2.3.0 (2026-04-25)

[2.3.0] - 2026-04-25

Added

• Runtime-enrichment context for pr_reviewer (sast-ai-app v5.7.0 / quality-plan §4 + §8.14): new --runtime-enrichment-context <file.json> CLI flag accepts a per-file production-incident summary from the parent app's runtimeEnrichmentService and injects it into the diff-review user prompt so the LLM can factor incident history into its severity + confidence calls. Shape: { default_branch_sha?, parsed_at?, files: [{ file, incident_count, last_seen_at? }], metadata? } with a max of 500 files (matches the import-graph cap; the parent-app contract caps the source list at 10k rows but only files-overlapping-the-PR are passed in). Fail-open on any parse/IO error — the authoritative gate override lives in the parent app (sast-ai-app/backend/src/routes/prScanProcessor.ts partitions findings into hot/cold and applies medium → low / 0.6 → 0.4 per file), so a bad payload only suppresses the LLM-side advisory hint.
• Schema module schemas/runtime_enrichment.ts: parseRuntimeEnrichmentContext (throws on structural errors; silently strips unknown extra fields per the §8.5 PHI gate so a future buggy backend revision can't leak incident bodies / stack traces / request payloads to the LLM) + formatRuntimeEnrichmentContextForPrompt (compact markdown table, sorted by incident_count desc to anchor the LLM's attention on the strongest-signal files when the prompt is truncated; advisory line surfaces both halves of the §4 transform — medium → low AND 0.6 → 0.4 — so the LLM and the post-LLM gate apply consistent thresholds) + RuntimeEnrichmentContext / RuntimeEnrichmentFileEntry types re-exported from the package entry.
• Role/flag compatibility check in bin/agent-run.ts: --runtime-enrichment-context only takes effect for -r pr_reviewer --diff-context <file>; other role combinations log a warning and ignore the flag (mirrors --import-graph-context).
• Tests: 16 unit tests in src/__tests__/schemas/runtime_enrichment.test.ts covering parse happy-path, last_seen_at inclusion + drop-on-empty/whitespace/non-string, fractional/negative count handling, the §8.5 PHI minimization invariant (extra fields silently dropped), metadata round-trip, structural error paths (non-object input, missing files array, 501-file cap, missing/empty/whitespace file string, non-numeric/NaN/Infinity count, non-object entries), and prompt formatter output (empty-list short-circuit, sort-order, §4 transform numbers surfaced, SHA omitted when absent). Plus 3 e2e tests in e2e/pr_reviewer_runtime_enrichment.e2e.test.ts exercising the full main() path: happy-path injection, fail-open on bad payload, and empty-files short-circuit. Total tests: 410 → 432.

Why a coordinated release with sast-ai-app@5.7.0

• Same cross-repo sequencing as v5.4.0 / v2.2.0: the agent PR lands and publishes first, then the backend PR merges pinning appsec-agent@^2.3.0. The backend's v5.7.0 wiring writes a .runtime-enrichment-context.json file and passes its path via --runtime-enrichment-context; shipping the backend first would degrade every diff-context scan with a flag the agent doesn't recognize (commander would error out before main() ran).

Scope boundaries (explicitly deferred)

• A live MCP-tool variant of the runtime-enrichment lookup (agent calls GET /api/projects/:id/hot-files during a turn rather than consuming a pre-computed summary) remains out of scope for the same reason as the import-graph variant: Claude Agent SDK doesn't expose a clean custom-tool channel for HTTP calls today; the v5.0.0 findings-history pattern (backend-composed narrative injected via env/CLI) is the ready-today mechanism and matches the cadence the plan has followed since v5.0.0.
• Auto-population of the parent app's hot-file list from production logs remains a v6.x sast-ai-app concern; this release accepts whatever the parent app emits.