The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires the installation of external tools including daijro/camoufox for browser automation and rebel0789/codexpro for MCP transport. These are utilized as the core infrastructure for the multi-agent workflow.
[COMMAND_EXECUTION]: The workflow orchestrates local environment operations through a series of npm run and uv commands to handle preflight checks, browser management, and task manifests.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the ChatGPT Web surface. This risk is mitigated by a mandatory evidence chain:
Ingestion points: Data is fetched from the ChatGPT composer and conversation history via the Camoufox loopback bridge (monitored in scripts/protocol.mjs).
Boundary markers: The protocol enforces a unique sentinel string generated at task creation; the skill ignores any completion that does not include this sentinel in the assistant's final turn.
Capability inventory: The local agent has file write and shell execution capabilities via the MCP server to integrate candidate code.
Sanitization: All remote suggestions are subjected to mandatory local verification including linting, type-checking, unit tests, and human-in-the-loop review before being marked as CODEX_VERIFIED.
[DATA_EXFILTRATION]: The skill implements a security-conscious 'Bundle' fallback for non-MCP environments, which uses an automated scanner to block the inclusion of credentials, private keys, and environment files before source code is uploaded to the external model.

chatgpt-pro-coding-agent