fix: skip invalid completions stream chunks

CHANGELOG.md

@@ -74,6 +74,7 @@ Docs: https://docs.openclaw.ai
 - Config/doctor: reject legacy `secretref-env:<ENV_VAR>` marker strings on SecretRef credential paths and migrate valid markers to structured env SecretRefs with `openclaw doctor --fix`. Fixes #51794. Thanks @halointellicore.
 - Providers/OpenAI: separate API-key and Codex sign-in onboarding groups, and avoid replaying stale OpenAI Responses reasoning blocks after a model route switch.
 - Providers/OpenAI-compatible: forward `prompt_cache_key` on Completions requests only for providers that opt in with `compat.supportsPromptCacheKey`, keeping default proxy payloads unchanged. Fixes #69272.
+- Providers/OpenAI-compatible: skip null or non-object streaming chunks from custom providers instead of failing the turn after partial output. Fixes #51112.
 - Providers/ElevenLabs: omit the MP3-only `Accept` header for PCM telephony synthesis, so Voice Call requests for `pcm_22050` no longer receive MP3 audio. Fixes #67340. Thanks @marcchabot.
 - Plugins/Voice Call: reap stale pre-answer calls by default, honor configured TTS timeouts for Twilio media-stream playback, and fail empty telephony audio instead of completing as silence. Fixes #42071; supersedes #60957. Thanks @Ryce and @sliekens.
 - Plugins/Voice Call: terminate expired restored call sessions with the provider and restart restored max-duration timers with only the remaining duration, preventing stale outbound retry loops after Gateway restarts. Fixes #48739. Thanks @mira-solari.

src/agents/openai-transport-stream.test.ts

@@ -466,6 +466,63 @@ describe("openai transport stream", () => {
     });
   });
 
+  it("skips null and non-object OpenAI-compatible stream chunks", async () => {
+    const model = {
+      id: "glm-5",
+      name: "GLM-5",
+      api: "openai-completions",
+      provider: "vllm",
+      baseUrl: "http://localhost:8000/v1",
+      reasoning: false,
+      input: ["text"],
+      cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+      contextWindow: 128000,
+      maxTokens: 4096,
+    } satisfies Model<"openai-completions">;
+    const output = {
+      role: "assistant" as const,
+      content: [],
+      api: model.api,
+      provider: model.provider,
+      model: model.id,
+      usage: {
+        input: 0,
+        output: 0,
+        cacheRead: 0,
+        cacheWrite: 0,
+        totalTokens: 0,
+        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+      },
+      stopReason: "stop",
+      timestamp: Date.now(),
+    };
+    const stream: { push(event: unknown): void } = { push() {} };
+
+    async function* mockStream() {
+      yield null as never;
+      yield "not-a-chunk" as never;
+      yield {
+        id: "chatcmpl-vllm",
+        object: "chat.completion.chunk" as const,
+        created: 1775425651,
+        model: "glm-5",
+        choices: [
+          {
+            index: 0,
+            delta: { role: "assistant" as const, content: "ok" },
+            logprobs: null,
+            finish_reason: "stop" as const,
+          },
+        ],
+      };
+    }
+
+    await __testing.processOpenAICompletionsStream(mockStream(), output, model, stream);
+
+    expect(output.content).toContainEqual({ type: "text", text: "ok" });
+    expect(output.stopReason).toBe("stop");
+  });
+
   it("keeps OpenRouter thinking format for declared OpenRouter providers on custom proxy URLs", async () => {
     const streamFn = buildTransportAwareSimpleStreamFn(
       attachModelProviderRequestTransport(

src/agents/openai-transport-stream.ts

@@ -1292,7 +1292,11 @@ async function processOpenAICompletionsStream(
     flushPendingPostToolCallDeltas();
     appendTextDeltaInternal(text);
   };
-  for await (const chunk of responseStream) {
+  for await (const rawChunk of responseStream as AsyncIterable<unknown>) {
+    if (!rawChunk || typeof rawChunk !== "object") {
+      continue;
+    }
+    const chunk = rawChunk as ChatCompletionChunk;
     output.responseId ||= chunk.id;
     if (chunk.usage) {
       output.usage = parseTransportChunkUsage(chunk.usage, model);