What this test proves
Inside the colima testbed, a real agent executes prompt AT9 using only public ATM CLI observations. It proves that atm task start writes the started line to the assigner, an out-of-order start reorders the queue, and an unstarted task remains assigned and keeps reminding.
Evidence layout
This procedure is one step of a colima integration run. The agent's unedited prompt-report-1 JSON is retained beside step.json; scripts/integration/render_colima.py renders the panel and aggregate. Historical payloads remain byte-identical and render through the same path.
Changes
The revision sections below list every runner revision that produced committed evidence, newest first. A run whose source revision is not an ancestor of any listed revision links to the revision in effect on its run date and is marked inferred.
Flow
Steps
| step | action | observable | evidence |
|---|---|---|---|
| 1 | Tell the fixture agent to assign and start a task | `atm list/read --json` shows the started operation to the assigner | `prompt-AT9.json` and `step.json` cases |
| 2 | Tell the agent to assign two tasks and start the second | `atm task list --json` shows the active task ahead of the assigned task | `prompt-AT9.json` and `step.json` cases |
| 3 | Tell the agent to leave an assigned task unstarted | bounded `atm task events --json` polls show repeated reminders while it stays assigned | `prompt-AT9.json` and `step.json` cases |