The #regate-churn fix bounded re-spend for disputed verdicts, but it only writes a cache row when runAiReviewForAdvisory returns something defined:
if (aiReview && aiReview.persistable !== false) { ... putCachedAiReview(...) } // processors.ts ~10422Both failure exits return undefined: if (result.status !== "ok") return undefined; (ai-review-orchestration.ts ~729) and the catch block (~935).
So any pass that throws inside the try (a DB hiccup in getRepository, a GitHub 5xx during grounding, an enrichment/REES failure) or returns non-ok writes no row at all. The next tick — 2 minutes later (wrangler.jsonc crons */2 * * * *; self-host setInterval default 120s) — misses the cache and re-executes the whole prologue: listPullRequestFiles → grounding (up to FILE_CONTENT_BUDGET = 96_000 chars fetched from GitHub) → RAG embeddings → impact-map embeddings → culture profile → the external REES POST → the LLM attempt.
This is the exact shape and cadence of the known 259-calls-in-24h incident, and the existing fix does not cover it.
Compounding: the budget ceiling sits AFTER the spend it bounds
AI_DAILY_NEURON_BUDGET is checked inside runLoopOverAiReview (src/services/ai-review.ts ~2504-2522), which is called at ai-review-orchestration.ts ~670 — after grounding, RAG, impact map, culture profile, and enrichment have all run (~549-662).
So on budget exhaustion: quota_exceeded → non-ok → no cache row → next tick re-runs all the context building again, then hits the quota again. And because every embedding is booked with estimatedNeurons: 0 (src/review/adapters.ts ~80, ~93), none of that spend increments the governor's counter — the ceiling can never converge and the loop never self-limits.
Fix
- Persist a non-durable "attempt failed" row (
cacheable: false, persistable: true) on the undefined paths, keyed on the same head+fingerprint, so the 30-minute cooldown applies to failures too.
- Hoist the budget/quota check to the top of
runAiReviewForAdvisory, before any context building.
- Charge embeddings a real non-zero estimate.
Acceptance
- A PR whose AI review fails repeatedly costs one prologue per cooldown window, not one per 2 minutes.
- With the budget exhausted, no grounding/RAG/enrichment work is performed at all.
The
#regate-churnfix bounded re-spend for disputed verdicts, but it only writes a cache row whenrunAiReviewForAdvisoryreturns something defined:Both failure exits return
undefined:if (result.status !== "ok") return undefined;(ai-review-orchestration.ts~729) and thecatchblock (~935).So any pass that throws inside the try (a DB hiccup in
getRepository, a GitHub 5xx during grounding, an enrichment/REES failure) or returns non-okwrites no row at all. The next tick — 2 minutes later (wrangler.jsonccrons*/2 * * * *; self-hostsetIntervaldefault 120s) — misses the cache and re-executes the whole prologue:listPullRequestFiles→ grounding (up toFILE_CONTENT_BUDGET = 96_000chars fetched from GitHub) → RAG embeddings → impact-map embeddings → culture profile → the external REES POST → the LLM attempt.This is the exact shape and cadence of the known 259-calls-in-24h incident, and the existing fix does not cover it.
Compounding: the budget ceiling sits AFTER the spend it bounds
AI_DAILY_NEURON_BUDGETis checked insiderunLoopOverAiReview(src/services/ai-review.ts~2504-2522), which is called atai-review-orchestration.ts~670 — after grounding, RAG, impact map, culture profile, and enrichment have all run (~549-662).So on budget exhaustion:
quota_exceeded→ non-ok → no cache row → next tick re-runs all the context building again, then hits the quota again. And because every embedding is booked withestimatedNeurons: 0(src/review/adapters.ts~80, ~93), none of that spend increments the governor's counter — the ceiling can never converge and the loop never self-limits.Fix
cacheable: false, persistable: true) on theundefinedpaths, keyed on the same head+fingerprint, so the 30-minute cooldown applies to failures too.runAiReviewForAdvisory, before any context building.Acceptance