---
requires: error message or stacktrace (via clipboard -c)
produces: fixed code
diff_files: false
action_style: procedural
---
Debug an error using the stacktrace or error message from clipboard.
If clipboard is empty or no -c flag, ask what error to debug.
## What makes a good fix
**Unblock first.** Ask: what would it take to unblock the person who wants this debugged? Sometimes that's a quick workaround or explanation before a deeper fix. Get them moving, then address the root cause.
**Loop until the root issue is addressed.** Don't just take the next step and stop. Fix, verify, see what happens. If a new error surfaces, keep going. The job is done when the original workflow succeeds.
**Minimal and targeted.** Fix the bug, not the neighborhood. Don't refactor, don't "improve while you're here."
**Grease the wheels.** If debugging was hard, add tooling that makes it easier next time—for both humans and LLMs. A well-placed log statement, a clearer error message, a helper function that surfaces state. Small improvements that compound.
## Input
Run with `-c` to include clipboard content:
```bash
lf debug -c
```
Parse the error/stacktrace. Identify file and line. Check if the file was changed on this branch:
```bash
git diff main...HEAD -- <file>
```
## Debugging strategy
**Follow the stack trace.** The deepest frame in your code (not library code) is usually where the problem originates. Start there.
**Check recent changes.** If the error is new, the bug is likely in the delta.
**Reproduce first.** Before fixing, understand how to trigger the error. A fix you can't verify isn't a fix.
**Write the causal prediction.** State the violated invariant, the leading
explanations, and what each predicts before editing. Preserve the reproduction
as observed evidence; keep guesses labeled as guesses.
**Discriminate before patching.** Run the smallest safe check whose outcomes
separate the leading causes. If the result contradicts the current explanation,
stop dependent edits and revise the model. Repeatedly patching the same
assumption is a signal that the representation is wrong.
**Replay the whole workflow.** After the candidate fix, run the original
reproduction, the relevant regression tests, and the user-visible path that
previously failed. The latest green check does not erase older counterexamples.
## Output
Fix the bug directly. If the cause isn't obvious from the fix, add a brief inline comment.
If you can't determine the cause, describe what you learned and what additional context is needed.