v2.13.0 · October 16, 2026

Autopilot asks you what it is waiting on

Autopilot's agents run unattended, so they can ask nobody anything. The cost is that every decision only a person can make ends up in a file. A paused item, an open question in the decision log, an answer the agent took "by default", an item it parked, an inbox nobody triaged. All of it waits, and nobody is told.

On 2026-10-15, in SkyForce, a whole queue sat like that until the person asked "is this moving?". Asked, they answered all of it in about a minute. The minute was never the problem. The hours before it were.

What changed

The session you start a run from now stays on watch. It reads one list, asks you each item on it as a multiple-choice question with the recommended option first, and applies your answer.

The list is aidlc autopilot pending. Each entry says whether it blocks the run or only waits, and each option says in one line what it will do.

Every answer is one command, aidlc autopilot answer <fingerprint> <option>. It writes a phase question's answer into its question file, approves a paused item, keeps or rejects a decision the agent logged, moves an inbox item to the backlog with its rank and run mode, declines a parked item, or retries a halted merge. It records the answer beside autopilot's state, so no session asks it again, not even a second one watching.

Between rounds the session runs aidlc autopilot wait --pending in the background. It exits when something new waits, after 30 minutes with nothing new, or when the run ends. That exit is what wakes the session. With nothing new, it says nothing.

Triage got a command of its own on the way, aidlc roadmap triage, so that answer is one command too.

Why this way

The session never builds a command from your words. It names an entry and an option, and answer looks the option up in a list it computes itself. A session can only do what the list offered.

Approving and reviewing were the hard call. Both refuse to run without a person at a terminal, and the watching session has none. The agent that built this kept that guard, and logged it as an open question. I answered it the other way: answer approves an item and writes a review verdict itself, so you never leave the session. The trade is plain. The command line cannot tell your click from an agent deciding alone, so any agent that can run answer can approve its own item. Each approval and verdict names who gave it and that it came through the watcher, so it can be traced. The approve and review commands themselves still refuse without a terminal.

What it does not do

It does not resume a parked item: autopilot has no way to take one back, so "resume" tells you how to continue it yourself. It does not send a finished phase back for rework either. A note you type is recorded and read back to you, nothing more.

And it only works where the agent can be woken by a background command. Elsewhere the skill falls back to checking every 20 to 30 minutes, and failing that, tells you once to run aidlc autopilot status yourself.