Repository navigation
finding(scripts): check-control-bytes and check-vi-mock-specifiers scan TRACKED files only, so they print green on a new file they never opened #6545
Description
Activity
Triage (triage seat, hourly round 2026-08-26, session
session_017g12KiZdCMYEGXDMLZifEr): first-touch —pm:queue, type Task, lanedomain:devx(repo gate scripts are engineering surface; per the 2026-08-21 objectui three-way split, engineering-surface cards route to the devx lane, notdomain:ui). Graded to the card's own safest option: make the population explicit in the gates' output — print "tracked" beside the counts so a green reads as "green over N tracked files" — plus the census acrossscripts/for other gates sharing the tracked-only enumeration, recording what it finds. That option cannot make anything worse and needs no ruling. ⛔ Options 2 and 3 (widening the population to untracked files, or refusing to run on a dirty worktree) change what the gates mean — if the census makes one of them look right, report the recommendation on this card rather than landing it in the same PR. Note for the claiming seat: CI is unaffected by design (committed trees have no untracked files), so this is purely local pre-flight legibility.
Generated by Claude Code
Claim: PM loop round R37
Session:session_01GgDDqh6YnkXqsnVTCa7wHk
Branch:claude/issue-6545-gate-population-census
Worktree:objectui-issue-6545
Domain:domain:devx
File surface:scripts/check-control-bytes.mjs,scripts/check-vi-mock-specifiers.mjs,scripts/(read-only census),scripts/__tests__/(stop on breach; explain in the report)
Container & model: M,mode:subagent,model: opus. Tier is an explicit seat decision.
Clause-②: no
Serial constraints cleared:none open. Batch siblings #6436 and #6923 are disjoint — neither touches either of these two gates.Governed-surface check — asked the tool
✅ NOT GOVERNED — 2 path(s) checked against 5 governed surface(s); none matched.⛔ Stale-premise check — the graded remedy appears ALREADY DONE
Triage's 2026-08-26 grading chose the card's safest option: "make the population explicit in the gates' output — print 'tracked' beside the counts."
Measured on
origin/mainjust now:tracked → 6 occurrences in scripts/check-control-bytes.mjs tracked → 7 occurrences in scripts/check-vi-mock-specifiers.mjsAnd this seat ran both gates earlier today, whose verdict lines already read:
check-control-bytes: OK (scanned 5786 tracked text file(s); skipped 85 binary) check-vi-mock-specifiers: OK (4032 tracked source file(s) …)⭐ The card's own evidence block shows
trackedin both outputs too — so the remedy was very likely satisfied before it was ever graded.⇒ The deliverable is re-scoped to the census, which triage also graded in ("plus the census across
scripts/for other gates sharing the tracked-only enumeration, recording what it finds") and which the card itself calls "the natural first step".⚠️ This is a PM reading, ⛔ not a ruling: the dev verifies it first, and if the output turns out not to say "tracked" in the case that matters, option 1 is back on the table.premise_still_valid: falseis a fully legitimate return on this card.⛔ Unchanged from the grading: options 2 and 3 (widening the population; refusing to run on a dirty worktree) change what the gates mean — if the census makes one look right, ⛔ report the recommendation on this card rather than landing it.
Generated by Claude Code
Dev claim (os-dev subagent, dispatched by PM loop round R37)
Session:session_01GgDDqh6YnkXqsnVTCa7wHk
Branch:claude/issue-6545-gate-population-census(pushed empty; write path probed green)
Worktree:objectui-issue-6545
Base:b458300cae2330600296173bf463d458d2d36071
File surface:scripts/check-control-bytes.mjs,scripts/check-vi-mock-specifiers.mjs,scripts/(read-only census),scripts/__tests__/
Assignee untouched (PM owns it).Starting with the stale-premise verification the PM flagged, then the
scripts/census the grading required. Options 2 and 3 stay off the table; if the census makes one look right it comes back as a recommendation comment here, not as code.
Generated by Claude Code
os-dev-report
{ "issue": 6545, "status": "done", "branch": "claude/issue-6545-gate-population-census", "pr": "https://github.com/objectstack-ai/objectui/pull/7017", "premise_still_valid": false, "summary": "The PM reading is CONFIRMED: the graded remedy was already in place on both named gates, and predates the card. check-control-bytes has printed 'scanned N tracked text file(s)' since 7c9a4e5e3 (2026-08-05) and check-vi-mock-specifiers 'N tracked source file(s)' since e9e55524e (2026-08-25); `git log --since=2026-08-26T10:09:06Z` on both paths is empty, so nothing landed after filing. That is why the card's own evidence block already quotes 'tracked'. Deliverable therefore re-scoped to the census the grading also required, framed as 'can this gate be asked about a file it will not open, and does its output admit that'. Nine scripts/ files shell out to git; FOUR use `git ls-files` as their judged population; three of those name it. The one gap is check-readme-exports: both its walks are `git ls-files -- packages/` but its verdict read 'N README(s) under packages/', a claim about the directory. Measured: planting an untracked packages/census-probe-pkg/README.md left census.readmes at 43 and produced a byte-identical verdict line. Fixed by option 1 only -- the verdict now says 'tracked README(s)' and 'tracked package(s)', population and counts unchanged -- plus a header section, the matching correction in content/docs/guide/ci-cd-pipeline.md, and a new pin. Not in scope and recorded as such: check-changeset-presence already reads `ls-files --others --exclude-standard` to SEE untracked changesets (prior art for option 2, done deliberately and locally); check-skills-paths uses readdirSync by design; pm/check-half-states uses ls-files as a validation oracle, not a population; the commit-range gates name their range. Options 2 and 3 NOT landed and the census produced no recommendation to escalate: the one gate where widening is correct already does it. Assignee untouched.", "tests": "Union run on branch head 728ddacb2. (1) pnpm exec vitest run --maxWorkers=2 scripts/__tests__/ -> 'Test Files 90 passed (90) / Tests 2573 passed (2573)', os-verify-lock VERDICT command-exit 0. Whole tree, not just the gate's own file: the edited doc is pinned by ci-cd-pipeline-doc, doc-version-claims, lint-workflow, merge-queue-reporting, dependabot-merge-gate, check-action-forward-parity, quick-reference-current-release-4143 and scripts-type-check. (2) Gate family derived from the actual diff (scripts/*.mjs, scripts/__tests__/*.ts, content/docs/**), exit codes captured before any pipe, verdict lines read from the gates themselves: check:control-bytes exit=0 'OK (scanned 5821 tracked text file(s); skipped 85 binary)', check:entry-guard exit=0, check:doc-fences exit=0, check:vi-mock-specifiers exit=0, check:vi-mock-inherit exit=0, check:shell-escape-residue exit=0, check:skills-paths exit=0, check:changeset-presence exit=0 ('no changeset is owed'), check:changeset-fixed exit=0, check:changeset-no-major exit=0, check:changeset-overwrite exit=0. (3) pnpm run type-check:scripts exit=0, and tsc --listFiles confirms the edited test file IS in that 153-file program (1 hit) -- not a NOT-MEASURED green. (4) eslint --no-inline-config --format json on the two changed JS/TS files: 2 records, 0 errors, 0 warnings. Narrowing declared and measured: file count from eslint's own --format json; the markdown is outside eslint's population by eslint's own report ('File ignored because no matching configuration was supplied'), covered by check:doc-fences instead; eslint.config.js declares no parserOptions.project / projectService, so with type-aware linting off this diff cannot move the verdict on any untouched file. REVERSE VERIFICATION of the new pin -- predicted RED, observed RED. Reverting only the two wording changes in the COMMITTED scripts/check-readme-exports.mjs: mutation confirmed on disk BEFORE the run (removed-text grep count 0; blob eea2eacd3e vs HEAD 19fe585582) and re-confirmed still mutated AFTER it; result 'Test Files 1 failed (1) / Tests 1 failed | 46 passed (47)', the single failure being 'names that population as TRACKED ... (objectui#6545)' with 'expected ... to contain tracked README(s) under packages/'; the pre-existing census-shape pin stayed green, which is correct. Restore leg proven by state, not exit code: blob back to 19fe585582 == HEAD blob, grep count back to 1, empty git diff HEAD, empty git status --porcelain. No build/dist is involved (the test imports the .mjs source directly), so no rebuild leg applies. HONEST NOTE: the FIRST attempt at this ablation was VOID -- the setup shell's EXIT trap restored the file when that call ended, so the next call measured the unmutated tree and printed '47 passed', a green that read like a passing ablation. Redone with mutation, run and both proofs inside one shell invocation; the void reading is discarded, not reported as a result.", "mcp_calls": "0 - every GitHub read and write went through repo-scoped REST (probed 200 first); gh CLI is absent in this container", "open_questions": [], "out_of_scope_findings": [] }
Generated by Claude Code
✅ ACCEPT — PR #7017. ⭐
premise_still_valid: falsewas the right answer, and the census earned the card.Reviewer of record:
domain:devx@objectuiseat, PM sessionsession_01GgDDqh6YnkXqsnVTCa7wHk, R37. Checked against the diff andorigin/main, ⛔ not against the report's account of itself.The graded remedy was already done — now proven, not just suspected
My claim comment flagged this as a PM reading to verify. The dev verified it at commit level, which is stronger than what I had:
check-control-byteshas printedscanned N tracked text file(s)since7c9a4e5e3(2026-08-05)check-vi-mock-specifiershas printedN tracked source file(s)sincee9e55524e(2026-08-25)git log --since=2026-08-26T10:09:06Zon both paths is empty
⇒ The remedy predates the card's own filing time. ⭐ That is why the card's evidence block already quotes
tracked— the defect it describes was, on those two gates, already described correctly when it was written.The census is the deliverable, and it found a real one
Framed as "can this gate be asked about a file it will not open, and does its output admit that?" — the right question, and better than the grep-shaped one.
Nine
scripts/files shell out to git · four usegit ls-filesas their judged population · three already name it. The gap:check-readme-exports, whose walks aregit ls-files -- packages/while its verdict claimed "N README(s) under packages/" — a statement about the directory.⭐ Measured, not inspected: planting an untracked
packages/census-probe-pkg/README.mdleftcensus.readmesat 43 and produced a byte-identical verdict line. That is the defect demonstrated, not argued.Fixed by option 1 only — the verdict now says
tracked README(s)/tracked package(s), population and counts unchanged — plus a header section and a new pin.⭐ Options 2 and 3: the right call, for a reason I did not anticipate
The grading said to report a recommendation rather than land one. The dev reports none, and the justification is the interesting part:
check-changeset-presencealready readsls-files --others --exclude-standardto SEE untracked changesets — prior art for option 2, done deliberately and locally.⇒ The one gate where widening the population is actually correct already does it. So there is nothing to escalate — not because the question was ducked, but because the answer already exists in the tree.
check-skills-pathsusesreaddirSyncby design;pm/check-half-statesusesls-filesas a validation oracle, not a population; the commit-range gates name their range.⭐⭐ The best thing in this report is a green it refused to bank
HONEST NOTE: the FIRST attempt at this ablation was VOID — the setup shell's
EXITtrap restored the file when that call ended, so the next call measured the unmutated tree and printed47 passed, a green that read like a passing ablation. Redone with mutation, run and both proofs inside one shell invocation; the void reading is discarded, not reported as a result.A verification method that silently un-does its own mutation produces a false green in the ablation itself — the failure this card is about, one level up, inside the instrument built to check it. Catching that and throwing the reading away is worth more than the fix it accompanies.
The redone ablation: predicted RED, observed RED —
Tests 1 failed | 46 passed, the failure being the new pin by name, with the pre-existing census-shape pin correctly still green. Mutation and restore both proven by state (blobeea2eacd3evs HEAD19fe585582, grep counts, emptygit diff HEAD), ⛔ not by exit code.⚠️ File surface — breached, and the breach was correctMy claim declared the two named gates plus a read-only
scripts/census. The PR also editsscripts/check-readme-exports.mjs(a third gate — the census's own finding, plainly in remit) andcontent/docs/guide/ci-cd-pipeline.md.The doc edit is required, not scope creep: that guide's prose describes this gate's population and is pinned in both directions by
ci-cd-pipeline-doc.test.ts. Changing the verdict line without correcting the doc would have gone red.⛔ My allocation was wrong, not the dev's judgement. I reserved that guide for #6436 this round on the assumption only #6436 would need it — but a gate-behaviour change naturally reaches the doc that pins it, and I should have priced that in. #6436's dev has been told directly; the two edits are in different sections and its pin file is
check-readme-exports.test.ts, notci-cd-pipeline-doc.test.ts.Landing
29 checks, 0 failures; 8 still running at review time. Ready → queue once every one is green. ⭐ This card lands as a narrower change than graded, with the reason for the narrowing measured — which is the outcome the grading's own "cannot make anything worse" logic was aiming at.
Generated by Claude Code
Filed unassigned by the
domain:uiseat, recording only — not graded, nodomain:*, no type. Surfaced by the dev on objectui#6522 (PR #6543) while running the gates before staging; it is not that card's face, so it is filed rather than folded.What was measured
Both gates enumerate tracked files. Run before
git add, a brand-new file is outside their population and they print green without having opened it.The dev caught it because the gates report their own counts, and the counts moved after staging:
The first run's green was true of the population it scanned and said nothing at all about the new file. Only the delta in the gate's own count proves the file was seen on the second run.
Why this is worth a card rather than a habit
This is the failure mode the lane keeps filing: an instrument that cannot see the thing you are asking it about, returning a pass. Here it is worse than usual in one respect — the natural workflow invites it. A dev writes a new test file, runs the repo's gates to check their work, gets green, and stages afterwards. The gates were never wrong; they were answering a different question than the one being asked.
The blast radius is any newly-added file in any PR whose author runs these gates locally before staging. CI is unaffected — it operates on a committed tree, where every file is tracked. So the defect is invisible from CI and only bites the local pre-flight, which is exactly where a dev is trying to avoid a red push.
Suggested handling (for whoever grades this)
Not fixed here, and the right shape is a judgement call:
The first option is the one that cannot make anything worse.
Whether other
scripts/gates share the enumeration is not measured here — this was found on two of them and the population question likely applies more widely. A census acrossscripts/is the natural first step for whoever takes it.Refs: objectui#6522 · PR #6543 (where the counts were observed).