feat: exclude GPT-5 models from 20% context window output token cap

[roomote]

This PR addresses a Slack request to exclude models containing "gpt-5" from the logic that caps the max output tokens to 20% of the context window.

Changes

• Modified getModelMaxOutputTokens() in src/shared/api.ts to detect GPT-5 models (case-insensitive check for "gpt-5" in modelId)
• GPT-5 models now bypass the 20% cap and use their exact configured max output tokens (e.g., 128k)
• Non-GPT-5 models continue to be capped at 20% of context window as before

Testing

• Added comprehensive test coverage for GPT-5 exclusion logic
• Tests verify that various GPT-5 model IDs bypass the cap
• Tests confirm non-GPT-5 models still have the cap applied
• All existing tests continue to pass

Impact

This change allows GPT-5 models to utilize their full 128k max output token capacity as configured, improving their ability to generate longer responses without artificial limitations.

---

Important

getModelMaxOutputTokens now allows GPT-5 models to bypass the 20% context window cap, using their full configured max tokens, with tests verifying this behavior.

• Behavior:
  • getModelMaxOutputTokens in api.ts now excludes GPT-5 models from the 20% context window cap, allowing them to use their full configured max tokens.
  • Non-GPT-5 models continue to be capped at 20% of the context window.
• Testing:
  • Added tests in api.spec.ts to verify GPT-5 models bypass the cap and use full max tokens.
  • Confirmed non-GPT-5 models still adhere to the 20% cap.
• Misc:
  • Comprehensive test coverage ensures existing functionality remains intact.

^{This description was created by for 133a953. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Reviewing my own code because apparently I trust no one, not even myself.

[roomote]

The pattern matching here could be more precise. Currently, modelId.toLowerCase().includes("gpt-5") would match unintended model IDs like "not-gpt-5-compatible" or "legacy-gpt-500".

Could we consider a more specific pattern? Perhaps:

Suggested change

	const isGpt5Model = modelId.toLowerCase().includes("gpt-5")
	// Check if this is a GPT-5 model (case-insensitive, more precise matching)
	const isGpt5Model = /^(openai\/)?gpt-5($|-|\.)/.test(modelId.toLowerCase())

This would match "gpt-5", "gpt-5-turbo", "openai/gpt-5-preview" but not "not-gpt-5" or "gpt-500".

[roomote]

This GPT-5 exception is a significant behavior change that should be documented in the function's JSDoc. Future maintainers might wonder why GPT-5 models get special treatment. Could we add a note to the function documentation explaining this exception?

[roomote]

Great test coverage! Though it might be worth adding edge case tests to ensure the pattern matching doesn't have false positives. For example, testing that "not-gpt-5", "gpt-500", or "legacy-gpt-5-incompatible" don't incorrectly bypass the cap.

[daniel-lxs]

LGTM