AgentFlow is still a private project.
I took AgentFlow back to a lightweight, file-based workflow. The separate runner experiment is retired. AgentFlow now works through an existing Codex or Claude Code session instead of adding another execution engine to maintain.
The runner had real benefits around isolation and recovery. But the workflow already has implementation, independent validation, and review, with repair and feature closeout instructions. I want to trial that properly before building more machinery around it.
The new addition is a thin Unattended launcher. It sends an explicitly approved, finite queue into the same workflow, with options to plan a run or resume one. Guided mode is still the default. Unattended doesn’t grant blanket permission to push or publish, and it isn’t a background service that keeps working after the client session ends.
I also fixed a smaller problem with closeout records. A committed progress file can still say a feature is waiting for merge after the PR has merged. The workflow now checks actual Git and PR state before treating that old checkpoint as unfinished work.
Both changes are merged. The verification suite passed 150 tests, and the hosted checks passed. The next step is a small trial: one Guided feature, then a two-feature Unattended queue. That practical trial is still ahead.
This supersedes the runner direction described in my October 2 workflow article.