fix(agents): support parallel_tool_calls config per model

SidQin-cyber · SidQin-cyber · commit 055753baeb17 · 2026-03-06T12:58:30.000+08:00
OpenClaw sends parallel_tool_calls:true to all OpenAI-compatible providers via the upstream pi-agent-core library. Models that don't support parallel tool calling (e.g. kimi-k2.5 on NVIDIA NIM) return 400, breaking all tool execution. Read parallel_tool_calls from model params and inject it into the LLM request payload via an onPayload wrapper, letting users set it to false per provider/model. Closes #37048 Made-with: Cursor
diff --git a/src/agents/pi-embedded-runner-extraparams.test.ts b/src/agents/pi-embedded-runner-extraparams.test.ts
@@ -1353,3 +1353,46 @@ describe("applyExtraParamsToAgent", () => {
     },
   );
 });
+
+describe("parallel_tool_calls wrapper", () => {
+  function capturePayload(provider: string, modelId: string, paramValue: boolean) {
+    let captured: Record<string, unknown> = {};
+    const fakeStreamFn: StreamFn = (_model, _context, options) => {
+      options?.onPayload?.(captured);
+      return undefined as never;
+    };
+    const agent = { streamFn: fakeStreamFn };
+    applyExtraParamsToAgent(
+      agent,
+      {
+        agents: {
+          defaults: {
+            models: {
+              [`${provider}/${modelId}`]: {
+                params: { parallel_tool_calls: paramValue },
+              },
+            },
+          },
+        },
+      },
+      provider,
+      modelId,
+    );
+    void agent.streamFn(
+      { api: "openai-completions", provider, id: modelId } as Model<"openai-completions">,
+      {} as Context,
+      {},
+    );
+    return captured;
+  }
+
+  it("injects parallel_tool_calls=false when configured", () => {
+    const payload = capturePayload("nvidia-nim", "moonshotai/kimi-k2.5", false);
+    expect(payload.parallel_tool_calls).toBe(false);
+  });
+
+  it("injects parallel_tool_calls=true when configured", () => {
+    const payload = capturePayload("openai", "gpt-4", true);
+    expect(payload.parallel_tool_calls).toBe(true);
+  });
+});
diff --git a/src/agents/pi-embedded-runner/extra-params.ts b/src/agents/pi-embedded-runner/extra-params.ts
@@ -963,6 +963,25 @@ function createZaiToolStreamWrapper(
   };
 }
 
+function createParallelToolCallsWrapper(
+  baseStreamFn: StreamFn | undefined,
+  parallelToolCalls: boolean,
+): StreamFn {
+  const underlying = baseStreamFn ?? streamSimple;
+  return (model, context, options) => {
+    const originalOnPayload = options?.onPayload;
+    return underlying(model, context, {
+      ...options,
+      onPayload: (payload) => {
+        if (payload && typeof payload === "object") {
+          (payload as Record<string, unknown>).parallel_tool_calls = parallelToolCalls;
+        }
+        originalOnPayload?.(payload);
+      },
+    });
+  };
+}
+
 /**
  * Apply extra params (like temperature) to an agent's streamFn.
  * Also adds OpenRouter app attribution headers when using the OpenRouter provider.
@@ -1069,6 +1088,13 @@ export function applyExtraParamsToAgent(
     }
   }
 
+  if (typeof merged?.parallel_tool_calls === "boolean") {
+    log.debug(
+      `applying parallel_tool_calls=${merged.parallel_tool_calls} for ${provider}/${modelId}`,
+    );
+    agent.streamFn = createParallelToolCallsWrapper(agent.streamFn, merged.parallel_tool_calls);
+  }
+
   // Guard Google payloads against invalid negative thinking budgets emitted by
   // upstream model-ID heuristics for Gemini 3.1 variants.
   agent.streamFn = createGoogleThinkingPayloadWrapper(agent.streamFn, thinkingLevel);
