feat(pm): close-cards --stateless acts on a card with no pm-state, and a half-written close resumes from the board

[objectstack-fleet]

Fixes #19954
Clause-②: no

scripts/pm/close-cards.mjs is the only file changed. It gains the two capabilities the card asks for, inside the existing one-card-one-dispatch shape. Maintainer, verbatim from the card: 「车队唯一的关单工具 close-cards.mjs 按设计只收带恰好一个 pm:* 状态标签的卡。 这个派卡处理。」 The spelling and the mechanism are the domain:skills seat's ruling at dispatch.

What changes

--stateless: a card that never carried a pm-state

• A mode FLAG, never an --expect-state value: --expect-state still accepts exactly the imported PM_EXCLUSIVE_STATE_LABELS, so --expect-state none is refused (its refusal now points at --stateless). The two flags together are usage exit 2 at the real door; --stateless=false is refused rather than read as ON; the default pm:queue is CLEARED under the flag rather than left beside it.
• Skip matrix: only the state row moves. Under --stateless a card carrying ANY pm-state (one or two; needs-user-decision included) is skipped, naming them; the assignee, pm:retriage, not-open and open-PR rows stand unchanged. An ordinary run's no-state skip now names --stateless.
• A stateless card packs comment → issue_patch: 2 actions, ONE dispatch.

Resume: only the steps the BOARD still owes

• Before anything is sent, every actionable card's thread is read (paged by number, like the timeline). A thread that cannot be read STOPS the run with zero writes on that card (exit 3 / 5), exactly as an unread timeline does.
• priorClosingComment: is THIS row's closing comment already on the card? The row's text is re-rendered by post-stamped's renderBody on the clock of each stamp the stored comment spells (protocolStamps), and judged against the stored bytes by post-stamped's pickRelayComment. No similarity of this file's own (not burstKey, which masks every digit run and would read two different rulings as one comment).
• residualOps: PACK_ORDER minus what the board shows. The comment unless this row's is there (never a second one); labels_remove while any label the close strips is still on the card; issue_patch always. packCard now demands exactly that residue; it still refuses a label plan that ADDS, and under an ordinary run still refuses a strip that does not take the expected state.
• Ordinary run: row 4 still requires the state, so the residue can only be the comment: labels_remove → issue_patch. Under --stateless both can be gone: the measured half-write shape (comment posted, state stripped, the close refused by the throttle) resumes as issue_patch alone.
• Record lines: RESUME — this row's closing comment is already on the board: comment ID (posted …, stamped … by its own act) … Still owed: …; the dry run prints WOULD CLOSE … (RESUME — comment ID already on the board); the summary counts N RESUMED.
• Step ④ is unchanged in kind: all three effects are read back every time. A resumed comment is read back off the board by id (the since window reaches back to it); a strip that was not owed reads as nothing surviving removal. "Nothing landed" (exit 5) is judged on the steps THIS run carried, so a failed resume or stateless pack is not misreported as a new half-write.
• The DIRECT transport follows the same residue: post-stamped is not spawned when the comment is on the board, and label-write is not called with nothing to strip.
• Still ONE card, ONE dispatch, ONE write on write-pace: a residual pack is fewer actions, never more writes. write-pace.mjs, check-half-states.mjs and package.json are untouched (grep: WIRED_WRITE_TOOLS names the file and nothing pins its usage or header).

One refinement the ruling's text does not spell, stated so it can be corrected: under --stateless the strip is still owed when a PM_RESIDUE_LABELS label that is not a state (today only pm:blocking) is on the card; that pack is comment → labels_remove → issue_patch. The header already defines the label step as "the pm-state plus its residue"; dropping the residue would leave a claim of work in flight on a closed card. Pinned by its own self-test row.

Evidence (tree ab2ee1ee)

Self-test

pnpm check:pm-close-cards → OK close-cards self-test: 227 cases pass across 20 batteries (base f7a3495a: 173 across 18). Two new pinned batteries (the stateless close, 13; the resume, 27); the CLI floor 14 → 22, the skip-matrix floor 14 → 21, the roster floor 18 → 20.

Three existing assertions were re-aimed, because the thread read is a new GET: the call sequence (now read, timeline, comments, read, comments); the read-back since count (filtered to since= calls); and the unreadable-read-back case, whose canned 502 now hits the SECOND comments read through a pass-through null in the fake's hook queue, with a new sent.length === 1 leg proving the dispatch happened. Without the re-aim that case went on passing while it tested the new thread-read stop instead.

The dispatch's self-test rows, each a case: stateless close = 2-action pack · half-write resume = 1-action pack · --stateless + a pm-state = skipped · two pm-states = skipped (both modes) · an assignee still skips · a dry run prints both packs and sends none · ordinary "comment there, state there" = labels_remove → issue_patch · --stateless + --expect-state = usage 2 (driven through main).

Ablations

One-shot, each through scripts/ablation-replace.mjs (anchor hit once, blob changed), each restored and proven: blob == HEAD c66f7a10, git diff HEAD empty.

leg	mutation	self-test
A1	the resume detector never matches	red, 16 cases
A2	the pack always carries the comment	red, 12
A3	"nothing landed" counts steps already on the board	red, 2
A4	the stateless skip row removed	red, 26
A5	an unread thread does not stop the run	red, 1
A6	the mode clash not refused	red, 3

Control after the six legs: 227 pass. Predicted direction for every leg: red; observed: red.

The live specimen, read-only

The resume detector against the stored bytes of the only measured half-write comment (5807380767 on #19334, read over REST; the row text rebuilt by putting {{NOW}} back where its stamp stands): found, stamp 2026-09-24T04:02Z. Control, the same text with one card number changed: not found.

Acceptance dry runs (read-only, zero writes; the probe comment files were never posted)

node scripts/pm/close-cards.mjs --repo objectstack-ai/objectstack --plan plan.txt --stateless --dry-run → exit 0:

#9180 ① 取现集 `tracking`, `domain:spec` → ② target `tracking`, `domain:spec` · strip none
#9180 ③ ONE dispatch, request fw-20260924T065047Z-7aa1d8: comment → issue_patch (2 actions · 352 of 65536 payload bytes)
#9180 → WOULD CLOSE `completed` — ⛔ nothing sent
#11592 ① 取现集 `tracking`, `domain:spec` → ② target `tracking`, `domain:spec` · strip none
#11592 ③ ONE dispatch, request fw-20260924T065048Z-158be8: comment → issue_patch (2 actions · 371 of 65536 payload bytes)
#11592 → WOULD CLOSE `completed` — ⛔ nothing sent
#15210 ① 取现集 `documentation`, `tracking`, `domain:spec` → ② target `documentation`, `tracking`, `domain:spec` · strip none
#15210 ③ ONE dispatch, request fw-20260924T065049Z-838228: comment → issue_patch (2 actions · 343 of 65536 payload bytes)
#15210 → WOULD CLOSE `completed` — ⛔ nothing sent
close-cards: DRY RUN — nothing was written. 3 read · 3 actionable · 0 skipped · 3 payload(s) packed, 0 sent

Control, the same plan without --stateless → exit 0, 3 read · 0 actionable · 3 skipped, each line SKIP carries no pm-state label, and this run acts only on pm:queue (a card carrying none closes under --stateless). --stateless --expect-state pm:queue at the real door → exit 2.

NOT MEASURED: live resume — reason: the only specimen was closed by hand at 05:22Z. The resume shape is proven by the self-test fixtures and by the read-only specimen probe above.

Gates

node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (no paths; change set from merge base f7a3495a4: 1 path) → 31 commands, every one run, every one exit 0; --ran → 31 derived famil(ies) accounted for — 31 run, 0 NOT-MEASURED. The dispatch's named gates are all in that set: check:pm-close-cards, check:pm-write-pace (113 cases), check:pm-dispatch-gates (1925 cases), check-scripts-symbol-anchors (and its self-test), check-self-test-wired (and its self-test), check-whole-set-label-write (and its self-test), check:nul-bytes.

Lint, narrowed and proven: eslint --no-inline-config --format json scripts/pm/close-cards.mjs → 1 file, 0 errors, 0 warnings. The file is in eslint's population (--print-config resolves 2 enabled rules for it, not an ignore), and no type-aware linting is configured (parserOptions.project is null), so this one-file diff cannot move any untouched file's verdict. Declared to CI: the path-scheduled Test Core job and the families the derivation prints outside its 31.

Acceptance notes

• .claude/skills/pm-dispatch/references/platform-readings.md still reads close-cards 每步一次 dispatch, stale since the one-dispatch-per-card landing. Not touched here: a governed surface outside this file surface. Carrier: none named.
• The comment above .github/workflows/lint.yml's Card-closure helper self-test step lists what the self-test instruments and does not mention the stateless mode or the resume: incomplete, not wrong. Carrier: none named.
• First use on [Ruling] The /meta type segment is ALWAYS singular — retire every plural spelling, no exception, in three steps #9180 · PageHeaderProps.actions declares z.array(z.string()) but objectui's canonical page:header renderer consumes ActionDef objects — satisfying the schema deletes every header button #11592 · record — the option-B reader by-shape sweep: method, derived 37-key collection set, triaged false positives, and the two packages it found #15210 is the seat's, after landing. This PR sent nothing to any card.
• No changeset: scripts/pm/** publishes nothing (skip-changeset).

维护者速读(草稿)

改了什么: 车队唯一的关单工具 close-cards.mjs 多了两种能力。① --stateless 模式:专门处理「从来没挂过 pm:* 状态标签」的记录卡(例如 tracking 卡),一张卡仍是一次 dispatch,只发「评论 → 关卡」两步。② 半写续接:动手前先读卡的评论线,如果这一行的关卡评论已经在板上(上次关到一半),就只补发还欠的步骤,永不再发第二条关卡评论。

为什么改: 本次召集里有五张维护者已批准关闭的卡,工具处理不了,只能手工补:#4606、#19334 已由维护者手工处理;#9180、#11592、#15210 还欠着。原因是工具按设计只收恰好带一个 pm:* 状态的卡,而且一次关卡半途失败后无法续接。

风险与代价(含回滚): 只改一个内部工具脚本,不发布任何包。新模式必须显式传 --stateless 才生效,默认行为不变;普通模式唯一新增的是先读评论线,读不到就停手、零写。自检 227 例全绿,6 次消融均按预期转红。回滚 = revert 本 PR。

席位意见:

你要做的: 无需操作。落地后由 skills 席用新工具处理 #9180、#11592、#15210 三张卡,并在卡上注明。

---

Generated by Claude Code