Skip to main content
Codex is the strongest harness on OpenAI models. paddock runs the real Codex CLI, unmodified, at the version it pins (currently 0.162.1), through its codex app-server protocol, in each agent’s sandbox.

Models

Codex runs OpenAI models over the Responses API:
  • The platform model gpt-5.5.
  • Project models of an openai or openai-compatible provider, named provider/model. An openai-compatible endpoint must serve the Responses API. See Models.
Codex chooses the tools it offers, among them apply_patch for editing files, from what it knows about the model name. A model it does not know runs without apply_patch, so the agent edits files through the shell. A project model whose ID Codex knows, such as acme/gpt-5.5, runs with the same tools as gpt-5.5. Codex has one model setting; small_model is Claude Code only and is rejected on Codex.

Capabilities

Plugins are installed with codex plugin commands and MCP servers with codex mcp, then taken up by the running app-server without reloading the conversation. Changes apply between turns; see Capabilities.

Configuration

Tools

Codex edits files with patches, so its file changes are file_patch calls rather than file_write and file_edit. Its reasoning arrives as summaries in reasoning messages. Items paddock does not map, such as a plan or a context compaction, are unknown messages with their raw JSON.

Harness status

Codex reports Context, Plugins, Skills, MCP, and Subagents; see Harness status. Its Context section has what Codex reports: the model, the usable window (95% of the model’s window) beside the model’s own, the auto-compact threshold, the tokens in use, the last request’s input, cached, output, and reasoning tokens, the conversation’s total, and the instruction files it loaded. Usage by category is not reported.

Background work

Codex keeps background terminals and subagent threads running after a turn ends, and may continue a goal with a turn of its own: paddock records it as a harness turn, which consumes no message. Stopping a session ends the current turn and terminates that turn’s own commands; background work started earlier keeps running.