dev-bugfix
Use for a concrete defect or regression that needs reproduction, root-cause diagnosis, a fix, and regression validation. Do not trigger for routine localized edits or primarily performance-focused work.
BlueSkyXN · AI review on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add BlueSkyXN/Codex-is-all-you-need --skill dev-bugfixAbout this skill
name: dev-bugfix description: Use for a concrete defect or regression that needs reproduction, root-cause diagnosis, a fix, and regression validation. Do not trigger for routine localized edits or primarily performance-focused work. metadata: version: "0.6" updated: "2026-07-26"
Bugfix workflow
Use this for defects where guessing is likely to waste time. Move through the phases in order unless you can state why a phase is already satisfied.
Before editing, read nearby repo instructions and any relevant local/sdlc
architecture or domain notes if they exist.
Phase 0 - Define the Observed Failure
First pin the observed failure:
- Error message or wrong output.
- Reproduction steps from the user or current evidence.
- Affected file, feature, endpoint, CLI command, or behavior.
Then map likely code paths, nearby tests, and the smallest command that could prove the failure.
Phase 1 - Build a Feedback Loop
The first job is a tight pass/fail loop that can go red for this exact bug.
Without that signal, code reading turns into speculation. Spend more effort
building the loop than feels comfortable.
Use references/feedback-loop-menu.md to choose the smallest loop that reaches
the symptom: test, HTTP check, CLI fixture, browser assertion, replayed payload,
temporary harness, property/fuzz run, bisect, differential comparison, or HITL
script.
Tighten the Loop
Improve the loop before relying on it. Narrow setup, assert the exact symptom, cache slow prerequisites, pin time/randomness, and isolate filesystem or network state. For flaky failures, the target is a high reproduction rate, not a perfect single repro. Raise the rate until it is debuggable: repeat the trigger in bulk, run copies in parallel, add load or stress, shrink timing windows, and inject delays at suspected races. A failure on half the runs is workable; one in a hundred is not yet a loop.
Completion Criterion - Red-Capable Tight Loop Evidence
Name one command you have already run and show its decisive output. It must be:
- Red-capable - reaches the real bug path and checks the user's symptom
- Repeatable - same verdict each run, or a pinned high repro rate for flaky bugs
- Fast - seconds, not minutes
- Agent-runnable - runs unattended, except for encoded HITL prompts
If you are forming a root-cause theory before this command exists, stop and build the loop first.
When You Genuinely Cannot Build a Loop
Say so directly, list the attempts, and ask for one missing input: access to a reproducing environment, a captured artifact such as logs/HAR/video/core dump, or permission for temporary instrumentation. Do not present a speculative fix as diagnosis.
Phase 2 - Reproduce and Minimise
Run the loop and confirm:
- It produces the user-reported failure, not an adjacent failure
- It is reproducible across runs, or flaky at a high enough rate to debug
- The exact symptom is captured: message, wrong output, state, or timing
Then reduce the scenario. Remove inputs, callers, config, and data one at a time, re-running after each cut. Minimisation is done when every remaining piece is load-bearing: removing any one of them turns the loop green. The minimised scenario shrinks the Phase 3 hypothesis space and becomes the Phase 5 regression test.
Phase 3 - Hypothesise
List 3-5 ranked hypotheses before probing. Each must predict what evidence or change would confirm or falsify it. Share the ranking before testing when the user can cheaply re-rank it; continue with your ranking if they are unavailable.
Phase 4 - Instrument
Each probe must test one prediction. Change one variable per run. Prefer a debugger or REPL when available; otherwise add targeted logs at the boundary that separates hypotheses. Tag temporary logs with a unique prefix for cleanup.
For slowdowns, spikes, latency regressions, or memory growth, use
dev-performance-diagnosis for baseline-first measurement and bisection.
Phase 5 - Fix and Regression Test
Add the regression test before the fix whenever there is a valid seam: a test
point that exercises the real failure pattern through the relevant call path.
Only write it alongside the fix when the first safe step is creating or exposing
that seam; record the constraint as a testability finding.
With a valid seam: watch the regression fail, apply the fix, watch it pass, then rerun the original loop.
Phase 6 - Cleanup and Post-mortem
- Initial repro command now returns the expected result for the reported symptom
- Regression test passes, or the absence of a correct seam is documented
- All
[DEBUG-...]instrumentation is removed by grep - Throwaway harnesses are deleted or clearly kept as intentional tests
- The correct hypothesis is stated in the commit, PR, or final report
After the fix, name the prevention lesson. If the lesson is architectural, such
as hidden coupling or no valid seam, record a follow-up and hand off to
dev-refactor-plan.
Output
- Feedback loop: the one command and its red/green evidence
- Root cause and the hypothesis that held
- Fix summary and files changed
- Regression test, or documented absence of a correct seam
- Residual risk and any architectural recommendation
Do not
- Do not assert a root cause before a red-capable loop exists.
- Do not blanket-log a whole path and search the noise afterwards; every probe maps to one hypothesis boundary.
- Do not weaken tests to make them pass.
- Do not hide failing or inconclusive validation.
- Do not leave
[DEBUG-...]tags or throwaway harnesses behind.