In progress
AI Code Review (R1a)
GitHub App for AI-assisted pull request review. Diff-only summary review with risk classification, posted back to the PR as Markdown. Per-repo configuration via .schneeai.yml (language, tone, path filters, custom instructions). Cost backed by credits with reserve / settle / release semantics.
- Diff-level PII scanning (17 categories) before sending content to the model
- Path filters with glob matching — skip vendored or generated code
- Idempotent webhook delivery with 24-hour deduplication
- Flash-tier models for summaries, premium for deep reviews
See the code review use case for the integration shape.
Recent
Phase 9 — JWT/JWKS authentication
Gateway and Admin BFF authentication moved from shared HMAC secrets to RS256 JWT issued by AuthenticationPlatformReplace (Ory Hydra). Signature verification via JWKS, with sub, active_tenant_id, tenant_role, and service_id claims mapped to platform scope. Admin endpoints require elevated tenant_role.
Phase 8.5 — PII Admin API
PII policy management and findings inspection via Admin BFF. GET/PUT/DELETE /internal/v1/pii/policies/:serviceId for policy, GET /internal/v1/pii/findings/:interactionId for inspecting findings on a specific interaction. Runtime scan behavior unchanged.
Phase 8.4 — Extended PII categories
Added 10 additional PII categories on top of the original 7 — phone (Japan, E.164), IP (v4, v6), API keys (Google, OpenAI, Anthropic, Stripe, Slack), and connection strings. Severity tiers and overlap resolution extend consistently across all 17 categories.
Phase 8.3 — Reasoning PII scanning
PII scanning applied to reasoning_content before Vault retention. Critical findings block the call; warning and info findings mask or flag per policy.
Phase 3 — PromptOps + Response Cache
Prompt Registry with versioning, canary rollouts, and instant rollback. Response cache with scope-aware keys (service / tenant / user / feature) and configurable invalidation.
More
For deeper technical detail on each release, browse the blog or email [email protected] with specific questions.