ce-debug: Systematic Bug Diagnosis & Fix
Systematically trace root causes of bugs, diagnose before fixing, and produce regression tests.
Evidence: SKILL.md requires checking uncommitted changes and default branch creation before editing, and follows least privilege. It asks the user to choose whether to fix, discloses data flow (reading git logs, logs, telemetry). The context script runs git commands but only outputs read-only info, no exfiltration. Dependencies like gh are disclosed. Deduction: publisher identity unverified, only MIT license metadata; context script executes shell commands, potential tampering; no rollback mechanism provided.
Evidence: SKILL.md provides detailed step-by-step process, references files, includes error handling (degradation if context script fails). Key path is guiding user through diagnosis, not independent reproduction; not execution-verified here. Deduction: no test suite or CI evidence to confirm reliability; failure feedback quality only defined in pipeline mode, not interactive.
Evidence: SKILL.md specifies triggers (errors, stack traces, regressions, test failures), defines input format and expected output (diagnosis, fix), has pipeline and interactive modes. No external services, uses local tools and git, suitable for mainland China networks. Deduction: non-fit boundaries (e.g., design problems) not explicitly excluded, potential false triggers; no Chinese translation, but English is acceptable for technical users.
Evidence: SKILL.md has clear hierarchy, references files, provides examples and common pitfalls, includes known limitations (divergent fixes in pipeline mode). License MIT, no version or changelog. Deduction: missing version and changelog, maintenance responsibility not explicit (who maintains), only implied via README as EveryInc plugin.
Evidence: The skill aims to diagnose and fix bugs, expected output is root cause analysis and fix suggestion, provides marginal value over manual debugging via systematic approach. However, output depends on user confirmation, actual fix not always automatic. Deduction: no verifiable representative output, not execution-verified, marginal value not quantified.
Evidence: Provides CI badge (build status) but no test suite or reproducible tests. Context script includes executable verification steps but no results. Deduction: static review cannot independently verify behavior; CI badge insufficient; no concrete test results.
- Publisher identity unverified; review before installation.
- Context script executes shell commands, potential security risk; review script content.
- Skill may be triggered for design problems; ensure clear boundaries.
- No version or changelog; update path unclear.
What does this skill do, and when should you use it?
ce-debug is an Agent Skill for diagnosing and fixing software bugs with a methodical, causal-chain investigation. It starts by reproducing the problem, tracing the code path, verifying the environment, and checking the issue tracker and PR history to avoid duplicating others' work. After establishing the root cause, it offers an interactive choice to fix now, diagnose only, or rethink the design. If fixing, it uses test-first discipline, follows a one-change-at-a-time principle, and includes workspace safety checks before committing to a PR. It also supports a pipeline mode for fully autonomous operation, ideal for CI babysitting. This skill is part of the Compound Engineering plugin collection and works across multiple AI coding tools.
This skill takes a bug description (e.g., error message, stack trace, test path, or issue reference) and executes a multi-phase workflow: it fetches issues from trackers in Phase 0, reproduces the bug and traces code paths in Phase 1, forms and tests hypotheses in Phase 2, optionally writes tests to drive the fix in Phase 3, and produces a structured summary in Phase 4. It reads local files, runs shell commands (e.g., git, tests), invokes browser tools if needed, and may use GitHub CLI for issue information. It also writes residual findings and compound learnings to designated artifact directories.
- A developer faces a stack trace or failing test and needs to find the root cause without guessing.
- An engineer works on a regression and wants to trace back to the original commit that introduced the bug.
- A developer is stuck on a bug and wants a disciplined approach to avoid shotgun debugging.
- A CI bot (like a PR babysitter) needs a non-interactive bug fix that can commit and push autonomously.
- A team wants to ensure every bug fix includes a regression test and captures lessons learned.
What are this skill's strengths and limitations?
- Enforces root-cause analysis, preventing symptom patching.
- Includes test-first discipline with regression test generation.
- Supports both interactive and autonomous pipeline modes.
- Integrates with other Compound Engineering skills for review and learning capture.
- Works across multiple AI tools (Claude Code, Cursor, Codex, etc.).
- Cognitive overhead; may be overkill for simple typos.
- Requires Node.js for the context script; functionality reduced without it.
- Depends on GitHub CLI for issue fetching; other trackers may need manual paste.
- No full automated test suite shown in docs; only internal scripts described.
- Phase 4's auto-commit/PR behavior requires user consent; may not fit all workflows.
How do you install this skill?
This skill is part of the Compound Engineering plugin. For Claude Code: run /plugin marketplace add EveryInc/compound-engineering-plugin and /plugin install compound-engineering. For Cursor, add from the plugin marketplace search. For Codex, follow the README steps. Existing users must refresh the marketplace before updating the plugin.
How do you use this skill?
After installation, invoke /ce-debug with a bug description in a supported tool, e.g., /ce-debug the checkout webhook sometimes creates duplicate invoices. The skill will guide you through investigation phases. In interactive mode, you can choose to fix now or diagnose only. In pipeline mode, it fixes and pushes automatically.