Use after nil and other errors in openai module

[rallyemax]

The OpenAI API was complaining about excessive tokens being requested, despite the fact that I specified :max-tokens 500 in my block. So I checked the code and saw this:

org-ai/org-ai-openai.el

Lines 720 to 723 in cc4a4eb

	;; o1 does not support max-tokens
	(when max-tokens
	(setq max-tokens nil)
	(setq max-completion-tokens (or max-tokens 128000))))

Is this AI-generated code? It doesn't look like a mistake a human would make. If max-tokens is non-nil, it sets it to nil, and then attempts to use the value. Of course it always falls back to the hard-coded default of 128000, which causes the API to complain about excessive tokens.

Is this what you meant?

(when max-tokens
        (setq max-completion-tokens (or max-tokens 128000))
        (setq max-tokens nil))

After changing that, I run into another issue when using the o4-mini model, which uses max_completion_tokens. Passing even a nil value for max_tokens causes the following error:

  Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.

It looks like the culprit is in the model check: it checks for an "o1" or "o3" prefix, but not an "o4" prefix, even though that is in the list of supported models.

Finally, when using a model that uses max_completion_tokens, the following error is returned:

  Unknown parameter: 'max-completion-tokens'. Did you mean 'max_completion_tokens'?

Looks like a simple typo. Was this even tested after c4b5b68?

(Edited due to confusing myself with the various errors.)

[rallyemax]

Created a narrow PR to fix these issues at #152

As noted in the comments there, it looks like the max-tokens parameter is being deprecated (but appears to still be supported for the older models) for all models. I did a quick check by setting the "o1 models..." condition to true for all openai models, which causes the max_completion_tokens parameter to be passed instead of max_tokens, and it looks like both old and new models work.

The larger change to switch completely from max_tokens to max_completion_tokens requires more comprehensive changes to both code and documentation and I don't have the bandwidth to take that on at the moment.