Agentic Video Receipts on PRs

Record a .webm of an agent-driven browser test via Playwright's page.screencast API, convert to .mp4 for inline playback, upload as a GitHub draft release asset, embed the URL in a PR comment, and garbage-collect the draft on PR close. The result is a short video artifact that reviewers can watch instead of mentally simulating a diff.

Terminology: Playwright 1.59 coined "agentic video receipts". Cursor Cloud Agents (Feb 2026) shipped the same idea under "Cloud Agents with Computer Use".

Two phases

When the user invokes one of the trigger phrases, run the Per-repo setup below. When a smoke-test skill is actually recording, follow Per-smoke-test usage.

Per-repo setup

Three changes. All idempotent.

1. .mcp.json — enable the devtools capability

Edit the repo's .mcp.json (or ~/.mcp.json) and append --caps=devtools to the args of whichever Playwright MCP server the smoke test uses. Without this flag, browser_start_video / browser_stop_video are not registered as tools.

Example (headed under xvfb):

"playwright-headed": {
  "command": "xvfb-run",
  "args": ["-a", "--server-args=-screen 0 1920x1080x24", "npx", "@playwright/mcp@latest", "--caps=devtools"]
}

Example (plain):

"playwright": {
  "command": "npx",
  "args": ["@playwright/mcp@latest", "--caps=devtools"]
}

If multiple Playwright MCP servers are configured (e.g. playwright-headed and playwright), add the flag to each one the smoke test invokes.

2. .github/workflows/playwright-pr-video-cleanup.yml — cleanup workflow

Create this workflow verbatim. It runs on every PR close (merged or not) and deletes the PR's draft video release if one exists.

name: Playwright PR video cleanup

on:
  pull_request:
    types: [closed]

permissions:
  contents: write

jobs:
  cleanup:
    runs-on: ubuntu-latest
    steps:
      - name: Delete draft video release
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          PR: ${{ github.event.pull_request.number }}
          REPO: ${{ github.repository }}
        run: |
          tag="pr-${PR}-videos"
          if gh release view "$tag" --repo "$REPO" >/dev/null 2>&1; then
            gh release delete "$tag" --repo "$REPO" --yes
            echo "Deleted $tag"
          else
            echo "No release $tag found"
          fi

The workflow uses the default GITHUB_TOKEN; permissions: contents: write is required for gh release delete.

3. Verify ffmpeg

Needed to convert .webm → .mp4 (GitHub PR comments render .mp4 inline; .webm shows up as a download link only). Check with ffmpeg -version. Install via apt install ffmpeg or brew install ffmpeg if missing.

Per-smoke-test usage

Use this inside an agent-driven smoke test (for example, a preview-smoke-test style skill that runs against a deployed PR preview env). All MCP tool names below use the playwright-headed server as an example — substitute the server name actually configured in .mcp.json.

Start recording

Right after the first browser_navigate, call:

mcp__playwright-headed__browser_start_video
  filename: smoke-pr-<N>.webm

A bare filename resolves relative to the MCP's output dir (default .playwright-mcp/ in cwd). Pass an absolute path to write elsewhere.

Optional — add chapter markers so the reviewer can scrub:

mcp__playwright-headed__browser_run_code
  code: async page => { await page.screencast.showChapter("Selecting medication", { duration: 2000 }); }

Optional — annotate actions on-screen:

mcp__playwright-headed__browser_run_code
  code: async page => { await page.screencast.showActions({ position: "top-right" }); }

Drive the test as normal

No changes to the rest of the smoke test. Click, fill, navigate, assert.

Stop recording

After the last assertion:

mcp__playwright-headed__browser_stop_video

The tool response reports the output path.

Convert, upload, comment

Only run this block if the smoke test passed. On failure, skip the upload entirely — no receipt means no false signal on the PR.

PR=<N>
REPO=$(gh repo view --json nameWithOwner -q .nameWithOwner)
WEBM=.playwright-mcp/smoke-pr-${PR}.webm
MP4=smoke-pr-${PR}.mp4
TAG="pr-${PR}-videos"

ffmpeg -y -i "$WEBM" -c:v libx264 -preset fast -crf 28 -movflags +faststart "$MP4"

gh release view "$TAG" --repo "$REPO" >/dev/null 2>&1 \
  || gh release create "$TAG" --repo "$REPO" --draft --title "PR #${PR} video receipts" --notes ""

gh release upload "$TAG" --repo "$REPO" "$MP4" --clobber

gh api "repos/${REPO}/issues/${PR}/comments" \
  --method POST \
  -f body="Smoke test passed. Video receipt:

https://github.com/${REPO}/releases/download/${TAG}/${MP4}"

The draft release is invisible in the public Releases tab and gets deleted by the cleanup workflow when the PR closes.

Fallback — browser_run_code + page.screencast

If --caps=devtools cannot be enabled (older MCP, no write access to .mcp.json), call page.screencast directly via browser_run_code. Same end artifact, different tool surface:

# start
mcp__playwright-headed__browser_run_code
  code: async page => { await page.screencast.start({ path: "/tmp/smoke-pr-<N>.webm", size: { width: 1280, height: 720 } }); }

# stop
mcp__playwright-headed__browser_run_code
  code: async page => { await page.screencast.stop(); }

Convert / upload / comment is identical.

Requirements

Why draft releases, not inline comment upload

GitHub's REST API does not support programmatic file upload to PR/issue comments. The web-UI drag-and-drop feature produces user-images.githubusercontent.com URLs via a session-cookie endpoint that isn't exposed by gh or the REST API. Draft releases are the supported automation path — invisible in the public Releases list, accessible at stable URLs, rendered inline in PR comments (for .mp4 / .gif), and disposable via the cleanup workflow.