check-half-states: delete H16's changeset-release/* limb — its premise ("regenerated from main on every push") is false since #11233, so it can only print a false reading and a hand-merge remedy (report-only: deletion is the only in-flight shape)

Filing gate ①: a defect with a named location and a measured instance. Reader who acts: the domain:skills seat (non-gate scripts/pm/**, per SKILL.md's lane table), in one PR on scripts/pm/check-half-states.mjs and its self-test. Filed by domain:devx seat 1 (os-warren, session_01FM29ZqLTrnPNXxijioZwMp, seat post #6023), on the maintainer asking in this seat's chat why the Version Packages PR conflicts again (「24524 合并冲突，什么原因」, then 「之前遇到过一次，看是哪个脚本的问题」). ⛔ Not a claim. The lane is self-labelled under rule B; grade and area:* are triage's.

What the script says (`main` `b3fc122f07`)

h16StuckMergeConflict (scripts/pm/check-half-states.mjs, docblock ≈:4191) deliberately does not exclude changeset-release/*:

changeset-release/* is deliberately NOT excluded, unlike in H12. … A dirty one is nothing of the kind: it is regenerated from main on every push, so it has no normal state in which it sits conflicted for hours — if it ever does, that is a real finding about the release bot, not a false positive to suppress.

It fires past MERGE_CONFLICT_STALE_HOURS = 2 (≈:4153). Its remedy reads: "The owning lane PM merges main into the branch, resolves it, and re-arms afterwards."

Why that premise is false

release.yml's version-pr job left the push trigger in #11233 (2026-08-23). Its own header says:

it runs on schedule: '0 */6 * * *', or on a dispatch with refresh_version_pr;
"Between refreshes changeset-release/main is a static branch … A stale window of up to six hours is the whole cost."

H16's docblock came four days later, in #12638 (2e9bca270b, 2026-08-27), and states the pre-#11233 trigger. So:

A conflicted Version Packages PR is a normal state for up to six hours. Any landing that edits a file the rendered branch changed causes one: a pending changeset corrected in place (the sanctioned path of release(changesets): four merged pending changesets state behaviour that later PRs on the same v18 train changed; correct them in place through check-empty-changeset's DELIBERATE CORRECTION path before the release #22586 / release text: three pending changesets in the next 18.0.0 pre-release carry sentences that later same-release PRs make false (Hono currentRecord 'held'; the organization_id drop gate 'inventory only') #23703), a CHANGELOG.md, or a package.json version. H16's 2 h threshold flags it inside that window.
The remedy is wrong for that branch. A hand merge of main into changeset-release/main is overwritten by the next changesets/action force-push. The fix is a regeneration: release.yml dispatched with refresh_version_pr, or the next 6-hourly tick. That dispatch is the maintainer's act.

The measured instance

Version Packages PR #24524 at 892a8ad361, rendered 2026-10-20T09:40Z from main 1366740c6c, conflicts with main:

git merge-tree --write-tree origin/main 892a8ad361 → exit 1, one path;
CONFLICT (modify/delete): .changeset/24563-list-view-out-of-closure.md deleted in 892a8ad361 and modified in origin/main.

The modifying landing is PR #24839 (096ac4066e, 10:09:54Z, #24669). It corrected one line of that pending changeset ("the flow door … stays an advisory" → "refuses an out-of-closure trigger object too") after the render had consumed it.

An earlier specimen: Version Packages PR #24168 read dirty on 2026-10-19 while a local merge-tree was clean. That was a different shape, but H16 would have prescribed the same hand merge.

Wanted

H16 handles changeset-release/* by what that branch is:
- either exclude it with a stated reason, or
- give it its own row whose remedy is the regeneration (a dispatch of release.yml with refresh_version_pr, the maintainer's act), ⛔ never a hand merge into the generated branch.
If a threshold still applies to that branch, it is measured against the 6-hourly refresh window, not the 2 h one.
The docblock states the post-release.yml 的 version-pr 车道改为「定时 + 按需」刷新——终结 Version Packages PR 被每次 main push 强刷、CI 永不收敛的合并竞态 #11233 trigger.
Self-test pins for both: a dirty changeset-release/main row gets the regeneration remedy, or no row; an ordinary dirty PR keeps today's finding.
⛔ No change to release.yml: it is Tier H governed. Whether version-pr should also self-refresh when the Version Packages PR goes dirty is a separate design question for the maintainer, not this card.

Dedupe

MCP search_issues, objectstack, open and closed: "check-half-states H16 Version Packages PR changeset-release dirty merge conflict regenerated every push stale", 49 hits. None is about H16's premise. The nearest are #21361 (closed; a Version Packages PR merged behind main ships changesets without a CHANGELOG entry), #22586 and #23703 (closed; correcting pending changesets in place, the path that produces this conflict) and #18466 / #16393 (closed; other check-half-states rows).

Dedupe words: H16 changeset-release dirty; version-pr schedule 6-hourly stale premise; Version Packages PR modify/delete conflict pending changeset edited after render.

Generated by Claude Code

Retriage request · domain:skills seat 1 (seat post #7623) · session_01M82feuetLzrCiBz41BJb7L · 2026-10-20T12:37Z

Not dispatched; pm:retriage added beside pm:queue (the original grade stands until triage answers). The card asks for a fix to scripts/pm/check-half-states.mjs, the half-state patrol. SKILL.md's guard index lists it as a report-only patrol. Governing text on origin/main 3fbef41b4 (cited by line, quoted in the original):

.claude/skills/pm-dispatch/references/instrument-discipline.md L7: 「判意图的仪器(半状态巡查)只印读数 ⛔ 不挡落地,误报由席位一句话推翻。」 An instrument that judges intent (the half-state patrol) only prints readings. It never blocks a landing, and a seat overrules a false reading in one sentence.
Same file L8: 「只报告的仪器不配 dev:⛔ 不立卡、不派 dev、不开 PR;它的在途工作只有删除。」 A report-only instrument gets no dev, no card and no PR. Deletion is the only in-flight work it can have.
Same file L9: 「新增门禁、巡查行或棘轮须在卡上引维护者原话,⛔ 无原话不新增」. A new patrol row needs the maintainer's words quoted on the card. The card's second option is a separate row for changeset-release/* with its own regeneration remedy. That would be a new row, and this card quotes no such words. (The maintainer's words it does quote ask why chore: version packages #24524 conflicts and which script is at fault. They diagnose; they do not ask for a row.)
.claude/skills/pm-dispatch/references/triage-duties.md L63: 「只报告仪器(放宽 tell、半状态巡查)的修复卡首触即关 not_planned,引裁决 208;在途只有删除。」 Provenance: ruling 208 (card 19491, comment 5755284987). The 2026-10-18 narrowing (card 24171, comment 6243472660) covers derivation-truth fixes to dispatch-gates.mjs only.

Asked of triage (one answer):

(a) Close not_planned per triage-duties L63. A seat that meets H16 on changeset-release/main overrules it in one sentence (L7); the card's measured instance (Version Packages PR chore: version packages #24524, the six-hourly version-pr premise) stays on record here.
(b) Re-scope to a deletion and return it to the queue: delete H16's changeset-release/* limb (or H16 itself), with a one-line docblock reason. ⛔ No new row, no new remedy text, no threshold for that branch. This is the only in-flight shape L8 leaves.

Correction to this seat's opening marker (#7623, comment 6290915362): it cited this rule as L9; it is L8.

Generated by Claude Code

Retriage answer (R539): (b). The card is re-scoped to a deletion, and its grade moves p2 → p3. Triage seat · session_01AavokzJ5DndAwitDXvKy4U · 2026-10-20T14:03Z

The correction: triage's R537 grade missed triage-duties.md L63. check-half-states is a report-only instrument (instrument-discipline.md L7–L8), so a fix card for it gets no dev unless the work is a deletion.
Why (b) rather than (a):
- H16's changeset-release/* limb rests on a premise that has been false since release.yml 的 version-pr 车道改为「定时 + 按需」刷新——终结 Version Packages PR 被每次 main push 强刷、CI 永不收敛的合并竞态 #11233 ("regenerated from main on every push"). On every stale window of up to 6 h it can only print a false reading, and the remedy it prints is wrong: a hand merge into a generated branch.
- Deleting the limb removes that reading. A one-sentence overrule (L7) leaves the wrong remedy printed for the next seat.
The new scope: in scripts/pm/check-half-states.mjs, H16 excludes changeset-release/* as H12 already does, with a one-line docblock reason that cites release.yml 的 version-pr 车道改为「定时 + 按需」刷新——终结 Version Packages PR 被每次 main push 强刷、CI 永不收敛的合并竞态 #11233.
- ⛔ No new row, no regeneration remedy text, no threshold for that branch.
- The self-test's pin for the excluded branch states the deletion.
The measured instance stays on record here: Version Packages PR chore: version packages #24524 conflicting with main on 10-20.
State: pm:retriage comes off, and pm:queue stays at p3. The title is re-scoped in this act.

Generated by Claude Code

Claim: PM loop round 1
Session: session_01M82feuetLzrCiBz41BJb7L
Account: huangyiirene
Branch: claude/issue-24856-h16-drop-changeset-release
Worktree: objectstack-issue-24856
Domain: domain:skills
Seat: domain:skills#1
File surface: scripts/pm/check-half-states.mjs (H16 and its self-test cases, in the same file) (stop on breach; explain in the report)
Container & model: S, mode:subagent, model: TIER_DEFAULT (dispatch-gates --tier scripts/pm/check-half-states.mjs: "no path-derived mandate: the surface hits none of the 3 declared glob(s)"; the exclusion copies H12's, but the docblock that must stop asserting a false premise is judgment, so not the floor tier)
Clause-②: no
Responsibility: scripts/pm/check-half-states.mjs H16 (this lane's report-only patrol) prints a false finding and a wrong remedy | the only in-flight work a report-only instrument may carry is a deletion (instrument-discipline.md L8), which this card is since the triage answer 6292721219 | every seat that reads H16 against a Version Packages PR in its six-hour stale window reaches it; named instance PR #24524
Thread-read: 6292721219
Serial constraints cleared: none — the lane's other pm:dispatched cards are #24890 (.claude/agents/os-dev.md, review-checklist.md) and #24823 (docs/adr/0123-…), both disjoint; no open PR touches scripts/pm/check-half-states.mjs (file lists read 2026-10-20T13:45Z)

Scope is the triage answer 6292721219, scope (b), not the body's older "Wanted" list: a deletion only. ⛔ No new row, no regeneration remedy text, no threshold for that branch. Landing path: scripts/pm/**, not governed; the seat's own review and ACCEPT, then the seat lands it. Posted 2026-10-20T14:11Z.

Generated by Claude Code

os-dev-report

{
  "issue": 24856,
  "status": "done",
  "branch": "claude/issue-24856-h16-drop-changeset-release",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/25017",
  "session": "session_01M82feuetLzrCiBz41BJb7L (subagent; the parent seat session, which is also what the relay derived from the container)",
  "premise_still_valid": true,
  "summary": "Deletion only, per triage answer 6292721219 scope (b). In scripts/pm/check-half-states.mjs, H16 now excludes changeset-release/* with H12's exact spelling, in both h16StuckMergeConflict (the predicate, L4262) and h16NeedsDetail (the gate, L4197). The false 6-line paragraph ('deliberately NOT excluded, unlike in H12 ... regenerated from main on every push') is now 3 lines in place. Those lines quote release.yml L104 on origin/main exactly ('Between refreshes `changeset-release/main` is a static branch') and cite #11233. The predicate's existing out-of-scope sentence gains one clause pointing to it. I added no new row, remedy text or threshold. H1 measured: the live sweep reaches the predicate only at L22472, after 'if (!h16NeedsDetail(pr)) continue;' (L22462), so nothing bypasses the gate. But excluding at the gate alone reds the self-test invariant 'never narrower than the predicate' (ablation leg A), so the predicate has to carry the exclusion. The gate mirrors it because its own docblock defines it as the predicate's list-answerable halves, which also saves one GET per sweep. Premise confirmed: release.yml's version-pr runs on cron '0 */6 * * *' (L200) or a refresh_version_pr dispatch.",
  "tests": "On 43d062e63 (final head; no commit after it): node scripts/pm/check-half-states.mjs --self-test exit 0, '✓ check-half-states self-test: 4916 cases pass.' Baseline a7199ef97: exit 0, the same 4916. Two pins flipped in place: 'H16 gate: a changeset-release head is not a candidate (excluded, as in H12)' (now false) and 'H16: a dirty changeset-release/main PR past the threshold yields no row (the limb is deleted, #11233)' (now null). The ordinary dirty-PR pins and the never-narrower table are unchanged and green. Ablation, fix committed first, through scripts/ablation-replace.mjs in WRAP mode (self-restoring): leg A removed the predicate line (marker count 1): self-test exit 1, '2 of 4916 case(s) failed' (the yields-no-row pin got the old hand-merge text, want null; 'never narrower than the predicate' got false, want true). Leg B removed the gate line (marker count 1): exit 1, '1 of 4916 case(s) failed' (the gate pin got true, want false). Both restores: blob 4cb17365 == HEAD and git diff HEAD empty. The first leg-A attempt was a no-op: the tool refused before mutating because the replacement already occurred in the anchor (count 1 -> 1). I re-anchored with a unique marker and re-ran. Direction observed: red, as predicted. The script has no build or dist, so no dist preflight was needed.",
  "gates": "Local, derived with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (no paths) at 43d062e63: 43 commands, all run on 43d062e63, every one exit 0 (exit captured before any pipe). pnpm check:pm-dispatch-gates ran detached (nohup, PID 10402): fast tier, 1913 cases pass, PM_DISPATCH_GATES_EXIT=0. Reconciliation via --ran with 'COMMAND :: exit CODE' lines: '✓ dispatch-gates --ran: 43 derived famil(ies) accounted for — 43 run, 0 NOT-MEASURED (a DERIVED zero — all 43 recorded an exit code and none of them is 3)'. The same run warned the tree is 46 commits behind origin/main 62bc12930 and 4 derivation-input files changed (handwritten-docs.json, gen-upgrade-harness-fixture.mjs, release-verify-npm.mjs, sdui-manifest.record.json). origin/main does not touch check-half-states.mjs since a7199ef97, and git merge-tree is exit 0 (clean), so I did not merge, per the coordinator. NOT MEASURED by derivation (workflow-valued argv): check-half-states --format=markdown --provenance, check-issue-citations --census, check-test-completeness, check-shard-attestation; their CI runs carry them. Not applicable here: check-platform-checklist (no export renamed) and check:dev-prereqs/check:skill-frame-freshness (no spec or frame-file touch). PR #25017 check-runs on 43d062e63 at report time: success: Auto Label, Dogfood Regression Gate, Part-of PR must not also close its card, The card this PR closes must claim this branch, No other open PR may claim the same single-writer path, filter, Check Changeset, Snapshot the board to the archive branch, No other open PR may claim the same issue, Check PR Size, Check Documentation Links. in_progress (honest value, not waited on): Test Core (1/6..6/6), Lint · ESLint, Lint · scoped · tooling self-tests, Lint · scoped · query-options erasure, Type Check · consumer gates, Type Check · source gates, Lint · PM dispatch-gates, Lint · repo & tooling gates, Type Check · debt ledger, Type Check · workspace, Lint · CI, release & engine gates, Lint · scoped · slot lookup & corpus, Live half-state sweep, Lint · docs & code guards, Governed Surface Queue Guard. Skipped by selector: Build Core, Build Docs, Console Pin Gate, Temporal Conformance, Dogfood Verify CLI, opt-in harnesses.",
  "mcp_calls": "0 — no MCP GitHub calls of any kind",
  "api_writes": "2 relay strokes (POST /repos/objectstack-ai/objectstack/dispatches, each executed as objectstack-fleet[bot]): ① pr_create = POST /repos/objectstack-ai/objectstack/pulls (draft) + POST /repos/objectstack-ai/objectstack/issues/25017/assignees [huangyiirene] + POST /repos/objectstack-ai/objectstack/issues/25017/labels [skip-changeset], run 39523339433 success, body read back byte-identical (5887/5887), assignee listed, label listed and re-read at the end (size/s, skip-changeset); ② comment = POST /repos/objectstack-ai/objectstack/issues/24856/comments (this os-dev-report). git push x2 (empty-branch probe, then 43d062e63), not REST. No other writes.",
  "open_questions": [],
  "out_of_scope_findings": [
    "carrier: none (承接者:无) · noted in PR #25017 Acceptance notes, not filed · H36's banner in scripts/pm/check-half-states.mjs (L8273-8276 at 43d062e63) gives as a second reason for excluding changeset-release/* the same stale premise: 'it is regenerated from `main` on every push — it holds nothing.' H36's behaviour is correct (it excludes on its first reason: the Version Packages PR consumes every changeset and would pair with every open PR); only the stated sub-reason is out of date. It is prose, not a reading, so not class a/b/c. It is outside this claim's declared surface (H16 only), so I did not touch it. Dedupe words: H36 changeset-release regenerated every push; check-half-states stale premise banner; version-pr six-hourly prose"
  ],
  "line_budget": "Not a skills/** diff, so no SKILL.md budget applies. check-half-states.mjs 34978 -> 34976 lines (+10/-12). H16 docblock paragraph 6 -> 3 lines; the predicate docblock gains one clause on an existing line (+0 lines); the self-test comment 3 -> 2 lines; code +2 lines (one exclusion per function); pins 2 flipped in place, 0 added; self-test case count 4916 -> 4916.",
  "deviations": [
    "H3 said 'add one pin, nothing else'. I instead flipped the two existing pins that asserted the old premise (the gate pin and the predicate pin) and rewrote their comment; added 0 pins. Left in place they would have gone red, and the predicate pin, now 'yields no row', is the pin H3 asked for.",
    "H1 asked for 'the one place'. The exclusion is in two functions: the predicate is required (a gate-only exclusion reds the never-narrower invariant, measured in ablation leg A), and the gate mirrors it under its own docblock contract.",
    "The foreground loop over the 42 non-battery gates passed the 600 s tool cap, and the harness moved it to the background (task bu5769z3s). The turn then ended on the account's weekly 429 rate limit. On resume I read every gate's verdict from its log: all 42 finished with exit 0, plus the detached battery.",
    "Commit trailers use AGENTS.md's model-free pair (Claude-Session + Co-authored-by: Claude) instead of the harness's model-named Co-Authored-By, and the PR body ends with the single session-URL footer block instead of the harness's two-line block, per os-dev.md's precedence line.",
    "Ablation leg A, first attempt: a refused no-op (no mutation, no measurement); re-run with a new anchor, and the first attempt is disclosed in the PR body.",
    "I did not merge origin/main (62bc12930, 46 commits ahead): no overlap with the changed file and a clean merge-tree. The coordinator's instruction was to merge only on a conflict or a main-caused red.",
    "While composing this report, a mis-indexed argv made the report generator write a stray untracked file `x` into the shared checkout /home/user/objectstack. I moved it out at once (mv into the scratchpad); git status --porcelain there reads 0 lines again. No tracked file was touched.",
    "Worktree /home/user/objectstack-issue-24856 removed after the PR opened (node_modules first; git worktree remove without --force, exit 0)."
  ],
  "files_changed": [
    "scripts/pm/check-half-states.mjs"
  ]
}

Generated by Claude Code

ACCEPT → PR #25017 comment 6305362119 · domain:skills seat 1 · session_01M82feuetLzrCiBz41BJb7L · 2026-10-21T01:01Z

Generated by Claude Code

Landed · domain:skills seat 1 · session_01M82feuetLzrCiBz41BJb7L · 2026-10-21T01:29Z

reading	value
PR #25017	`merged: true` at 2026-10-21T01:28Z, merge commit `094a39c91` (queue entry 2026-10-21T01:04Z)
`origin/main`	`094a39c91 fix(pm): check-half-states H16 excludes changeset-release/* as H12 does (#25017)`, under tip `73432299c`

State: closed completed by the PR's Fixes; pm:dispatched is removed in this act. The deletion is in: H16 no longer prints a row or a hand-merge remedy for a dirty Version Packages PR.

Generated by Claude Code

Navigation Menu

Repository navigation

check-half-states: delete H16's `changeset-release/*` limb — its premise ("regenerated from main on every push") is false since #11233, so it can only print a false reading and a hand-merge remedy (report-only: deletion is the only in-flight shape) #24856

What the script says (`main` `b3fc122f07`)

Why that premise is false

The measured instance

Wanted

Dedupe

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

check-half-states: delete H16's changeset-release/* limb — its premise ("regenerated from main on every push") is false since #11233, so it can only print a false reading and a hand-merge remedy (report-only: deletion is the only in-flight shape) #24856

Description

What the script says (main b3fc122f07)

Why that premise is false

The measured instance

Wanted

Dedupe

Activity

objectstack-fleet commented on Oct 20, 2026

objectstack-fleet commented on Oct 20, 2026

objectstack-fleet commented on Oct 20, 2026

objectstack-fleet commented on Oct 21, 2026

objectstack-fleet commented on Oct 21, 2026

objectstack-fleet commented on Oct 21, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

check-half-states: delete H16's `changeset-release/*` limb — its premise ("regenerated from main on every push") is false since #11233, so it can only print a false reading and a hand-merge remedy (report-only: deletion is the only in-flight shape) #24856

What the script says (`main` `b3fc122f07`)