Local learning

The Coach uses evidence from individual calls to inform local routing. It does not retrain your language model, and it does not need a hosted Roster service.

Useful before a model download

Lexical retrieval uses SQLite FTS5 over capability names, descriptions, and skill bodies. It works without an embedding model. When direct matches are sparse, the draft can fill remaining positions from a local rating-based fallback.

This is a small shortlist for the agent to judge, not a promise that every candidate is relevant. State your immediate need clearly and re-draft when that need changes.

What a call teaches the Coach

The local outcome record includes a call class, latency, capability identity, and hashes used for attribution. Raw prompts, call arguments, and results are not persisted in the outcome log.

A call returning without detected failure is classified success. Transport faults, protocol errors, tool-reported errors, and output-schema problems are separate classes. Error subtypes use heuristics, so a label is not infallible diagnosis.

Caller-side input-validation failures are excluded from attributable ratings under the current policy. Skill instruction delivery is also excluded. Local call statistics are not certified League results.

Maintenance, not instant retraining

At router startup, opportunistic maintenance checks whether it is due, with an approximately 20-hour debounce. It recomputes local ratings and can refine stored capability vectors through OATS when suitable need vectors and outcome evidence exist.

OATS uses a trailing 90-day window, bounded evidence, and a minimum positive-signal requirement. It adjusts routing vectors; it does not change the underlying embedding model’s weights. A single click in the homepage example does not trigger model training.

The current approach has limits, including single-centroid learning and imperfect fallback behavior. The repository records these as measured limitations, rather than hiding them behind an accuracy promise.

Optional dense retrieval

Check whether the optional runtime can execute a local backend probe:

macOS / Linux

npx --yes @npmmo/roster@0.0.4 dense status

Windows

npx.cmd --yes @npmmo/roster@0.0.4 dense status

Add the optional runtime only if you want semantic matching:

macOS / Linux

npx --yes @npmmo/roster@0.0.4 dense enable

Windows

npx.cmd --yes @npmmo/roster@0.0.4 dense enable

The runtime is approximately 385 MB in the recorded clean-install measurement. A model download is additional and happens in the background on first use. Network conditions and cache state affect the actual footprint and first warmup.

A fresh packaged installation does not automatically install this optional peer dependency. The enabled runtime is stored under ~/.roster/runtime. A runtime already available through a source workspace can also be used.

If you previously set embeddings to "off", dense enable does not change that field. After consenting to model use, set it to "auto" in the existing config, preserve the other fields, and restart your client’s router process. dense status runs a bounded local backend probe without fetching model weights. READY means the reported backend can execute; it does not mean a model is warm or configuration permits embeddings.

Native first, portable fallback: Roster uses native inference when available and a local WASM worker when native bindings cannot load. This covers the verified Alpine/musl and Intel macOS cases that failed in 0.0.2. Worker runtime assets come from the installed package, not a CDN. The owned runtime’s WASM model cache lives under ~/.roster/runtime/cache/wasm; it stores public model artifacts, not embedding inputs. Once the model files are present, fresh-process offline inference is supported in the tested configurations. Missing or unusable backends preserve lexical routing. See the Alpine and Intel macOS notes.

The default model selection uses EmbeddingGemma on machines with at least 8 GiB RAM and MiniLM below that. Drafts keep using lexical results while warmup is incomplete or unavailable. These are local embedding models, not a choice of which AI agent model handles your conversation.

Stay lexical

Set embeddings to "off" to disable the dense path persistently. There is no dense disable subcommand. ROSTER_NO_FETCH=1 also disables the dense path for that process, including cached-model use.

Source builds deserve special care: their workspace includes the runtime, so declining the installation question alone does not prevent first-use fetching. The first-run guide explains the explicit lexical setup.

Next: Read about failures and drift.

Retrieval and maintenanceRuntime setupRecorded live inference checkNative compatibility checks