/platform-readings.md: +11 (314→395, less 70 lines of ordinary ruled raise) against a net source decrease of 20; authorised by #14685 item 5 (comment 5520452691).

pnpm --filter @objectstack/lint run check:doc-formula-expressions first exited 3 (PREREQUISITE NOT MET — @objectstack/formula unbuilt), which is NOT MEASURED rather than red; its dependency closure was built under the shared verify lock and it then exited 0. Repo-wide pnpm lint (eslint . --no-inline-config) exits 0 on the same head, run under the lock rather than narrowed. node scripts/pm/check-governed-merges.mjs --test on the two paths exits 3 (GOVERNED, one path on the register). A driver-less bare clone (merge.os-regen.driver absent, its config read exits 1) merge-trees this head against origin/main cleanly, and no open PR touches either file.

维护者速读(草稿)

改了什么 —— 平台读数事实表新增七行规则(一事一行),另有一行被它自己的实测推翻、原地替换(行数不增);棘轮上限 388 → 395(落地行数),并按常设例外在 ruledRaises 里记一条引用裁决原文的记录。没有删任何既有内容付账,没有折行重排,was 与其它文件上限一律未动。

为什么改 —— 七条都是席位这周真实踩过的坑:把截断的第一页读成"不存在"(差点让一个已完成的实现被重做)、单字段更新把 assignees 清空、控制词命中就以为空结果可信、指派静默失败、评论里的标记读不到就误判成写侧被删。事实表是这些教训唯一的家;写不进去,下一个席位照样踩。第八行更值钱:它把一条"未实测的猜想"换成了实测结论,那条猜想本身就会误导人。

风险与代价(含回滚) —— 代价是 skills 语料每席会话多读七行。风险低:全是描述性事实行,不改任何流程判据、不改门禁行为。回滚 = revert 本 PR 两个提交,账本行与棘轮行同笔回到 388,无残留状态。真正的判断点是那八行是否真的不重复 —— 上面的去重表逐条给了理由,请按理由核,不按标题核。另一处要拍板的是 AGENTS.md 的编辑侧那句话现在与实测冲突,改不改在本 PR 面外。

席位意见 —— (留空,待复核席位填写)

你要做的 —— 逐条核实去重表(尤其四张 already-present 与一张 hold-not-met 的判定)、确认抬上限走的是常设例外而不是需要新决策卡、决定 AGENTS.md 那句是否另开一卡,然后人工合并。⛔ 本 PR 不翻 ready、不入队、不挂 auto-merge。

Generated by Claude Code

…m facts, ceiling 388 -> 395 (ruled) Seven deduplicated readings land as one rule line each in the readings ledger, and the ceiling row moves to the landed count under the STANDING one-file exception (pm-dispatch SKILL.md), recorded as a `ruledRaises` record rather than a decision card, exactly as that exception directs. The lines, each written in the file's voice and placed in the section that owns its subject: - MCP's read limiter and write limiter are independent of each other, measured in both polarities, so one side's refusal does not predict the other's. - A control word's hit does not license a different term's zero: the same document is reachable by one term and empty by another in the same batch, so the defect is term-shaped, not document-shaped. - A listing that returns a bare array and no total cannot announce its own truncation, so its zero is not a reading — the neighbouring row already covers the full-first-page case, this one states the response-shape property. - An existence question goes to a single-point 200/404 read (`GET /repos/{o}/{r}/branches/BRANCH`) or `git ls-remote --heads`, never to a listing page. - `issue_write` clears every field it is not passed, assignees included — the labels half was already on the ledger, this is the other half. - Assignability is a repo+account property that moves mid-session, with its 204/404 probe and the read-back rule. - A standalone marker line is dropped by the COMMENT read path with the bytes still in storage, so its absence is not evidence of a write-side strip. Ceiling 388 -> 395 at the landed count, headroom 0, same convention; `was` is untouched and no other ceiling moves. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018dxq7YqsLDMeZDZ5AzsgJX

…place the untested row with the reading The row proposed a mechanism for the attribution-footer variation and marked it 未实测 with an explicit refusal to compress the observation into a law: the tool synthesises a footer only for a body that carries none. Measured twice now, on two different PRs, in the direction that mechanism predicts: a body edit sent through the MCP `update_pull_request` wrapper with no footer of its own reads back, through the same channel, with NO footer at all — no synthesised block, nothing appended. The mechanism is false, so the row that proposed it is replaced BY the reading, at the same line count. The neighbouring rows are untouched, including the one that refuses to generalise any single form across channels and actions, which still governs this line. ⚠️ This disagrees with `AGENTS.md`'s edit-side reading (session-URL footer survives, the platform appends a bare block). AGENTS.md is outside this flight's file surface; the conflict is reported rather than silently resolved. Ceiling unchanged at 395: this correction is line-neutral and buys nothing. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018dxq7YqsLDMeZDZ5AzsgJX

os-musk · 2026-09-07T14:21:59Z

维护者速读(终稿 · skills 席,2026-09-07T14:2xZ)

改了什么:references/platform-readings.md 新增七行平台事实(一事一行),另一行原地替换——那条标着「未实测」的猜想「MCP 只为无页脚正文合成页脚」换成两次实测的结论:它不合成;上限 388 → 395 = 落地行数,scripts/pm/check-skill-line-ratchet.mjs 的 ruledRaises 记一条引用常设例外原文的记录。七行是:MCP 读/写限流互相独立;控制词命中只证通道活着、换个词仍会回零;裸数组无总数的 listing 自报不了截断;存在性问题改单点读(GET …/branches/BRANCH 200/404 或 git ls-remote --heads);issue_write 清空未传字段(assignees 在内);可指派性是仓+账号属性且会话中可变(204/404 探针);评论里独立成行的标记被读路径吃掉、缺失不证写侧剥除。
为什么改:七条都是本周席位亲自踩过并已在卡上量过的坑(第 3、4 条差点让一个已完成的实现被重做;第 5 条今天让一张卡的 assignee 被清空)。事实表是这些读数唯一的家。抬上限走的是 SKILL.md 的常设例外(增量抬到落地行数、免决策卡、记 ruledRaises),条件是席位逐条核实与去重计数——本席已核。
风险与代价(含回滚):每次席位会话多读七行;全是描述性事实,不改任何流程判据或门禁行为。回滚 = revert 两个 commit,账本与棘轮同笔回到 388。真正的判断点是去重:席位逐条核了 13 条候选——落地 7(每条关键短语在 origin/main 账本 grep 零命中,被替换那行的短语作正控命中 1)、已有 4(邻行在 origin/main 第 194/199/196/298 行实读)、拒收 1([finding] A contract-review PASS leaves ZERO trace on the PR itself — verdict on the card, label cleared, no review object — so a legitimate un-park is indistinguishable from an intruder from the PR side #13417 的事实属 contract-review.md,且已在其第 22–23 行)、hold 未满 1([Decision] The shared GitHub identity's GraphQL quota is being burned to 2× — MCP writes go through GraphQL, so seats are silently write-blocked while reads keep working #11742 最新转换要求的 /rate_limit 判别读数尚不存在)。八行每行 ≤116 字节、零 issue 编号,ratchet 与 id-lint 亲跑绿,对 origin/main 8341ed21 merge-tree 干净,CI 33 项全部完成、零红。
席位意见:ACCEPT。一处保留意见不阻塞:AGENTS.md 编辑侧那句「平台会在你送的正文后追加裸页脚」现与三次实测(PR docs(skills): teach the calendar-hours rule where a hook stamps a deadline #16591、本 PR、PR pm gates: refuse a widening diff whose claim declares Clause-②: no (the mechanical half of the directional ruling) #16604)冲突,在本 PR 面外,席位另立一张 finding 交本车道自分诊。
你要做的:确认抬上限走常设例外无异议,然后人工合并(受管面),或批准后由席位入队。

Generated by Claude Code

github-actions Bot added size/s documentation Improvements or additions to documentation labels Sep 7, 2026

os-musk added the skip-changeset PR has no user-facing published change; bypasses the changeset gate label Sep 7, 2026 — with Claude

os-musk mentioned this pull request Sep 7, 2026

[PM seat] domain:skills — ⚪ vacant · last shift hotlong · session_01P58euzUXCVJNwmhuPC9DXY · closed 2026-09-08T10:25Z · landed 14 cards · queue 9 · decision box 2 · awaiting human merge 1 #7623

Open

os-musk requested review from hotlong and os-zhuang September 7, 2026 14:21

os-musk added the needs-user-decision label Sep 7, 2026 — with Claude

os-musk mentioned this pull request Sep 7, 2026

[finding] a branch-existence check with per_page=100 reads a truncated first page as absence — this repo has 1063 branches over 11 pages, and a PM takeover brief asserted "pushed no branch" about a branch that existed #16566

Closed

os-zhuang approved these changes Sep 7, 2026

View reviewed changes

os-zhuang marked this pull request as ready for review September 7, 2026 14:23

os-zhuang enabled auto-merge September 7, 2026 14:23

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

Merged via the queue into main with commit 4d72868 Sep 7, 2026
39 checks passed

os-zhuang deleted the claude/issue-16566-readings-increment-six branch September 7, 2026 14:50

hotlong mentioned this pull request Sep 8, 2026

docs(agents): PR-body footer rule states channel/action dependence, not a form #16771

Merged

Navigation Menu

Uh oh!

docs(pm-skill): the sixth readings increment — seven measured platform facts, ceiling 388 → 395 (ruled) - #16615

docs(pm-skill): the sixth readings increment — seven measured platform facts, ceiling 388 → 395 (ruled)#16615
os-zhuang merged 2 commits into
mainfrom
claude/issue-16566-readings-increment-six

os-musk commented Sep 7, 2026 •

edited

Loading

Uh oh!

os-musk commented Sep 7, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

os-musk commented Sep 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Dedup ledger

The measurement candidate 3 required

The line #11742 would carry, if the seat rules its hold met

Gates

维护者速读(草稿)

Uh oh!

os-musk commented Sep 7, 2026

维护者速读(终稿 · skills 席,2026-09-07T14:2xZ)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

os-musk commented Sep 7, 2026 •

edited

Loading