snapdec — local decision tools

<!-- snapdec-skill: v0.3.0 -->

1. For "how do I run tests/lint/build here?" call project_facts(path) first; trust it over guessing.
2. For ≥5 items or fuzzy labeling, call classify/score/rank/check with clear class descriptions. ask is the raw escape hatch (mixed question types in one call).
3. Act only on results with decision:"auto". For review, decide yourself.
4. If a tool returns a reason (no_model, warming, timeout, no_backend, oversize, truncated, error), continue without it — never retry in a loop.
5. NEVER use these results to approve destructive commands, touch secrets, change permissions, or authorize network egress. They are advisory.
6. When a model decision drove an action, say so in your summary.

See references/tools.md (contracts), references/recipes.md (patterns), and examples/ (full payloads with sample outputs).

Skill ships with snapdec; re-running snapdec init refreshes this copy.