Phase 3.5 — Tracks

Track A — Worker Infrastructure

Celery + Redis is the preferred starting worker stack for bounded extraction tasks. Alternatives (Temporal, Arq, Dramatiq) remain open for later investigation if workload shape changes.

Preferred sequencing within track: 3.5.A.1 → 3.5.A.2

Track B — Extraction Pipeline

LLM-powered memory extraction with structured output, cost-tiered execution, incremental turn tracking, and quality eval harness.

Preferred sequencing within track: 3.5.B.1 → 3.5.B.2 → 3.5.B.3 → 3.5.B.4 after Track A.

Track C — Context Assembly Engine

Token budget calculation, parallel retrieval, relevance-gated scoring, packing, context formatting, and gRPC service with degradation contract. Packer/scorer choices are planning defaults and should be validated against latency budgets.

Preferred sequencing within track: 3.5.C.1 → 3.5.C.2 → 3.5.C.3 → 3.5.C.4 → 3.5.C.5 → 3.5.C.6

Track D — Proxy Integration

Go gRPC client, handler wiring with fail-open semantics, response metadata, extraction enqueue, config/feature flags, and integration tests.

Preferred sequencing within track: 3.5.D.1 → … → 3.5.D.6 after Track C.

Track E — MCP Tool Surface

MCP tools backed by the same memory substrate as the proxy path, with observability and rate limiting.

Preferred sequencing within track: 3.5.E.1 → 3.5.E.2 → 3.5.E.3 → 3.5.E.4

Track F — Exit Gate

End-to-end learning-loop tests, cross-tenant isolation matrix, benchmark sign-off, and ADR index consolidation.

Preferred sequencing within track: 3.5.F.1 → 3.5.F.2 → 3.5.F.3 after Tracks A–E.