⚪ RELEASED · last seat session_019aCUUSwWefnbCJ4Xk1vqQW · R41+R42 closed 2026-09-03T16:50Z · 6 landed (#6467 #6600 #7185 + #5174 b9/b10/b11) · UNGATED_DOCS 32→26 · #3240 delivered, AWAITING HUMAN MERGE (PR #7523, governed) · queue: #5174 next batch, #6653 p2 · findings #7446 #7448 #7460 #7466 #7483 #7524 #7525

[os-zhuang]

Seat registration — domain:devx @ objectui. Lane scope and standing duties live in the pm-dispatch skill (references/lanes/devx.md); this post carries current state only. Readers: body + comments later than the body's last edit.

1 · Current PM

> ## ⚪ RELEASED — last held by PM session session_019aCUUSwWefnbCJ4Xk1vqQW (GitHub login os-sales), R41 + R42, 2026-09-03T04:31Z → 16:50Z
>
> Released on the maintainer's instruction (「当前任务处理完，合并后就下班」). The shift-end brief is the comment posted immediately after this edit; it is the release marker.

Rounds	R41 and R42 closed. R41 released at 07:40Z; the seat was re-taken at 07:50Z on the maintainer's instruction (「你应该自主继续派发」) because #5174 was sitting dispatchable in the queue — releasing with dispatchable work in the queue was the wrong call and is recorded as such
Dispatched	10 dispatches on 7 cards, all mode:subagent — #6467 (patch round + merge round), #6600 (×2, after a fable-quota kill), #7185 (diagnosis, then the clause), #5174 b9/b10/b11, #3240 (resumed once after a server 500)
Landed (merged)	6 PRs: #7332 (#6467) · #7449 (#6600) · #7452 (#7185, governed, human-merged) · #7458 (b9) · #7484 (b11) · #7485 (b10). Each verified on main by content, never by the merge event
Delivered, ⛔ NOT merged	#3240 → PR #7523, 106 files, CI 27 green / 0 red, draft, both approvers requested. Governed (AGENTS.md), so a human merges it. ⛔ Never nag
UNGATED_DOCS	32 → 26, and the .mdx half is now empty — all 26 remaining are package READMEs plus the root
Filed	#7446 · #7448 · #7460 · #7466 · #7483 · #7524 · #7525
Evidence added, ⛔ no grading	#5250 (three consumers, three mechanisms, discriminators for each) · #6653 (14/14 measurement; byte-exact body-write counter-measurement) · #7035 (wider metric-card measurement, recorded rather than filed as a near-duplicate)
Queue at close	#5174 pm:queue — next ledger batch, scope measured below · #6653 pm:queue p2 (triage re-judged it out of pm:retriage today) · #4853 pm:retriage · #6342 another identity's claim

2 · Inherited ledger

⚠️ objectui now has a central triage seat (maintainer ruling 2026-09-03, 「objectui 也纳入,分仓报数」). It re-judged #6653 out of pm:retriage and moved #5250 into needs-user-decision during this shift. The execution seat's "evidence only, ⛔ never grade" posture is now correct and has a counterpart — keep it.

Awaiting a human:

• PR test(repo): one vitest config, one verdict — delete the 17 per-package configs (#3240) #7523 (17 个包各自带 vitest.config.ts 与根 vitest.config.mts 行为分叉:先回答「哪些包真的需要本地配置」 #3240) — governed, draft, both approvers requested.
• A skip-changeset label exists on this repo, but no workflow reads it — the repo's own test records that neither the workflow nor the label was ever real #4912 — pm:awaiting-maintainer: delete the skip-changeset label. Independently re-confirmed dead twice this shift (two devs checked at the gate; a pin test forbids applying it).
• The half-state patrol needs its anchor issue created and HALF_STATE_ANCHOR_ISSUE set — a maintainer action outside GitHub's API surface for agent seats #5986 — pm:awaiting-maintainer: the patrol anchor is unset for this repo, so half-state-patrol.yml has nowhere to land and produces no H-lines.

Next UNGATED_DOCS batch (re-take the census on main, ⛔ do not inherit): takeable by density are plugin-chatbot, auth, plugin-calendar, react (~10 blocks each). ⛔ plugin-gantt (18, the largest) and plugin-kanban stay owned by #7302 / PARTIAL_EXCERPTS — plugin-gantt being biggest is exactly what a density-first instinct reaches for.

Lane facts — corrections and additions from this shift:

1. ⛔ "REST is 403, MCP only" is FALSE. Repo-scoped REST reads and writes return 200/201; only /search/* is 403, and that 403 is the local egress proxy, not GitHub ("sessions are bound to their configured repositories"). Three devs ran entire deliveries at zero MCP calls. ⚠️ The PM seat cannot do REST writes — the permission classifier blocks curl to the GitHub API — so PM writes remain MCP-only and the PM is the quota bottleneck, not the devs.
2. ⛔ MCP issue_read mangles bodies on the READ side; the stored body is intact. Byte-confirmed in two forms (angle-bracket fragments eaten, titles HTML-entity-escaped).
3. A PM seat CAN safely write a body line — measured twice, byte-exact — provided it never round-trips through the read channel: read raw over REST, inspect for tag-shaped fragments first, write, read back and diff. (Counter-measurement to objectstack#14898's stronger claim; recorded on finding(pm): 17 of 24 domain:ui pm:blocked cards carry no machine-readable Blocked-by: line — they are invisible to the unlock scan's reverse index #6653.)
4. Governed reviewer requests: pass draft: true explicitly in the same update_pull_request call and read back. Verified working on test(repo): one vitest config, one verdict — delete the 17 per-package configs (#3240) #7523 — draft survived. docs(agents): record how the Actions workflow registry is keyed, and the paginated-listing rule #6183 is the incident where it did not.
5. Shallow clone: git merge-base returns an empty string rather than failing; git rev-list --left-right --count then reads plausibly and wrongly. git fetch origin --deepen=500 first.
6. node --check passes a file whose block comment was truncated by a stray */. Verify config edits with a real import().
7. enable_pr_auto_merge prints both a populated and an empty confirmation for the same call — the output is not a signal. The queue ref is the receipt, it is transient, and auto_merge: not armed on a queued PR is the normal post-enqueue state.
8. Console eager-closure margin is ~9.5 KB (3181.9 / 3191.4). It moves with main, so a docs-only PR showing a delta is the base moving, not the diff.

3 · Hot-file serial queue

file / family	holder	note
scripts/check-doc-snippet-types.mjs (UNGATED_DOCS)	— free	Batches 10 and 11 ran in parallel and merged as a clean union. The licence is removals only — verified +0/-2 on both before release, and merge-tested against each other beforehand. ⛔ Removal-vs-insertion is not the case that fact was measured on (batch 9 was serialised behind #7449 for exactly that reason).
the three doc gates' roots	— (#7449 landed)	all three now walk apps/*/docs/**; the fences/snippets walk-equality pin still couples two of them
AGENTS.md (governed and hot)	PR #7523 (draft, awaiting human)	#7452 landed earlier this shift
every packages/*/vite.config.ts + vitest.config.*	PR #7523	the whole config surface is in one PR; ⛔ nothing else should touch it until that merges or closes

4 · Notes

#3240 is the shift's most consequential delivery and is not merged. Direction A was ruled 2026-08-06, and the card sat held from 2026-08-25 behind a scheduling condition — "board drained or near-drained, single dev, no other work in flight" — that existed only as prose in a comment. Nothing evaluates prose, so the card stayed held through periods when its own condition was true. R42 wrote that condition into the body as a machine-readable Restart-when:, measured it true, and dispatched into the ruled exclusive window.

⚠️ And immediately got it wrong once: the first version required "zero other pm:queue cards", which dropped the ruling's "or near-drained" and made the condition unreachable while #5174 cycles batches. An invisible condition replaced by an unsatisfiable one is the same defect in a machine-readable coat. Corrected the same day. Whoever backfills the other twelve should expect exactly that failure mode: transcribing prose into a predicate is where meaning gets narrowed.

The recurring discipline, and it was mostly the PM failing it. A control proves your probe is alive; it does not prove your scope is right. Four instances this shift, three of them mine: a filename-shaped grep that missed a pin test which exists; a guessed workflow filename returning zero with a control that also returned zero; a 403 read as GitHub's when its body said "local proxy"; and a count differenced against a stale comment's count, where the comment had simply miscounted. Devs caught three of the four.

The devs modelled the inverse repeatedly: three consecutive batches predicted in writing that a probe would come back green, then used that green to bound where a gate's checking starts — producing discriminators for three distinct "green about nothing" mechanisms (BaseSchema's index signature, a bare-any props parameter, and the labelled-block parse trap, whose tell turns out to be key count). #3240's dev went further: executing the ruling literally would have moved 14 packages onto the vite.config fallback route and re-opened the hole #5406 closed, so it closed that route first.

Standing constraints. Governed surface (.claude/**, docs/adr/**, skills/**, AGENTS.md, CLAUDE.md): draft only, request review from os-zhuang and hotlong with draft: true in the same call, ⛔ never ready / enqueued / auto-merged / approved by this seat. Release is human-only. Entry to the merge queue requires every check green, not the required subset. Closing strips pm:* on the card being closed; ⛔ no bulk cleanup of leftover pm:* on already-closed cards.