version: "1.0.0" name: skill-compiler description: Generate and run validation cases for a trusted skill, then inspect a saved report of results and assertion provenance. metadata: shelllm: requires: bins: ["bash", "python3", "jq", "llm", "shellm", "skills"]

skill-compiler

Use when developing a trusted skill or checking its documented behavior. Requires Bash, Python 3, jq, llm, shellm, and skills. Model calls use normal llm configuration and can incur charges. This produces validation reports, not trained adapters or automatically loaded examples.

sc="${SKILLS_DIR:-./.skills}/skill-compiler/skill-compiler"
"$sc" compile --skill shellm --num-tests 3 --max-iterations 5
"$sc" show --skill shellm
"$sc" pool --skill shellm
"$sc" clean --skill shellm

This skill requires explicit installation; default agents do not receive it. The executable is not on PATH. See README.md for installation and source-checkout usage. Use --teacher-model only to override the teacher's configured model. Defaults are up to eight cases and fifteen iterations per case.

Both compile and pool can execute a skill script with no arguments and with --help. Probes and tests use disposable state and work directories. Known messaging/live-store skills and aliases are refused. This is not an OS sandbox: only run trusted skills, since credentials and filesystem access are not revoked.

Inspect the JSON report's summary.fail as well as the exit status. Runner or setup failures return nonzero. Assertions require a successful runner and the named evidence from its actual root run. VERIFIED means the passing literals are attested in the offered vocabulary, not that the test proves useful behavior. Review generated cases and include negative examples.

Reports persist under <state-home>/skill-compiler/compiled, outside the skill installation. See README.md for outcome semantics, report fields, isolation limits, and optional pool/cache controls.