v0.4.0: Fix correctness bugs, add lossless reconnect, outbound track…

[ajit-zer07]

Summary

• Fix premature run finalization: Commitment messages no longer synthesize SESSION_STATE_RESOLVED — runs only complete
  when the runtime reports resolution via session-snapshot or GetSession reconciliation
• Lossless reconnect: Stream cursor persisted after each event; recovery resumes from max(persisted_cursor, last_event_seq)
• Outbound message tracking: New run_outbound_messages table tracks all kickoff/signal/context messages with lifecycle
  status
• Operational resilience: Durable webhook outbox, distributed recovery locking via advisory locks, batch recovery reporting,
  6 new Prometheus counters
• New endpoints: POST /runs/validate (preflight), POST /runs/:id/context, POST /runs/:id/projection/rebuild, GET /runs/:id/messages, GET /runs/:id/export/stream (JSONL)
• Horizontal scaling: Redis pub/sub StreamHubStrategy for multi-instance SSE fan-out

Changes by Phase

Phase 0: Critical Correctness Fixes

Fix	File(s)
clientVersion from package.json	app-config.service.ts, run-executor.service.ts
Remove premature Commitment → RESOLVED	event-normalizer.service.ts
Remove decision.finalized auto-finalize	stream-consumer.service.ts
Atomic finalization via finalizingPromise	stream-consumer.service.ts
findByIdOrThrow → NotFoundException	run.repository.ts
console.error → NestJS Logger	main.ts

Phase 1: Runtime Integration Fidelity

• POST /runs/:id/context — send context updates to running sessions
• POST /runs/:id/projection/rebuild — rebuild projection from persisted events
• Persist runtime capabilities from Initialize response (4 new columns on runtime_sessions)
• schema_version column on run_events_canonical
• Fix webhooks.active from integer to boolean

Phase 2: Lossless Reconnect & Outbound Tracking

• Stream cursor persistence for crash recovery
• run_outbound_messages table + repository + GET /runs/:id/messages
• OutboundMessageSummary in RunStateProjection

Phase 3: Signal & Progress Enrichment

• signalType, severity fields on SendSignalDto
• TaskUpdate/TaskComplete/TaskFail → additional progress.reported events

Phase 4: Operational Resilience

• Recovery batch result reporting { recovered, failed }
• Durable webhook outbox with webhook_deliveries table
• Advisory lock–based distributed recovery (pg_try_advisory_lock)
• 6 new Prometheus counters (outbound, inbound, signals, reconnects, recovery, webhooks)

Phase 5: New Features & Scalability

• POST /runs/validate — preflight validation without creating a run
• RedisStreamHubStrategy for horizontal SSE scaling
• GET /runs/:id/export/stream — streaming JSONL export
• src/runtime/grpc-types.ts — typed gRPC interfaces

Migrations

File	Purpose
0006_capabilities_and_stream.sql	capabilities + stream cursor on runtime_sessions
0007_canonical_schema_version.sql	schema_version on run_events_canonical
0008_webhook_active_boolean.sql	integer → boolean for webhooks.active
0009_outbound_messages.sql	run_outbound_messages table
0010_webhook_deliveries.sql	webhook_deliveries table

New Files

File	Purpose
src/storage/outbound-message.repository.ts	CRUD for run_outbound_messages
src/webhooks/webhook-delivery.repository.ts	CRUD for webhook_deliveries
src/events/redis-stream-hub.strategy.ts	Redis pub/sub StreamHubStrategy
src/runtime/grpc-types.ts	Typed interfaces for gRPC shapes

Test plan

• 253 tests passing (up from 248), 20 suites
• TypeScript compiles cleanly (tsc --noEmit)
• 0 lint errors
• Run npm run drizzle:migrate against a live database
• POST /runs with decision-mode request → verify run completes only via runtime session-snapshot
• Kill and restart control plane during active run → verify recovery resumes from persisted cursor
• POST /runs/validate with unsupported mode → verify error response
• POST /runs/:id/context during running session → verify context update flows through