Debug with an AI agent. Keep the proof.

Start with the smallest failing case

Write down the command or user actions, the expected result, and the actual result. Include the relevant error and environment, with secrets removed. A report such as 'checkout is broken' invites guessing; a cart total that stops updating after a quantity change gives the agent a specific path to investigate.

The cart starts with one item at 20. Increasing its quantity
to 3 still displays 20; expected total: 60. Find the calculation
and its callers. Reproduce the failure, add a regression test,
make the smallest fix, and report the test command and result.
Do not change pricing rules or remove existing tests.

Trace data across the boundary

Have the agent identify where the input enters the system, where state is updated, and where the displayed value is derived. Use references when a language server is available and text search for events, configuration, and templates. Ask for the proposed cause before a broad rewrite.

An exception disappearing is not enough. A swallowed error, skipped assertion, or disabled validation can hide the symptom while leaving the underlying problem. Review the explanation against the actual code path.

A regression test should fail before the patch

Keep the input, expected output, and assertion visible. If the agent cannot establish the pre-fix failure, label the result as a proposed fix awaiting reproduction. Do not count a skipped, deleted, or rewritten-to-pass test as evidence.

1. Run the existing relevant checks to establish a baseline.
2. Add a test that reproduces the reported failure against the original behavior.
3. Confirm the test fails for the intended reason.
4. Apply a focused patch and rerun the regression test.
5. Run adjacent tests and the relevant full suite, type check, or build.

Check the user flow and the final diff

In LoopCode desktop, browser tools can exercise a reachable local web app. Check the corrected cart action and an adjacent case such as removal or an empty cart. For terminal or remote editor work, use the project's available automated checks and state any browser step that remains unverified.

Review every changed file, dependency, and configuration setting. Ask for commands, exit codes, remaining failures, and assumptions in the final report. Keep unrelated failures visible so a green-looking summary does not hide an incomplete verification.