Bug investigation

A bug report rarely arrives with a stack trace attached. This guide shows how to hand Mel a vague symptom and get back a reproduction, a root cause with file and line references, and a fix you approved — with every step recorded as a navigable block.

1. Describe the symptom

Press ctrl+enter in the prompt (anything you've already typed carries over as the first message) to start an agent conversation. Describe what's wrong in plain English. A prompt that works well:

One of the tests is failing. Run the test suite to reproduce the exact
failure, read the failing test and the code under test, and trace the root
cause — do not guess. Explain the cause with file/line references, then
propose the minimal fix and wait for my approval before editing.

The two instructions that matter: reproduce first and wait for approval. They keep the agent honest — it has to show you the failure before it touches anything.

2. Watch it reproduce

The agent works with real tools, and each one shows up as a card in the transcript:

• run_command executes your test suite or script and reads the full output.
• grep and find_files locate the code that produced the error.
• read_file pulls in the failing test and the code under test.

Commands the agent runs need your go-ahead the first time — Allow, Always allow for the rest of the conversation, or Deny. See Tools & approvals.

3. Dig through the output yourself

You're not a spectator. Any command's output — in the conversation or in the plain terminal — is a block you can work with:

• Select a block and open its filter bar to search the output live, with regex (.*) and case-sensitive (Aa) modes and highlighted matches. This works on streaming output too, so you can filter a long test run while it's still printing.
• File paths in output are links: ctrl+click one to open it. In the code review panel or editor, ctrl+click on an identifier jumps to its definition via the language server.

If the bug lives on a remote machine, ssh in and press ctrl+enter — the agent's read, grep, and run tools then execute on that host. See SSH & remote sessions and the incident response guide.

4. Approve the fix

Once the agent has the cause pinned, it proposes the minimal edit. Every applied change appears as a diff card — the path, +A −D counts, and red/green rows — so you see exactly what changed, in context, before moving on.

5. Verify

Ask the agent to re-run the failing test (or it will offer to — verify-after-change is built into how it works). The passing run lands as one more block in the transcript. When you're done, esc collapses the conversation into your scrollback; ctrl+shift+Y resumes it later, and the whole investigation — commands, output, diffs — is saved under your conversation history, ready to paste into the ticket.