The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The script scripts/bili-subs.sh uses yt-dlp --cookies-from-browser to access user credentials stored in local browser cookie databases. This access is intended to allow the skill to fetch login-gated Bilibili subtitles. The skill documentation explicitly describes this requirement and instructs the agent to ask for user permission before execution.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted external content from Bilibili, including video titles, tags, and danmaku (bullet comments), which are then processed by the agent.\n
Ingestion points: scripts/bili-fetch.sh (API metadata) and scripts/bili-danmaku.sh (danmaku text).\n
Boundary markers: Absent. The scripts return data directly to the agent's context.\n
Capability inventory: The skill can execute shell commands, perform network operations, and read browser cookies.\n
Sanitization: No explicit sanitization or filtering of the fetched text content is performed to prevent prompt injection.\n- [COMMAND_EXECUTION]: The skill executes several shell scripts (.sh) and external CLI utilities such as curl, jq, and yt-dlp to fetch and process data.\n- [REMOTE_CODE_EXECUTION]: In scripts/bili-danmaku.sh, data retrieved from api.bilibili.com is piped to a python3 command: curl ... | python3 -c \"import sys,zlib; ...\". This pattern involves passing remote data directly to a code interpreter for decompression.\n- [EXTERNAL_DOWNLOADS]: The skill fetches data from Bilibili's public and authenticated APIs (api.bilibili.com, b23.tv) and uses yt-dlp to interact with remote media and subtitle servers.

bilibili-source