The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the codex CLI to process discussion rounds. It includes strong validation rules for user-provided parameters like model names and reasoning effort to prevent command injection attacks. All user overrides are checked against a strict regex allow-list before being passed as discrete arguments to the shell command.
[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data (user plans, ideas, and files) which are interpolated into prompts passed to the Codex CLI. It mitigates this risk through several layers of defense.
Ingestion points: User-provided topic content, stated goals, and referenced file paths are read from the conversation context or local filesystem and inlined into prompts.
Boundary markers: The skill requires the use of fresh, random-suffix delimiters (e.g., <<<UNTRUSTED[k9x2]>>>) and explicit instructions to the AI to treat delimited content as data rather than directives.
Capability inventory: The skill executes reasoning tasks via the codex exec utility, which is confined by a read-only sandbox (-s read-only) to prevent unauthorized file modifications.
Sanitization: CLI argument overrides provided by the user are validated against alphanumeric patterns and predefined enums to ensure no shell metacharacters are passed.
[EXTERNAL_DOWNLOADS]: The skill supports an optional web search feature for fetching current data. This is implemented as a managed tool within the Codex CLI rather than unconstrained network access, and it requires explicit user opt-in or confirmation before activation.

codex-discuss