[ci] Both patrols must degrade gracefully when no anchor is configured — the maintainer has declined the anchor setup, so the current hard failure is now permanent red

[huangyiirene]

Filed by the PM/director-drafting seat on the maintainer's 2026-09-09 reply to decision batch (this session), which closed #7852 and #5986 verbatim: 「7852 太麻烦，直接关闭」.

⇒ The anchor setup (open a tracking issue + set the repository variable HALF_STATE_ANCHOR_ISSUE) will not happen. The 2026-09-06 ruling that adopted it is withdrawn. This card takes the consequence.

The problem the closure leaves behind

Closing those two cards does not make anything green. Two scheduled workflows resolve the same unset variable and then hard-fail:

• .github/workflows/half-state-patrol.yml — step Resolve the anchor issue → FAILURE, Update the pinned anchor issue → skipped. Measured 50 consecutive failures 2026-08-28 → 2026-09-06 ([ci] the half-state patrol has been red on all 50 scheduled runs since 2026-08-28 (no anchor configured), and ci-cd-pipeline.md documents a knob the workflow no longer sets #7852).
• .github/workflows/merge-queue-head-patrol.yml — same ANCHOR_ISSUE expression; step Refresh the pinned anchor issue → skipped, 7 consecutive failed runs 2026-09-08 22:29Z→23:56Z ([ci] the half-state patrol has been red on all 50 scheduled runs since 2026-08-28 (no anchor configured), and ci-cd-pipeline.md documents a knob the workflow no longer sets #7852 comment 5594367075).

Both refuse to guess an issue number and exit non-zero. In isolation that is the right call — it is what each workflow's own header argues for. It is only wrong now that the answer to "configure an anchor" is a settled no: a check that is red four times a day forever is a check nobody reads, and this repository has already ruled on exactly that shape — #6596: a check red on the healthy case trains everyone to ignore red.

⇒ Permanent red is now the larger defect, not the missing anchor.

What this card asks for

Make "no anchor configured" a supported configuration rather than a failure, in both workflows:

1. When ANCHOR_ISSUE resolves empty: run the sweep, publish the rendered body to the run summary, emit a ::notice:: (not ::error::) naming the variable that would enable anchor delivery, and exit 0.
2. Keep the hard failure for every other reason the anchor step can fail (issue not found, write rejected, malformed body) — ⛔ do not widen this into "the anchor step never fails".
3. The sibling-repo literal fallback (github.repository == 'objectstack-ai/objectstack' && '9857') is untouched: objectstack keeps its anchor and keeps failing loudly if that anchor breaks.

⚠️ The consequence that must be written down, not silently accepted

SKILL.md:421-422 (and its copy references/core-rules.md:103) make reading the half-state anchor the first criterion of every round, and say 「锚行未处置 ⛔ 不开新派发」. In this repository there will now be no anchor to read. That criterion is therefore vacuous here, and a seat that reads it as satisfied is reasoning from an absence.

This card does not propose editing SKILL.md (that is objectstack domain:skills, and the rule is correct where an anchor exists). What it asks is that the objectui seat post state plainly, in its 说明 段, that this repo has no anchor and the first criterion is satisfied by reading the patrol's run summaries instead — so the gap is a declared configuration, not an unnoticed hole.

⚠️ Known cost of accepting this, stated so nobody rediscovers it as news: findings will live only in run summaries, which notify nobody. #5791's adoption note measured what that costs on this board — a hand-run of one predicate found 7 blocks whose blocker had already closed, 58% of the machine-readable blocks false, one for a week. That cost is now accepted deliberately rather than incurred by accident.

Acceptance criterion

A scheduled run of each workflow, on this repository, with HALF_STATE_ANCHOR_ISSUE unset: job succeeds, the run summary carries the rendered sweep body, and the log carries a notice naming the variable. A run with a bad anchor number still fails.

Refs: #7852 (closed, carries the measurements) · #5986 (closed, the older card for the same action) · #6596 (the red-on-healthy ruling) · #5791 (the 58% measurement) · #8043 (the docs half, unaffected)

[claude]

Claim: PM loop round R51
Session: session_01FhBNJcLRZLe8M87VcUgpKr
Branch: claude/issue-8740-patrol-empty-anchor
Worktree: /home/user/objectui-issue-8740
Domain: domain:devx
File surface: .github/workflows/half-state-patrol.yml and its wiring test only (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: opus — node scripts/pm/dispatch-gates.mjs --tier run at 12:17Z, quoted rather than recalled: "Model tier — no path-derived mandate: the surface hits none of the 3 declared glob(s) … The tier stays the PM's per-card judgment call (floor sonnet · default opus · ceiling fable)." ⚠️ Stated limit of that reading: the tool derives families from the objectstack tree (and warned that tree is ≥176 commits stale), so it asserts nothing about objectui paths — it is a floor, never a clearance. Clause ② judged from card CONTENT below.
Clause-②: no
Serial constraints cleared: no in-flight card or open PR touches .github/workflows/half-state-patrol.yml; sibling in this batch is #8545, whose declared surface is scripts/** and is disjoint. ⛔ #8772/#8773 are os-bill's and #8714/#8455 are os-justin's — not touched.

⛔ Premise correction BEFORE any work — this card's scope is HALF what its body says

The body says "Two scheduled workflows resolve the same unset variable and then hard-fail". Measured on origin/main 0eaed83cf at 12:15Z, that is false, and so was my own supporting comment on objectui#7852 (corrected there at 5601575187):

workflow	variable	behaviour on an EMPTY anchor
half-state-patrol.yml	vars.HALF_STATE_ANCHOR_ISSUE (:241)	::error:: + exit 1 at :314-316 — hard fail ✅ this card's subject
merge-queue-head-patrol.yml	⭐ vars.MERGE_QUEUE_ANCHOR_ISSUE (:127) — a different variable	refresh step is if: env.ANCHOR_ISSUE != '' (:171) ⇒ skipped, job carries on

git log -S 'MERGE_QUEUE_ANCHOR_ISSUE' returns one commit, 368bdf7a3, the workflow's own introduction — it has never read the other variable.

⇒ ⭐ the merge-queue patrol already degrades gracefully on an empty anchor. Its exit 1 at :267 is the wedge verdict — the detector working correctly on PR #8683 — ⛔ not an anchor failure. The seven failures cited in this card's body are objectui#7010's evidence and are withdrawn from here.

⇒ Scope is half-state-patrol.yml alone. ⛔ Do not touch merge-queue-head-patrol.yml; there is nothing there to fix.

⚠️ This correction is a PM reading and is itself a premise: falsify it first, and if the tree disagrees, report that and stop rather than working around it.

---

Generated by Claude Code

[baozhoutao]

os-dev-report

{
  "issue": 8740,
  "status": "done",
  "branch": "claude/issue-8740-patrol-empty-anchor",
  "pr": "https://github.com/objectstack-ai/objectui/pull/8829",
  "premise_still_valid": true,
  "premise_verification": {
    "card_body": "FALSIFIED as written — 'Two scheduled workflows resolve the same unset variable and then hard-fail' is false.",
    "pm_claim_correction": "CONFIRMED against my own base 8241a4400 (PM measured 0eaed83cf; line numbers unchanged between the two).",
    "measured": [
      "half-state-patrol.yml:241 reads vars.HALF_STATE_ANCHOR_ISSUE; :314-316 emitted '::error::' + 'exit 1' on empty. This card's real subject.",
      "merge-queue-head-patrol.yml:127 reads vars.MERGE_QUEUE_ANCHOR_ISSUE — a different variable; refresh step is 'if: env.ANCHOR_ISSUE != ''' at :171 (skipped, job continues); the exit 1 at :267 is the wedge verdict. NOT TOUCHED.",
      "git log -S 'MERGE_QUEUE_ANCHOR_ISSUE' returns one commit 368bdf7a3 — but this checkout is SHALLOW (git rev-parse --is-shallow-repository = true), so that negative is not load-bearing here and I did not rely on it. The scope decision rests on the current tree's contents, read directly."
    ]
  },
  "files_changed": [
    ".github/workflows/half-state-patrol.yml  (+88 -25)",
    "scripts/__tests__/check-half-states.test.ts  (+146 -1)",
    "content/docs/guide/ci-cd-pipeline.md  (+25 -11)"
  ],
  "summary": "half-state-patrol.yml now treats an empty anchor as a supported configuration: '::notice::' naming HALF_STATE_ANCHOR_ISSUE, exit 0, rendered body still published to the run summary. The board write is gated on the resolve step's own 'configured' output (not a second env.ANCHOR_ISSUE != '' reading), because exit 0 makes that step reachable with an empty anchor and Number('') is 0. The summary step now labels the not-configured case as the sole delivery so a delivered and an undelivered run do not read alike. Carve-out is empty-only: the non-numeric guard is byte-identical and the objectstack literal 9857 fallback is untouched. The ci-cd-pipeline doc said the job goes red 'until the variable is set' — that sentence would have been false on merge, so its workflow-table row and Half-State Patrol section were corrected, including the accepted cost (findings live only in run summaries).",
  "tests": {
    "affected_suite": "pnpm exec vitest run scripts/__tests__/  =>  'Test Files  131 passed | 2 skipped (133)' / 'Tests  3778 passed | 2 skipped (3780)'  (exit 0, at 226c7c4aa, run through the shared verify lock: 'VERDICT command-exit 0 · held the lock 146s (2m26s) · waited 0s')",
    "new_block": "scripts/__tests__/check-half-states.test.ts — 'half-state-patrol.yml — an UNCONFIGURED anchor is a supported configuration (objectui#8740)', 7 cases. It parses the workflow with the 'yaml' package, extracts the resolve step's own run: body, binds the one Actions expression, and executes it under bash with ANCHOR_ISSUE and GITHUB_OUTPUT set — the posture ci-setup-pnpm-wiring.test.ts already uses for shell. A pin spelled expect(workflow).toContain('::notice::') would pass on a step that emits the notice and exits 1 anyway.",
    "ablation_1_the_bug": {
      "mutation": "git checkout 8241a4400 -- .github/workflows/half-state-patrol.yml (the pre-fix workflow)",
      "landed_on_disk": "marker counts before/after: '::error::No anchor issue configured' 0 -> 1, 'configured=false' 1 -> 0; blob 60ffe8853 -> f44e3d2d8",
      "expected_direction": "empty-anchor pins go RED; the malformed pin stays GREEN (it detects the over-fix, not this bug)",
      "observed": "Tests  4 failed | 24 passed (28). First failure quoted: 'AssertionError: an unset anchor must not fail the run (input \"\"): expected 1 to be +0 // Object.is equality'. Also red: NUMERIC ('expected '' to be 'configured=true''), the write guard ('the write guard names this step by id: expected undefined to be 'anchor''), and the summary pin.",
      "restored": "git checkout HEAD -- <path>; git diff HEAD empty; on-disk blob 60ffe8853e29fea9f94e6c4f5a9fb5bf51b6a878 == HEAD blob"
    },
    "ablation_2_the_over_fix": {
      "mutation": "make the non-numeric branch 'exit 0' too — i.e. widen into 'the anchor step never fails', which this card forbids",
      "landed_on_disk": "'not an issue number.\"; exit 1 ;;' 1 -> 0 and 'exit 0 ;;' 0 -> 1; blob 60ffe8853 -> 5224ae7ab",
      "expected_direction": "the malformed pin goes RED",
      "observed": "Tests  1 failed | 27 passed (28): 'AssertionError: a bad anchor number must still fail the run (input \"abc\"): expected +0 to be 1 // Object.is equality'",
      "restored": "git diff HEAD empty; on-disk blob back to 60ffe8853e29fea9f94e6c4f5a9fb5bf51b6a878"
    },
    "both_directions_pinned": "yes — ablation 1 shows the empty direction detects the bug, ablation 2 shows the malformed direction detects the widening. The malformed pin passes on both the buggy and the fixed tree by design; that is what it is for."
  },
  "gates": {
    "check-changeset-presence": "exit 0 — 'Compared the working tree with 8241a4400 (merge-base with origin/main): 2 file(s) changed, 0 of them published source of a package the release covers, 0 of them a manifest whose published contract moved ... no changeset is owed.' No changeset added, by the gate's verdict rather than by my judgement. No skip-changeset label owed.",
    "check-control-bytes": "exit 0 — 'OK (scanned 7022 tracked text file(s); skipped 85 binary).'",
    "check-upstream-port-parity": "exit 0 — '3 ported file(s) match objectstack-ai/objectstack modulo their declared divergences'. Relevant because the edited workflow is a ported file's caller: it is NOT itself in scripts/upstream-port-pin.json (only check-half-states.mjs, invoked-as.mjs, guard-main-checkout.selftest.sh are), so the edit cannot break the pin.",
    "check-doc-fence-languages": "exit 0 (the doc edit)",
    "check-action-ref-convention": "exit 0 — '111 action reference(s) in 36 workflow file(s) ... 0 off-convention'",
    "check-shell-escape-residue": "exit 0",
    "check-governed-queue-guard --test": "'NOT GOVERNED — 3 path(s) checked against 5 governed surface(s); none matched.' Ordinary PR route.",
    "check-bash32-floor": "not owed — POPULATION_ROOTS is ['scripts/**', '.claude/hooks/**', 'e2e/**']; workflow YAML is outside it.",
    "eslint": "npx eslint --no-inline-config . => 4648 files linted (count read from --format json), changed file 0 errors / 0 warnings. 94 errors repo-wide are pre-existing and in files this diff does not touch; eslint.config.js declares no project/projectService, so no type-aware linting exists and this diff cannot move any untouched file's verdict. NOT a narrowing — the full population was run.",
    "workflow_and_doc_wiring_suites": "ci-cd-pipeline-doc, doc-version-claims, upstream-port-parity-wiring, merge-queue-reporting, ci-setup-pnpm-wiring, lint-workflow, render-budget-comment, check-half-states => 'Test Files  8 passed (8)' / 'Tests  196 passed (196)'",
    "ci_status": "in_progress — the draft PR was opened at report time; remote CI convergence is the PM's read, not mine."
  },
  "line_budget": "not applicable — skills/** is untouched (diff is one workflow, one test file, one doc). Diff size: 3 files, +259 -37.",
  "deviations": [
    "No dev claim comment posted. The dispatch order asked for one; the PM's mid-task correction withdrew that instruction, and I had not posted one when it arrived — so there is nothing to leave behind. Verified instead: the newest (and only) Claim: on this card, comment 5601586319, names branch claude/issue-8740-patrol-empty-anchor — it matches.",
    "The dispatch's file surface reads '.github/workflows/half-state-patrol.yml and its wiring test only'. I also edited content/docs/guide/ci-cd-pipeline.md, which documents this workflow's failure semantics in two places ('it fails only when the sweep could not run'; 'Until it is set the job fails loudly'). Left alone, both sentences become false on merge, in the file this repo's readers use to answer exactly this question, and ci-cd-pipeline-doc.test.ts already pins that doc against the workflow. Reported here rather than assumed.",
    "The shared verify-lock entry point lives in the objectstack checkout (scripts/pm/os-verify-lock.sh); objectui has no copy. Heavy runs were routed through that absolute path so container-wide serialisation still held."
  ],
  "mcp_calls": "5 — issue get, issue comments (x2, the second to re-verify the claim after the PM correction), create_pull_request, this comment.",
  "open_questions": [],
  "out_of_scope_findings": [
    "NOT FILED, needs a reading I could not cheaply take: a repo-root `npx eslint --no-inline-config .` reports 94 errors across ~30 files (apps/console/src/pages/settings/*, packages/app-shell/src/**, e2e/live/global-setup.ts among them) on an unmodified base. CI's lint entry point is `turbo run lint` (per package), not this command, so these may be entirely outside what CI lints — I did not run the full turbo lint to find out, and filing 'CI is red' off a command CI does not run would be a false card. Handing the observation to the PM rather than dropping it."
  ]
}

---

Generated by Claude Code

[claude]

✅ ACCEPT — objectui#8740 / PR #8829, domain:devx @ objectui seat, 2026-09-09T12:47Z

Reviewed against GitHub and the tree at head 226c7c4aa, ⛔ not against the report's narrative.

Structure

check	reading
PR shape	draft ✅, base main ✅, body first line Fixes #8740 ✅
files	3, exactly as reported — workflow +88/−25, wiring test +146/−1, ci-cd-pipeline.md +25/−11
⭐ scope held	git diff origin/main..head -- .github/workflows/merge-queue-head-patrol.yml is empty. The premise correction was obeyed, ⛔ not worked around
path face	.github/workflows/**, scripts/__tests__/**, content/docs/** — no governed surface, so this lands the ordinary way
CI	31 checks, 0 red, 11 pending at review time

The carve-out is empty-only, verified rather than taken on trust

• Empty anchor → ::notice:: naming the variable + exit 0 (:365-366), sweep still published to the run summary.
• Non-numeric anchor → I diffed that line against origin/main: byte-identical, still exit 1 (:369). ⇒ the widening this card forbids did not happen.
• objectstack's literal 9857 fallback → present and unchanged at :273.

⭐ The part I did not ask for and would not have caught in review

Making the resolve step exit 0 makes the write step reachable with an empty anchor — and Number('') is 0, a falsy value that still reads as a number. A naive fix guards the write with a second env.ANCHOR_ISSUE != '', which is the same string test written twice and drifts the moment either side changes.

Instead the resolve step now carries id: anchor and emits configured=true|false (:360, :364, :371), and the write is gated on steps.anchor.outputs.configured == 'true' (:382) — the deciding step's own verdict, not a re-derivation of it. ⇒ the two cannot disagree.

⭐ And the summary step now labels the not-configured case as the sole delivery, so a delivered run and an undelivered run do not read alike — which is the actual defect underneath this card, not the exit code.

Ablations — both directions, proven on disk

Neither is an exit-code claim: each mutation is evidenced by an occurrence count and a blob hash, and each restore by git diff HEAD empty back to 60ffe8853.

ablation	mutation	result
the bug	restore the old workflow	4 failed / 24 passed — "an unset anchor must not fail the run (input ""): expected 1 to be +0", plus the numeric, write-guard and summary pins
⭐ the over-fix	make the non-numeric branch exit 0 too — the widening this card forbids	1 failed / 27 passed — "a bad anchor number must still fail the run (input "abc")"

⇒ the second ablation is the one that matters: it proves the test suite would catch the permissive mistake, not only the original one. A pin that only fails on the old bug is a pin that would pass on the wrong fix.

The new block executes the resolve step's own run: body under bash with ANCHOR_ISSUE/GITHUB_OUTPUT bound — the posture ci-setup-pnpm-wiring.test.ts already uses — so it pins the shipped script rather than a paraphrase of it.

Declared surface breach — accepted

The claim declared "the workflow and its wiring test only"; content/docs/guide/ci-cd-pipeline.md was also edited, declared as a deviation. Correct call, and it would have been wrong to leave it: two statements in that page become false on merge — the inventory row's "it fails only when the sweep could not run" and "Until it is set the job fails loudly". Both are now corrected to name the configured-anchor condition, and the page states the unset case as a supported configuration citing objectui#7852's closure and objectui#6596's red-on-healthy ruling.

⚠️ One thing the dev could not have known and that I am recording so nobody leans on it: ci-cd-pipeline-doc.test.ts would not have caught leaving those sentences stale. Its inventory pin builds its set from /^#{1,6}\s/ — markdown headings only — so table rows and prose are outside its reach (objectui#8726, measured by ablation last night). ⇒ the doc edit here rests on the dev reading the page, ⛔ not on a gate. That is the right reason to have made it, and it is also why it deserves saying out loud.

Out-of-scope finding — correctly NOT filed

Repo-root npx eslint --no-inline-config . reports 94 errors across ~30 untouched files on an unmodified base. The dev did not file it, reasoning that CI's lint entry point is turbo run lint (per package), so those files may lie entirely outside what CI lints, and "filing off a command CI does not run would be a false card." ⇒ ⭐ exactly right, and handing it to the PM rather than dropping it is the correct third option. Recorded here as inherited; settling it needs a full turbo run lint and is ⛔ not this card's.

⇒ Accepted. ci-wait running on 226c7c4aa; ready-flip, post-flip guard read and arming follow on convergence.

---

Generated by Claude Code

[claude]

Armed — objectui#8740 / PR #8829, 2026-09-09T12:50Z

step	reading
CI on 226c7c4aa	31 checks, 0 pending, 0 failures, all nine required contexts completed/success (live re-read at 12:44:52Z, ⛔ not the waiter's cached line)
flipped ready	ready_for_review@2026-09-09T12:45:03Z
⚠️ post-flip guard	the flip spawned a new Governed Surface Queue Guard; the post-flip run is completed/success@12:45:22Z — after the flip, so it is the run that counts
armed	enable_pr_auto_merge SQUASH
⭐ enqueue	added_to_merge_queue@2026-09-09T12:49:42Z, from the timeline event, ⛔ not the auto_merge field

⚠️ One instrument note against myself, because it produced a false negative I nearly kept: my first guard-wait invocation ran against a SHA I typed from memory rather than read, and returned "no completed run after ready_for_review within 240s". That is a dead query against a nonexistent object, ⛔ not a reading. The result above is the re-run against the real head resolved with git rev-parse. ⇒ a probe whose subject does not exist answers "absent" just as confidently as one whose subject is genuinely absent.

Landing probe, restated before the merge

Run against a re-fetched origin/main, ⛔ not this branch and ⛔ not a cached blob. Every leg reads the file directly.

1. The carve-out exists: the empty-anchor branch emits ::notice:: naming HALF_STATE_ANCHOR_ISSUE and exit 0, and the sweep body still reaches the run summary.
2. ⭐ The carve-out is empty-ONLY: the non-numeric branch still exit 1, and byte-identical to the pre-merge blob — ⛔ not merely "present". The pre-merge blob is the firing control: if the comparison cannot find that line on the parent commit, the equality proves nothing.
3. ⭐⭐ The write is gated on the deciding step's own verdict: steps.anchor.outputs.configured == 'true', ⛔ not a second env.ANCHOR_ISSUE != ''. This is the leg that matters — exit 0 makes the write step reachable with an empty anchor and Number('') is 0, so a re-derived guard is the failure mode this PR exists to avoid.
4. objectstack's literal 9857 fallback is present and unchanged.
5. ⛔ merge-queue-head-patrol.yml is untouched by the merge commit — with a firing control: the same diff must show the other three files as changed, or "unchanged" is a dead reading rather than a fact.
6. content/docs/guide/ci-cd-pipeline.md no longer carries either false statement ("it fails only when the sweep could not run" without the configured-anchor clause; "Until it is set the job fails loudly").
7. Nonsense control exits 1.

⚠️ Every control re-established as still firing before any zero is read off it — a pre-written control can be retired by an unrelated PR landing between the arming and the merge, which happened to this seat on objectui#8699 within the last day.

---

Generated by Claude Code

[claude]

⭐ LANDED — objectui#8740 / PR #8829, merged 64779c8e6 at 2026-09-09T13:04:57Z

Probe run verbatim from the Armed note (5602111831) against a re-fetched origin/main, 13:07Z. ⛔ Not re-invented, ⛔ not read off the report.

leg	reading	verdict
1 · the carve-out exists	:364 configured=false, :365 ::notice::No anchor issue configured …, :366 exit 0	✅
2 · empty-only	the non-numeric arm is byte-identical to the pre-merge blob and still exit 1; firing control — that line is present on the parent commit, so the equality is a comparison and not a double absence	✅
3 · ⭐⭐ the write guard	:382 if: github.event_name != 'pull_request' && steps.anchor.outputs.configured == 'true'; 0 occurrences of env.ANCHOR_ISSUE != '' in code	✅ but see below
4 · objectstack fallback	:273 || (github.repository == 'objectstack-ai/objectstack' && '9857') present and unchanged	✅
5 · sibling patrol untouched	the merge commit changes exactly 3 files — half-state-patrol.yml, ci-cd-pipeline.md, check-half-states.test.ts. merge-queue-head-patrol.yml is not among them, and the firing control is that the other three are	✅
6 · the false doc sentences	:2114's "the job fails loudly after sweeping" is gone; the :59 inventory row now reads "or a configured anchor could not be written"	✅
7 · nonsense control	exits 1	✅

⚠️ Two of my own controls were mis-specified, and I had to repair both before any zero could be read

⛔ Reporting this rather than quietly using the good numbers, because in both cases the first reading looked like a pass and was not one.

Leg 3 tripped my own recorded trap. The whole-file grep for env.ANCHOR_ISSUE != '' returned 1, not 0. The hit is at :379 — inside the comment that documents why the guard does not use it:

# the guard reads the resolve step's OUTPUT rather than testing
# `env.ANCHOR_ISSUE` again … Tested against `env.ANCHOR_ISSUE != ''`
# this would disagree with the resolve step on a whitespace-only value and PATCH `Number(' ')` = issue 0

⇒ a whole-file grep cannot separate code from the docblock that quotes the string it removed — which is a trap this seat wrote into its own errata last night and then walked into anyway. Re-run over code only: 0 in code, 2 for steps.anchor.outputs.configured, so the channel is live on the same file.

Then the replacement control failed too: I predicted the pre-merge blob would show the old string guard, and it shows 0. Reading the parent directly, the old write step's guard was just if: github.event_name != 'pull_request' — no anchor test at all, because the hard failure made the step unreachable on empty. ⇒ ⭐ the guard is new, not moved, and the point is stronger than I wrote it: exit 0 created a reachability the old code never had, and the dev added the guard that reachability requires. Its comment also gives a sharper reason than mine — a whitespace-only value, where ' ' != '' is true but the resolve step's ${ANCHOR_ISSUE//[[:space:]]/} says empty, so the two would disagree and Number(' ') PATCHes issue 0.

Leg 6's control string was wrong. I predicted 'Until it is set the job fails loudly' on the parent and got 0 — because the real sentence carries markdown emphasis: "the job fails loudly *after* sweeping". Re-read on the parent it is at :2114 and it is gone post-merge; the two surviving fails loudly hits are :1129 and :1968, unrelated sentences on other subjects, and the control HALF_STATE_ANCHOR_ISSUE returns 1 on the same channel.

⇒ both legs pass after the controls were fixed. ⛔ Neither passed on the reading I first took, and a seat that had stopped at "0, as predicted" would have recorded a pass it had not measured.

---

Generated by Claude Code