The Agent Skills Directory

[COMMAND_EXECUTION]: The skill wraps the Codex CLI using scripts/run_bundle.py, which invokes the executable through subprocess.run with shell=False. This design prevents shell injection and ensures that only validated, structured arguments are passed to the external tool.
[INDIRECT_PROMPT_INJECTION]: While the skill ingests untrusted data from project plans and source files, it implements strong mitigation strategies. These include the use of strict Markdown templates for isolation, mandatory pre-flight validation of placeholders via scripts/validate_bundle.py, and clear instructions for the agent to verify all Codex findings against the repository state before synthesizing them for the user.
[DATA_EXFILTRATION]: The skill accesses local conversation logs to build its context bundles. Analysis of the Python scripts confirms that this data is processed locally and stored within the project's own .codex-bridge directory for use by the Codex tool, with no evidence of unauthorized network transmission to third-party domains.
[SAFE]: The project demonstrates a mature security posture by including explicit permission policies, detailed trust reports, and comprehensive test suites that verify sandbox integrity and path safety.

codex-bridge