The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes session JSONL files and task artifacts which are external inputs, creating a surface for indirect prompt injection if the logs contain malicious instructions intended for the analyzing agent. However, the skill's operations are strictly limited to data extraction, metrics reporting, and comparison of existing logs. It includes a specific safety warning (Forensic Trap) to prevent self-poisoning from its own documentation during log analysis.
Ingestion points: JSONL session files via _pi_session_lib.py:iter_entries_with_lines and task metadata via analyze_run.py.
Boundary markers: Relies on standard JSONL structure and turn-based indexing; instructions explicitly advise excluding the skill-loading turn from searches to avoid false positives.
Capability inventory: Read-only access to localized session logs, internal module imports, and text/JSON console output. No network or file-write capabilities were detected.
Sanitization: The skill consistently uses json.loads for parsing external data.
[DYNAMIC_EXECUTION]: The initialization script _project.py dynamically locates the project root and modifies the Python path (sys.path) to import required modules from the local environment (ego_bench). This is a standard and safe architectural pattern for project-specific developer tools and does not involve runtime code generation from untrusted sources.

pi-session-analyzer