autonomous-bug-fix · git:20260519.40bb110 · 2026-05-19 · sha256 5ae0ef35cdb84ec2
autonomous-bug-fix git:20260519.40bb110A
Immutable. This exact content never changes and is served at /api/v1/blob/5ae0ef35cdb84ec2.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_b6eciv6n52qv46ty/5ae0ef35cdb84ec2a9a2985efeac52acec980da3d43dfb23b47042f078d8e07d.json
Then /plugin install autonomous-bug-fix@markdownregistry-b6eciv6n52qv46ty in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: autonomous-bug-fix description: Reproduces, diagnoses, fixes, and verifies bugs unaided. tier: core category: discipline created_by: human platforms: [windows, macos, linux] tags: [debugging, autonomy, root-cause] author: Andreas Wasita (@andreaswasita) --- # Autonomous Bug Fix Skill Walks the agent through the full reproduce → diagnose → fix → verify cycle on its own, without asking the user for navigation help, log-reading instructions, or test commands. Does NOT cover net-new feature work — for that use `plan-before-code` and `executing-plans`. ## When to Use - User reports a bug, regression, or unexpected behavior. - CI is red on a branch you own. - An error or warning appears in logs. - A previously working feature stopped working. - The user says "this is broken," "it doesn't work," or similar. ## Prerequisites - Read access to the failing component (use `glob` / `grep` to locate it). - The `powershell` tool to run tests, `git`, and reproduction commands. - A clean working tree on a feature branch (use `using-git-worktrees` if needed). - `verify-before-done` skill loaded — you'll invoke it at the end. ## How to Run ```text 1. Reproduce → locate the code, run the failing path, confirm with evidence. 2. Diagnose → trace from symptom to root cause; check recent changes. 3. Fix → minimal change at the root cause + a regression test. 4. Verify → re-run reproduction + full suite via `verify-before-done`. ``` ## Quick Reference | Phase | Action | Tool | |---|---|---| | Find code | Search by symbol or file pattern | `grep`, `glob` | | Read code | Inspect the failing path | `view` | | Reproduce | Run failing command/test | `powershell` | | Diagnose | Recent changes / blame | `powershell` → `git log -20 --oneline`, `git blame <path>` | | Edit | Minimal, root-cause fix | `edit` | | Add test | Regression test | `create` or `edit` under `tests/` | | Verify | Re-run reproduction + full suite | `powershell`, then `verify-before-done` | ## Procedure ### Step 1: Reproduce Find the relevant code yourself — `grep` for the symbol, `glob` for the path. Do not ask the user "which file is this in?" Run the failing path via the `powershell` tool. If no failing test exists, write a minimal one that captures the bug. Confirm reproduction with concrete evidence (stack trace, exit code, log line). If you can't reproduce: - Check for environment-specific behavior (OS, Node/Python version). - Look for race conditions or state-dependent paths. - Ask the user for clarification only as a last resort, and only with specific questions. ### Step 2: Diagnose Root Cause Read the error message and stack trace before doing anything else — they exist for a reason. Trace from the user-visible symptom back to the root cause: ```bash # via the `powershell` tool git log -20 --oneline git blame path/to/suspect.ts git diff HEAD~5 -- path/to/suspect.ts ``` Stop at the root cause, not the first thing that looks suspicious. "Why did this break?" — not "what broke?" ### Step 3: Fix at the Root Edit with precision via the `edit` tool: - Change only what needs to change. Bigger diffs hide bigger bugs. - Fix the root cause, never the symptom. Swallowing an exception is not a fix. - Add a regression test that fails without the fix and passes with it. - Check whether the same root cause exists elsewhere (`grep` for the pattern). If yes, fix those too or log them in `tasks/lessons.md` for follow-up. ### Step 4: Verify Hand off to the `verify-before-done` skill: - Re-run the reproduction. The error must be gone. - Run the full test suite. No regressions. - Record evidence in `tasks/todo.md` under the task. ### Step 5: Rollback If Verification Fails If verification fails after your fix: 1. Reset the working tree (`git checkout -- .` or `git stash` via `powershell`). 2. Log the failed attempt in `tasks/lessons.md` with the root-cause analysis you had. 3. Re-plan with a different approach via `plan-before-code`. 4. **Never** push broken code hoping someone catches it later. ## Pitfalls - **DO NOT** ask the user "which file is it in?" or "how do I run the tests?" Use `grep`/`glob`/`view`/`powershell` to discover it. - **DO NOT** apply a band-aid (catch-and-swallow, ignore the warning). Fix the root cause. - **DO NOT** fix the symptom. A login button that fails because of a missing provider is a provider bug, not a button bug. - **DO NOT** skip the regression test. Without one, the bug comes back in 3 months. - **DO NOT** push without verification. An unverified fix is often worse than no fix. - **DO NOT** abandon failed attempts silently. Log them in `tasks/lessons.md`. ## Verification - [ ] Reproduction was confirmed with concrete evidence before any fix. - [ ] Root cause is identified in writing (in commit message or `tasks/todo.md`). - [ ] Regression test added that fails on `main` and passes on the branch. - [ ] `verify-before-done` checklist passed. - [ ] If similar patterns exist elsewhere, they are fixed or logged.