Review & governance

How RobOS reviews code that an agent wrote: where the work is checked, what counts as evidence, and which parts of governance are not built yet. This section has four pages, listed at the end of this one.

The delivery loop

An agent does the implementation. A person decides whether the result is good enough to publish and merge. The recording on the RobOS home page shows the whole loop; each step below names the app that owns it.

Step App What happens
1. Plan Task Explorer (task-planner) An agent turns a goal into tracked issues with acceptance criteria.
2. Implement Task Implementer An agent creates a feature branch (codex/<task-key> by default), implements and tests the task, commits and pushes. It does not open a pull request.
3. Review locally Review window, local review The owner reads the diff on the Changes tab, checks the Evidence tab and takes the agent-led Walkthrough.
4. Publish Review window, Pull Request tab The owner reviews the description and creates the PR, as a draft if they choose.
5. CI Review window The CI banner shows check results. A draft can move to ready only when CI has passed.
6. Approve Review window, reviewer mode A reviewer reads the diff and submits Approve, Request Changes or a comment through gh pr review.
7. Merge GitHub The author merges. The review window has no merge button.

For a step-by-step version of this loop, see the Build apps guides: on your own or as a team.

The apps are described on their own pages: Task Planner, Task Implementer and Agent Code Review & PR Review Theater.

How review and proof of work fit together

The review window does not ask you to trust the agent’s summary. It shows three kinds of evidence next to the diff:

  • An evidence bundle recorded at one commit. When a new commit lands, the bundle is marked stale.
  • Walkthrough checkpoints where an agent runs the changed behavior and pauses so you can try it.
  • CI results read from GitHub and, for job logs, Buildkite.

Each is a signal for a person to weigh. The tool does not decide that a change is correct. Reaching the last walkthrough checkpoint does not approve or merge anything, and a PR can be created without completing the walkthrough.

Known limits

  • In reviewer mode, several tabs show built-in sample content rather than content from the PR, and by default the Changes tab and the Approve step stay locked until a fixed three-question sample quiz is passed. PR Review Theater lists which tabs use real data.
  • The walkthrough and evidence agents are Codex CLI sessions, configured as demoAgent in the review’s launch manifest. Review branch… sets it only when it finds a codex binary. Without one, those tabs have no agent to run.
  • Recorded demos in the repository use sandbox data. The Walkthroughs and proof of work page separates what the tests assert from what a recording only shows.
  • Automatic selection of a model by task difficulty is not built. See Agent tiers.

In this section

Page What it covers
PR Review Theater The review window: tabs, how to open a review, creating and readying a PR, the walkthrough, and the launch manifest.
Walkthroughs and proof of work The review lifecycle recording and its test, how the demo recorder works, and a catalog of the recorded demo scripts.
Agent tiers What the Agent Tiers settings do today and what is not built yet.
Workspace preflight How an agent gets a clean checkout before a task that needs one.

Table of contents