The Agent Skills Directory

[DATA_EXFILTRATION]: Accesses sensitive user session history stored locally.
Evidence: The skill is configured to read raw interaction transcripts from ~/.codex/sessions and ~/.codex/archived_sessions as documented in SKILL.md and references/session-analysis-rubric.md.
Context: This data is used as the primary source for identifying repeated user corrections and steering signals.
[COMMAND_EXECUTION]: Executes a bundled Python script to automate data analysis.
Evidence: SKILL.md instructs the agent to run scripts/transcript_signal_scan.py to perform deterministic session analysis.
Behavior: The script parses JSONL files and applies regular expressions to extract specific interaction patterns.
[EXTERNAL_DOWNLOADS]: Fetches live documentation from a well-known technology provider.
Evidence: agents/openai.yaml includes an MCP tool dependency pointing to https://developers.openai.com/mcp for OpenAI developer documentation.
Context: This is used to verify product behavior against current official guidance.
[PROMPT_INJECTION]: Analyzes data that may contain untrusted instructions (Indirect Prompt Injection surface).
Ingestion points: Processes user-role messages from historical session files (~/.codex/sessions).
Boundary markers: The skill lacks formal boundary markers for processed snippets, although references/session-analysis-rubric.md instructions suggest using compact paraphrases.
Capability inventory: Uses local script execution (scripts/transcript_signal_scan.py) and project-wide file inspection.
Sanitization: Relies on regex filtering in the scanning script to identify 'noise' and instruction blocks.

codex-productivity-auditor