Use evidence to locate the failing boundary and fix the cause while preserving business intent.
Delegation
Run in the main conversation by default. Delegation can increase usage: obtain explicit approval for the proposed agent count and scope before using subagents. Reuse that approval within its bounds; ask again before expanding the approved count or scope.
Investigate
- Capture the failing command or flow, input, expected behavior, actual result, and relevant environment. Read the error and nearby source before expanding the search.
- Reproduce the smallest useful case when practical. If reproduction is unavailable, distinguish what logs or source prove from what remains a hypothesis.
- Trace the symptom through the affected boundaries. Choose checks that distinguish plausible causes; independent observations can run together when safe. Avoid stacking speculative fixes.
- For a UI failure, inspect the supplied image or rendered state. Annotate a copy only if that makes a material ambiguity easier to resolve.
Read debugging-playbook.md for deeper isolation, flaky failures, or temporary instrumentation. Keep sensitive data out of logs and remove temporary debug code.
Fix and verify
Make the smallest coherent fix, preserving unusual business rules unless evidence shows they are the defect. Avoid unrelated refactors. Add a regression test when it provides durable protection, ideally demonstrating failure before the fix.
Rerun the original failure and affected checks. Complete repository-required checks; broaden only when a remaining risk warrants it and authorization permits. Reuse results that remain valid.
If measurement or test design becomes the main work, consult the relevant specialist guidance directly when available.
Report
Explain the cause, fix, and evidence that the failure is resolved. Include reproduction limits or remaining uncertainty. Record a difficult bug in existing durable docs only when future work would benefit.