The Agent Skills Directory

[COMMAND_EXECUTION]: Executes several external binaries via subprocess to perform core tasks:
uv is used for virtual environment and dependency management during setup (ensure_deps.py).
ffmpeg and ffprobe are utilized to extract audio clips and verify media durations against transcript timelines (verify_queue_audio.py, fetch_minute_audio.py).
curl is used to download audio files from meeting platforms (fetch_minute_audio.py).
lark-cli is invoked to retrieve authenticated media URLs from the Feishu/Lark platform (fetch_minute_audio.py).
[EXTERNAL_DOWNLOADS]: Interacts with remote endpoints to fetch tools and data:
Communicates with the Zhipu AI (GLM) API at open.bigmodel.cn to perform automated text corrections (ai_processor.py).
Downloads audio recordings from the Feishu/Lark platform to facilitate human verification of ASR errors (fetch_minute_audio.py).
Setup scripts and documentation recommend installing the uv utility from its official source at astral.sh (ensure_deps.py).
[OBFUSCATION]: Uses specific Unicode Private Use Area characters (\ue000, \ue001) as internal sentinels. These are employed to 'mask' immutable transcript spans (like speaker labels) before they are sent to the AI model, ensuring structural markers are not altered during the correction process (protected_spans.py).
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcript data which is subsequently interpolated into model prompts.
Ingestion points: Loads content from .md and .txt transcript files (fix_transcription.py).
Boundary markers: Employs double-newline delimiters to separate transcript chunks within the correction prompt (ai_utils.py).
Capability inventory: Includes filesystem write access, network communication (via httpx), and external command execution (via subprocess).
Sanitization: Implements comprehensive path validation to prevent traversal attacks (path_validator.py) and strictly validates correction inputs and domain names used in SQL queries (domain_validator.py).
[SAFE]: Demonstrates several security best practices:
Centralized configuration management that automatically enforces restrictive filesystem permissions (0o600 for configuration files and 0o700 for the configuration directory).
Proactive sanitization of sensitive data, including a utility to mask API keys and other secrets in application logs (security.py).
Robust validation of all user-supplied file paths and domain names before they are used in operations.

transcript-fixer