CLI

haiku.skills provides a haiku-skills command-line interface.

`validate`

Validate skill directories against the Agent Skills specification using the reference implementation:

haiku-skills validate ./skills/web ./skills/calculator

Prints VALID or INVALID with error details for each path. Exits with code 1 if any skill is invalid.

`sign`

Sign a skill directory with sigstore:

haiku-skills sign ./skills/my-skill

Writes a SKILL.sigstore bundle alongside the skill's SKILL.md. Each bundle carries exactly one signer; signing again overwrites the previous bundle. See Signing and verification for details.

`verify`

Verify a signed skill directory against trusted identities:

haiku-skills verify ./skills/my-skill \
    -i author@example.com --issuer https://accounts.google.com

Multiple identities can be provided (each --identity/-i needs a corresponding --issuer). Verification succeeds if the bundle's signer matches any of them:

haiku-skills verify ./skills/my-skill \
    -i author@example.com --issuer https://accounts.google.com \
    -i https://github.com/org/repo/.github/workflows/sign.yml@refs/heads/main \
    --issuer https://token.actions.githubusercontent.com

To verify cryptographic integrity without checking signer identity, pass --unsafe:

haiku-skills verify ./skills/my-skill --unsafe

Prints VERIFIED (or INTEGRITY OK with --unsafe) on success, FAILED on failure, and exits with code 1 on failure.

`list`

List discovered skills with name and description:

# From filesystem paths
haiku-skills list -s ./skills

# From entrypoints
haiku-skills list --use-entrypoints

# Both
haiku-skills list -s ./skills --use-entrypoints

If any skill directories have validation errors, they are printed as warnings to stderr while valid skills are still listed.

`chat`

A debug/development chat TUI built with Textual. Requires the tui extra:

uv add "haiku.skills[tui]"

Point it at a directory of skills for filesystem discovery:

haiku-skills chat -s ./skills -m openai:gpt-4o

Or use entrypoint discovery:

haiku-skills chat --use-entrypoints -m openai:gpt-4o

Filter to specific skills by name:

haiku-skills chat --use-entrypoints -k web -k codeexecution -m openai:gpt-4o

Set the model to use for skill sub-agents (overrides HAIKU_SKILL_MODEL env var):

haiku-skills chat -s ./skills -m openai:gpt-4o --skill-model ollama:llama3

The chat TUI uses the AG-UI protocol adapter for event streaming, making it useful for debugging skills with per-skill state:

State deltas are displayed inline as JSON Patch operations whenever a skill modifies state
Full state snapshot is available via the "View state" modal in the command palette

Environment variables

Variable	Description
`HAIKU_SKILLS_MODEL`	Default main agent model for `chat` (fallback when `-m` is not provided, defaults to `ollama:gpt-oss`)
`HAIKU_SKILL_MODEL`	Model to use for skill sub-agents (overridden by `--skill-model` or per-skill `model` in SKILL.md)
`HAIKU_SKILLS_PATHS`	Colon-separated skill directory paths (fallback when `-s` is not provided)
`HAIKU_SKILLS_USE_ENTRYPOINTS`	Set to `1`, `true`, or `yes` to enable entrypoint discovery by default