Fix a bug

Bugs go through the skill diagnose. It builds a feedback loop that goes red on the bug before it forms any theory, so a fix is proven by a command, not by reading code.

Example prompt

Fix issue #57: the invoice total is wrong when a discount code is applied twice.

or without an issue: Diagnose: the export job returns an empty file when the date range spans a year boundary. The agent selects diagnose for words like bug, broken, failing or debug. Paste the first error line and the steps to reproduce, not the whole log; secrets and personal data are redacted before anything is quoted.

Expected steps

1. Feedback loop. diagnose finds one command that you can run unattended and that asserts the exact symptom: a failing test at the seam that reaches the bug (command from <lang>-test), else a curl script, a CLI run against a fixture, a replay of a captured payload, a throwaway harness, a property loop or git bisect run. It tightens the loop until it is fast and deterministic, and shows the command and its red output. If no loop is possible it stops and asks for a reproducing environment, a redacted artifact or permission for temporary instrumentation; it does not guess.
2. Minimal repro. It runs the loop, confirms the red result is your symptom and cuts inputs, callers, config and steps until each remaining element is load-bearing.
3. Ranked hypotheses. It writes 3 to 5 hypotheses, each with a prediction ("if X is the cause, changing Y makes the bug disappear"), and shows the ranking.
4. Instrument. One probe per prediction, one variable at a time, preferably a debugger over logs. Debug lines carry a unique prefix such as [DEBUG-a4f2].
5. Fix with regression test. The minimal repro becomes a failing test at a seam that exercises the real bug pattern. You see it fail for the right reason, the fix, the test passing, and the original loop green again. Tests follow implementation and <lang>-test.

Then continue as in Work an issue by hand: Review the current diff., then Create a PR for this branch.