[duncanleung]

Summary

• normalize top-tier reasoning effort from max to xhigh across Codex-facing translation paths
• fix OpenAI Chat -> Responses and Claude -> OpenAI bridges so Claude Code fallback traffic no longer sends an unsupported effort value upstream
• add regression coverage for Codex executor and translator paths

Why

Claude Code can emit or imply top-tier effort as max. Codex now expects xhigh, so max leaks cause 400 errors like:
Invalid value: 'max'. Supported values are: 'none', 'minimal', 'low', 'medium', 'high', and 'xhigh'.

Testing

• npx --yes tsx --test tests/unit/translator-openai-responses-req.test.ts
• live local verification against OmniRoute with POST /v1/chat/completions using reasoning_effort: "max" after patching the running install

[gemini-code-assist]

Code Review

This pull request standardizes the "max" reasoning effort tier by mapping it to "xhigh" across the Codex executor, Claude-to-OpenAI translator, and OpenAI Responses translator. It also updates the thinking budget logic and includes comprehensive unit tests for these changes. The reviewer suggested adding string trimming during normalization in the translators to ensure consistency with the Codex implementation and improve robustness against whitespace in client input.

[gemini-code-assist]

For robustness and consistency with the implementation in open-sse/executors/codex.ts, it is recommended to include .trim() when normalizing the effort string. This ensures that any leading or trailing whitespace provided by the client doesn't bypass the normalization logic.

Suggested change

	function normalizeOpenAIReasoningEffort(effort: unknown): string | undefined {
	if (typeof effort !== "string") return undefined;
	const normalized = effort.toLowerCase();
	if (normalized === "max") return "xhigh";
	return normalized || undefined;
	}
	function normalizeOpenAIReasoningEffort(effort: unknown): string | undefined {
	if (typeof effort !== "string") return undefined;
	const normalized = effort.trim().toLowerCase();
	if (normalized === "max") return "xhigh";
	return normalized || undefined;
	}

[gemini-code-assist]

To ensure consistency across the codebase and improve robustness against client input variations, consider adding .trim() to the normalization logic here as well. This aligns with the behavior in open-sse/executors/codex.ts.

Suggested change

	function normalizeResponsesReasoningEffort(value: unknown): string {
	const effort = toString(value).toLowerCase();
	return effort === "max" ? "xhigh" : effort;
	}
	function normalizeResponsesReasoningEffort(value: unknown): string {
	const effort = toString(value).trim().toLowerCase();
	return effort === "max" ? "xhigh" : effort;
	}

[diegosouzapw]

Thanks for this great contribution! 🎉 It's now officially merged into the release/v3.7.4 branch and will be part of the next release. We appreciate your effort!