Skills

The catalogue of the agent's skills is imported with:

php artisan griglia:skills-import --file=skills.json   # or JSON on stdin: [{name, description, source, agents}, …]

The task modal shows it as an accordion (with live search); the skills you tick for a task are printed by griglia:check (skills to activate for this task: …) so the agent invokes them. For generated plans, PlanBuilder assigns pertinent skills task by task, considering only the catalogue available to the plan's default agent. You do not have to write the JSON: the package ships sync-skills.py, which reads the Claude Code, Codex and Gemini skill folders of the machine the agent runs on and imports them.

php artisan vendor:publish --tag=griglia-scripts   # → scripts/ in your project
python3 scripts/sync-skills.py                             # host: reads the folders and imports (--print to just look)

These helpers run on the host, not in the container: they read files that only exist there (skills, agent credentials, transcripts). See the scripts.

One catalogue, several agents

The SKILL.md format is portable — frontmatter plus instructions in markdown, nothing tied to a particular CLI — but a skill only exists for the agent that finds it on disk: what is installed in ~/.claude/skills is invisible to Codex CLI, and the built-in skills of a CLI cannot be installed anywhere at all. Each entry therefore carries agents, the keys (from griglia.agents) of the agents that can use it, and the modal of a task only offers the skills of its agent plus the ones with no agents at all — shared, or imported before this field existed. A skill already ticked stays visible even if the task changes agent, so you can untick it.

sync-skills.py fills the field from the folder it read the skill from: ~/.claude/skills, project .claude/skills, plugins and built-ins → claude; ~/.codex/skills and project .codex/skills → codex; ~/.gemini/skills → gemini; the shared ~/.agents/skills → no constraint. The same skill found in two folders is available to both agents.

The JSON is a plain list:

[
  { "name": "tdd", "description": "Test-driven development", "source": "user", "agents": ["claude"] },
  { "name": "code-review", "description": "Review a branch", "source": "plugin", "agents": ["claude", "codex"] },
  { "name": "commit-style", "description": "How we write commits", "source": "agents" }
]

See also

• The agent side — how the chosen skills reach the agent.
• Agent context — the instructions file behind them.