videoscrub 0.2.0

Ask your agent about a video without choosing sampling settings yourself. This release adds a short video-inspection skill, native Codex plugin metadata, and more reliable visual evidence.

What's new

• Scan an entire video with just its path or URL. Inspect smaller regions with cropping and higher-resolution frames.
• Read source timestamps without covering the video image. Portrait clips and narrow crops now fit the image budget.
• Inspect high-frame-rate footage without duplicated frames. Small motion changes no longer round down to zero.
• Read embedded text subtitles and page through long transcripts without losing lines.
• Fetch remote metadata before downloading media. Download video and audio only when needed, reuse cached files, and cancel work cleanly.
• Get shorter errors when a dependency is missing or a command fails.
• Follow a simpler README with install and update commands. Detailed parameters live in the tool reference.

Update

Codex:

codex plugin marketplace upgrade videoscrub
codex plugin add videoscrub@videoscrub

Claude Code:

claude plugin marketplace update videoscrub
claude plugin update videoscrub@videoscrub

For a Git clone, run git pull --ff-only and bun install --frozen-lockfile. Restart your agent session after updating.

Changes for existing integrations

The four tool names are unchanged. frames now accepts source alone, and an omitted fps adapts to the range. Each call is limited to three image sheets. At widths of 320, 640, and 1280 pixels, the limits are 48, 12, and 3 frames. width bounds both image dimensions.

Transcripts are paged at 12,000 characters. Follow the returned continuation offset to read more. The new cache layout does not reuse v0.1.0 downloads or transcripts, so the first request may download or transcribe again. Old cache files are left untouched.

No new dependencies or API keys are required. Speech transcription remains optional and may incur provider charges. Music and other sounds are not supported.

Verification

33 local tests pass, including real MCP calls, frame rendering, cancellation, subtitle extraction, and manifest version checks. A clean packaged copy starts and returns timestamped frames. GPT-5.6-terra correctly located a status change and four indicator activations in two fresh synthetic clips. These are focused checks, not a general accuracy benchmark.

GitHub CI passes on Linux. A fresh clone of the published commit also starts and returns frames through MCP.