v2.23.0 ·
The autopilot watcher stops asking me to say keep
On November 1 I answered eleven autopilot questions in four minutes. One was a real decision: approve a phase. The other ten were calls the agents had already made on their own, and I said keep to every one of them. I had to stop the run's chat to ask what the questions meant, because they read like "Keep what was built?" with nothing around it.
The rule behind that was mine. The session watching an autopilot run was told never to answer for me. So every time an agent took the option its own question file recommended, I was asked to bless it.
What changed
The watching session now keeps those calls itself. A call counts as routine in two cases: the agent answered a question with the option that question recommended, or it logged how it read an unclear requirement. Everything else still comes to me: phase approvals, open questions an agent was unsure about, an answer that went against the recommendation, and any answer the CLI cannot place.
When it keeps something, it tells me in one line which item it was and the command that overturns it:
aidlc autopilot review <instance>
The questions I do get name the roadmap item by its title and say what the agent chose, instead of only the heading of a log entry.
Why this way
The watcher does not decide what is routine. The CLI does, and marks each pending entry routine or not. The watcher can only run an option the CLI listed, and the watcher's own keep is a separate option that appears only on routine entries. A model comparing an answer to a recommendation in prose cannot be tested. A parser can.
A keep the watcher made is written as the watcher's, never in my name, so the decision log still tells me which calls a person actually read. Review asks those again, and my answer replaces the watcher's. A watcher keep on an open question never counts toward an automatic merge.
What it does not do
The parser fails toward me. Run over the 136 answers agents have defaulted in this repo, it classed 99 as routine and none that named a different option than the recommended one. The other 37 still reach me, mostly questions whose options are written inline or without labels.
The part I cannot test is the watcher itself. The instruction is there and a test checks it is there. Whether the session follows it shows on the next run I sit and watch.