The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/keep_codex_fast.py uses subprocess.check_output to execute system commands (ps on Unix and powershell on Windows). This behavior is used to monitor running processes (identifying the agent app and Node.js) to ensure maintenance is only performed when the application is not actively writing to its database, preventing corruption.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses local agent state files, including SQLite databases (state_5.sqlite, logs_2.sqlite) and configuration files (config.toml), to analyze storage usage. This access is limited to the local file system and is necessary for the skill's primary function of reporting on and managing session size and log growth. No network exfiltration patterns were found.
[DYNAMIC_EXECUTION]: The script generates a Python recovery script (restore-sessions.py) within the user's local backup directory. This script is intended for manual user intervention to roll back changes and is not executed automatically by the agent.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface where it reads potentially untrusted data (chat titles) from the local database.
Ingestion points: Reads thread titles and paths from state_5.sqlite in scripts/keep_codex_fast.py.
Boundary markers: The skill defaults to a pseudonymous report mode (using labels like session_001) unless the user explicitly requests raw details with the --details flag.
Capability inventory: File movement (shutil.move), file creation (backups), and configuration pruning.
Sanitization: The script performs internal parsing of SQLite data and uses parameterized queries when updating the database.

keep-codex-fast