The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a bundled Python script (scripts/codex_insights.py) to process local session logs. This is a standard and safe implementation for the skill's primary purpose of log analysis.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses local Codex history files located in ~/.codex and similar paths. However, it implements a robust sanitize() function in scripts/codex_insights.py that uses regex patterns to detect and redact API keys, tokens, bearer headers, and passwords before writing them to the report artifacts. This mitigates the risk of accidental data exposure.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from local Codex history files. It implements boundary markers by instructing the agent to 'Treat this as the Codex version of Claude Code’s /insights' and provides structured output templates. The ingestion is coupled with a read-only script that focuses on metadata (timestamps, tool counts) and sanitized message excerpts, reducing the surface area for injection attacks.
[SAFE]: All operations are performed locally on the user's machine. There are no external network calls, remote code downloads, or attempts to escalate privileges detected in the scripts or instructions.

codex-insights