Known limits
Every gateway has a list like this. Most of them do not publish it.
Codex CLI
Codex speaks OpenAI's Responses API, served here at /v1/responses — see the Codex quickstart. previous_response_id is refused, because continuing a stored response would mean keeping your conversation; Codex sends the full input each turn, so nothing is lost.
Server-side tools
Anthropic's hosted tools, web search among them, have no chat-completions equivalent, so they cannot be offered here regardless of model.
Text-only input
The catalogue accepts text requests only: image, audio or file structures in messages, system or tools are rejected outright rather than silently dropped. The vision column describes the model itself; the API surface remains text.
Default caps
Requests without max_tokens (max_output_tokens for Responses) are forwarded with a 2048 cap and held accordingly — which also rules out the unbounded-output bill. Explicit caps go up to 64000. Reasoning counts as output: too tight a cap on a reasoning model can return an empty message with finish_reason: "length".