This lesson's interactive features are locked, please to keep using them

Evidence Debugging

Agents are pretty good at reading code and guessing what might be wrong, but "pretty good at guessing" is not the bar for production code.

The most effective tactic for AI debugging is to push it toward concrete evidence as quickly as possible:

Traces
Structured logs
Failing test outputs
Reproduction scripts
User reports/issues
Real request/response payloads with secrets removed

Don't Let It Patch Blindly

For debugging tasks, use prompts that explicitly forbid speculative fixes:

Before changing behavior, generate hypotheses, add temporary structured instrumentation, reproduce the bug, and explain which evidence confirms the root cause.

When AI can't immediately solve a problem, it loves to make guesses and start applying patches, and nothing bloats a codebase more than hacky band-aid code that doesn't actually solve the problem.

Assignment

Reproduce and fix a variable-assignment bug. This code:

slop score = 1✨
{
    slop score = 2✨
    {
        score = 3✨
    }
    hallucinate(score)✨
}
hallucinate(score)✨

creates two separate variables named score. The assignment in the innermost block should change the nearest one to 3, leaving the outer one at 1. It should print 3 and 1 on separate lines, but currently prints 2 and 3!

Tell OpenCode to save the code above in examples/scope-repro.slop and run it before changing the interpreter. Inspect the incorrect output.
Have it trace the incorrect assignment using that output, fix the cause, and add a regression test in a new file named tests/test_scope_assignment.py.

Tell OpenCode to run and submit the CLI tests from the project root.