fix(pm): check-half-states takes the checkout that SERVES the board, and refuses a foreign one

[os-tesla]

Fixes #19191

Clause-②: no

scripts/pm/check-half-states.mjs took three readings off the LOCAL checkout with plain
execFileSync git calls and no cwd, so each inherited process.cwd() while the board it reports on
is named by PM_SWEEP_REPO. The filer measured both directions on the same board with the same script
(#19191, comment 5740947564): with cwd in objectstack, H17's trigger-file index validated objectui's
candidate paths against objectstack's 8888 tracked files (objectui had 7945) and H57 read
NOTHING for the whole run — and the row it never read was hiding a scheduled lane dead through five
consecutive fires (objectstack-ai/objectui#10010). Internally consistent, externally wrong.

维护者速读(草稿)

改了什么 —— 给这个巡检工具加了一个环境变量 PM_SWEEP_CHECKOUT,用来回答「哪个本地检出服务于这块
board」。设了它,三处本地 git 读取(H17 的 tracked-file 判据、H57 的 workflow 文件、origin)就都在那棵
树上读;那棵树的 origin 不是被扫的仓时,整轮巡检响亮拒绝(退出码 3),⛔ 绝不静默换一棵树读。不设它,
行为与今天逐字一致 —— 只多了一句:H17 页脚现在写明它到底在哪棵树上读的(路径 + origin)。

为什么改 —— 这不是「跑错目录」的操作失误,而是一个没有症状的读数缺陷。跨仓巡检是常态(工具住在
objectstack,board 可以是 objectui),而 H17 的页脚正是在告诉派发席位「派发前请拿你的文件面和这张清单求
交集」。清单是拿另一个仓的文件列表验过的,只在另一个仓存在的触发文件被静默丢弃 —— 席位读到的是一份干净
清单,而那份干净是假的。H57 那半反而是诚实的:它明说自己什么都没读,代价是那一轮真有红行没人看见 ——
本 PR 的实测里那是两行,其中一行连卡都没提到。

风险与代价(含回滚) —— 风险面很窄:变量不设时是逐字旧行为(自测里有一条专门盯这件事),巡检
workflow 本身不设它,所以线上那条 lane 的行为不变。新增的唯一失败模式是「变量设错路径」,而那正好是本卡
要的那个响亮拒绝。回滚 = revert 这两个 commit,没有数据迁移、没有已发布面(scripts/pm/** 不随任何包
发布,故无 changeset)。

席位意见 ——

你要做的 —— 无需维护者动手。若希望巡检 workflow 把这棵树写明(目前不需要,因为 runner 的检出就是被扫
的仓),那是 .github/workflows/half-state-patrol.yml 的一行 env,已写在下面的 Acceptance notes 里,留给
单独一个由人合的改动 —— 本 PR ⛔ 不碰 workflow(#19259、#19225 正在改它)。

What changed, mechanically

• PM_SWEEP_CHECKOUT — one knob, beside PM_SWEEP_REPO and PM_SWEEP_CLOSED_FLOOR in --help,
  carrying the PATH of the checkout that serves the board. resolveSweepCheckout(env) resolves and
  trims it; whitespace is unset.
• One git read site. All three readers (readTrackedFiles, readRepoRoot, readOriginUrl) now go
  through a single gitRead(args, extra) helper that passes that path as cwd. That is the only
  execFileSync git call left in the file, and the self-test pins the count at one — so a fourth
  reader cannot be added later without the cwd. That is the card's mechanism assumption turned into a
  measured property instead of a belief.
• The refusal. checkoutPrerequisite(sweepRepo, env, originUrl) is a pure verdict in the shape
  reportPrerequisiteNotMet prints: null when the knob is unset (today's behaviour is not a
  prerequisite) or when the named tree really serves the board, otherwise the file's own exit-3
  PREREQUISITE NOT MET, named. It is answered FIRST in sweep() — ahead of the transport probe — so a
  foreign checkout costs zero requests and reads nothing.
• localCheckoutServes resolves the same way: the knob leg first (it names the tree, so the
  origin read IN that tree is the authority), then GITHUB_REPOSITORY (which names the RUNNER's
  tree, i.e. exactly not the knob's), then the checkout's origin. H57 therefore judges when the knob
  names the right tree, instead of refusing for the whole run.
• The H17 footer names the tree it read — path plus origin — beside its oracle size, in BOTH
  oracle states (read, and EMPTY BY FAILURE), knob set or not. A wrong-tree reading is internally
  consistent; the tree it names is the only thing that tells it from a right one.

Readings

reading	before	after
scripts/pm/check-half-states.mjs lines	36,176	36,256 — net +80, budget +80
--self-test cases	5,063	5,081 (+18), exit 0
battery roster	6 batteries, each above its pin	unchanged, each above its pin

Derived gates: node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack at this
branch → 39 runnable commands, all run locally with the exit code captured BEFORE any pipe, all 0. The
per-command verdicts are in this card's os-dev-report comment. Control-character self-scan over the
changed file: no match. No package contains scripts/pm/**, so no dependency-closure build and no
package test suite is owed; repo-wide scans stay CI's.

The firing pair, live

Both legs ran at this branch's head against the objectui board, from the objectstack tool.

Knob = the tree that serves the board — PM_SWEEP_REPO=objectstack-ai/objectui with
PM_SWEEP_CHECKOUT=/home/user/objectui, detached under the shared heavy-verify lock (held 398s, waited
0s — shared-box seconds, not idle-box figures): exit 0, 502 half-states over 374 open
pm-/p0-labeled issues. The H17 footer:

(read on 75 of 75 open `pm:on-hold` card(s); 8142 tracked file(s) in the oracle.) Read in `/home/user/objectui` (`origin` `https://github.com/objectstack-ai/objectui`).

8142 is objectui's LIVE tracked count today (git -C /home/user/objectui ls-files | wc -l = 8142); the
filer read 7945 on 2026-09-19 and the board has moved since. This worktree reads 9031, so the two trees
are still 889 apart — the number that never used to appear is now the one printed, with the tree it came
from beside it.

H57 is judged rather than UNRESOLVED: 10 workflow(s) on the swept repo declare a schedule; 8 were judged against their latest event=schedule run and 0 are UNJUDGED rather than clean because that read failed. It filed two red rows, neither reachable by any wrong-cwd sweep:

• .github/workflows/changeset-release.yml — latest scheduled run 35493590744, started
  2026-09-20T06:11:39Z, concluded failure. This is the lane the card's re-run traced to
  finding(release): changeset-release.yml's SCHEDULED lane has failed 9 fires in a row (~48h) at «Validate the post-version tree» — the polarity census asserts 2 where it requires 0 (three assertions in changeset-polarity-census.test.ts), so publish and the version-PR refresh are skipped while every push fire stays green because that job is SKIPPED there objectui#10010, still dead today.
• .github/workflows/check-links.yml — latest scheduled run 35489310168, started
  2026-09-20T04:30:22Z, concluded failure. Not named anywhere in the card.

Knob = a foreign tree — same board, PM_SWEEP_CHECKOUT=/home/user/objectstack: exit 3, nothing
swept, nothing spent:

check-half-states: PREREQUISITE NOT MET — PM_SWEEP_CHECKOUT="/home/user/objectstack" does not serve `objectstack-ai/objectui`

  this checkout's `origin` is `objectstack-ai/objectstack` while the sweep reads `objectstack-ai/objectui`.

  H17 validates every on-hold trigger path against that checkout and H57 classifies its workflow
  files there, so a foreign tree renders an index whose paths were checked against another repo.

  Fix:  point PM_SWEEP_CHECKOUT at the checkout whose `origin` IS the swept board,
        or unset it and run the sweep from inside that checkout.

Reverse verification — both new pins can fail

Each ran from the COMMITTED state, mutated through scripts/ablation-replace.mjs (which proves the
write landed on disk and restores byte-exactly), and each turned exactly ONE case red out of 5,081.
The committed blob is f49d566ed276.

ablation	mutation landed	case that went red	restore
strip the cwd from the one git-read site	anchor 1 to 0, blob f49d566ed276 to 994882b2fcfc	#19191 knob: …and that one site takes the knob as its cwd — got false, want true	blob back to f49d566ed276 == HEAD, git diff HEAD empty
re-spell readRepoRoot with its own execFileSync git call — a fourth unrouted reader	anchor 1 to 0, blob f49d566ed276 to ecd24ab63593	#19191 knob: ONE git read site in the file, so a fourth reader cannot skip the cwd — got 2, want 1	same

The first is the one that mattered most: before that pin existed, stripping the cwd left the whole
5,080-case suite green, i.e. the fix's central wiring had no test that could fail. The second is what
makes the card's "the three readers are the ONLY local-tree reads" a property the suite enforces.

Acceptance notes

• Knob name: PM_SWEEP_CHECKOUT. It joins the established 12-member PM_* PM-tooling family
  (PM_SWEEP_REPO, PM_SWEEP_CLOSED_FLOOR, PM_STATE_LABELS, PM_TOKEN, …) rather than the
  product's OS_{DOMAIN}_{NAME} runtime family of Prime Directive [WIP] Create a new release version #9 — the card asked for a knob
  "beside PM_SWEEP_REPO", and --help groups the three under one heading where a reader looks.
  Flagged because that directive's wording is categorical; a maintainer who prefers
  OS_PM_SWEEP_CHECKOUT gets it for the price of one rename.
• The patrol workflow needs no env line, and this PR writes none. On every real patrol fire the
  runner's own checkout IS the swept board (PM_SWEEP_REPO: ${{ github.repository }}, per-repo
  installs), so the knob would be a no-op there and GITHUB_REPOSITORY already answers
  localCheckoutServes. If it should be stated explicitly anyway, the exact line for the sweep step's
  env: block in .github/workflows/half-state-patrol.yml is
  PM_SWEEP_CHECKOUT: ${{ github.workspace }} — a separate, human-merged change, since ci: wire the issue-citation verdict (blocking) and its census (report-only), plus the merged-result probe #19259 and
  ci(pm): make the half-state patrol callable instead of copied #19225 are open on that file.
• origin/main was NOT merged into this branch. It moved to e6a03e6 while this ran and touched
  scripts/pm/check-half-states.mjs in none of those commits (verified against a fetch into a ref this
  worktree owns, not against the shared origin/main pointer), so the line budget's 36,176 baseline
  still holds and no serial writer conflict exists. The merge queue rebuilds the PR onto current main
  and re-runs the required contexts there, which is where a jointly-wrong merge would surface.
• Noted, not filed: sweep() and sweepScheduledWorkflows each perform their own
  readOriginUrl() / readRepoRoot() — two duplicate local git reads per sweep (microseconds, zero
  requests). Threading one reading through sweepInto would widen that function's signature, which is
  the kind of change this card was told to keep out of. Successor: whoever lands finding(scripts/pm): H59 calls a deliberately REOPENED card a FALSE OPEN — state_reason: 'reopened' is on the payload it already reads, and H22 in the same file reads it #19177 (H59) or finding(scripts/pm): H52 has no clearing act on a pm:on-hold card — the only three null-exits are close, needs-user-decision (which the row itself forbids here) and a newer dev report a held card can never get #19160
  (H52) is next on this writer.
• Nothing else in the file was touched: check-half-states: the anchor's trim notice says the omitted rows are 「in the workflow run log」, but they are only in the step summary — unreachable from a seat, so trimmed H rows naming a lane go unworked #19230, finding(scripts/pm): H59 calls a deliberately REOPENED card a FALSE OPEN — state_reason: 'reopened' is on the payload it already reads, and H22 in the same file reads it #19177, finding(scripts/pm): H52 has no clearing act on a pm:on-hold card — the only three null-exits are close, needs-user-decision (which the row itself forbids here) and a newer dev report a held card can never get #19160, finding(scripts/pm): check-half-states.mjs reports its own MEASURED_MERGES_PER_DAY pin has drifted to a factor of 0.40 (observed ~54.7/day against a pinned 137.5/day, measured 28d ago), so every page-capped window OVER-states the days it covered #19203, [finding] a pm:blocked card whose Blocked-by target is a needs-user-decision card can never be released — the unlock predicate is "closed", and Unlock-action recognizes exactly one PR-shaped spelling #19255 and [finding] a claim whose Seat: line is not line-initial is silently read as seat 1 — three claims in one round misattributed, producing a false H38 stale row on another seat post #19108 queue
  behind this card on the same writer.
• scripts/pm/** publishes nothing from any released package, so this diff owes no changeset
  (Clause-②: no). No label writes.
• Tier S: this stops at the draft PR. The owning seat writes the ## Contract review record, reads
  --pair, and lands it.

---

Generated by Claude Code

[os-tesla]

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 6b1ff2cdc4f507e880d686a542f718b3831e5e6c

In-seat review by the domain:skills seat at the contract-review tier (served tier read at seating, #7623 marker 5738863635) — the lane's record for a scripts/pm/** landing (Tier S: the seat lands through the queue on this record + --pair 0 + every check green; the register lists no scripts/pm/** row — check-governed-merges --pr 19288: 「NOT governed — ordinary queue landing」 — so the record is the lane's own discipline). Clause-②: no on the claim (5748809579) and at line start in the PR body. Verified against GitHub (GET /pulls/19288/files: one file, scripts/pm/check-half-states.mjs +93 / −13; two commits ed059414 · 6b1ff2cd, trailer pairs model-free) and the diff read in full, ⛔ not against the report (5749138366).

① Derived judgments

• One knob, one read site: PM_SWEEP_CHECKOUT (beside PM_SWEEP_REPO / PM_SWEEP_CLOSED_FLOOR in --help) is resolved by resolveSweepCheckout(env); the three local reads (ls-files -z for H17's oracle, rev-parse --show-toplevel and remote get-url origin for H57) now go through a single gitRead(args) that hands the knob as cwd, and the self-test pins the site count at one (execFileSync('git' ×1, spawnSync('git' ×0) so a fourth unrouted reader cannot return. Unset ⇒ cwd: undefined, today's behaviour byte for byte.
• Refusal, never a silent read: checkoutPrerequisite(sweepRepo, env, originUrl) answers first in sweep(), ahead of the transport probe, with the file's exit-3 「PREREQUISITE NOT MET」 shape when the knob names a tree whose origin (read IN that tree) is not the board — headline names the checkout, detail names the repo it actually is, fix says point it right or unset it; a knob path with no origin at all refuses too. localCheckoutServes gains the knob leg ahead of GITHUB_REPOSITORY (the runner's tree, not the board's), so H57 judges cross-repo instead of reading nothing.
• The wrong-tree read becomes visible even without the knob: the H17 footer names the tree it read (Read in <root> (origin <url>)) beside the oracle size, in both oracle states, and says 「an unnamed tree」 when it cannot — the card's 「internally consistent and externally wrong」 hazard now prints its own discriminator.
• Firing pair measured live, both directions: PM_SWEEP_REPO=objectstack-ai/objectui with the knob = /home/user/objectui ⇒ exit 0, H17 「8142 tracked file(s)… Read in /home/user/objectui」 (objectui's live count today; the filer read 7945 on 2026-09-19; this worktree reads 9,031 — 889 apart), H57 judged (10 scheduled workflows, 8 judged, 0 UNJUDGED) and filing two red rows (changeset-release.yml, the objectui#10010 lane; check-links.yml, a second dead scheduled lane the card did not name); the knob = /home/user/objectstack ⇒ exit 3, nothing swept, no request spent. Self-test 5,063 → 5,081; ablation from the committed state — stripping the cwd reds exactly the pin written for it (the second commit exists because the first left the wiring untested), re-spelling a fourth reader reds the count pin; both restored by blob equality.
• Correct per the card and the grading: the three readers read the board's tree or refuse; nothing else in the file moved (check-half-states: the anchor's trim notice says the omitted rows are 「in the workflow run log」, but they are only in the step summary — unreachable from a seat, so trimmed H rows naming a lane go unworked #19230 / finding(scripts/pm): H59 calls a deliberately REOPENED card a FALSE OPEN — state_reason: 'reopened' is on the payload it already reads, and H22 in the same file reads it #19177 / finding(scripts/pm): H52 has no clearing act on a pm:on-hold card — the only three null-exits are close, needs-user-decision (which the row itself forbids here) and a newer dev report a held card can never get #19160 / finding(scripts/pm): check-half-states.mjs reports its own MEASURED_MERGES_PER_DAY pin has drifted to a factor of 0.40 (observed ~54.7/day against a pinned 137.5/day, measured 28d ago), so every page-capped window OVER-states the days it covered #19203 / [finding] a pm:blocked card whose Blocked-by target is a needs-user-decision card can never be released — the unlock predicate is "closed", and Unlock-action recognizes exactly one PR-shaped spelling #19255 / [finding] a claim whose Seat: line is not line-initial is silently read as seat 1 — three claims in one round misattributed, producing a false H38 stale row on another seat post #19108 untouched); no workflow file (the patrol's env line, if wanted, is stated in the PR body for a separate human-merged change).

② Semver level

None — scripts/pm/** publishes nothing; skip-changeset applied by the seat (precedent PRs #19173 / #19280) so this repo's Check Changeset exempts.

③ Boundary flags

• Net +80 lines in scripts/pm/check-half-states.mjs (36,176 → 36,256) — exactly the budget; the dev wrote the new docblocks tighter than the file's usual density to fit and says which facts were kept.
• Knob spelled in the PM_* PM-tooling family (12 members) rather than Prime Directive 9's OS_{DOMAIN}_{NAME} runtime family — accepted: the directive names runtime configuration; the sweep's siblings are PM_SWEEP_REPO / PM_SWEEP_CLOSED_FLOOR, and a third spelling family for one knob would be the defect.
• origin/main not merged into the branch (it moved 7ec8534 → e6a03e6 without touching this file — checked against a fetched ref, not the shared pointer); the queue rebuilds on current main.
• Gates: 39 derived families all exit 0 (codes captured before any pipe), including pnpm check:pm-dispatch-gates and pnpm check:pm-half-states; repo-wide pnpm lint left to CI. CI on the head at this record: 25 success · 7 skipped · 1 failure (Check Changeset, for want of the label above; re-read before the ready flip).
• out_of_scope_findings carried: sweep() and sweepScheduledWorkflows each read origin / root once more (microseconds, zero requests; the next dev on this writer — finding(scripts/pm): H59 calls a deliberately REOPENED card a FALSE OPEN — state_reason: 'reopened' is on the payload it already reads, and H22 in the same file reads it #19177 / finding(scripts/pm): H52 has no clearing act on a pm:on-hold card — the only three null-exits are close, needs-user-decision (which the row itself forbids here) and a newer dev report a held card can never get #19160 — is in that function); objectui's check-links.yml is a second dead scheduled lane, objectui's H57 now reports it every sweep (the domain:ui seat owning objectui#10010). open_questions: none.

Implemented-by: claude/issue-19191-check-half-states-cwd
Reviewed-by: session_01W5y9kRg1YtYaMQYExVLRc2

VERDICT: PASS

---

Generated by Claude Code