Working in Frost

Frost recommends a model and execution profile for a natural-language task. Profiles use a harness or API interface, with effort when supported. Slices 1, 1a, and 2 are implemented: cmd/frost over internal/cli, the deterministic selection and capacity policy in internal/router, neutral snapshot parsing in internal/capacity, the TypeSafe adapter in internal/analyzer/typesafe, operator config in internal/config, the catalog reader in internal/catalog, and the separate public-data producer in tools/catalog-build. The controlling plan is docs/plans/active/model-router.md; completed slice plans live in docs/plans/implemented/. experiments/probe is the original feasibility probe and its recorded evidence, not the CLI. Run make check before handing off.

Follow the project-standards skill when available. Use Go, a small shared core, and replaceable adapters for external judgment providers. Keep the CLI non-interactive with structured output. Source-attributed model evidence, operator profiles, and transient capacity have separate owners. Personal rankings are optional. Public-catalog and usage producers stay outside the router. The current public catalog has no applicable task-family measurements for the configured models, latency classes remain priors, and paired outcome evidence is still missing; preserve those limits in user-facing claims. Do not launch a recommended model as a side effect of asking for a recommendation.

Read td usage --new-session -q in a new context. Track substantive work in this repository's td workspace, and record review honestly. Preserve unrelated changes. Never operate on the default tmux server in tests.

Use the TypeSafe skill when working on its adapter. The experimental local copy is /Users/marcus/code/clara-home/skills/typesafe-ai/SKILL.md on Marcus's machine; the portable upstream is https://github.com/typesafe-ai/skills/tree/main/skills/typesafe-ai. Treat live vendor documentation as API authority, and retain exact question sets and responses for reproducible experiments.

Keep credentials in the environment. Do not print keys, commit secrets, or automatically source a user's shell files. Synthetic eval fixtures are shareable; real tasks and run artifacts are local by default. Keep Markdown paragraphs unwrapped.