ce-debug

Reproduce the failure first, then trace the code path backward from the symptom to where valid state first became invalid. Form hypotheses grounded in observed values, test one change at a time, and confirm the full causal chain before editing code.

Fetch the full issue thread with gh issue view including all comments, since later comments often contain narrowed reproduction steps or prior failed attempts. Extract symptoms and environment details, then reproduce the bug before tracing the code path.

Use intermittent-bug techniques: add logging traps at the suspected failure point, run the scenario in a loop to establish a reproduction rate, isolate environment variables, and check for test-order pollution from leaked state between tests.

A fix that works despite a wrong prediction means you patched a symptom, not the root cause. The real cause is still active. Re-examine the causal chain, explicitly invalidate the prior hypothesis with evidence, and form a new one with its own prediction.

Use git bisect when the bug is a regression where the code previously worked. Mark the current commit as bad and a known-good commit as good, then let bisect binary-search the breaking commit, optionally automated with a test script.

Apply it when the root-cause pattern appears in three or more other files or the bug would have been catastrophic in production. Add guards at distinct layers such as entry validation, invariant checks, environment guards, and diagnostic breadcrumbs.