Add --cookies-from-browser flag for login-walled sources

[asayeed95]

Summary

Surfaces yt-dlp's --cookies-from-browser flag through /watch so login-walled sources work without dropping out to a manual yt-dlp invocation. Instagram Reels, private YouTube videos, and some X posts currently fail with Requested content is not available, rate-limit reached or login required — this PR makes the documented yt-dlp workaround available as a first-class flag.

Off by default. Public sources are byte-identical to current behavior.

Changes

• scripts/watch.py — new --cookies-from-browser BROWSER argparse flag, passed through to download().
• scripts/download.py — download_url() and download() accept a cookies_from_browser kwarg and inject --cookies-from-browser <value> into the yt-dlp argv. The kwarg is rejected if it starts with - to defend against accidental flag injection. The CLI entry point accepts an optional 3rd positional arg for parity.
• SKILL.md — documents the new flag, updates the "Download fails" failure-mode recipe to point at it as the documented retry path, and discloses in the Security & Permissions section that cookies are read transiently by yt-dlp (never written to disk, logged, or sent anywhere except the URL's host).
• CHANGELOG.md — [Unreleased] entry.

Security notes

• The value is appended to the yt-dlp argv as a single argument (not shell-interpolated), so subprocess injection isn't a concern.
• Values starting with - are rejected before invocation, so a caller can't sneak in something like --config-location /etc/passwd as the "browser name."
• The flag forwards BROWSER[+KEYRING][:PROFILE][::CONTAINER] verbatim — yt-dlp does its own validation of browser names and rejects unknown ones with a clear error.
• yt-dlp reads cookies from the browser's existing cookie store transiently. They're not copied into out_dir, not written to logs, not exfiltrated. The user remains signed in to the platform afterwards.

Test plan

• python3 scripts/watch.py --help shows the new flag with full help text
• download_url(..., cookies_from_browser='--config-location') raises SystemExit before invoking yt-dlp (flag-injection guard)
• End-to-end against https://www.instagram.com/reel/DX9k8WDPZ2D/: default path fails with "login required" (current behavior); retry with --cookies-from-browser chrome succeeds and produces an mp4 + audio that the rest of the watch pipeline consumes normally (60 frames extracted, Whisper/Groq transcription returned 28 segments)
• Public source without the flag: unchanged behavior

Why this is in scope for the skill

The existing "Download fails" recipe in SKILL.md currently tells the agent to give up on login-required content:

If it's a login-required or region-locked video, tell the user plainly; do not keep retrying.

That's the right behavior when there's no escape hatch — but yt-dlp itself documents --cookies-from-browser as the canonical workaround. Promoting it to a /watch flag means the agent has the same retry option a human running yt-dlp directly would have, with the security tradeoffs spelled out upfront.

🤖 Generated with Claude Code