Coding agents can drive, debug, and record live browser sessions

donobu session start lets an external coding agent (or a human) drive a real browser one step at a time and leave behind a deterministic @donobu/test spec, with replay-until-failure and amend to close the debug loop.

The donobu CLI now has a live-session mode. An external coding agent, or a human, drives a real browser one primitive at a time through donobu session start, and the byproduct is a deterministic, evidence-backed @donobu/test spec. No Donobu-side AI key is required in the loop.

Drive verbs cover observe, click, type, goto, act, caller-supplied asserts, dialog arming, route mocking, and console/network logs. Sessions record through the same path autonomous runs use, so they show up in donobu results and donobu files unchanged.

Two commands close the debug loop: session start --replay <result-id> --until-failure rebuilds a failed run and stops just before the recorded failure, and save --amend splices corrected steps into an existing spec by title while seeding the assert cache so emitted page.ai.assert lines replay without calling an LLM again.

  • Live session mode for coding agents and humans via the donobu CLI
  • Deterministic @donobu/test specs as the byproduct of driving a real browser
  • Replay-until-failure rebuilds a broken run and stops before the failure
  • Amend splices corrected steps into an existing spec by title