pr-media

pr-media

The script lives next to this file: scripts/pr-media.sh in this skill's directory. Run it from a checkout of the PR's repo.

1. Estimate the capture cost. Quick means a page reachable in a running dev server, or frames you already captured. Slow means auth or fixture setup, a real third-party API run, a multi-step flow, or several viewports. If it's slow, ask the user (AskUserQuestion) before capturing: shots now, shots later, or none. If it's quick, just do it.

2. Delegate the capture to a Haiku subagent. Screenshots are a mechanical browser-driving job, so run them on the cheap tier: Agent with subagent_type: "general-purpose" and model: "haiku".

   • The model must be Haiku 5.5 (claude-haiku-5-5) or newer. Older Haiku models drive the browser unreliably. If you know the harness maps haiku to an older model, omit model so the subagent inherits yours instead.
   • This is deliberate, and it holds even where the user's instructions otherwise avoid Haiku or small models for subagents: this capture job is the sanctioned exception.

   Pass it: the diff summary (git diff <base>...HEAD --stat, where <base> is the PR's base branch, plus the user-facing files), the dev server URL, an output folder (the gitignored scratch folder the project names in AGENTS.md/CLAUDE.md, else mktemp -d), and these instructions verbatim:

   • Decide which screens the diff touches and capture one PNG per screen, before/after where a baseline exists. Name files for what they show (settings-after.png), because the name becomes the caption.
   • Drive the browser with whatever browser tools you have: the t3 preview tools (preview_navigate, preview_click, device_screenshot) when preview_status reports available, else the chrome-devtools MCP (navigate_page, take_screenshot with a file path). Never read screenshot pixels back into context; verify the page with the text snapshot instead.
   • Video only when motion is the change (preview_recording_start/stop or equivalent), clips ≤15 s.
   • Crop out hostnames, tokens, API keys, and user emails.
   • Return only the list of file paths and a one-line description each.

3. Spot-check one file by reading it. It should show the change, not a login wall, spinner, or blank page. If it doesn't, recapture that one yourself.

4. Open the PR first, then run scripts/pr-media.sh <pr> <files...>. Done when it prints attached … to PR #<N>. If it fails, report the error; don't paste local file paths into the PR.

Facts the script relies on:

• Default backend is gh pr edit --attach (gh ≥ 2.99.0): files go to GitHub's user-attachments CDN, the same as drag-and-drop. In a private repo they stay private. Videos become an inline player. Each file must be ≤10 MB (videos up to 100 MB on paid plans: PR_MEDIA_ATTACH_MAX_MB=100).
• With an older gh, or a file over the limit, it falls back to the orphan branch pr-assets/pr-<N>, linked by commit SHA. Videos become a GIF that links to an MP4 (GitHub strips <video> that points at repo files). Each file must be ≤25 MB; shrink a GIF with PR_MEDIA_GIF_WIDTH=540 PR_MEDIA_GIF_FPS=8. Force a backend with PR_MEDIA_BACKEND=attach|branch.
• Treat uploads as permanent either way.
• The PR body gets one block between <!-- pr-media --> markers. Re-runs replace items with the same file name and keep the others.
• --no-edit prints the block instead of editing the PR (branch backend).