The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/codex-goal-report.py script executes system commands using subprocess.run. It specifically calls rg (ripgrep) to search for goal patterns in log files and ssh to facilitate connection to fleet machines. These calls are constructed using shlex or static lists to manage argument safety.
[REMOTE_CODE_EXECUTION]: The skill implements a fleet-wide reporting feature that runs its own source code on remote targets. The collect_remote function in scripts/codex-goal-report.py reads its own file content and pipes it into a Python interpreter on remote machines via SSH (ssh ... python3 -). While intended for distributed monitoring, this represents a remote code execution capability.
[INDIRECT_PROMPT_INJECTION]: The skill processes history logs (.jsonl) and databases (.sqlite) which contain historical goal objectives and statuses. Because this data originates from past user or agent interactions, it presents an attack surface where malicious instructions embedded in the logs could influence the agent's current analysis or reporting behavior.
Ingestion points: Reads files from ~/.codex/sessions/ and SQLite databases goals_1.sqlite and state_5.sqlite.
Boundary markers: None explicitly implemented; the script relies on semantic mining instructions in the skill prompt.
Capability inventory: Subprocess execution (rg, ssh), SQLite read-only access, and file writing for reporting.
Sanitization: The script uses shlex.join when building remote command strings to prevent simple command injection.
[DATA_EXFILTRATION]: The skill performs network operations using ssh to connect to machines defined in the user's fleet policy. The instructions advise the user to scrub sensitive data such as credentials and personal paths before sharing generated reports externally.
[DYNAMIC_EXECUTION]: The test file scripts/codex-goal-report-test.py uses the importlib.util module to dynamically load and execute the main reporting script for unit testing purposes.

codex-goal-mining