fix(scripts): route check-closing-target-claim through the proxy, and make a transport failure speak

[os-elon-musk]

Fixes #18947

Clause-②: no

scripts/check-closing-target-claim.mjs read GitHub through node's global fetch
with no proxy route. In an agent container api.github.com is reachable only through
the session proxy — which is also where the credential is injected — and fetch
does not read the proxy variables, so every request arrived uncredentialled and
GitHub answered 401. This gate folded that 401 into a per-target UNDETERMINED at
exit 0: honest in prose, green in the one thing a derived-gate sweep records.

Two commits, in the order the triage asks for: the route, then the negative control.

Before / after / control — this container, transcripts

Exit codes captured with redirect-then-$?; no token is printed by any leg.

⚠️ One substitution, declared: where a probe's PR_BODY bound #18844, the closing
keyword is written CLOSING below. The real command used the ordinary spelling; it is
masked here so that quoting a probe does not make this PR body bind a second card. The
keyword next to #18947 is left literal — that is this PR's own target, already bound on
line 1.

BEFORE (d8b12fca9, the seat's own probe re-run at this base):

$ GITHUB_REPOSITORY=objectstack-ai/objectstack PR_NUMBER=18921 \
    PR_HEAD_REF=claude/issue-18844-probe PR_BODY='CLOSING #18844' \
    node scripts/check-closing-target-claim.mjs
::warning::UNDETERMINED - #18844 was not judged: its comment thread could not be read. ...
✓ check:closing-target-claim: PR #18921 closes no card this run could hold against a claim; ...
  1 closing target(s) could not be judged - see the UNDETERMINED warning(s) above.
exit=0

AFTER, same invocation — the re-exec line, then a real per-target verdict:

ℹ️  re-exec with --use-env-proxy: HTTPS_PROXY is set (http://127.0.0.1:37101) and node's fetch does not read it.
✓ check:closing-target-claim: PR #18921 closes no card this run could hold against a claim; ...
  #18844 was not judged: the card is already closed, so the merge closes nothing.
exit=0

⭐ The reading, reported and not steered: #18844 resolves to TARGET_CLOSED, not to a
finding. The classifying read was really spent (it is the second read, taken only on
the path about to go red), and it answered state != open. The declined-number policy
this file already documents is what declines it — the merge closes nothing, so there is
no second implementation to prevent.

AFTER, live acceptance pair against the real board, which is the stronger evidence
that the read reaches the thread at all — this card, this branch, and this card with a
branch nobody claimed:

$ ... PR_HEAD_REF=claude/issue-18947-closing-target-claim-proxy-route PR_BODY='Closes #18947'
✓ check:closing-target-claim: PR #18921 closes #18947, and each carries a `Claim:` whose
  `Branch:` line names `claude/issue-18947-closing-target-claim-proxy-route`.
exit=0

$ ... PR_HEAD_REF=claude/issue-18947-someone-else PR_BODY='Closes #18947'
::error::PR #18921 closes #18947 with no `Claim:` on the card naming `claude/issue-18947-someone-else`
  - `Closes #18947` - the governing `Claim:` on #18947 names
    `claude/issue-18947-closing-target-claim-proxy-route`, not `claude/issue-18947-someone-else`.
exit=1

CONTROL A — the negative control the triage requires: same command, HTTPS_PROXY /
https_proxy unset, so the request really fails at the transport:

❌ check:closing-target-claim: PREREQUISITE NOT MET - the board was not read.
   GitHub API 401 for /repos/objectstack-ai/objectstack/issues/18844/comments?per_page=100&page=1 - ⛔ not a verdict.
   ⛔ NOTHING was judged. ...
exit=3

⛔ No longer exit 0. 「修好之后,一次真实的传输失败必须能让门禁出声」 — it speaks.

CONTROL B — the network-error leg, live: HTTPS_PROXY=http://127.0.0.1:1 (a dead
port). This one also proves the hand-off is invisible in the exit codes: the parent
re-execs, the child refuses, and 3 comes back verbatim.

ℹ️  re-exec with --use-env-proxy: HTTPS_PROXY is set (http://127.0.0.1:1) ...
❌ check:closing-target-claim: PREREQUISITE NOT MET - the board was not read.
   GitHub API unreachable for /repos/.../issues/18844/comments?...: ECONNREFUSED - ⛔ not a verdict.
exit=3

The boundary between transport and truncation, and the line that draws it

The card's own docblock said a transport failure is 「declared UNDETERMINED rather than
an unhandled rejection」. That sentence is what this card turns around, for the
whole-board case only. Where the catch sat: three sites inside collect — apiOrNull
(the queue leg's PR read and the classifying issue read) and the try around the
comment-thread walk. All three swallowed everything.

The line that draws the new boundary is a single predicate, read at all three sites:

export function isBoardNotRead(error) {
  if (!(error instanceof GitHubApiError)) return false;
  if (error.networkError) return true;
  return error.status !== 404;
}

• Typed, never a message match. githubApi now throws GitHubApiError carrying the
  HTTP status, or networkError: true when fetch itself rejected. A plain Error from
  anywhere else — a fake transport in the self-test, a bug — is not a transport
  failure and keeps its old per-target reading, so no untyped throw can promote itself to
  a whole-board refusal. Every pre-existing self-test case that threw new Error('boom')
  is green unchanged.
• 404 stays per-target. It is the one status carrying information about the resource
  the path names: that number does not resolve for this token. So does a thread walked
  past MAX_COMMENT_PAGES. Both remain exit 0 + ::warning::UNDETERMINED — 「never
  clean, never an accusation」 — and both are pinned.
• Everything else is about the run. 401/403 (credential or permission), 429 (rate
  limit), 5xx (GitHub), or no connection at all. Those propagate out of collect, so
  nothing at all is judged, and the dispatch turns them into boardNotReadRefusal(error)
  — pure, { exit, lines }, driven offline by the self-test.
• 429 is a declared superset of the set named on the card (401/403/5xx/network). It
  is the same class — the board was not read — and routing it to UNDETERMINED would be
  exactly today's silence under a different number. Called out here rather than folded in
  silently. The card's claim comment also named 404 in a parenthetical; the ruling section
  names 401/403/5xx/network and explicitly keeps 「a target whose number cannot be
  read」 as UNDETERMINED, so 404 is per-target. That tension is resolved in the ruling's
  direction, and stated rather than discovered.
• ⛔ No ::error:: annotation on the refusal path, matching NOT MEASURED: the exit
  code is what makes the job red, and an annotation would name an author who caused none
  of it.
• Only a transport failure is refused. The dispatch re-checks isBoardNotRead and
  re-throws anything else, so a genuine bug in this gate stays an unhandled rejection
  where it is visible as one, instead of being filed under "GitHub was unreachable".
• Residual, stated in the header rather than discovered later: a token that can see
  nothing in this repository gets 404 on every path, so every target reads UNDETERMINED
  at exit 0. That is the same shape as a board of numbers that do not resolve and no
  status distinguishes them. What this gate can tell apart is what it does tell apart.

The route

The shape main carries today, not PR #18980's unlanded guard parameter: the shared
proxyRearmPlan is imported, proxyPlanEnv() folds this file's own guard name
(OS_CLOSING_TARGET_CLAIM_PROXY_REARMED) onto the shared name, and rearmThroughProxy()
is called once, only by a wired run about to read the API — never on --self-test,
never on a NOT WIRED run. A sibling instrument's guard must not suppress this one's,
which is pinned. On an Actions runner there is no proxy in the environment, the plan
never re-arms, and nothing about this gate changes there.

Exit register — before / after

code	before	after
0	judged, clean	unchanged
1	judged, finding	unchanged
2	NOT MEASURED — no usable context; "a usage, wiring or transport failure"	NOT MEASURED — no usable context; "a usage or wiring failure"
3	—	PREREQUISITE NOT MET — a request failed at the transport, the board was not read, nothing judged

3 is EXIT_PREREQUISITE_NOT_MET, imported from the half-state module exactly as
check-issue-citations.mjs spells it, ⛔ not re-declared here.

The CI consumer needs no change — measured, not assumed.
.github/workflows/closing-target-claim-guard.yml invokes the script with a bare
run: node scripts/check-closing-target-claim.mjs: no continue-on-error, no || true,
no if: on the step. A non-zero exit fails the step and reds the job, so exit 3 is a red
job with no workflow edit. That exit is also unreachable on a runner for a routing
reason: there is no proxy there, and the token is the job's own.

Self-test

	before	after
cases	80	105
batteries	10	12
SELF_TEST_BATTERY_FLOOR	10	12

The floor rose; no existing battery's pin moved. Two batteries added — the routing
decision (9 cases, offline, pinned exactly as the landed mapper pins it: set ⇒ route,
absent ⇒ none, flag present ⇒ none, own guard set ⇒ none, the patrol's guard set ⇒ still
routes, plus the two structural pins that the plan is imported and the hand-off is called
once before the first network read) and the transport boundary (16 cases: each refused
status, the 404 and untyped-throw exclusions, the fleet phrase, the quoted error, ⛔ not a verdict, exit 3 being none of 0/1/2, no ::error::, no ✓, the route named when a
proxy is configured, and both UNDETERMINED legs still at exit 0).

Reader test

A derived-gate run in a container reads a real per-target verdict from this gate; with
no proxy it reads PREREQUISITE NOT MET and exit 3, never a ✓ with a warning.

Held, by CONTROL A above and by the live acceptance pair.

Changeset

skip-changeset, measured rather than assumed. Every published package in this repo
declares files[], and every entry resolves inside that package's own directory
(dist, README.md, CHANGELOG.md; @objectstack/spec adds generated dirs and
src/**/*.zod.ts). scripts/check-closing-target-claim.mjs is at the repo root, under
no package.json at all (there is none anywhere under scripts/), so it is inside no
publishable package directory and no files[] entry can name it. The root manifest is
private: true. Positive control: a file that is published —
packages/spec/src/data/object.zod.ts — sits inside a published package directory and is
matched by that package's src/**/*.zod.ts entry. Nothing published moves.

Gates

Derived from this worktree with
node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack — 29
commands, the derivation's own first line naming this repo and commit. Every one was
run, exit code captured before any pipe, and the record reconciled back with --ran in
the COMMAND :: exit CODE form the reconciler asks for:

✓ dispatch-gates --ran: 29 derived famil(ies) accounted for — 29 run, 0 NOT-MEASURED
  (a DERIVED zero — all 29 recorded an exit code and none of them is 3).

All 29 exit 0, including this gate's own pnpm check:closing-target-claim (105 cases)
and pnpm check:pm-dispatch-gates (1849 cases, 756.6s on this box, run detached behind an
exit-capturing wrapper and waited on with tail --pid). origin/main moved during the
run (f347c793e); the derivation was re-run after a fetch and the 29-command set is
byte-identical.

Repo-wide lint, not a narrowing. eslint . --no-inline-config was run in full at
c1e174eba, after the final commit: 6858 files linted (eslint's own count from
--format json), 0 files with problems, exit 0. So no narrowing needs declaring here.

Local verification scope

• (1) No package is touched, so the dependency closure is empty and there is no build leg.
• (2) The changed file is a repo-root CI instrument in no package; its own test surface is
  its --self-test, run above.
• (5) The diff edits a gate script, so that script's own suite is owed on top of the
  derived families — here the two are the same thing: pnpm check:closing-target-claim
  is node scripts/check-closing-target-claim.mjs --self-test, and it is in the derived
  list.
• pnpm lint is a repo-scan CI owns; it fitted in the foreground here and was run whole
  (above), so there is nothing narrowed to declare.

Files changed

• scripts/check-closing-target-claim.mjs (only)

---

Generated by Claude Code

[os-elon-musk]

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: c1e174eba1fba07f9c03a545b3d15cbbc60ae0f4

① Derived judgments

• Surface: scripts/check-closing-target-claim.mjs only (+336 / −20), two commits (the route, then the typed transport boundary). Off the governed register; no published package moves (root scripts/ ships in no package's files[]); no contract's accept/reject set moves. No landing on the file since the claim (git log origin/main --since 2026-09-18T09:56Z -- <path>: none; origin/main at f347c793e).
• Route. The landed mapper shape of scripts/check-single-claim-paths.mjs (PR fix(scripts): route check:single-claim-paths through the session proxy so a seat can run it #18935) reproduced symbol for symbol: own guard OS_CLOSING_TARGET_CLAIM_PROXY_REARMED, proxyPlanEnv folding it under the shared name, rearmThroughProxy called once at the entrypoint and only for a wired run, the child's exit forwarded verbatim, --disable-warning=UNDICI-EHPA where the node can take it, every status line on stderr. proxyRearmPlan is imported in main's three-argument signature — no dependency on PR fix(pm): the proxy re-exec guard is the CALLER’s name, and it speaks when it suppresses #18980's unlanded guard parameter, which is the right call while that PR waits at the channel. The structural cases pin one call site and the call sitting before the first await collect(.
• Boundary. GitHubApiError (status, or networkError) is thrown by githubApi alone; isBoardNotRead = typed ∧ (network ∨ status ≠ 404); the predicate is read at the three swallow sites in collect (apiOrNull and the thread-walk catch) and re-thrown; the entrypoint refuses through boardNotReadRefusal → EXIT_PREREQUISITE_NOT_MET (imported, not re-declared), no ::error::, no ✓; an untyped throw stays an unhandled rejection, so a gate bug cannot file itself under 「GitHub unreachable」 — correct. 404 and the page-cap truncation stay per-target exit 0 + ::warning::UNDETERMINED, both pinned. 429 as a declared superset of the card's 401 / 403 / 5xx / network: accepted, same class (the board was not read); the claim comment's parenthetical 404 resolved in the ruling's direction.
• Seat probes on a scratch worktree at this head (the agent's worktree was already retired): the seat's own probe (PR_NUMBER=18921, a PR_BODY binding [finding] scripts/check-single-claim-paths.mjs reads GitHub through node's global fetch with no proxy re-exec, so with real PR context it exits 1 on GitHub API 401 inside an agent container — the gate can be read only in CI, while its sibling post-stamped already re-execs with --use-env-proxy #18844) routed → exit 0 with a real per-target verdict ([finding] scripts/check-single-claim-paths.mjs reads GitHub through node's global fetch with no proxy re-exec, so with real PR context it exits 1 on GitHub API 401 inside an agent container — the gate can be read only in CI, while its sibling post-stamped already re-execs with --use-env-proxy #18844 is closed ⇒ the merge closes nothing) after the re-exec line; live pair on [finding] check-closing-target-claim.mjs reads GitHub through bare fetch with no proxy route, so in an agent container every closing target reads UNDETERMINED behind exit 0 — #18844's defect one sibling over (scripts/check-closing-target-claim.mjs :660) #18947 → exit 0 naming the claimed branch, exit 1 ::error:: for claude/issue-18947-someone-else; CONTROL A (all four proxy variables unset) → exit 3, PREREQUISITE NOT MET — the board was not read, GitHub API 401; CONTROL B (all four set to http://127.0.0.1:1) → the re-exec line, then exit 3 ECONNREFUSED, the child's 3 forwarded verbatim. On main, CONTROL A reads ::warning::UNDETERMINED behind ✓ at exit 0 — the defect, reproduced before the fix was read.
• Self-test 80 cases / 10 batteries on main → 105 / 12 on this head (both seat-measured); roster floor 10 → 12; the two new batteries carry 9 + 16 cases exactly as the roster declares. Ablation (dev, report 5728764997 §H): collapsing return error.status !== 404 fails 4 of 105 and reverts CONTROL A to the old ✓ + warning at exit 0; restore proven by blob hash 0cb88e863edb, which the seat re-measured on the head (git hash-object = the diff's index line).
• Seat re-runs on the head: --pair 18998 exit 0; dispatch-gates --commands = 29, byte-identical to the dev's set; --ran over the dev's exit-coded results → 29 / 29, 0 NOT-MEASURED (a derived zero, none is 3); eslint on the file exit 0; closingKeywordTargets(body) = {18947 → Fixes} only, so the transcripts' masked CLOSING #18844 binds nothing.
• Seat's own probe slip, recorded: the first CONTROL B overrode HTTPS_PROXY alone and read exit 0 through the still-live lowercase https_proxy — a flaw of the probe, not of the PR; it surfaced the shared-plan reading under ③.

② Semver level

• skip-changeset is correct: root scripts/ ships in no package's files[] (the [finding] check-single-claim-paths 违反自己声明的两条契约:传输失败走 exit 1(该文件把 1 定义为「指控」),且 self-test 握手旗标在异步结论块之前同步置位 ⇒ 永远报不出「没走到结论」 #18940 review's 70-package scan holds; no package path is in the diff). Clause-②: no on the claim (5728366689) and at body line 3; PM_SWEEP_REPO=objectstack-ai/objectstack node scripts/pm/check-clause2-carriers.mjs --pair 18998 at 2026-09-18T10:33Z: exit 0, one live claim, no widening tell.

③ Boundary flags

• Dev flags: no open questions. Three declared readings (429 as a superset; 404 stays per-target; no ::error:: on the refusal) — each agreed above. Four out_of_scope items: (i) the no-access-token residual (every path 404 ⇒ every target UNDETERMINED at exit 0), stated in the header — noted, not a card: no status distinguishes it from numbers that do not resolve; (ii) check-half-states.mjs's own re-exec writing the shared guard — already [finding] 共享 re-exec 守卫名互相压制:一个继承来的 OS_HALF_STATES_PROXY_REARMED=1 让兄弟工具静默不走代理并回 401 —— 评论 5715705458 记的那次「凭据死了」1 小时事故,可被一个变量机械复现 #18939 / PR fix(pm): the proxy re-exec guard is the CALLER’s name, and it speaks when it suppresses #18980, agreed; (iii) the 11 unrouted root scripts/ readers all refuse non-zero — a negative reading, nothing to file, agreed; (iv) the dev asks the seat to append the ablation subsection to the PR body — declined: the body is the dev's once-written carrier, the ablation transcript already has a durable carrier in the report on the card (5728764997), which this record and the ACCEPT cite, and a seat PATCH of a PR body is a write no rule asks for.
• Seat-found, filed by the seat (number in the ACCEPT): the shared plan's reason names HTTPS_PROXY, read uppercase-first (check-half-states.mjs :19569), while node 22's --use-env-proxy honours the lowercase https_proxy first — measured on this head: uppercase live + lowercase dead ⇒ ECONNREFUSED; uppercase dead + lowercase live ⇒ routed. Imported text, not this PR's; p3.
• Gates: 29 derived / 29 run / 0 NOT-MEASURED at this head; check:pm-dispatch-gates 756.6 s detached with its exit captured; 0 MCP calls; 4 REST writes, the PR body written once.
• CI on this head at 2026-09-18T10:36Z: 21 success · 11 skipped · 2 in_progress — nothing red; the landing record reads it again. ⚠️ The landing channel is closed (this session's auto-mode classifier, record 5726459594 on PR fix(pm): a size refusal is a refusal — post-stamped stops reporting a real 422 as a missing route #18934) — this PR joins the six waiting for it.

Implemented-by: claude/issue-18947-closing-target-claim-proxy-route
Reviewed-by: session_01BTeBejoPUvRHN8WdAJC6oF

VERDICT: PASS

---

Generated by Claude Code

[os-elon-musk]

Landing record — every pre-check met, blocked at the channel (skills seat, session_01BTeBejoPUvRHN8WdAJC6oF) · 2026-09-18T10:52Z

Every check on c1e174eba1 completed with none red (35 runs: 24 success, 11 rostered skips; the last, Lint & Repo Gates, completed at 2026-09-18T10:51Z, read at 2026-09-18T10:52Z); the review of record is 5728825257 (PASS); --pair 18998 exit 0; ACCEPT 5728850738 on #18947. One file, scripts/check-closing-target-claim.mjs, and no other open draft on it. The landing act (POST …/pulls/18998/ccr/ready_for_review then PUT …/pulls/18998/ccr/auto_merge {"merge_method":"SQUASH"}) is not attempted here: the same act was refused by this session's auto-mode classifier on PR #18934 at 2026-09-18T07:03Z (record 5726459594 there) and a classifier refusal is 「no channel」 per references/rest-channel.md. The PR stays draft, ACCEPTED, and lands the moment a channel exists — this seat's once the maintainer allows the two CCR calls, or the maintainer's own hand. ⛔ Nothing here is a finding against the PR.

---

Generated by Claude Code

[os-elon-musk]

Landing record — landed on the seat's record through the CCR route (skills seat, session_01BTeBejoPUvRHN8WdAJC6oF) · 2026-09-18T14:29Z

Channel: the two allow rules of PR #19047 (67624b78d, merged 2026-09-18T14:20Z) on disk, and the session's permission mode switched by the maintainer from auto to manual confirmation at 2026-09-18T14:25Z (the auto-mode classifier had refused the same act four times, the last at 14:24Z with the rules already on disk); the act typed in the rules' literal-prefix shape. Pre-checks re-run at 2026-09-18T14:23Z: NOT governed (scripts/check-closing-target-claim.mjs); --pair 18998 exit 0; the review of record is 5728825257 (PASS) on the head c1e174eba1; CI on the head 21 success · 10 rostered skips · 0 red; mergeable_state clean; ACCEPT on #18947.

Acts: POST …/pulls/18998/ccr/ready_for_review HTTP 200 at 2026-09-18T14:27Z (timeline ready_for_review, draft: false read back); PUT …/pulls/18998/ccr/auto_merge {"merge_method":"SQUASH"} HTTP 200 at 2026-09-18T14:27Z (timeline auto_merge_enabled; the repository allows squash only — allow_squash_merge true, allow_merge_commit false — so the queue's merge is a squash whatever the response's merge_method field prints); armed and waiting for the queue's admission (timeline auto_merge_enabled; added_to_merge_queue follows when the queue admits it). The queue rebuilds the PR onto the current main and re-runs the subscribing workflows; the merge is the queue's, ⛔ not this seat's click. Residue on the merge: #18947 → pm:done, unassigned.

---

Generated by Claude Code