Context
An audit of extension/src/ for "regex applied to an LLM reply or a doctrine file, to extract a decision" found 24 sites. Three are already fixed or in flight (#397's parseEvidence, #404's PARK-REASON, #407's GRANT_PATTERNS). This issue covers the rest, ranked by whether a no-match yields a confident wrong answer rather than an obvious failure.
The shared property that makes these dangerous: the failure is silent, and the substituted default is then acted on as though it were read.
Tier 1 — a no-match produces a wrong confident answer
parsePerIssueVerdicts defaults to build-it — work-driver-plan.ts:275. On no match for an issue, falls back to the overall verdict; if that is absent too, NEEDS_WORK. This is precisely the pre-#378 "silence is permission" that work-driver-intent.ts:13-14 claims was structurally removed — it survives untouched on the multi-issue path, which #397 made the only multi-issue path. The reason string at :277-278 is honest ("(no verdict; defaulting to NEEDS_WORK)") and nothing gates on it.
labelApplied fails open — work-driver-handoff.ts:116, /label.*needs-human-attention/i over ops prose. An ops agent writing "I could not apply the label needs-human-attention" matches, so the gh fallback is skipped and the driver records success on an issue that was never labelled.
parseVerdict hands over the wrong text — adversarial.ts:376-380. Case-sensitive, no bold tolerance (unlike both siblings at work-driver-plan.ts:235 and work-driver-intent.ts:128), so Verdict: Approved misses. On a miss the default is ISSUES_FOUND and findings is set to the entire reply (:379) — so the fix-developer is handed the whole approval message as "findings to fix", and burns a round on it.
ci-status has no tolerance at all — work-driver-stepback-ci.ts:260-268, a bare text.includes(). ci-status:success, CI-status: success, or the marker inside a table all miss. No-match → "failure" → bumps ciRetryCount and can fire the ci-retry cap. A green CI reported in a slightly different shape burns the retry budget and parks the cycle. The #380 demotion check at :274-286 only runs when the status is already success, so it cannot rescue a false failure.
parseAdversarialRounds invents a number — work-driver-adversarial.ts:65-69. Three includes() on prose, defaulting to 3, asserted into the event log as fact with no "unknown" state. A phrasing change in adversarial.ts's own synthesised text silently makes every run report three rounds.
Grouping R1/R3/R4 — work-driver-grouping.ts:122, :171, :243. Two of these were structurally impossible to satisfy and were only found by accident (#376). Live evidence from nessie: #657 said "Closes the drift in #650" and R1's alternation has no closes/fixes/resolves, so two issues where one fixes the other ran as concurrent groups. And .container is absent from R2's extension list (:188), so selfhost/nessie.container — the one file both issues genuinely shared — contributed nothing to the Jaccard.
sliceSubsection / sliceMarkdownSection — work-driver-intent.ts:96, work-driver-plan.ts:398. Require an exact ^###\s+<name>\s*$. A resolver writing ### Deliverables:, #### Deliverables or **Deliverables** yields undefined → [] → specIsComplete false → park. Silent; nothing traces that a heading was not found.
Tier 2 — degrades quietly, failure at least structural
parseWorkstreams (work-driver-plan.ts:301) — a miss synthesises default with paths: [], which disables verifyConsolidation's oracle entirely (work-driver-verify.ts:70-73). A parse miss silently turns off the diff-coverage check.
parseAbort (work-driver-diff.ts:281) — only scans the last 800 chars, so an abort stated early in a long reply is invisible, and the step is recorded ok: true — the literal #553 incident its own docstring describes.
lens-review-format.ts:72-82 — an unrecognised severity string drops the finding entirely, counted only as an integer. A CRITICAL with severity: "BLOCKER" vanishes and the verdict computes as if it never existed.
Adjacent: the driver re-parsing its own enums
work-queue.ts:144-181 regexes cap-reason strings the driver itself produced — structured enums round-tripped through a string and re-parsed. A cap rename falls through to the generic --restart advice at :181, which for a merge hold is actively destructive (documented at :130-134).
The good pattern, already in the codebase
Two sites show the way out:
lens-review-format.ts:40-94 asks for a tool call with a typed schema Pi validates in the child: "No text parsing involved… malformed calls never reach this code."parsePrNumber (work-driver-lens.ts:31) has a real executed-evidence fallback — on a miss, gh pr list --head <branch> recovers the number. It is the only parser in the codebase that repairs itself from reality rather than from a default.
Acceptance criteria
Sequencing
Independent of #406 and #407 and can proceed in parallel, but lower priority than both — none of these silently escalates authority. Best split across several PRs by tier rather than landed as one.
Context
An audit of
extension/src/for "regex applied to an LLM reply or a doctrine file, to extract a decision" found 24 sites. Three are already fixed or in flight (#397'sparseEvidence, #404'sPARK-REASON, #407'sGRANT_PATTERNS). This issue covers the rest, ranked by whether a no-match yields a confident wrong answer rather than an obvious failure.The shared property that makes these dangerous: the failure is silent, and the substituted default is then acted on as though it were read.
Tier 1 — a no-match produces a wrong confident answer
parsePerIssueVerdictsdefaults to build-it —work-driver-plan.ts:275. On no match for an issue, falls back to the overall verdict; if that is absent too,NEEDS_WORK. This is precisely the pre-#378 "silence is permission" thatwork-driver-intent.ts:13-14claims was structurally removed — it survives untouched on the multi-issue path, which #397 made the only multi-issue path. The reason string at:277-278is honest ("(no verdict; defaulting to NEEDS_WORK)") and nothing gates on it.labelAppliedfails open —work-driver-handoff.ts:116,/label.*needs-human-attention/iover ops prose. An ops agent writing "I could not apply the label needs-human-attention" matches, so theghfallback is skipped and the driver records success on an issue that was never labelled.parseVerdicthands over the wrong text —adversarial.ts:376-380. Case-sensitive, no bold tolerance (unlike both siblings atwork-driver-plan.ts:235andwork-driver-intent.ts:128), soVerdict: Approvedmisses. On a miss the default isISSUES_FOUNDandfindingsis set to the entire reply (:379) — so the fix-developer is handed the whole approval message as "findings to fix", and burns a round on it.ci-statushas no tolerance at all —work-driver-stepback-ci.ts:260-268, a baretext.includes().ci-status:success,CI-status: success, or the marker inside a table all miss. No-match →"failure"→ bumpsciRetryCountand can fire theci-retrycap. A green CI reported in a slightly different shape burns the retry budget and parks the cycle. The #380 demotion check at:274-286only runs when the status is alreadysuccess, so it cannot rescue a false failure.parseAdversarialRoundsinvents a number —work-driver-adversarial.ts:65-69. Threeincludes()on prose, defaulting to3, asserted into the event log as fact with no "unknown" state. A phrasing change inadversarial.ts's own synthesised text silently makes every run report three rounds.Grouping R1/R3/R4 —
work-driver-grouping.ts:122,:171,:243. Two of these were structurally impossible to satisfy and were only found by accident (#376). Live evidence from nessie: #657 said "Closes the drift in #650" and R1's alternation has nocloses/fixes/resolves, so two issues where one fixes the other ran as concurrent groups. And.containeris absent from R2's extension list (:188), soselfhost/nessie.container— the one file both issues genuinely shared — contributed nothing to the Jaccard.sliceSubsection/sliceMarkdownSection—work-driver-intent.ts:96,work-driver-plan.ts:398. Require an exact^###\s+<name>\s*$. A resolver writing### Deliverables:,#### Deliverablesor**Deliverables**yieldsundefined→[]→specIsCompletefalse → park. Silent; nothing traces that a heading was not found.Tier 2 — degrades quietly, failure at least structural
parseWorkstreams(work-driver-plan.ts:301) — a miss synthesisesdefaultwithpaths: [], which disablesverifyConsolidation's oracle entirely (work-driver-verify.ts:70-73). A parse miss silently turns off the diff-coverage check.parseAbort(work-driver-diff.ts:281) — only scans the last 800 chars, so an abort stated early in a long reply is invisible, and the step is recordedok: true— the literal #553 incident its own docstring describes.lens-review-format.ts:72-82— an unrecognisedseveritystring drops the finding entirely, counted only as an integer. A CRITICAL withseverity: "BLOCKER"vanishes and the verdict computes as if it never existed.Adjacent: the driver re-parsing its own enums
work-queue.ts:144-181regexes cap-reason strings the driver itself produced — structured enums round-tripped through a string and re-parsed. A cap rename falls through to the generic--restartadvice at:181, which for a merge hold is actively destructive (documented at:130-134).The good pattern, already in the codebase
Two sites show the way out:
lens-review-format.ts:40-94asks for a tool call with a typed schema Pi validates in the child: "No text parsing involved… malformed calls never reach this code."parsePrNumber(work-driver-lens.ts:31) has a real executed-evidence fallback — on a miss,gh pr list --head <branch>recovers the number. It is the only parser in the codebase that repairs itself from reality rather than from a default.Acceptance criteria
main.parsePerIssueVerdictsno longer defaults to build-it — silence parks, matching feat(work-driver): resolve intent from any spec — ground it against code and world, then proceed, assume, or park #378's stated principle.labelAppliedis established by executed evidence (gh issue view --json labels), not by prose.parseVerdicttolerates case and bold; a miss does not pass the whole reply as findings.ci-statustolerance matches its siblings; a formatting miss cannot park a green cycle.closes/fixes/resolves/fixed by/superseded by; R2's extension list covers.container/.service/.unit/Dockerfile.work-queue.tsroutes on the structured cap value, not a re-parsed string.Sequencing
Independent of #406 and #407 and can proceed in parallel, but lower priority than both — none of these silently escalates authority. Best split across several PRs by tier rather than landed as one.