CLI Reference

skillci <command> [options]

Invoke via the built binary (node dist/cli/index.js …), or npm run demo for the offline run. After npm link / global install, the skillci binary is on your PATH.

`skillci run`

Run a baseline-vs-candidate evaluation (or the offline demo) and print the report.

Flag	Default	Description
`--agent <kind>`	`claude-code`	Target agent: `claude-code` \| `cursor` \| `codex`.
`--baseline <dir>`	—	Directory holding the baseline (trusted) config.
`--candidate <dir>`	—	Directory holding the candidate (proposed) config.
`--tasks <dir>`	bundled samples	Directory of task definitions.
`--demo`	`false`	Fully-offline run over the sample tasks via the mock adapter.
`--open-pr`	`false`	Open a real PR when promotable (otherwise dry-run).
`--no-color`	—	Disable colorized output.

Behavior notes

If --baseline or --candidate is omitted, run falls back to demo mode.
The real agent adapter is used when available; otherwise it falls back to the offline mock (and says so).
Judge backend: Anthropic SDK when ANTHROPIC_API_KEY is set, else the claude -p CLI judge when judging Claude and its binary is present, else skipped.

Exit codes

Code	When
`0`	Verdict `improved` or `neutral`.
non-zero	Verdict `regressed` (use as a CI gate).

Examples

# Offline, deterministic
skillci run --demo

# Live, on a Claude subscription (no API key)
skillci run --agent claude-code --baseline ./cfg/base --candidate ./cfg/cand

# Custom task suite, open a PR if promotable
skillci run --agent claude-code --baseline ./base --candidate ./cand \
  --tasks ./my-tasks --open-pr

`skillci validate <dir>`

Discover and validate the config artifacts for an agent in a directory — a quick sanity check that your skills/hooks/rules parse before you evaluate them.

Flag	Description
`--agent <kind>`	Agent to interpret the directory as (default `claude-code`).
`--baseline <dir>`	Optional baseline to diff against (reports added/removed/modified artifacts).

skillci validate ./my-config --agent claude-code
skillci validate ./pr-config --agent claude-code --baseline ./main-config

`skillci tasks`

List the available task definitions (bundled samples by default).

Flag	Description
`--tasks <dir>`	List tasks from a custom directory instead of the samples.

skillci tasks
skillci tasks --tasks ./my-tasks

Output shows each task's id, target agent, title, check count, and whether it's judged.

Global

skillci --help / skillci <command> --help — usage.
skillci --version — version.