feat: fetch providers over DRPC - #26650

Clean, well-scoped refactor. The DB-to-DRPC migration is symmetric across embedded and standalone paths, the proto surface is minimal, dead code is cleanly removed, and the concurrency test against real Postgres advisory locks is solid. Test density is healthy at 44%.

As Knov put it: "Good perimeter: disabled providers withhold keys and settings at the server, not the client."

Severity breakdown: 2 P2, 3 P3, 6 Nit.

The two P2s: (1) the embedded daemon can now start serving with an empty provider pool if the pubsub subscription or initial reload fails, losing the old boot-time fallback that guaranteed providers were loaded before the daemon accepted requests; (2) the DRPCServer interface was not updated to include DRPCProviderConfiguratorServer, so the compile-time assertion and any future mock targeting that interface silently omits the new service.

Observations not posted as findings: the advisory lock uses pg_advisory_xact_lock (exclusive) where pg_advisory_xact_lock_shared would let concurrent reads proceed without serializing against each other. At current call rates this has no impact, but worth noting before AIGOV-465 adds refresh-on-reseed. Server-side row skipping for corrupt settings is an intentional tradeoff (one corrupt blob does not break the fleet); the observability gap (gateway metrics cannot see the skip) is acceptable since the server-side Error log is the correct signal.

coderd/aibridged/server.go:5

P2 [CRF-2] DRPCServer embeds DRPCRecorderServer, DRPCMCPConfiguratorServer, and DRPCAuthorizerServer, but not DRPCProviderConfiguratorServer. The client-side DRPCClient (client.go:17) was updated to embed DRPCProviderConfiguratorClient, but the server-side union was not.

"Every consumer that type-asserts against DRPCServer silently misses the new service. If a future mock or alternative implementation targets DRPCServer, it will silently omit GetAIProviders." (Ryosuke)

The compile-time assertion at aibridgedserver.go:58 (var _ aibridged.DRPCServer = &Server{}) does not verify ProviderConfigurator.

Fix: add proto.DRPCProviderConfiguratorServer to this interface.

🤖

🤖 This review was automatically generated with Coder Agents.

9 of 11 findings from Round 1 were addressed in c3a38c2. Good progress.

Further review is blocked until the remaining 2 silent findings receive a response (fix, acknowledge, or contest):

CRF-2 (P2) coderd/aibridged/server.go:5: DRPCServer interface still embeds only 3 services; DRPCProviderConfiguratorServer is missing. The compile-time assertion at aibridgedserver.go:58 does not verify the new service, and any future mock targeting DRPCServer silently omits GetAIProviders. Fix: add proto.DRPCProviderConfiguratorServer to the interface.

CRF-5 (P3) coderd/aibridgedserver/providers_test.go:155: The negative assertion in TestGetAIProvidersBlocksOnSeedLock relies on time.After(testutil.IntervalMedium) without verifying the fetch goroutine reached AcquireLock. On a loaded CI runner the assertion can pass vacuously. Fix: verify the goroutine entered the blocking call (e.g., query pg_stat_activity) or acknowledge the known limitation.

🤖 This review was automatically generated with Coder Agents.

All 14 findings from rounds 1 and 3 are resolved. The author addressed every finding with substantive fixes, not band-aids: subscription failure made fatal (CRF-1), DRPCServer interface aligned with DRPCClient (CRF-2), pg_locks polling replaced timing-based inference (CRF-5), test helper replaced with production code (CRF-17), retry-succeed test coverage added and moved to stacked PR #26605 (CRF-19).

The refactor is clean. coderd owns the database; the gateway fetches provider config over DRPC. Both embedded and standalone daemons share an identical data path through NewPoolRPCReloader. The advisory lock serializes reads against the env seed. Disabled providers withhold keys and credentials at the server. Dead code is fully removed. Test density is healthy at 39%.

NewWebsocketDialer has no in-tree callers in this PR; its consumer lives in the stacked PR #26605.

🤖 This review was automatically generated with Coder Agents.