Severity never determines consequence anywhere on the decision path. severityToAnnotationLevel (src/rules/advisory.ts ~430-434) and formatCheckRunOutput (~573) are its only readers, both purely presentational. The two disagree in both directions:
| finding |
severity |
actual consequence |
slop_risk_above_threshold (~1300-1307) |
warning |
pushed straight into blockers (~681) → failure → CLOSE |
missing_linked_issue (~859-867) |
warning |
block mode → CLOSE, and breaker-exempt |
self_authored_linked_issue (~895-901) |
warning |
breaker-exempt CLOSE |
duplicate_pr_risk (~878-884) |
warning |
defaults to block → breaker-exempt CLOSE |
ai_consensus_defect |
critical |
defaults to advisory (~1051) → no effect at all |
oversized_pr / guardrail_hold (~602, ~620) |
warning |
hold only |
A contributor reading a ⚠️ on the finding that is about to close their PR is being actively misinformed.
Taxonomy drift found alongside it
slop_risk_above_threshold is in neither CONFIGURED_GATE_BLOCKER_SIGNAL_CODES (~192-209) nor resolveConfiguredGateMode. It gates real closes but records reversals under a different id (slop_gate_score, ~1241), so the per-rule precision check in downgradeCloseToHold (Reason B) can never apply to it.surface_lane_reject is in CONCRETE_EVIDENCE_BLOCKER_CODES but not in resolveConfiguredGateMode — live, but invisible to recordConfiguredGateBlockerSignals and therefore to the entire calibration corpus.- Near-miss pair:
repo_not_registered holds the gate (~1010) while repo_unregistered (~786) does not. Two codes one character apart with opposite consequences.
- Confirmed dead/inert:
authorMergedPrCount, authorClosedUnmergedPrCount, confirmedContributor (~117-126, self-documented INERT), reconcileGateEvaluationForGreenCi (~281-285, voids both args), aiCiRefutationEnabled, ciRequiredContextsVerified, assignLinkedIssues.
- Dead branch:
conclusion === "action_required" in manualHoldReason (agent-actions.ts ~1131) is unreachable — the core only returns success/neutral/failure, and other callers pass "skipped" which returns earlier.
Also: confidence defaults to maximum on the gate side
advisory.ts ~242 (blocker.confidence ?? 1) and ~1055 finding.confidence ?? 1 default missing confidence to 1.0, suppressing the low-confidence hold and yielding a straight one-shot close. That directly contradicts CONFIDENCE_WHEN_UNSTATED = 0.5 (src/services/ai-review.ts ~756), whose own doc says "silence is not certainty". Currently masked because producers always set confidence — but the cache path (repositories.ts ~5098 parses AdvisoryFinding[] with no schema validation) would land straight on the wrong default.
Fix
- Make severity load-bearing (or rename it to something honestly presentational) — a
warning must not be able to close a PR.
- Register
slop_risk_above_threshold and surface_lane_reject in the blocker-signal taxonomy so the breaker and calibration can see them.
- Default absent confidence to
CONFIDENCE_WHEN_UNSTATED, not 1.
- Rename one of
repo_not_registered / repo_unregistered.
Refs #9074, #9040, #9003.
Severity never determines consequence anywhere on the decision path.
severityToAnnotationLevel(src/rules/advisory.ts~430-434) andformatCheckRunOutput(~573) are its only readers, both purely presentational. The two disagree in both directions:slop_risk_above_threshold(~1300-1307)blockers(~681) →failure→ CLOSEmissing_linked_issue(~859-867)self_authored_linked_issue(~895-901)duplicate_pr_risk(~878-884)ai_consensus_defectadvisory(~1051) → no effect at alloversized_pr/guardrail_hold(~602, ~620)A contributor reading a⚠️ on the finding that is about to close their PR is being actively misinformed.
Taxonomy drift found alongside it
slop_risk_above_thresholdis in neitherCONFIGURED_GATE_BLOCKER_SIGNAL_CODES(~192-209) norresolveConfiguredGateMode. It gates real closes but records reversals under a different id (slop_gate_score, ~1241), so the per-rule precision check indowngradeCloseToHold(Reason B) can never apply to it.surface_lane_rejectis inCONCRETE_EVIDENCE_BLOCKER_CODESbut not inresolveConfiguredGateMode— live, but invisible torecordConfiguredGateBlockerSignalsand therefore to the entire calibration corpus.repo_not_registeredholds the gate (~1010) whilerepo_unregistered(~786) does not. Two codes one character apart with opposite consequences.authorMergedPrCount,authorClosedUnmergedPrCount,confirmedContributor(~117-126, self-documented INERT),reconcileGateEvaluationForGreenCi(~281-285,voids both args),aiCiRefutationEnabled,ciRequiredContextsVerified,assignLinkedIssues.conclusion === "action_required"inmanualHoldReason(agent-actions.ts~1131) is unreachable — the core only returns success/neutral/failure, and other callers pass"skipped"which returns earlier.Also: confidence defaults to maximum on the gate side
advisory.ts~242(blocker.confidence ?? 1)and ~1055finding.confidence ?? 1default missing confidence to 1.0, suppressing the low-confidence hold and yielding a straight one-shot close. That directly contradictsCONFIDENCE_WHEN_UNSTATED = 0.5(src/services/ai-review.ts~756), whose own doc says "silence is not certainty". Currently masked because producers always set confidence — but the cache path (repositories.ts~5098 parsesAdvisoryFinding[]with no schema validation) would land straight on the wrong default.Fix
warningmust not be able to close a PR.slop_risk_above_thresholdandsurface_lane_rejectin the blocker-signal taxonomy so the breaker and calibration can see them.CONFIDENCE_WHEN_UNSTATED, not 1.repo_not_registered/repo_unregistered.Refs #9074, #9040, #9003.