The persistence + signaling layer for v5. No integration with the
existing in-flight inference path yet — that's Stage 2. This commit
just lands the durable queue underneath, with a state machine that
mcpd's HTTP handlers, the worker result-POST route, and the GC sweep
will all build on.
Schema (src/db/prisma/schema.prisma + migration):
- New `InferenceTask` model + `InferenceTaskStatus` enum
(pending|claimed|running|completed|error|cancelled).
- Routing fields stored at enqueue time so a later rename of
`Llm.poolName` doesn't reroute already-queued work: `poolName`
(effective pool key), `llmName` (pinned target), `model`, `tier`.
- Worker tracking: `claimedBy` (providerSessionId) + `claimedAt`,
cleared on revert.
- Bodies as `Json`: requestBody (always set), responseBody (set at
completion). Streaming chunks are NOT persisted — too expensive at
delta granularity. The final assembled body lands once per task.
- Lifecycle timestamps: createdAt, claimedAt, streamStartedAt,
completedAt. Plus ownerId (RBAC + audit) and agentId (null for
direct chat-llm calls).
- Indexes for the hot paths: (status, poolName) for the dispatcher's
drain query, claimedBy for the disconnect revert, completedAt for
the GC retention sweep, owner/agent for the async API listing.
Repository (src/mcpd/src/repositories/inference-task.repository.ts):
- CRUD + state transitions as conditional CAS via `updateMany`. Two
workers racing to claim the same row both run the UPDATE; whichever
the DB serializes first sees affected=1 and gets the row, the loser
sees 0 and falls through to the next candidate. No application-
level locking required.
- findPendingForPools(poolNames[]) for the worker drain on bind.
- findHeldBy(claimedBy) for the unbindSession revert.
- findStalePending + findExpiredTerminal for the GC sweep.
Service (src/mcpd/src/services/inference-task.service.ts):
- Owns the in-process EventEmitter that wakes blocked HTTP handlers
when a worker POSTs results. The DB row is the source of truth for
*state*; the EventEmitter just signals "go re-read row X" so we
don't have to poll. Single-instance assumption for v5; pg
LISTEN/NOTIFY is the v6 swap when scaling horizontally — no schema
change needed, just replace the emitter wakeup.
- waitFor(taskId, timeoutMs) returns { done, chunks }: the terminal
promise + an async iterator of streaming deltas. Throws on cancel
(clear message) or error (worker's errorMessage propagates) or
timeout. Polls the row once at subscribe time so an already-
terminal task resolves immediately without waiting for an event
that's never coming.
- gcSweep flips stale pending rows to error (with a clear message
about the timeout) and deletes terminal rows past retention.
Defaults: 1h pending timeout, 7d terminal retention; both
configurable.
Tests:
- 6 db-level schema tests (defaults, json roundtrip, drain query
shape, claimedBy filter, GC predicate, agentId nullable).
- 13 service tests covering enqueue, the CAS race on tryClaim,
complete/fail/cancel, idempotent terminal transitions, revertHeldBy
on disconnect, and the full waitFor signal lifecycle (immediate
resolve, wake on event, chunk streaming, cancel/error/timeout
paths). Plus a gcSweep test with a fixed clock.
mcpd 881/881 (was 868; +13). db pool-schema 14/14, +6 new
inference-task-schema. Pre-existing failures in models.test.ts
(Secret FK fixture issue, also fails on main HEAD) are unrelated.
Stage 2 (next): VirtualLlmService rewires through this — remove the
in-memory pendingTasks map; enqueue creates a row, dispatch picks an
active session, the result-route updates the row + emits the wakeup.
Worker disconnect reverts; worker bind drains.
