Diagnosing Bugs is for hard bugs and performance regressions. Phase 1 builds a feedback loop, a single command that goes red on this exact bug, and the skill states that no theorising may start before that command exists. The next phases reproduce and minimise, hypothesise, instrument, fix with a regression test, and clean up.
It triggers when the user says diagnose or debug this, or reports something broken, throwing, failing or slow.
What You Get
- A loop that is red-capable, deterministic, fast and runnable by the agent without supervision
- A minimised repro in which every remaining element is load-bearing
- Three to five ranked, falsifiable hypotheses shown to you before any is tested
- A regression test written before the fix at a correct seam, or a note that no correct seam exists
- A cleanup checklist that removes tagged debug logs and states the confirmed hypothesis in the commit message
Common Use Cases
- Chasing an intermittent flake by raising its reproduction rate
- Finding which commit introduced a regression by automating a bisection run
- Measuring a performance regression against a baseline before changing anything
What Makes It Unique
The skill treats the feedback loop as the whole job and says the rest is mechanical. It forbids hypothesising without a loop, asks you to review the ranked hypotheses at a cheap checkpoint, tags every debug log with a unique prefix so cleanup is one grep, and treats a missing test seam as a finding about the architecture.
Under The Hood
The skill is Markdown instructions plus a bash template, scripts/hitl-loop.template.sh. The template defines a step helper that waits for Enter and a capture helper that reads an answer into a variable, then prints the captured values as KEY=VALUE for the agent to parse. The skill tells the model to run commands, tests, curl requests, debuggers and git bisect, and to redact secrets before showing any output. It also reads GLOSSARY.md and ADRs when they exist.