fix(ollama): reject garbled Kimi symbol output

CHANGELOG.md

@@ -55,6 +55,7 @@ Docs: https://docs.openclaw.ai
 - CLI/channels: list configured chat channel accounts from read-only setup metadata even when the standalone CLI has not loaded the runtime channel registry, so `openclaw channels list` shows Telegram accounts before auth providers. Fixes #73319 and #73322. Thanks @mlaihk.
 - CLI/model probes: keep `infer model run --gateway` raw by skipping prior session transcript, bootstrap context, context-engine assembly, tools, and bundled MCP servers, so local backends can be tested without full agent-context overhead. Fixes #73308. Thanks @ScientificProgrammer.
 - CLI/image describe: pass `--prompt` and `--timeout-ms` through `infer image describe` and `describe-many`, so custom vision instructions and slow local model budgets reach media-understanding providers such as Ollama, OpenAI, Google, and OpenRouter. Addresses #63700. Thanks @cedricjanssens.
+- Providers/Ollama: reject long non-linguistic Kimi/GLM symbol runs as provider failures instead of storing them as successful visible assistant replies, so fallback or error handling can recover from garbled cloud output. Fixes #64262; refs #67019. Thanks @Kloz813 and @xiaomenger123.
 - CLI/model probes: reject empty or whitespace-only `infer model run --prompt` values before calling local providers or the Gateway, so smoke checks do not spend provider calls on invalid turns. Fixes #73185. Thanks @iot2edge.
 - Gateway/media: route text-only `chat.send` image offloads through media-understanding fields so `agents.defaults.imageModel` can describe WebChat attachments instead of leaving only an opaque `media://inbound` marker. Fixes #72968. Thanks @vorajeeah.
 - Gateway/Windows: route no-listener restart handoffs through the Windows supervisor without leaving restart tokens in flight, so failed task scheduling can be retried and successful handoffs do not coalesce later restart requests. (#69056) Thanks @Thatgfsj.

docs/providers/ollama.md

@@ -1062,6 +1062,18 @@ For the full setup and behavior details, see [Ollama Web Search](/tools/ollama-s
 
   </Accordion>
 
+  <Accordion title="Kimi or GLM returns garbled symbols">
+    Hosted Kimi/GLM responses that are long, non-linguistic symbol runs are treated as failed provider output instead of a successful assistant answer. That lets normal retry, fallback, or error handling take over without persisting the corrupted text into the session.
+
+    If it happens repeatedly, capture the raw model name, the current session file, and whether the run used `Cloud + Local` or `Cloud only`, then try a fresh session and a fallback model:
+
+    ```bash
+    openclaw infer model run --model ollama/kimi-k2.5:cloud --prompt "Reply with exactly: ok" --json
+    openclaw models set ollama/gemma4
+    ```
+
+  </Accordion>
+
   <Accordion title="Cold local model times out">
     Large local models can need a long first load before streaming begins. Keep the timeout scoped to the Ollama provider, and optionally ask Ollama to keep the model loaded between turns:
 

extensions/ollama/src/stream-runtime.test.ts

@@ -1297,6 +1297,69 @@ describe("createOllamaStreamFn streaming events", () => {
     );
   });
 
+  it("emits an error instead of accepting garbled Kimi visible text", async () => {
+    const garbled =
+      '$$"##"%#"##"####""$""""##""$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$' +
+      '#"$"$"""$""""#$"""$"""%"%###"""#%""""&"#"""$"""#"#""""%#""""&"#"""$"""$"""#%"""';
+    await withMockNdjsonFetch(
+      [
+        JSON.stringify({
+          model: "kimi-k2.5:cloud",
+          created_at: "t",
+          message: { role: "assistant", content: garbled },
+          done: false,
+        }),
+        '{"model":"kimi-k2.5:cloud","created_at":"t","message":{"role":"assistant","content":""},"done":true,"prompt_eval_count":20,"eval_count":40}',
+      ],
+      async () => {
+        const stream = await createOllamaTestStream({
+          baseUrl: "http://ollama-host:11434",
+          model: { id: "kimi-k2.5:cloud", provider: "ollama" },
+        });
+        const events = await collectStreamEvents(stream);
+
+        const types = events.map((e) => e.type);
+        expect(types).toEqual(["start", "text_start", "text_delta", "error"]);
+        const errorEvent = events.at(-1);
+        expect(errorEvent).toMatchObject({
+          type: "error",
+          error: expect.objectContaining({
+            errorMessage: expect.stringContaining("garbled visible text"),
+          }),
+        });
+      },
+    );
+  });
+
+  it("does not reject punctuation-heavy text from unrelated Ollama models", async () => {
+    const punctuationHeavy =
+      '$$"##"%#"##"####""$""""##""$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$' +
+      '#"$"$"""$""""#$"""$"""%"%###"""#%""""&"#"""$"""#"#""""%#""""&"#"""$"""$"""#%"""';
+    await withMockNdjsonFetch(
+      [
+        JSON.stringify({
+          model: "qwen3:32b",
+          created_at: "t",
+          message: { role: "assistant", content: punctuationHeavy },
+          done: false,
+        }),
+        '{"model":"qwen3:32b","created_at":"t","message":{"role":"assistant","content":""},"done":true,"prompt_eval_count":20,"eval_count":40}',
+      ],
+      async () => {
+        const stream = await createOllamaTestStream({ baseUrl: "http://ollama-host:11434" });
+        const events = await collectStreamEvents(stream);
+
+        expect(events.map((e) => e.type)).toEqual([
+          "start",
+          "text_start",
+          "text_delta",
+          "text_end",
+          "done",
+        ]);
+      },
+    );
+  });
+
   it("emits a single text_delta for single-chunk responses", async () => {
     await withMockNdjsonFetch(
       [

extensions/ollama/src/stream.ts

@@ -41,6 +41,54 @@ const log = createSubsystemLogger("ollama-stream");
 
 export const OLLAMA_NATIVE_BASE_URL = OLLAMA_DEFAULT_BASE_URL;
 
+const GARBLED_VISIBLE_TEXT_MODEL_RE = /\b(?:glm|kimi)\b/i;
+const GARBLED_VISIBLE_TEXT_MIN_CHARS = 80;
+const GARBLED_VISIBLE_TEXT_SYMBOL_RE = /[$#%&="'_~`^|\\/*+\-[\]{}()<>:;,.!?]/gu;
+const LETTER_OR_DIGIT_RE = /[\p{L}\p{N}]/gu;
+
+function countMatches(text: string, re: RegExp): number {
+  re.lastIndex = 0;
+  return Array.from(text.matchAll(re)).length;
+}
+
+function maxCharacterFrequency(text: string): number {
+  const counts = new Map<string, number>();
+  let max = 0;
+  for (const char of text) {
+    const count = (counts.get(char) ?? 0) + 1;
+    counts.set(char, count);
+    max = Math.max(max, count);
+  }
+  return max;
+}
+
+function isKnownOllamaGarbledVisibleTextModel(modelId: string): boolean {
+  return GARBLED_VISIBLE_TEXT_MODEL_RE.test(modelId);
+}
+
+function isLikelyGarbledVisibleText(params: { text: string; modelId: string }): boolean {
+  if (!isKnownOllamaGarbledVisibleTextModel(params.modelId)) {
+    return false;
+  }
+  const compact = params.text.replace(/\s+/g, "");
+  if (compact.length < GARBLED_VISIBLE_TEXT_MIN_CHARS) {
+    return false;
+  }
+
+  const letterOrDigitCount = countMatches(compact, LETTER_OR_DIGIT_RE);
+  const symbolCount = countMatches(compact, GARBLED_VISIBLE_TEXT_SYMBOL_RE);
+  const maxFrequency = maxCharacterFrequency(compact);
+  const letterOrDigitRatio = letterOrDigitCount / compact.length;
+  const symbolRatio = symbolCount / compact.length;
+  const dominantCharacterRatio = maxFrequency / compact.length;
+
+  return (
+    letterOrDigitRatio < 0.08 &&
+    symbolRatio > 0.6 &&
+    (dominantCharacterRatio > 0.22 || /[$#%&="'_~`^|\\/*+\-[\]{}()<>:;,.!?]{12,}/u.test(compact))
+  );
+}
+
 export function resolveOllamaBaseUrlForRun(params: {
   modelBaseUrl?: string;
   providerBaseUrl?: string;
@@ -1129,6 +1177,12 @@ export function createOllamaStreamFn(
             throw new Error("Ollama API stream ended without a final response");
           }
 
+          if (isLikelyGarbledVisibleText({ text: accumulatedContent, modelId: model.id })) {
+            throw new Error(
+              `Ollama returned non-linguistic garbled visible text for ${model.id}; retry or switch models`,
+            );
+          }
+
           finalResponse.message.content = accumulatedContent;
           if (accumulatedThinking) {
             finalResponse.message.thinking = accumulatedThinking;