Human in the loop: where people still sign off

Human in the loop means people approve key steps of an automated process; DevFlow keeps 2 such gates on every task, plan approval and mark-done, even with autopilot on.

The principle

Automation is fastest when nobody waits, and safest when somebody checks. This design picks the few points where a person's judgment is worth the delay and automates the rest. For coding agents, the hard part is choosing those points well.

The two gates in DevFlow

The first is plan review. Before any code exists, you read the proposal and approve, edit or reject it.

The second is mark-done. Squashing, pushing and opening the pull request start only when a person marks the work done.

Between them, the pipeline runs unattended. Decomposition, implementation, verification and the summary need no clicks unless the workspace turns off automatic implementation start.

Questions during planning

The planner can also pause to ask. Questions arrive in batches; you answer inline and planning resumes. Work waiting for answers is shown as Awaiting answer and survives a server restart.

Through chat

You do not need the web app for every decision. With the Slack or Mattermost bridge, a thread reply such as approve, lgtm or ship it approves the proposal, and go starts new work.

Recorded actors

Every status change is written with a typed actor: a user email, an agent role or a system poller. The merge evidence report lists who approved the plan and who signed off at the end.

The same trail appears in the History tab, newest first, so you can see who changed a status and when.

With autopilot

Autopilot can open tasks from analyzed issues and start planning by itself, but nothing past planning is automated. Both sign-offs stay with people.

FAQ

Can I approve a DevFlow plan from Slack?

Yes. With the DevFlow chat bridge connected, replying approve, lgtm or ship it in the task's Slack thread approves the plan.

Is any DevFlow step after planning automatic?

In DevFlow, decomposition, implementation, verification and the summary run on their own after plan approval. Publishing the pull request waits for mark-done.