From f2d94957c8291967be091bce56cb73a07f2528b6 Mon Sep 17 00:00:00 2001
From: Chewji <126886556+Chewji9875@users.noreply.github.com>
Date: Thu, 13 Aug 2026 10:48:35 +0700
Subject: [PATCH] fix(ollama-cloud): map xhigh reasoning effort to max (#10160)

---
 .../providers/registry/ollama-cloud/index.ts  | 15 ++++++-
 open-sse/executors/base/reasoningEffort.ts    | 15 ++++++-
 .../base-executor-sanitize-effort.test.ts     | 41 +++++++++++++++++++
 3 files changed, 69 insertions(+), 2 deletions(-)

diff --git a/open-sse/config/providers/registry/ollama-cloud/index.ts b/open-sse/config/providers/registry/ollama-cloud/index.ts
index 37f560fa0d..05b28df65f 100644
--- a/open-sse/config/providers/registry/ollama-cloud/index.ts
+++ b/open-sse/config/providers/registry/ollama-cloud/index.ts
@@ -27,7 +27,20 @@ export const ollama_cloudProvider: RegistryEntry = {
     { id: "deepseek-v4-pro", name: "DeepSeek V4 Pro", supportsReasoning: true },
     { id: "deepseek-v4-flash", name: "DeepSeek V4 Flash", supportsReasoning: true },
     { id: "kimi-k2.6", name: "Kimi K2.6" },
-    { id: "glm-5.1", name: "GLM 5.1" },
+    // Ollama Cloud accepts low|medium|high|max|none and rejects xhigh, so the
+    // explicit supportsXHighEffort:false makes the sanitizer map xhigh → max.
+    {
+      id: "glm-5.1",
+      name: "GLM 5.1",
+      supportsReasoning: true,
+      supportsXHighEffort: false,
+    },
+    {
+      id: "glm-5.2",
+      name: "GLM 5.2",
+      supportsReasoning: true,
+      supportsXHighEffort: false,
+    },
     // #3110: MiniMax M3 via Ollama
     { id: "minimax-m3", name: "MiniMax M3", contextLength: 1048576, supportsVision: true },
     { id: "minimax-m2.7", name: "MiniMax M2.7" },
diff --git a/open-sse/executors/base/reasoningEffort.ts b/open-sse/executors/base/reasoningEffort.ts
index 5edf1188ae..8356e6e131 100644
--- a/open-sse/executors/base/reasoningEffort.ts
+++ b/open-sse/executors/base/reasoningEffort.ts
@@ -154,7 +154,8 @@ export function supportsMaxEffortForProvider(provider: string, model: string): b
   // upstream. Scoped to opencode-go deliberately: OpenRouter's DeepSeek path
   // (pi#4055) is the documented inverse and expects xhigh, not max.
   // Ollama Cloud also accepts literal max (for example GLM 5.2 supports
-  // low|medium|high|max|none) and rejects xhigh.
+  // low|medium|high|max|none) and rejects xhigh; xhigh is mapped to max by the
+  // provider guard in sanitizeReasoningEffortForProvider.
   const isOpencodeGoDeepSeek =
     (provider === "opencode-go" || provider === "opencode-zen") &&
     resolvedModelId.toLowerCase().includes("deepseek");
@@ -284,6 +285,18 @@ export function sanitizeReasoningEffortForProvider(
     return writeEffortValue(b, "max", c);
   }
 
+  // Ollama Cloud accepts low|medium|high|max|none and rejects xhigh. Map
+  // xhigh → max (its literal top tier) before the generic xhigh handling so
+  // passthrough (unregistered) models are covered too — the registry opt-out
+  // only covers known models.
+  if (provider === "ollama-cloud" && effortStr === "xhigh") {
+    log?.info?.(
+      "REASONING_SANITIZE",
+      `${provider}/${modelStr}: mapped reasoning_effort xhigh → max`
+    );
+    return writeEffortValue(b, "max", c);
+  }
+
   // Native DeepSeek (api.deepseek.com) — V4 thinking mode uses the native
   // {low, high, max} vocabulary on Flash and {high, max} on Pro. OmniRoute's
   // internal top tier xhigh maps to DeepSeek's literal max. Pro's unsupported
diff --git a/tests/unit/base-executor-sanitize-effort.test.ts b/tests/unit/base-executor-sanitize-effort.test.ts
index 2a712106c5..ba7ce28955 100644
--- a/tests/unit/base-executor-sanitize-effort.test.ts
+++ b/tests/unit/base-executor-sanitize-effort.test.ts
@@ -118,6 +118,47 @@ test("sanitizeReasoningEffortForProvider: Ollama Cloud preserves nested max", ()
   assert.equal((result as Record<string, unknown>).reasoning.summary, "auto");
 });
 
+test("sanitizeReasoningEffortForProvider: Ollama Cloud maps registry model xhigh → max", () => {
+  const log = makeLog();
+  const body = {
+    model: "glm-5.2",
+    reasoning_effort: "xhigh",
+    messages: [{ role: "user", content: "hi" }],
+  };
+  const result = sanitizeReasoningEffortForProvider(body, "ollama-cloud", "glm-5.2", log) as Record<
+    string,
+    unknown
+  >;
+  assert.notEqual(result, body, "must return a new object when mutating");
+  assert.equal(result.reasoning_effort, "max");
+  assert.equal(result.model, "glm-5.2", "other fields preserved");
+  assert.ok(
+    log.messages.some(([tag, m]) => tag === "REASONING_SANITIZE" && /xhigh → max/.test(m)),
+    "logs the xhigh → max mapping"
+  );
+});
+
+test("sanitizeReasoningEffortForProvider: Ollama Cloud maps passthrough unknown model xhigh → max", () => {
+  const log = makeLog();
+  const body = {
+    model: "some-future-glm-model",
+    reasoning_effort: "xhigh",
+    messages: [{ role: "user", content: "hi" }],
+  };
+  const result = sanitizeReasoningEffortForProvider(
+    body,
+    "ollama-cloud",
+    "some-future-glm-model",
+    log
+  ) as Record<string, unknown>;
+  assert.notEqual(result, body, "must return a new object when mutating");
+  assert.equal(result.reasoning_effort, "max");
+  assert.ok(
+    log.messages.some(([tag, m]) => tag === "REASONING_SANITIZE" && /xhigh → max/.test(m)),
+    "logs the xhigh → max mapping"
+  );
+});
+
 test("sanitizeReasoningEffortForProvider: OpenRouter DeepSeek passes max through (new default)", () => {
   const log = makeLog();
   const body = {