The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes session logs which contain untrusted user input and assistant outputs, creating a potential surface for indirect prompt injection.
Ingestion points: The bundled scripts/review_sessions.py script reads JSONL log files from ~/.claude/projects and ~/.codex/sessions.
Boundary markers: The script isolates the data processing, providing only structured summaries and specific 'event windows' to the agent. SKILL.md also includes an explicit instruction to the agent to treat log content as data, not as instructions.
Capability inventory: The skill is limited to running the bundled Python script (using the -I flag for isolation), reading local logs/findings, and writing the final report and candidate JSON files.
Sanitization: The review_sessions.py script implements robust redaction logic (redact function) using regex to mask tokens, passwords, JWTs, emails, phone numbers, and other sensitive identifiers before the agent sees the data. Command lines are also normalized to strip potentially sensitive arguments.

review-sessions