What it does
A normal agent guesses at a fix from source code alone. Debug Mode refuses to. It first reads the codebase and generates several competing hypotheses about the cause, then instruments the relevant code with logging statements designed to tell those hypotheses apart.
You then go and reproduce the bug in your running application while the agent collects the logs. With variable states, execution paths, and timing in hand, it identifies the root cause and proposes a targeted change, often two or three lines instead of the hundreds of speculative lines a blind agent produces.
The loop closes with human verification. The agent asks you to reproduce the bug once more with the fix in place. If it is gone, you mark it fixed and the agent strips out every logging statement it added, leaving a clean diff. If it is not gone, the agent adds more instrumentation and goes round again.
Using it
- Describe the bug. Select Debug Mode from the mode dropdown and describe the symptom in as much detail as you can. More detail means better hypotheses.
- Let it instrument. The agent adds logging aimed at distinguishing its hypotheses. Do not clean these up yourself; the agent tracks what it added.
- Reproduce. Trigger the bug in your app. The agent reads the runtime logs that come back.
- Verify and clean up. Reproduce once more with the proposed fix applied. Marking it fixed makes the agent remove all instrumentation.
Worth knowing
- Debug Mode needs a bug you can actually trigger on demand. Intermittent failures that you cannot reproduce give the agent nothing to read.
- Because instrumentation is written to your working tree, run Debug Mode on a clean branch so the temporary logging never lands in an unrelated commit.
Checked against Introducing Debug Mode on 25 September 2026.