feat(service): Enable hosted memory and source context

[gricha]

Enable the production service to derive, promote, embed, and recall repository memory through Vercel AI Gateway. Finding detail responses now also expose commit-pinned GitHub links and render bounded source evidence inline when the run used the code profile.

Historical and new runs already request memory, but the hosted app previously had no model or embedding providers. This adds OIDC-backed providers with an API-key fallback, conservative three-run promotion, durable embedding work, optional pgvector indexing, and operation cost attribution.

Repeated evidence now updates an existing candidate transactionally instead of returning before it can be promoted. Vector persistence falls back only when pgvector is unavailable, so unrelated database failures remain visible.

The defaults use Luna and text-embedding-3-small to keep backfill costs bounded. Ingestion remains fail-open and write-only credentials still cannot read findings or memory.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
warden-prod	Ready	Preview	Aug 14, 2026 3:49pm

[sentry-warden]

Hosted memory recall has no rate or cost controls

Enabling AI Gateway embedding and relevance on POST /api/v1/memory/recall lets any read-scoped service token spend gateway quota on each unique clientRecallId, and createVercelWardenService still does not wire a rateLimit hook.

Evidence

• This hunk wires memoryRecall: { embedding: memory.embedding, relevance: memory.relevance } into createWardenService(), so hosted recall now calls paid Vercel AI Gateway embed + classify paths.
• createVercelWardenService() still constructs the app without a rateLimit hook; the shared middleware only enforces limits when that optional hook is provided.
• POST /api/v1/memory/recall is authorized for any read role, and recallMemories() skips the paid path only on an existing (tenantId, clientRecallId) batch—fresh IDs always embed/classify.
• Personal PATs are blocked from this POST, but long-lived read service tokens are not, and there is no quota/idempotency throttle beyond client-chosen IDs.

_{Identified by Warden · security-review · BJL-XZR}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 2e5dffc. Configure here.}

[cursor]

Recall backfill breaks without pgvector

Medium Severity

The recall backfill INSERT now filters on embedding_vector outside the vector savepoint. When that column is absent, the query throws after lexical ranking and is swallowed by the outer catch, so embed jobs are not enqueued and embedding usage is dropped. Write-path JSON fallback cannot recover those memories through recall.

Additional Locations (1)

• packages/warden-service/src/memory/store.ts#L482-L540

 

^{Reviewed by Cursor Bugbot for commit 2e5dffc. Configure here.}