This is a deterministic, local recording plan for AgentDock. It uses the repository fixture, LLM_PROVIDER=fake, a local Git remote, and the GitHub API stub. It is intended to show real application state, not fabricated outcomes.
Use a small task that produces a visible code and test change:
Add request validation to the
/usersendpoint and add tests.
The expected path is:
Connect repository → enter task → inspect → edit → validate → review diff → approve → create PR
- Landing: AgentDock identity, security promise, workflow preview, and trust-boundary diagram.
- Overview: connected repository count, active/completed runs, and recent workflow outcome.
- Active run: repository/task header, durable execution timeline, live operational activity, and separate command output.
- Validation passed:
Validatingcompleted with the result visible before approval. - Diff review: changed-file navigation, additions/deletions, bounded-preview warning where applicable, artifact status, and base SHA.
- Approval: repository, base commit, changed-file count, validation, and artifact integrity appear beside Approve & Create PR and Reject.
- Publication result: branch, commit, and pull-request link after the trusted publication worker completes.
| Time | Product state |
|---|---|
| 0–8s | Landing page or overview |
| 8–15s | Select the fixture repository and enter the task |
| 15–30s | Agent inspects, edits, and streams operational activity |
| 30–40s | Validation completes successfully |
| 40–52s | Navigate changed files and review the diff |
| 52–58s | Approve the immutable artifact |
| 58–70s | Show branch/commit and the created PR request |
Keep the browser at a developer-console width so the timeline and diff remain readable. Do not include terminal secrets or internal exception details in the recording.