feat(ai-review): config-as-code provider/model + maintainer self-serve BYOK (backend, closes #662)

[JSONbored]

Backend for #662 — makes BYOK AI review usable end-to-end by maintainers. (UI is #663.)

Config-as-code provider/model

• RepositorySettings gains aiReviewProvider (anthropic|openai|null) + aiReviewModel (migration 0029), settable from .gittensory.yml:

  gate:
    aiReview: { mode: block, byok: true, provider: anthropic, model: claude-3-5-sonnet-latest }

  The secret key is never in the yml (world-readable) — only the non-secret provider/model.
• The engine resolves the effective provider/model: the stored key's provider is authoritative; a declared provider must match it or BYOK is skipped (→ free Workers-AI fallback); a declared model overrides the stored/default model.

Maintainer self-serve (the missing enabler)

Keys can't live in repo secrets — a GitHub App can't read them — so maintainers configure them through gittensory:

• PUT /v1/repos/:owner/:repo/ai-review — set mode/byok/provider/model (merges onto current settings; preserves unrelated settings).
• POST/GET/DELETE /v1/repos/:owner/:repo/ai-key — write-only encrypted key; GET returns only {configured, provider, last4, model} (never the key).
• All session-authenticated and scoped to repos the maintainer owns/maintains (requireRepoMaintainer helper, mirroring the focus-manifest routes).

Real bug fixed

The new repo subpaths had to be added to canSessionAccessPath's allowlist — without that, the global protected middleware 403 insufficient_roles every maintainer before the handler runs (caught by a session-auth test). The internal-token routes from #652 remain for operator/backend use.

Tests

Config-as-code parse + round-trip + resolve; engine provider-match / mismatch-skip / model-override; the maintainer routes (operator-token + session owner success + non-owner 403 on every route + cross-repo 403 + 401 unauth + invalid-body 400 + 503 when key storage is unconfigured). Coverage holds at the 97% gate; alignment + OpenAPI + Workers checks pass.

Part of #525.

[ghost]

Note

Gittensory Gate skipped

PR closed before full evaluation. No late first comment was created.

Signal	Result	Evidence	Action
Gate result	⚠️ Skipped	#664 is no longer open.	No action.

---

💰 Earn for open-source contributions like this. Gittensor lets GitHub contributors earn for the work they already do — register to start earning →.

Checked by Gittensory, a quiet PR intelligence layer for OSS maintainers.

[ghost]

gittensory · advisory review

Reviewed 12 changed file(s) — two independent AI reviewers.

Suggested action: ✅ Safe to merge — both reviewers found no blocking issues.

Reviewer A · gpt-oss-120b — recommends ✅ merge
This PR introduces configurable BYOK provider and model settings for AI review, adds maintainer self‑serve routes for configuring those settings and managing encrypted keys, updates DB schema and focus‑manifest parsing, and adds comprehensive tests. The changes are well‑scoped, respect the public/private boundaries, and include proper authorization and error handling.

Suggestions

• Update any API documentation or OpenAPI spec to include the new /ai-review and /ai-key endpoints.
• Consider adding a migration test to verify the new columns are added without data loss on existing databases.
• Document the expected shape of the providerKey object in runGittensoryAiReview to avoid future mismatches.

Worth double-checking

• Migration must be applied in production; ensure the migration runs before any code that expects the new columns.
• The upsert logic resets provider/model to null when omitted – verify this is the intended behavior for all callers.

Reviewer B · nemotron-3-120b-a12b — recommends ✅ merge
The PR implements config-as-code for AI review provider/model and maintainer self-serve BYOK backend, adding new settings fields, routes, DB schema updates, and tests. Changes respect the public/private boundary by keeping secret keys encrypted and non-secret in public outputs. Implementation is consistent with existing patterns and includes comprehensive unit tests.

Suggestions

• Add a comment in normalizeAiReviewProvider to clarify it returns null for non-matching values.
• In upsertRepositorySettings, ensure aiReviewModel trimming aligns with Zod schema validation (though internal callers benefit from the safeguard).

Worth double-checking

• Verify encryption secret handling in production (503 when missing) aligns with deployment docs.
• Ensure provider/model validation in focus manifest parsing doesn't reintroduce deprecated values.

[superagent-security]

Superagent didn't find any vulnerabilities or security issues in this PR.