Description

Add integration tests covering the end-to-end vMCP horizontal scaling scenario implemented by TASK-001 through TASK-007 in the vMCP epic (stacklok/stacklok-epics#262). These tests verify that session reconstruction, cross-pod hijack prevention, LRU eviction, and Redis metadata lifecycle work correctly together, using miniredis to avoid a live Redis dependency.

Parent epic: stacklok/stacklok-epics#262

Dependencies: #283 (RC-9: Reconstruct vMCP Sessions from Redis on Cache Miss), #285 (RC-10: LRU Eviction), #276 (RC-16: Update Redis Metadata When Backend Sessions Expire)

Context

Individual task issues (TASK-001 through TASK-007) include unit-test acceptance criteria for their respective packages. This task adds higher-level integration tests that exercise the full session lifecycle across the components introduced by those tasks, analogous to TASK-008 in the operator epic (stacklok/stacklok-epics#284).

Acceptance Criteria

• Integration test: session created on "pod A" (Redis-backed manager) can be reconstructed and served on "pod B" (second Redis-backed manager pointing at the same miniredis instance) via GetMultiSession cache-miss path
• Integration test: hijack prevention is enforced cross-pod — a request with a different token cannot use a session restored from Redis
• Integration test: LRU eviction on "pod A" causes session to be evicted from memory; subsequent GetMultiSession call triggers RestoreSession from Redis and succeeds
• Integration test: when all backends fail to reconnect during RestoreSession, session is returned with empty routing table (not an error), consistent with makeSession partial-failure behavior
• Integration test: when a backend session expires (RC-16), the corresponding Redis metadata keys are removed; a subsequent RestoreSession does not attempt to reconnect that backend
• Integration test: when Redis is not configured (nil RedisConfig), vMCP operates with in-memory-only storage and the above multi-pod scenarios are not exercised (verifies R-VMCP-4 regression guard)
• All tests use miniredis.RunT(t) — no live Redis required
• All tests pass: go test ./pkg/vmcp/... ./pkg/transport/session/...

Technical Approach

Test Location

Add a new file pkg/vmcp/server/sessionmanager/horizontal_scaling_integration_test.go (or equivalent) containing table-driven tests that:

1. Spin up a miniredis instance
2. Create two transportsession.Manager instances both backed by the same RedisStorage
3. Use the first manager to create a session (simulating pod A)
4. Use the second manager's GetMultiSession to reconstruct the session on a cache miss (simulating pod B)
5. Verify the reconstructed session responds correctly to tool calls with the original caller identity

Patterns & Frameworks

• github.com/alicebob/miniredis/v2 — in-process Redis for tests; follow pattern in pkg/authserver/storage/redis_test.go
• Table-driven, t.Parallel() style — follow pkg/transport/session/storage_test.go
• Fake backend connector pattern — follow pkg/vmcp/session/connector_integration_test.go

Code Pointers

• pkg/vmcp/server/sessionmanager/session_manager.go — GetMultiSession with Redis fallback (added in TASK-005)
• pkg/vmcp/session/factory.go — RestoreSession (added in TASK-005)
• pkg/vmcp/session/internal/security/security.go — RestoreHijackPrevention (added in TASK-004)
• pkg/transport/session/storage_redis.go — RedisStorage (added in TASK-001)
• pkg/transport/session/manager.go — NewManagerWithRedis (added in TASK-002)
• pkg/authserver/storage/redis_test.go — miniredis helper pattern to follow

Out of Scope

• E2E tests against a real Kubernetes cluster
• Load or performance testing
• Operator-level integration tests (covered in stacklok/stacklok-epics#284)

References

• Parent epic: stacklok/stacklok-epics#262
• RFC THV-0047: RFC: Horizontal Scaling for vMCP and Proxy Runner toolhive-rfcs#47
• Operator equivalent: stacklok/stacklok-epics#284 (TASK-008 Unit Tests)
• miniredis: https://github.com/alicebob/miniredis