Delivery roadmap

Phase 0 — Evidence and decisions

Deliverables:

Verify OpenAI and Ollama request/response and streaming contracts from primary documentation.
Compare the relevant local-first gateways and extract concrete differentiation, not feature lists.
Select runtime, config format, package/distribution strategy, and supported operating systems.
Write ADRs for each consequential choice.

Exit criteria: a reviewed MVP API contract and a reproducible local development setup.

Deliverables:

GET /health and non-streaming POST /v1/chat/completions.
Ollama discovery/configuration and a local-only auto router.
OpenAI SDK contract tests plus real-provider integration tests where Ollama is available.
Content-free structured logs and a default-deny remote-provider policy.

Exit criteria: offline chat request succeeds against a local Ollama model with documented setup.

Deliverables:

Streaming, timeouts, bounded retries, circuit breaking, and explicit fallback chains.
Per-key rate limits and an opt-in cache with documented privacy semantics.
Read-only local usage/decision dashboard.

Exit criteria: deterministic fallback and observable error behaviour under integration tests.

Deliverables:

Explicitly configured cloud/OpenAI-compatible adapters.
Policy-controlled routing based on locality, capability, latency, and optional cost metadata.
Credential-store integration and configuration validation.

Exit criteria: no cloud request is possible without a user-visible allow rule and tests prove it.

Deliverables:

Stable provider-adapter extension contract.
MCP integration design and a constrained experimental implementation, if the threat model permits it.

Exit criteria: extension APIs do not weaken the default privacy or routing guarantees.