You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
#8159 registered ai_review_close_confidence as a REPORT-ONLY loosenable knob: backtest-cleared proposals (from the real ai_consensus_defect corpus) surface in the advisor with full evidence, but nothing can apply them — no override consumer exists. Wiring one is a gate-authority change (a loosened close confidence means more auto-closes), so it was deliberately excluded from the registry PR.
Requirements
⚠️ Required pattern. Mirror the satisfaction floor's consumption chain exactly: a validated, bounds-checked read (loosening-only: below DEFAULT_AI_REVIEW_CLOSE_CONFIDENCE, at/above the knob's 0.85 hard minimum; anything else ⇒ shipped default) resolved at ONE point, threaded as a defaulted parameter — never a mutable global.
Consumption point: the policy.aiReviewCloseConfidence ?? DEFAULT_AI_REVIEW_CLOSE_CONFIDENCE fallbacks (isConfiguredGateBlocker's sub-floor disposition path and resolveAiReviewLowConfidenceHold, host + engine twin). An explicit per-repo aiReviewCloseConfidence setting must ALWAYS win over the global override — the knob only moves the default.
Flip the registry entry to applyMode: "live" ONLY in this PR, together with the generic apply path (per-knob system_flags key, audit event, one-shot alert — reuse the satisfaction machinery generalized, not duplicated).
Bounds-validated override read + threading through both gate-decision twins (parity checks green)
Generic apply path + registry flip to live, with the same double flag-gating as the satisfaction loop
Status-surface coverage + 99%+ branch-covered tests on every touched src path
Links & Resources
#8159 (the registry), #8121 (the discipline), src/services/loosening-knobs.ts, src/rules/advisory.ts + packages/loopover-engine/src/advisory/gate-advisory.ts (the twins)
Boundaries
Per-repo explicit settings always beat the global override. No other knob goes live in this PR. Trust-critical: gate authority changes only through the full backtest gate.
Problem
#8159 registered
ai_review_close_confidenceas a REPORT-ONLY loosenable knob: backtest-cleared proposals (from the realai_consensus_defectcorpus) surface in the advisor with full evidence, but nothing can apply them — no override consumer exists. Wiring one is a gate-authority change (a loosened close confidence means more auto-closes), so it was deliberately excluded from the registry PR.Requirements
policy.aiReviewCloseConfidence ?? DEFAULT_AI_REVIEW_CLOSE_CONFIDENCEfallbacks (isConfiguredGateBlocker's sub-floor disposition path andresolveAiReviewLowConfidenceHold, host + engine twin). An explicit per-repoaiReviewCloseConfidencesetting must ALWAYS win over the global override — the knob only moves the default.applyMode: "live"ONLY in this PR, together with the generic apply path (per-knobsystem_flagskey, audit event, one-shot alert — reuse the satisfaction machinery generalized, not duplicated).Deliverables
Links & Resources
#8159 (the registry), #8121 (the discipline), src/services/loosening-knobs.ts, src/rules/advisory.ts + packages/loopover-engine/src/advisory/gate-advisory.ts (the twins)
Boundaries
Per-repo explicit settings always beat the global override. No other knob goes live in this PR. Trust-critical: gate authority changes only through the full backtest gate.
maintainer-only — gate-authority consumption plumbing.