The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and summarize local Claude transcript files (.jsonl), which contain historical conversation data. This data is untrusted and can contain malicious instructions meant to override the agent's system prompt or current tasks.
Ingestion points: The scripts/claude-session.py script reads JSONL files from the ~/.claude/projects/ directory.
Boundary markers: Instructions in SKILL.md advise the agent to treat transcript content as untrusted, but the script output does not use explicit structural delimiters to isolate this content from the agent's instructions.
Capability inventory: The skill is restricted to reading local files and lacks any network, file-write, or privilege-escalation capabilities.
Sanitization: The script provides options to truncate output and limit the number of processed messages, reducing the risk of context-window manipulation.
[COMMAND_EXECUTION]: The scripts/test_claude_session.py test suite uses the subprocess module to execute the primary skill script for functional validation.
Evidence: The run_script method in scripts/test_claude_session.py calls subprocess.run with a list of arguments to execute the local Python script.

claude-session