Develop with an AI agent

This tutorial gives an agent a bounded task: add a Main.CalculateTotal procedure that writes the total of Input!B2:B10 to Result!B2, then prove the result.

Prompt

You are working in an xlflow project. Read xlflow.toml, inspect src/, and run
xlflow status --json before editing. Add Main.CalculateTotal and a TestCalculateTotal
test. Use xlflow fmt, lint, and analyze before touching Excel. Start one session,
push with --session --no-save, run the test and macro with --diagnostic, inspect
Result!B2, export an image of the result, then save the session and show git diff.
If a command fails, parse error.code and fix the source; never wait for a GUI dialog.

Expected loop

xlflow status --json
xlflow lint --json
xlflow analyze --json
xlflow test --json
xlflow session start --json
xlflow push --fast --session --no-save --json
xlflow test --session --json
xlflow run Main.CalculateTotal --diagnostic --session --json
xlflow inspect range --sheet Result --address B2 --session --json
xlflow export-image --sheet Result --range A1:C5 --session --json
xlflow save --session --json
xlflow session stop --json
git diff -- src

The first test should fail if the procedure is missing. The agent uses the structured failure, edits source, and reruns the same session. --no-save keeps the experiment reversible; save --session is the explicit persistence boundary. If execution returns workbook_recovery_required, stop the loop and use recovery, not --wait.