The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/review-with-codex.js executes the codex CLI to process the review packet. This is protected by a strict validation layer that pins the CLI version to exactly 0.146.0 and verifies that all security-sensitive feature toggles (such as shell_tool, browser_use, and multi_agent) are present and stable before execution. The subprocess is run in an ephemeral, read-only temp directory with a restricted environment whitelist.
[DATA_EXFILTRATION]: The skill transmits data to OpenAI Codex for review. This is handled safely through mandatory user consent (the --consent-to-openai flag) and instructions that require the agent to redact secrets and private context from the data packet before the transfer occurs.
[PROMPT_INJECTION]: The workflow incorporates defenses against indirect prompt injection by using explicit boundary markers (<BEGIN_UNTRUSTED_DRAFT>, etc.) in the review packet template. These markers are accompanied by specific instructions to the reviewing model to treat the enclosed content as data and ignore any embedded instructions.
[SAFE]: The skill follows security best practices for cross-model communication, including detailed labeling of provider relationships and failing closed if the expected security boundaries of the external adapter cannot be verified.

council-multi-model