Java engineering skills that help coding agents make smaller, safer, better-proven changes.
JAIPilot is a skills-only plugin for ChatGPT, Codex, and Claude Code. It gives the coding agent three focused workflows for reviewing Java diffs, generating tests, and cleaning Java code.
There is no JAIPilot runtime, MCP server, dashboard, background process, installer, or automatic hook. Installing the plugin adds Markdown instructions and brand assets—nothing starts, downloads, or scans a repository.
Large Java repositories contain behavior, integrations, exceptions, and architectural constraints that no agent can hold perfectly in context. Agents drift sooner when code has hidden coupling, duplicated logic, weak tests, and unclear boundaries.
JAIPilot does not try to replace the agent or build another analysis platform. It gives the agent a repeatable engineering checklist:
- establish the exact requested scope;
- preserve unrelated work;
- inspect the complete Java and build diff;
- use the repository's existing Maven or Gradle wrapper;
- use configured tests, JaCoCo, PIT, ArchUnit, OpenRewrite, and static analyzers when applicable;
- keep the smallest coherent change; and
- report what was actually run, what passed, and what remains unknown.
codex plugin marketplace add JAIPilot/jaipilot
codex plugin add jaipilot@jaipilotRun inside Claude Code:
/plugin marketplace add JAIPilot/jaipilot
/plugin install jaipilot@jaipilot
No Java version is required to install JAIPilot itself. The Java repository still needs its normal JDK, build wrapper, dependencies, and configured engineering tools.
| Skill | Use it for |
|---|---|
| jaipilot-review-diff | Review a Java Git diff, find risk or unnecessary code, and run the repository's applicable verification. |
| jaipilot-generate-tests | Raise meaningful per-class Java coverage with bounded parallel workers and fresh configured JaCoCo or PIT evidence. |
| jaipilot-clean-java | Safely remove unused code, consolidate logic, modernize Java/dependencies, and optimize measured performance. |
Example requests:
Use JAIPilot to review and verify my current Java diff.
Use JAIPilot to raise every eligible class in the orders module toward at least 80% line coverage.
Use JAIPilot to remove everything the repository can prove unused, retaining anything uncertain.
Use JAIPilot to consolidate equivalent services and reduce classes without changing behavior.
Use JAIPilot to upgrade this project to the newest stable JDK and dependencies it can prove compatible.
Use JAIPilot to optimize this workload from profiler and benchmark evidence.
The skills can also trigger naturally when the request clearly matches their descriptions.
For a coverage campaign, the test skill assigns one production class to each available worker and runs bounded waves in isolated worktrees or build outputs. It does not run concurrent Maven or Gradle processes in one checkout, assume shared tests are runtime-independent, or turn missing JaCoCo data into a pass. After integration, the agent refreshes one aggregate report and lists the baseline and final coverage of every eligible class, including honest blockers below 80%.
The clean skill has four composable modes: fail-closed unused removal, behavior-locked consolidation, verified stable modernization, and measured performance optimization. It can invoke the test skill for characterization and regression coverage and the review skill for final proof. It does not treat fewer lines, larger version numbers, virtual threads, or one passing test suite as success by themselves. Anything uncertain stays in place and is reported.
Every workflow ends with a concise record:
Scope: module and files reviewed
Changes: what the agent changed and why
Commands: exact builds, tests, and analyzers executed
Evidence: passed, failed, skipped, or unavailable results
Limitations: anything that was not measured
Missing JaCoCo, PIT, ArchUnit, OpenRewrite, or another analyzer is reported as unavailable. JAIPilot does not silently add tooling, lower thresholds, or describe an unmeasured property as passing.
JAIPilot is procedural guidance, not an enforcement engine. It cannot guarantee correctness, business intent, coverage, architecture, or agent compliance. Its value is helping the agent use the real repository toolchain consistently and explain the resulting evidence.
The plugin has no telemetry and uploads nothing. Repository commands may still execute project build scripts and resolve dependencies according to that project's configuration. Treat untrusted builds as executable code.
The previous runtime-based implementation—including MCP, metrics, and the local dashboard—is preserved at v4.0.8.
- JAIPilot Cloud campaign results — a separate bounded GitHub App experiment
- Contributing
- Support
- Security
- Privacy
- Terms
- Changelog
Licensed under the MIT License.