AI failure analysis¶
When the claude CLI is installed, the pane of a failed package shows an ✦ ask claude button. It starts a headless claude run in the workspace, with the tail of the failed log as the prompt. A new ✦ pane streams the investigation live: the files it reads, the commands it runs, and the answer.
- Nothing runs by itself. Each analysis starts with a click, and it spends your Claude usage.
- The run gets read tools approved and no edit tools. Headless claude refuses actions that need more permission.
- The transcript persists next to the logs of the package, so it survives a server restart and reopens instantly.
- The input box under the transcript asks follow-up questions in the same session, through
claude --resume.