Roadmap and Phases

Roadmap & Phases

UASL's work is organized into phases, from the stable pre-phase baseline through router intelligence, an MCP service, a feedback loop, composition, and distribution.

Source docs: docs/PHASE_STATUS.md · docs/ADAPTER_DEPRECATION.md · docs/DISTRIBUTION_STRATEGY.md

---

Phases

Phase	Theme	Status
PRE-0	Baseline corpus + working router	✅ Stable
0	Foundations — canonical router verbs, the Go CLI, layer decoupling, the Node → Go migration, goreleaser packaging, schema + routing-eval harness	⏳ In progress
1	Semantic routing — close the recall gap on paraphrased prompts	Planned
2	MCP service — the MCP server surface over the shared engine	Planned / underway
3	Feedback loop — telemetry + reranker learning from routing decisions	Planned / underway
4	Composition — multi-step DAG compose --pipeline	Planned / underway
5	Distribution — signed per-skill packages	Planned

Note

Phase numbers describe capability tracks, not a strict sequence — several later-phase capabilities (MCP serve, reranker, compose pipeline) already exist in the engine while Phase 0 foundations are still being finalized. Treat docs/PHASE_STATUS.md as the live source of truth.

Phase 1 target (routing quality)

Phase 1 must improve routing recall while regressing nothing:

• Recall@5: +≥15 points over the Phase-0 baseline (0.6879).
• P@1: +≥10 points over baseline (0.5319).
• Abstention: zero regressions (baseline 1.0000).

These are measured by the routing-eval harness — see Performance and Testing & CI.

---

Adapter deprecation

The integration model changed fundamentally:

	Old model	New model
How clients get skills	Physical-copy propagation — a copy of the corpus in each downstream agent	One canonical engine (skillservice) reached via CLI or MCP
Drift	Each copy drifts independently	Single source of truth; no drift
Client footprint	Full corpus per client	A compact wrapper per client

The old copy-propagation adapters are deprecated in favor of the router + wrapper approach described in Architecture.

---

Distribution strategy

Distribution is a two-layer model:

Layer	What ships	How
A — Binary	The skill-router binary	goreleaser builds macOS (universal), Linux, and Windows artifacts; the release workflow signs checksums with cosign (keyless / sigstore OIDC) and can publish a Homebrew tap.
B — Corpus	The skills themselves	Today: whole-repo. Phase 5: per-skill signed packages (minisign + cosign).

See Testing & CI for the release workflow, and the Node → Go migration for how the registry generator consolidates into the single binary.

---

Related pages

• Architecture — the structure these phases build on
• Node → Go Migration — Phase-0 foundation work
• Performance & Benchmarks — the metrics phases are measured against