A day with RobOS

One task, taken from a plan to a merged pull request and a follow-up task, using the RobOS apps. This page is a written version of the recording on the home page, including which parts of that recording are real and which are scripted.

The setup

The recording uses a small fixture project, acme-corp/pawfinder-api, a pet-adoption search API. The task is “Let adopters filter pet search by species, and show the best matches first”. The whole run is about 143 seconds long and has 11 chapters.

It is produced by a script, not by hand: review-lifecycle-hero-demo.js drives the real Electron apps, with helpers in packages/robos-test/lib/lifecycle/. The chapter list is in site/_data/hero.json.

Part Real Scripted stand-in
Repository and tests The git repository, its local “remote”, the branch, commits, pushes and test runs Nothing
GitHub The apps’ issue, PR and CI screens and flows The gh command (lib/lifecycle/bin/gh) is a stateful fake. CI check results are set by the script.
Task planning The Task Explorer screens and the issue-creation flow The AI that drafts the tasks (lib/lifecycle/plan.js)
Coding agents The commits, pushes and test runs they perform bin/claude (Task Implementer) and bin/codex (review chat and walkthrough) are scripts. The edits they make are prewritten in the fixture (the feature, the removal of six comments, the species chips, the extra evidence case). No AI model writes code in the recording.
Evidence Before and after results from calling searchPets on main and on the branch Nothing
Merge The merged state shown in the app The merge itself: gh pr merge --squash against the fake forge

Nothing is deployed in the recording. The last slide shows the sample app started locally with node src/server.js in a browser window.

Walkthrough

1. Plan the task in Task Explorer

Chapter: Task created. App: RobOS Task Explorer, package task-planner.

You type the goal into the prompt box and click Generate with AI. The planner returns an epic and a task. The task body holds acceptance criteria plus a happy-path, a negative and an edge case. Click Sync All to Server to file them. In the recording the epic becomes issue #41 and the task #42.

2. Implement it in Task Implementer

Chapter: Task implemented. App: Task Implementer.

Pick issue #42 from the list. The agent persona comes from the issue’s role: label. Click Start Agent. The agent creates the branch codex/42, makes the change, runs the project’s tests, commits and pushes. It does not open a pull request. The Review branch… button next to Start Agent hands the branch to the review app.

3. Review as the author in PR Review

Chapters: Owner reviews the changes, Owner checks the evidence, Owner takes the walkthrough. App: Agent Code Review Platform, using the stages described on PR Review Theater.

  • Changes tab: a file tree and the full diff of the pushed branch, in unified or split view. Suggest changes opens a chat with an agent. In the recording the owner asks it to remove comments that only restate the code. The agent edits the file, runs the tests and commits, and the diff refreshes.
  • Evidence tab: before and after results for each scenario, tied to one commit. The agent’s new commit makes the existing evidence stale, so the owner asks for fresh evidence plus a new case that combines a text search with a species filter. Scenarios appear as the agent captures them.
  • Walkthrough tab: click Start. An agent explains one checkpoint at a time and you click Next step to move on. Checkpoint 2 opens the running sample app in its own window. While it is open, the owner asks for species chips instead of a dropdown. The page reloads and the change is committed.

4. Open a draft pull request and wait for CI

Chapters: Draft PR created, Owner watches CI, Moved to ready for review.

In the Pull Request tab, tick Create as draft and click Create PR. The tab label shows the PR state, here “In Draft”. Refresh CI reads the checks. The script moves them from pending to passing for build, unit-tests, lint and contract-check.

Move PR to Ready is refused unless the PR is a draft whose head matches the reviewed commit and its CI checks have passed. The dialog pre-fills reviewers from the project’s review settings. Afterwards the tab shows “In Review”.

5. Approve, merge, and start the loop again

Chapters: Reviewer approves, Owner merges, Owner watches CI and acts.

  • A second window opens the same PR in Reviewer mode. The reviewer reads the diff and approves with notes. The reviewer-mode knowledge check holds placeholder content, so the recording answers it off camera.
  • The script merges with gh pr merge --squash. After Reload PR, the tab shows “Merged”.
  • The script then sets the post-merge CI result on main: contract-check fails. The owner goes back to Task Explorer and files a follow-up, “Document the species query parameter in the OpenAPI spec”, which becomes issue #44. That is chapter 1 again.

Record it yourself

The recording needs xvfb, ffmpeg, Electron installed under packages/robos-test/node_modules, and a Chromium binary for the sample app window. Point FAKE_CHROME at that binary; the default path in appwin.js is /opt/pw-browsers/chromium-1194/chrome-linux/chrome.

xvfb-run -a -s "-screen 0 1920x1080x24" \
  node packages/robos-test/demos/review-lifecycle-hero-demo.js /tmp/hero

lib/lifecycle/README.md describes the next two steps, compose.js (cut dead time, add captions, encode) and publish-hero.js (copy into the front site).