claude-harnesses

Monitor a pull request after opening it or resume a stalled PR, fix CI failures and reviewer feedback, push updates, reply to and resolve review conversations, and comment with the outcome.

Consult a fresh, review-only Claude subagent as a fallback for environments where the native Advisor tool is unavailable. Use when the user asks for an adviser, advisor, second opinion, stronger-model review, or /advisor-like behavior; before committing to a consequential multi-step approach; when work is stuck or changing direction; and before declaring substantial work complete.

Compatibility alias for pr-guardian. Use pr-guardian for stalled pull requests, CodeRabbit/Codex review feedback, CI failures, unresolved review conversations, and merge-ready PR follow-up.

Run an end-to-end Claude Code delivery workflow from user request to merge-ready pull request, including planning, implementation, tests, self-review, ready-for-review PR creation, CI/review follow-up, and final status.

Analyze recent Claude Code work and package repeated patterns as the smallest useful skill, custom subagent, slash command, hook, or automation. Use when asked to mine prompts, sessions, shell history, or recent repeated work for reusable Claude Code assets.

Fetch public web pages with Jina AI Reader API and convert them into LLM-friendly Markdown. Use when an agent needs to read, summarize, quote briefly, extract facts from, or inspect a public URL that normal browsing, scraping, search snippets, or direct fetch tools cannot read cleanly, including JavaScript-heavy pages, reader-hostile pages, X/Twitter posts and threads, articles, GitHub pages, and PDFs. Trigger when the user asks to use Jina, use r.jina.ai, read an otherwise unreadable site, recover readable text from a URL, read an X/Twitter post by URL, or turn a webpage into Markdown.

Manage durable task goals for long-running Claude Code work using the project ledger, including objective writing, success criteria, resumption, and completion decisions.

Run a continuous product improvement loop: evaluate a product or codebase, present prioritized recommendations for approval, then implement only approved improvements through verification and PR follow-up.

Direct high-quality frontend UI work by using image-generated mockups when useful, translating the chosen direction into code, and verifying the rendered UI in a browser.

Teach the human everything that happened in this session until they can demonstrably explain it, using an incremental checklist, restate-first probing, and AskUserQuestion quizzes. Use after a non-trivial change, debugging session, or hand-off when the goal is the human's understanding, not just a working result.

Iteratively improve a skill or prompt by having a fresh subagent execute it against fixed scenarios, scoring two-sided (executor self-report + a pre-frozen requirements checklist), and looping until improvements plateau. Use after authoring or revising any skill, slash command, or CLAUDE.md addition.

Fix bugs with reproduction, minimal changes, regression coverage, and targeted verification.

Remove AI-generated boilerplate (vague comments, unused try/except, defensive validation, restated docstrings) from changed code without changing behavior.

Update documentation after code or behavior changes while keeping docs concise and relevant.

Implement a feature safely with planning, focused edits, tests, docs, and verification.

Diagnose and repair failing CI checks on the current PR by parsing logs, reproducing locally when possible, and pushing focused fixes in batches.

Run multi-hour or multi-session work safely by combining the ledger, checkpoints, cost-ceiling guard, and resumable progress notes so context loss does not erase progress.

Refactor code without behavior changes using small mechanical steps and verification.

Check whether a change is ready to release by reviewing diff, tests, docs, risks, and migration notes.

Review code changes with a bug-first, risk-focused workflow before merge or release.

Run a multi-pass security review over the current diff using parallel specialist subagents focused on injection, authn/authz, secrets, supply-chain, and infrastructure-as-code risks.

Review changed code on the current branch for reuse, clarity, and dead complexity, then propose or apply minimal cleanup without changing behavior.

Drive development with red-green-refactor using isolated test-writer, implementer, and refactorer subagents so tests reflect intent rather than anticipated implementation.