Skip to content

Agent skills

DejaVu's four skills use the open Agent Skills format. Each is a portable folder containing SKILL.md and any required references. They teach an agent how to use DejaVu; your project still needs its normal Gradle, Android SDK or multiplatform test environment to run the resulting tests.

Skill Use it for
dejavu-onboarding Install DejaVu and prove tracking in a first test
dejavu-test-writer Add recomposition assertions to an existing Android or KMP test setup
dejavu-error-triage Diagnose failures using source evidence
dejavu-perf-loop Reduce unnecessary application recompositions within an agreed budget

The skills preserve deliberately inefficient accuracy fixtures and distinguish diagnosis from authorized optimization.

Install with the Skills CLI

From your application's repository, with Node.js/npm installed:

npx skills add mttmcknn/dejavu

Choose the skills and agents interactively. The Skills CLI supports Codex, Claude Code, Cursor, GitHub Copilot, OpenCode and other agents. To select agents explicitly:

npx skills add mttmcknn/dejavu --skill '*' --agent codex cursor

Use the agent identifiers claude-code, github-copilot or opencode for those clients. Add --global to install across projects, or --copy when symlinks are unavailable. Install only one skill with --skill dejavu-test-writer.

To preview available skills without installing:

npx skills add mttmcknn/dejavu --list

To test unpublished changes from a local checkout, replace mttmcknn/dejavu with the absolute path to that checkout. GitHub installation uses the repository's default branch; an unmerged PR is not part of that installation.

Claude Code marketplace

The existing plugin remains supported. Run these commands inside Claude Code:

/plugin marketplace add mttmcknn/dejavu
/plugin install dejavu@dejavu

Use either the plugin or standalone skills for a given agent, to avoid duplicates. The plugin reads the same skills/ folders; no Claude-specific instruction fork is maintained.

Manual installation and other clients

Download or clone the repository and copy the complete desired folders from skills/ into your client's documented skill directory. Include references/, not just SKILL.md. The canonical folders contain real files, so copying them does not depend on symlinks or a retained DejaVu checkout.

For example, from the DejaVu checkout, copy the four folders into a fresh Codex project skill directory (replace the destination with your app's path):

mkdir -p /path/to/your-app/.agents/skills
cp -R skills/dejavu-* /path/to/your-app/.agents/skills/

Codex discovers project skills in .agents/skills/. Other clients may use different paths or import commands; follow their current documentation. If skill folders with these names already exist, replace the old folders deliberately rather than merging stale references into a newer bundle. For agents without native skill support, explicitly ask the agent to read the chosen SKILL.md and its referenced files; automatic discovery is unavailable there.

Verify and update

For CLI installs, run npx skills list (add --global for global installs). Reload the agent's skill catalog or start a new session, then check that the four names appear. Ask it to use dejavu-test-writer on a small existing test. Successful installation establishes discovery; run the actual test to verify its output.

Use npx skills update for CLI-managed installs, the plugin updater for marketplace installs, or replace manually copied folders from the chosen source revision. For reproducible manual installs, check out a reviewed commit before copying and record that revision. Skill bundle 0.3.0 is independent of DejaVu library versions; follow each skill's compatibility guidance for the library you use.

Maintaining the shared source

Canonical files live in skills/. In the DejaVu checkout, .agents/skills/ and .claude/skills/ contain discovery links to those files. On checkouts without working symlinks, use the CLI's copy mode in your consumer project or manual copying.

Packaging checks verify complete bundles and discovery links. The evaluation guide describes behavioral comparisons. Its current execution adapter is Codex; portable distribution does not imply that live evaluations have run on every supported client.