The Agent Skills Directory

[Command Execution]: The skill uses local shell utilities including git, jq, sed, and sqlite3 to process session data and reconstruct run trajectories. This is a standard and necessary approach for analyzing logs and project history.
[Indirect Prompt Injection]: The skill ingests external session records (JSONL and SQLite) from various AI agent hosts which may contain untrusted content. While this presents a potential attack surface, the risk is mitigated as the skill focuses on generating a static markdown report for human review rather than executing actions based on the ingested content.
[Dynamic Execution]: The skill utilizes a Python script to query an OpenCode SQLite database. The script is provided statically within the instructions and correctly implements security best practices by using mode=ro and immutable=1 to ensure the database is accessed in a read-only, non-destructive manner.
[Sensitive Data Access]: The instructions involve reading local AI agent session directories (e.g., ~/.claude/, ~/.pi/, ~/.codex/). This access is explicitly required for the skill's primary purpose of performing a post-mortem audit of an AI agent's performance and methodology.

investigate-run