fix(combo/validator): accept reasoning_content as valid output (#2341)

`validateResponseQuality` flagged any response with `content: null` as
empty and triggered a false-positive 502 combo fallback, even when the
upstream returned the entire answer in `reasoning_content`. This
affected every reasoning model that omits `content` by design:
moonshotai/Kimi-K2.5-TEE, zai-org/GLM-5-TEE, zai-org/GLM-4.7-TEE,
DeepSeek-R1 and Qwen-thinking variants via Chutes.ai, Nvidia, and other
OpenAI-compatible gateways.

Treat a non-empty `reasoning_content` (or its legacy `reasoning` alias)
as valid content. Empty/whitespace-only strings still fall through to
the existing empty-content rejection so we don't weaken the guard.

Backward compat verified: regular content-only and tool_calls-only
responses still validate without change.

open-sse/services/combo.ts

@@ -211,7 +211,15 @@ export async function validateResponseQuality(
 
   const content = message.content;
   const toolCalls = message.tool_calls;
-  const hasContent = content !== null && content !== undefined && content !== "";
+  // Issue #2341: Reasoning models (Kimi-K2.5-TEE, GLM-5-TEE, etc.) emit their
+  // output in `reasoning_content` (or `reasoning`) with `content: null`. The
+  // validator used to flag those as empty and trigger a false-positive 502
+  // fallback. Count a non-empty reasoning_content as valid output too.
+  const reasoningContent = message.reasoning_content ?? message.reasoning;
+  const hasReasoningContent =
+    typeof reasoningContent === "string" && reasoningContent.trim().length > 0;
+  const hasContent =
+    (content !== null && content !== undefined && content !== "") || hasReasoningContent;
   const hasToolCalls = Array.isArray(toolCalls) && toolCalls.length > 0;
 
   if (!hasContent && !hasToolCalls) {

tests/unit/combo-quality-validator-reasoning.test.ts

@@ -0,0 +1,107 @@
+/**
+ * Issue #2341 — `validateResponseQuality` must treat a response carrying
+ * `reasoning_content` (Kimi-K2.5-TEE, GLM-5-TEE, etc.) as valid even when
+ * `content` is null. The previous implementation only inspected `content`
+ * and `tool_calls`, so reasoning models triggered a false-positive
+ * "empty content" 502 and an unnecessary combo fallback.
+ */
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const { validateResponseQuality } = await import("../../open-sse/services/combo.ts");
+
+function makeResponse(body: unknown, contentType = "application/json"): Response {
+  return new Response(JSON.stringify(body), {
+    status: 200,
+    headers: { "content-type": contentType },
+  });
+}
+
+const silentLog = { warn: () => {} };
+
+test("#2341 reasoning_content with null content is treated as valid", async () => {
+  const res = makeResponse({
+    choices: [
+      {
+        message: {
+          content: null,
+          reasoning_content: " The user simply said 'Say OK'. OK. ",
+        },
+      },
+    ],
+  });
+  const out = await validateResponseQuality(res, false, silentLog);
+  assert.equal(out.valid, true, `expected valid, got reason: ${out.reason}`);
+});
+
+test("#2341 legacy `reasoning` field is also recognized", async () => {
+  // Some upstream variants use `reasoning` (no `_content` suffix).
+  const res = makeResponse({
+    choices: [
+      {
+        message: {
+          content: null,
+          reasoning: "Step-by-step deduction body here.",
+        },
+      },
+    ],
+  });
+  const out = await validateResponseQuality(res, false, silentLog);
+  assert.equal(out.valid, true, `expected valid, got reason: ${out.reason}`);
+});
+
+test("#2341 empty reasoning_content + empty content + no tool_calls still rejected", async () => {
+  // Regression guard: the new branch must not weaken the empty-response check.
+  const res = makeResponse({
+    choices: [
+      {
+        message: {
+          content: null,
+          reasoning_content: "   ",
+        },
+      },
+    ],
+  });
+  const out = await validateResponseQuality(res, false, silentLog);
+  assert.equal(out.valid, false);
+  assert.match(out.reason ?? "", /empty content/i);
+});
+
+test("#2341 normal content-only response remains valid (backward compat)", async () => {
+  const res = makeResponse({
+    choices: [{ message: { content: "Hello world." } }],
+  });
+  const out = await validateResponseQuality(res, false, silentLog);
+  assert.equal(out.valid, true);
+});
+
+test("#2341 tool_calls-only response remains valid (backward compat)", async () => {
+  const res = makeResponse({
+    choices: [
+      {
+        message: {
+          content: null,
+          tool_calls: [{ id: "c1", type: "function", function: { name: "x", arguments: "{}" } }],
+        },
+      },
+    ],
+  });
+  const out = await validateResponseQuality(res, false, silentLog);
+  assert.equal(out.valid, true);
+});
+
+test("#2341 reasoning_content as non-string is ignored (defensive)", async () => {
+  const res = makeResponse({
+    choices: [
+      {
+        message: {
+          content: null,
+          reasoning_content: { unexpected: "object" },
+        },
+      },
+    ],
+  });
+  const out = await validateResponseQuality(res, false, silentLog);
+  // Non-string reasoning_content shouldn't count as content; still rejected.
+  assert.equal(out.valid, false);
+});