JAIPilot/jaipilot

Make Java smaller, newer, and faster without guessing

Optimize for verified value, not fewer lines, newer version numbers, or fashionable constructs. Never claim universal safety: Java reflection, frameworks, configuration, external consumers, and unmeasured workloads make that impossible. Fail closed when required evidence is unavailable.

Select the requested modes

Read only the references needed for the request:

• unused.md: remove repository-proven unused code, dependencies, or resources;
• consolidate.md: merge genuinely equivalent logic and reduce classes, methods, and lines without creating a generic abstraction;
• modernize.md: upgrade the JDK, build, framework, plugins, and public dependencies to verified compatible stable releases; and
• performance.md: improve measured algorithms, allocation, I/O, and concurrency, including virtual threads when the workload proves they fit;
• architecture-and-rewrites.md: use ArchUnit for a stable architectural invariant or OpenRewrite for a bounded deterministic transformation.

Use only the modes requested by the user or controlling workflow. For a comprehensive request, use this order unless repository constraints justify another: remove proven unused leaves, consolidate equivalent behavior, optimize measured workloads, then modernize justified build or runtime paths in isolated batches. Evaluating a mode does not require editing in it.

Establish the shared boundary

1. Confirm that the selected root contains a Java Maven or Gradle project. Otherwise report that this skill is not applicable and stop.
2. Read repository instructions, build files, modules, source sets, generated-source rules, profiles, tests, resources, packaging, supported runtimes, and public API policy.
3. Record Git status and save the tracked diff plus an inventory and content snapshot of relevant untracked files for later comparison. Preserve edits outside the explicit scope. Do not assume every dirty line is unrelated: an in-scope candidate belongs to the requested cleanup, while unclear ownership stays untouched and is reported.
4. Define whether the boundary is a closed application or a published library, plugin, SDK, framework extension, or service with downstream consumers.
5. Identify the repository's focused checks and normal verification command. Record known existing failures and never make a failing check look green by weakening measurement.

Work in reversible evidence-backed batches

1. When an architecture rule or automated rewrite would materially improve proof, follow architecture-and-rewrites.md. Prefer configured tools; never add ArchUnit, OpenRewrite, recipes, plugins, or build configuration without approval.
2. Use an isolated worktree or equivalent reversible copy when relevant state can be reproduced without hiding dirty work. Never reset, clean, or stash unrelated work.
3. Apply one coherent leaf, consolidation seam, performance hypothesis, or upgrade axis at a time. Compile and run the narrow relevant tests after every material batch.
4. Reverse only a failed batch. Recompute references, compatibility, and measurements before the next batch because one accepted change can alter later evidence.
5. Use the jaipilot-generate-tests skill when the request includes tests or coverage, or when a retained change exposes a specific regression gap. Do not add hollow tests merely to permit a rewrite.
6. After integration, use the jaipilot-review-diff skill to inspect the complete Java and build diff, then run the repository's applicable final clean proof.
7. Use the jaipilot-remote-java skill when a build, analyzer, profiler, or benchmark materially benefits from disposable remote hardware and the exact state is already a GitHub-available commit. Do not treat committed remote evidence as proof of dirty local edits.

Shared acceptance rules

Accept a change only when:

• the observable contract, errors, ordering, transactions, security, resource ownership, and framework lifecycle remain intentional;
• public API and downstream compatibility match the declared boundary;
• configured tests, architecture checks, static analysis, packaging, and relevant profiles pass;
• no dependency, exclusion, suppression, warning, timeout, threshold, or test was weakened to pass; and
• the final benefit is measured in the mode's terms and outweighs added indirection or risk.

Prefer zero net change when proof is incomplete. A passing suite is evidence, not a universal guarantee.

Report

Return:

• modes, boundary, initial worktree state, baseline, and unavailable evidence;
• candidate or hypothesis ledger with accepted, rejected, and retained items;
• ArchUnit rules and violations, plus OpenRewrite recipes and proposed edits accepted or rejected;
• exact commands and pass, fail, skipped, or unavailable results;
• before/after code shape, resolved versions, or performance measurements as applicable;
• the final diff relative to the saved starting state, not only relative to HEAD;
• external-consumer, runtime, workload, and profile assumptions; and
• remaining risk plus how to reverse the uncommitted patch.