run-skill-tunerlisted

skill-tuner is a **CLI**, not a server or a GUI — `skills/skill-tuner/scripts/tune.py` with seven subcommands. There is nothing to click and no window to screenshot. The thing that shapes how you drive it: **the eval subcommands spend real money.** Every one shells out to `claude -p`. But most of the surface is reachable for free — `verify` and `compare` make zero model calls by design, the 177-test suite runs on a fake adapter, and the refuse-to-spend guards are observable precisely because they fire *before* the first call. All paths below are relative to the repo root. ## Run (agent path) — start here ```bash ./.claude/skills/run-skill-tuner/smoke.sh ``` 12 checks, ~30 s, **$0.00**, no network. Covers the environment preflight, the unit suite, the stdlib-only guard, all three spend guards, `verify`, `compare`, the too-few-cases refusal, and the differential check against skill-creator's own reader. Exits non-zero if any check fails. Output from a verified run (verdict lines such as `DRIFTED` are point-in-time — they depend on the banked runs' current state): ``` environment ok doctor.sh: environment ready unit suite (fake adapter — never calls a model) ok Ran 177 tests ok shipped runner imports stdlib only (R7) refuse-to-spend guards (observable because they fire before the call) ok unmetered run refused before any call (AE3) ok prints a cost estimate before spending (R8) ok aborts non-interactively without --yes ok a refused pre-flight