The Agent Skills Directory

[DATA_EXFILTRATION]: The skill scans local session logs from Claude Code, OpenCode, and Codex platforms (e.g., ~/.claude/projects, ~/.local/share/opencode). These logs contain full conversation history and potential sensitive user data. Extracted clusters and 'evidence' snippets are sent to remote LLM APIs (Claude and Codex CLIs) for clustering and scoring.
[EXTERNAL_DOWNLOADS]: Fetches skill content from raw.githubusercontent.com using curl. It also performs searches on external registries via npx (e.g., @floomhq/skills and skills package).
[COMMAND_EXECUTION]: Uses subprocess.run to execute multiple external CLI tools including npx, curl, claude, and codex. It also uses os.popen within its test suite to run local scripts.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcript data from user sessions and interpolates it into prompts for LLM clustering and scoring.
Ingestion points: Session logs found in ~/.claude/projects, ~/.local/share/opencode, and ~/.codex/sessions (implemented in sources.py).
Boundary markers: Uses basic delimiters and quotes to wrap evidence, but lacks rigorous separation between data and instructions.
Capability inventory: System command execution via subprocess.run and file writing capabilities in draft.py and score.py.
Sanitization: Implements basic truncation for snippets (MAX_SNIPPET = 220 chars) and inputs (MAX_INPUT = 2000 chars).
[DYNAMIC_EXECUTION]: The draft.py script dynamically generates new agent instructions (SKILL.md) based on output provided by a remote LLM.
[PRIVILEGE_ESCALATION]: Documentation in mine.py suggests using sudo to access session logs in restricted system directories (e.g., /root/.claude/projects).

skill-miner