Use custom/freeform tool format for OpenAI Responses API

[nyxkrage]

The OpenAI Responses API supports custom tools, which take raw text instead of a JSON arguments object. I propose making it so python_repl is one too, so OpenAI models write Python cells directly instead of JSON-escaping every quote, backslash and newline in a {"code": "..."} string. That saves output tokens and removes a common source of escaping mistakes.

Codex uses custom tools for its apply_patch tool and its code-mode exec tool.

Quick benchmark measuring output tokens per tool call, 3 runs per model, official OpenAI API:

Model	Reasoning	Function tool	Custom tool	Change
gpt-6-luna	medium	336	280	-17%
gpt-6-luna	xhigh	448	325	-27%
gpt-6.1-terra	medium	212	137	-36%
gpt-6.1-terra	xhigh	435	465	+7%
gpt-6.1-sol	medium	357	346	-3%
gpt-6.1-sol	xhigh	622	350	-44%
All		383	306	-20%

One thing to handle carefully: when a session switches between a model that gets the custom tool and one that gets the plain function tool, earlier calls have to be replayed in the shape the current request uses, including the item ids the API expects for each shape.

I have a working implementation on my fork (https://github.com/nyxkrage/prime-agent/tree/feat/freeform-tools). It's on by default for the openai and openai-codex providers; enabling it for another provider only changes how its Responses API requests are sent.
Happy to open a PR if wanted.