claude-video

This happens because Windows consoles often use cp1252 instead of UTF-8, while the report prints characters such as →, …, or —.

Temporary user workaround:

setx PYTHONUTF8 1

Close and reopen the terminal after running it. For only the current PowerShell session, use:

$env:PYTHONUTF8 = "1"

Permanent application fix: reconfigure stdout and stderr to UTF-8 at program startup, before any report output:

import sys

for stream in (sys.stdout, sys.stderr):
    try:
        stream.reconfigure(encoding="utf-8")
    except (AttributeError, ValueError):
        pass

Call this helper from every entry point, such as watch.py, frames.py, whisper.py, transcribe.py, setup.py, and download.py. TextIOWrapper.reconfigure() is available in Python 3.7+; the affected project requires Python 3.10+. Do not merely remove the non-ASCII characters, because other report output can trigger the same cp1252 error.

The slash command fails because Claude Code 2.1.126 does not parse the YAML inline-list format for allowed-tools. Edit ~/.claude/plugins/cache/claude-video/watch/0.1.2/commands/watch.md and replace:

allowed-tools: [Bash, Read, AskUserQuestion]

with:

allowed-tools: Bash, Read, AskUserQuestion

You can apply this directly with:

sed -i '' 's/^allowed-tools: \[Bash, Read, AskUserQuestion\]$/allowed-tools: Bash, Read, AskUserQuestion/' ~/.claude/plugins/cache/claude-video/watch/0.1.2/commands/watch.md

Fully quit and reopen Claude Code, then run /watch or select /watch:watch from autocomplete. The permanent fix is for the plugin maintainer to make this one-line change in commands/watch.md on main; adding frontmatter validation can prevent regression.

Upgrade to watch v0.1.3 or later. The bug occurs because scripts/download.py calls Path.read_text() without encoding="utf-8"; on Windows Python uses cp1252, so non-ASCII characters in yt-dlp's video.info.json cause UnicodeDecodeError. The broad except then silently leaves only the URL, causing title and uploader to be omitted. To patch older versions, change:

raw = json.loads(info_path.read_text())

to:

raw = json.loads(info_path.read_text(encoding="utf-8"))

Also log parse failures instead of silently ignoring them, for example:

except Exception as exc:
    print(f"[watch] info.json parse failed: {exc}", file=sys.stderr)
    info = {"url": url}

As a temporary workaround, re-read video.info.json from the output directory explicitly as UTF-8. If Windows console output fails on emoji or other non-ASCII title characters, set PYTHONUTF8=1 or PYTHONIOENCODING=utf-8 before running the command.

Insert "--" immediately before the URL in the yt-dlp argv array so its parser stops processing options. Also validate the URL and reject inputs starting with -.

cmd = [
    "yt-dlp",
    "-N", "8",
    "-o", output_template,
    "--",
    url,
]
subprocess.run(cmd, check=True)

Harden URL validation as defense in depth:

from urllib.parse import urlparse

def is_url(source: str) -> bool:
    if source.startswith("-"):
        return False
    parsed = urlparse(source)
    return parsed.scheme in ("http", "https") and bool(parsed.netloc)

For ffmpeg/ffprobe, do not rely on -- because those tools do not treat it as end-of-options; resolve user-controlled paths to absolute paths with Path.resolve(). This issue was fixed in claude-video v0.1.3/commit c333c22.

There is currently no one-click option that writes a corrected VTT/SRT file, but a two-pass workflow works:

1. Run /watch <url> --detail transcript and ask Claude for the raw timestamped transcript plus suspicious segments, such as proper nouns, homophones, and abrupt context changes. For already-captioned URLs, the tool uses the VTT pulled by yt-dlp and skips downloading the video.
2. Download the video and recheck only the questionable timestamps:

/watch <local-video> --detail transcript --timestamps 1:23,4:56,7:08 --no-whisper

Claude extracts frames only at those timestamps, compares them with the transcript, and can suggest context-aware corrections. Apply those corrections manually to the VTT/SRT file.