Skip to main content
Skills let you inject context files, prompts, rules, and reference documentation into agent trials. A skill is a directory containing a SKILL.md file.

​
How skills reach the agent

For each trial, Harbor:
  1. Resolves local and Git skill sources on the host into individual skill directories.
  2. Uploads each directory to /harbor/skills/<skill-name> in the agent sandbox. If the task declares environment.skills_dir, Harbor uses that directory instead.
  3. Passes the sandbox skills directory to the agent integration as skills_dir before agent setup.
  4. Lets the integration register the skills in its native location or pass the directory directly to the agent runtime.
For example, the Codex integration copies them into $HOME/.agents/skills, while Claude Code uses $CLAUDE_CONFIG_DIR/skills. Custom agents receive the same skills_dir constructor argument and should register its contents during setup() or run(). Job-provided skills are uploaded after the environment healthcheck and before agent setup. When multiple supplied skills have the same directory name, the last one wins. See Tasks → Skills to bundle skills directly into a task environment.

​
Local skills

Pass a local directory path:
The path can identify one skill containing SKILL.md, or a root whose immediate child directories each contain SKILL.md.

​
Git skills

Instead of checking skills into every project, you can reference a git repository. Harbor clones the repo, resolves a commit SHA, and caches the result locally.

​
Shorthand syntax

Shorthand always resolves to github.com and uses the repository’s skills/ directory.

​
Full URL syntax

Plain repository URLs use the repository’s skills/ directory. The /tree/<ref>/<subdir> format lets you point at a specific subdirectory of a monorepo.

​
Multiple skills

Use --skill multiple times to inject several skills. If two skills share the same directory name, the last one wins:

​
Caching

Git skills are cached locally at ~/.cache/harbor/skills/<host>/<org>/<name>/<sha>/. Once a commit is cached, subsequent runs skip the clone. Different subdirectories from the same repository and commit are handled automatically. To clear the cache:

​
Reproducibility

When a job runs, Harbor records each skill’s provenance in the job lock file:
  • name — the skill directory name
  • source — local path used during the run
  • digest — SHA-256 content hash of all files in the skill
  • git_url — the source repository (for git skills)
  • git_commit_id — the resolved commit SHA (for git skills)
This ensures that job results are fully traceable back to the exact skill content that was used, even if the branch has since moved.