[Gi99lin]

Problem

The common input sanitization in chatCore.ts unconditionally normalizes max_output_tokens to max_tokens (#994). While this works for Chat Completions targets, it breaks OpenAI Responses API passthrough when sourceFormat === targetFormat === "openai-responses".

Reproduction

1. Configure a client (e.g. Codex CLI, Claude Code) to use openai-responses API format
2. Send a request with max_output_tokens to an OmniRoute instance that routes to a Responses API provider (e.g. Codex)
3. The upstream rejects the request with: 400 Unsupported parameter: max_tokens

Root Cause

The flow is:

Client sends: { max_output_tokens: 4096, input: [...] }
  → chatCore.ts:914: max_output_tokens → max_tokens (normalization)
  → translateRequest(): sourceFormat === targetFormat → SKIP
  → Upstream receives: { max_tokens: 4096 } → ❌ "Unsupported parameter"

When source and target formats are both openai-responses, translateRequest() skips translation entirely (line 122-123 of translator/index.ts). The translator from PR #1245 — which correctly maps max_tokens back to max_output_tokens — is never invoked.

Fix

Make the normalization conditional: skip it when targetFormat is openai-responses, where max_output_tokens is the canonical field.

- if (body.max_output_tokens !== undefined) {
+ if (body.max_output_tokens !== undefined && targetFormat !== FORMATS.OPENAI_RESPONSES) {

This preserves the existing behavior for all non-Responses paths (Chat Completions, Claude, Gemini etc.) while fixing the Responses→Responses passthrough.

Why this is safe

• Chat Completions targets: Normalization still runs → max_tokens is used ✅
• openai → openai-responses path: The translator from PR fix: translate max_tokens/max_completion_tokens → max_output_tokens in Chat→Responses translator #1245 handles max_tokens → max_output_tokens conversion ✅
• openai-responses → openai-responses path (FIXED): max_output_tokens is preserved → upstream receives the correct field ✅
• Codex executor: Already deletes both max_tokens and max_output_tokens in transformRequest (lines 512-513), so no behavioral change ✅

Test

Added a test that verifies max_output_tokens is not normalized to max_tokens when routing to a Responses API target via the /v1/responses endpoint.

Related

• [BUG] Unsupported parameter: max_output_tokens — OpenAI SDK clients rejected #994 — Original normalization PR
• fix: translate max_tokens/max_completion_tokens → max_output_tokens in Chat→Responses translator #1245 — max_tokens/max_completion_tokens → max_output_tokens translator fix (openai → openai-responses path)

[gemini-code-assist]

Code Review

This pull request modifies the input sanitization in chatCore.ts to skip the normalization of max_output_tokens to max_tokens when the target format is the OpenAI Responses API, preventing errors in passthrough requests. A unit test was added to verify this change. The review feedback suggests further improving the sanitization by also handling the reverse normalization—converting max_tokens or max_completion_tokens to max_output_tokens—for Responses API targets to ensure universal compatibility.

[gemini-code-assist]

While preserving max_output_tokens for Responses API targets is correct, we should also handle the reverse normalization for passthrough requests. If a client sends max_tokens or max_completion_tokens to a Responses API endpoint, it will fail upstream because the translator is skipped for same-format routing. Adding this normalization here ensures "universal compatibility" for passthrough as well, which is the stated goal of this sanitization block.

Suggested change

	// #994: Normalize max_output_tokens to max_tokens for universal compatibility.
	// Skip for Responses API targets where max_output_tokens is the canonical field;
	// normalizing it to max_tokens there causes "Unsupported parameter: max_tokens"
	// errors because Responses→Responses passthrough skips the translator that would
	// convert it back (see openaiToOpenAIResponsesRequest).
	if (body.max_output_tokens !== undefined && targetFormat !== FORMATS.OPENAI_RESPONSES) {
	// #994: Normalize between max_output_tokens and max_tokens for universal compatibility.
	// For Responses API targets, max_output_tokens is the canonical field. For others,
	// max_tokens is preferred. We handle normalization here to support passthrough
	// paths where the translator is skipped.
	if (targetFormat === FORMATS.OPENAI_RESPONSES) {
	if (body.max_output_tokens === undefined) {
	if (body.max_completion_tokens !== undefined) {
	body.max_output_tokens = body.max_completion_tokens;
	delete body.max_completion_tokens;
	} else if (body.max_tokens !== undefined) {
	body.max_output_tokens = body.max_tokens;
	delete body.max_tokens;
	}
	}
	} else if (body.max_output_tokens !== undefined) {