4 months ago
8884568The reason must quote the code it accuses (`secrets.GITHUB_TOKEN`, `pull_request`) so it reads as evidence, not prose — a material prompt change, so a version bump. - instructions.md (@1) untouched, byte-for-byte; @1 stays registered so stored runs remain interpretable (versioning law). @2 uses instructions-v2.md (= @1 + one output rule about backticks). rule.ts is now a defineAiReview(version, instructions) factory registering both. - RULE_CATALOG pins ai-review to @2 — a repo enabling it runs the current version; existing @1 configs keep working. - the findings renderer already parses inline backticks → sanitized mono <code> chips, handling @2 backticked and @1 plain notes identically. replay unchanged: `bun run replay --corpus` = 15 runs · 13 unchanged · 2 flips · 0 skipped (replay reuses stored snapshots + step envelopes, never re-invokes the model — a prompt change can't move a verdict). Checks: typecheck all, biome, boundaries, 204 tests.
Parent3bb7675