Find And Fix Bugs

Entrypoint

• $AGENT_HOME/skills/automation/find-and-fix-bugs/scripts/render_issues_pr.sh
• Legacy wrapper paths are not supported; keep docs and callers pinned to this script.

Setup

• Record the starting branch/ref (for return after PR creation): start_ref="$(git symbolic-ref --short HEAD 2>/dev/null || git rev-parse --short HEAD)"

Trigger

Use this skill when the user asks to find or fix bugs, or when no concrete issue is provided and you are asked to proactively discover issues.

Intake rules

• If the user provides a bug report: ensure reproduction steps, expected vs actual, and environment. Ask only for missing details.
• If the user provides no input: do not ask; proceed autonomously.
• When using GitHub issues as intake targets, you must read the full issue context before selection: description/body, reproduction steps, expected vs actual behavior, labels, comments/discussion, and linked references.
• Skip immediately and move to the next issue when either condition is true: the issue has any comments/discussion, or the issue has any opened linked PR (including draft PRs).

Discovery

• Scope scanning to tracked files only (ignore untracked files).
• Use rg to scan for bug-prone patterns (TODO, FIXME, BUG, HACK, XXX, panic, unwrap, throw, catch, console.error, assert).
• Exclude generated, vendor, or codegen directories when present (node_modules, dist, build, vendor, .git, gen, generated, codegen).
• Keep scan rules general; do not add repo-specific patterns.
• Do not rely on grep results alone; use LLM analysis to confirm plausibility and impact.
• Produce an issues list using references/ISSUES_TEMPLATE.md.
• Use the ID format PR-<number>-BUG-## (example: PR-128-BUG-01). If the PR number is not known yet, use PR-<number> as a placeholder and update after PR creation.
• For project-specific skills, consider adding a minimal repro script requirement; see references/REPRO_GUIDE.md.

Context window management

• Prioritize scanning changed files and adjacent code before wider searches.
• If rg results are large, process in batches (for example: 20-50 hits per batch).
• Stop after enough high-confidence issues are identified to proceed with a fix.

Selection

• If user input exists, fix that issue.
• If autonomous, fix the single most severe or highest-confidence issue.
• Only fix multiple issues when they share the same root cause and the diff remains small.
• Severity levels are fixed: critical, high, medium, low.

Severity rubric

• critical: security exploit, auth bypass, data loss/corruption, or production outage
• high: frequent crash, major feature broken, or incorrect core outputs
• medium: incorrect behavior with workaround, edge cases, or performance regression
• low: minor bug, UX issue, or noisy logs without functional impact

Confidence rubric

• high: clear reproduction or strong evidence, root cause identified, fix is validated
• medium: likely issue with partial evidence, root cause inferred, fix is plausible
• low: speculative issue, weak evidence, or no repro path

High-risk guardrails

• Do not auto-fix changes involving auth, permission/authorization, payment/billing, migration, or infrastructure/deployment.
• If autonomous and the top issue is high-risk, record it and move to the next eligible issue. If all issues are high-risk, stop after reporting.

Fix workflow

1. Create a new branch: fix/<severity>-<slug> using the fixed severity levels.
2. Implement the fix with minimal scope; avoid refactors.
3. Add or update tests when possible; set up the repo’s test/build environment per its docs, then run lint/test/build commands (see Validation commands fallback). Treat validation as a gate: if validation fails, do not commit/open a PR; follow Retry policy.
4. Update the issues list with status.

Validation commands

• Prefer the repo’s documented commands (README/DEVELOPMENT/CONTRIBUTING/CI). Use the below as fallback heuristics.
• package.json scripts: lint, test, build (npm, pnpm, yarn, or bun)
• Makefile targets: lint, test, build
• Justfile targets: lint, test, build
• Taskfile targets: lint, test, build
• Language defaults when applicable: go test ./..., pytest, cargo test, mvn test, gradle test, dotnet test

Retry policy

• If validation fails, fix based on error output and retry up to 2 times.
• After 2 failed attempts, stop and report the failure with the last error output.

Commit

• Use semantic-commit-autostage for end-to-end automation (it stages changes); use semantic-commit only when the user has explicitly staged a reviewed subset.
• Prefer a single commit unless there is a clear reason to split.

PR

• Use gh pr create and write the body using references/PR_TEMPLATE.md.
• Set the PR title to the primary issue or a short summary of the fix. Do not reuse the commit subject. Capitalize the first word.
• Replace the first H1 line in references/PR_TEMPLATE.md with the same PR title.
• The PR must include:
  • Problem description (expected vs actual, impact)
  • Reproduction steps (or why repro is not feasible)
  • Issues found (including those not fixed)
  • Fix approach
  • Testing results or "not run"
• Include the issues list in the PR body.
• Use $AGENT_HOME/skills/automation/find-and-fix-bugs/scripts/render_issues_pr.sh --pr (or --issues) to generate templates quickly.
• After the PR is created, return to the original branch/ref: git switch "$start_ref" (or git switch - if you stayed on the fix branch the whole time).

Output and clarification rules

• Use references/ASSISTANT_RESPONSE_TEMPLATE.md as the response format.
• The response must include, in order:
  1. Issues list
  2. git-scope output
  3. Tests run (commands and results)
  4. PR link