Agentcompat

agentcompat

Which agent extension features actually work on which agent.

Your `SKILL.md` sets `model: opus`. On Claude Code that switches the model. On Grok Build the field parses fine, does nothing at all, and nobody tells you. Your `allowed-tools` list pre-approves tools on one agent and grants nothing on the other, which is a security difference, not a cosmetic one.

Spec linters cannot catch this. The file is valid. It just does not do what you think on the agent someone else is running.

npx agentcompat lint .

.claude/skills/demo/SKILL.md
     4  ignored     `model` on Grok Build
        Accepted by the parser and never applied. The session model is used instead, with no warning.
     5  ignored     `allowed-tools` on Grok Build
        Parsed, but it neither grants nor restricts tools. A skill relying on it for a permission grant silently gets nothing.

.claude/hooks/start.sh
     2  unsupported `CLAUDE_PROJECT_DIR` on Grok Build
        Not set. The nearest equivalent carries a GROK_ prefix, so a hook written for one agent reads an empty value on the other.

Usage

agentcompat lint .                    # warn where this project relies on ignored behaviour
agentcompat lint . --agent grok-build # only care about one agent
agentcompat lint . --json             # machine-readable
agentcompat table                     # print the matrix as markdown
agentcompat missing                   # every cell nobody has verified
agentcompat agents                    # the agents being tracked
agentcompat check                     # validate the data files

`lint` exits 1 when it finds something an agent silently ignores, so it can gate a pull request.

The matrix

Value	Means
yes	Works as documented
no	Not implemented. The agent errors, or there is no equivalent
partial	Works, but not fully or not the same way. See the note
ignored	Accepted without compl