The Agent Skills Directory

[PROMPT_INJECTION]: The skill is subject to Indirect Prompt Injection (Category 8) risks because it modifies authoritative agent memory files (such as CLAUDE.md) based on external file content.\n- Ingestion points: Local files are read in scripts/compress.py using Path.read_text() and sent to the LLM.\n- Boundary markers: The prompt in scripts/compress.py instructs the model to only return the compressed body and not interpret the text as instructions, although it lacks strict data/instruction separators.\n- Capability inventory: The skill utilizes Path.write_text() to overwrite the target memory file with the compressed output.\n- Sanitization: Logic in scripts/validate.py ensures technical elements like code blocks, URLs, and headings are preserved to prevent malicious or accidental context corruption.\n- [COMMAND_EXECUTION]: The script scripts/compress.py executes the claude CLI as an alternative authentication path.\n- Evidence: subprocess.run(["claude", "--print"], ...) is used as a fallback when the ANTHROPIC_API_KEY is not configured.\n- Assessment: The command is executed with a fixed argument list and user input is passed via stdin, which effectively mitigates shell command injection risks.\n- [DATA_EXFILTRATION]: The skill transmits the contents of local memory files to the Anthropic API.\n- Context: This behavior is the intended function of the skill for processing context via a trusted service provider. The script scripts/detect.py includes explicit safeguards to skip sensitive files like .env and .json and enforces a 500KB file size limit.

caveman-compress