fix: Improve model compatibility in parsing and logging

[abinashkarki]

• Handle None model name in sub-call logging (verbose.py)
• Accept both repl and python code block tags (parsing.py)
• Handle special tokens around FINAL statements (parsing.py)

These changes improve compatibility with various LLM backends that use different output formats (e.g., Gemini, some OpenRouter models).

[alexzhang13]

So I don't enable python tags because 1) it affects the prompting and 2) I wanted it to be distinct from when the model wants to output python code for whatever reason, e.g. in a task that requires outputting Python code.

The None call was handled in a separate PR, thanks though!

I can merge this if you revert the above 2, "Handle special tokens around FINAL statements (parsing.py)" is fine. I will think on it though if it's necessary. Thanks!

[abinashkarki]

Thanks for the feedback! I've reverted the python code block tag support and the None model handling (didn't realize it was already fixed in #5).

The updated PR now only includes the special token handling for FINAL statements, which helps with models that wrap output in tokens like <|begin_of_box|>FINAL(...)<|end_of_box|>.

[alexzhang13]

@abinashkarki Returning to this, the only problem now is if the model is thinking / doing some kind of CoT and is saying "I will then do FINAL(...)...". We don't want to accept these cases -- I think we should just be stricter about what exactly we accept, e.g. maybe just some trivial special tokens in case this is something people are observing.

[abinashkarki]

Thanks for the feedback @alexzhang13 Updated the regex to be stricter.

Why this matters: zai-org/glm-4.6v-flash wraps FINAL in special tokens like <|begin_of_box|>FINAL(4)<|end_of_box|>. Without handling this, benchmarks fail.

What changed: FINAL is now only accepted at:

• Start of line (with optional whitespace)
• After special tokens <|...|>

This rejects CoT false positives like "I will do FINAL(42)" while still accepting valid cases.