# Recorded interaction driver

Environment for every run: Ubuntu 24.04, strict Snap glab 1.129.0, Chrome web client, 1280×800 CSS viewport, light theme. Each commit ran its own `T3CODE_DEV_SEED=0 vp run dev` from a separate worktree. Snap glab was first on server PATH; no temporary-directory override was set. The existing GitLab credential was inherited without recording it. `snap debug confinement` returned `strict`; `glab mr list` succeeded before the actions.

1. Create a private synthetic GitLab repository initialized with a README. Use seed.py to prepare identical clones at the same fixture commit.
2. Pair the browser with each isolated server. Add the clone using Add project → Local folder → enter path → Enter.
3. In Settings → Source control → Advanced enable the separate source control writer, keeping the default selection. Apply the same setting on both sides.
4. In the new thread draft for the fixture, start the Browser panel recording. Click `button:has-text('Create MR')`. Retain the recording through the settled error or Created MR result. Stop recording. Check the real GitLab API for MR existence and description.
5. In both clones, `git switch -c evidence-commit-push origin/main`. Set synthetic author identity. Write example.txt containing `Synthetic uncommitted content for Commit, push & MR.` plus a newline. Reload the same draft so version control refreshes.
6. Start recording. Click `button:has-text('Commit, push & MR')`. Retain the recording through the settled error or success. Stop recording. Check local/remote commit and GitLab MR.
7. Between base and head step 6, delete only the synthetic fixture's remote evidence-commit-push branch created by the base action. This restores the same absent-remote-branch input for head.

The native Browser panel tools drove the UI. No mock glab, mock GitLab API, unit test, or replaced server code was used. Videos have only the environment caption, normalization from device-pixel recording to the CSS viewport, and privacy masks over the environment name and unrelated composer model selector. All action results remain visible. render-video.py records that transformation.

AI-assisted; verified with the tests and tools listed above.
