diff --git a/src/agents/pi-embedded-runner-extraparams.test.ts b/src/agents/pi-embedded-runner-extraparams.test.ts
index a6d3e9191e8..3f5189a40ea 100644
--- a/src/agents/pi-embedded-runner-extraparams.test.ts
+++ b/src/agents/pi-embedded-runner-extraparams.test.ts
@@ -202,6 +202,58 @@ describe("applyExtraParamsToAgent", () => {
     return calls[0]?.headers;
   }
 
+  it("does not inject reasoning when thinkingLevel is off (default) for OpenRouter", () => {
+    // Regression: "off" is a truthy string, so the old code injected
+    // reasoning: { effort: "none" }, causing a 400 on models that require
+    // reasoning (e.g. deepseek/deepseek-r1).
+    const payloads: Record<string, unknown>[] = [];
+    const baseStreamFn: StreamFn = (_model, _context, options) => {
+      const payload: Record<string, unknown> = { model: "deepseek/deepseek-r1" };
+      options?.onPayload?.(payload);
+      payloads.push(payload);
+      return {} as ReturnType<StreamFn>;
+    };
+    const agent = { streamFn: baseStreamFn };
+
+    applyExtraParamsToAgent(agent, undefined, "openrouter", "deepseek/deepseek-r1", undefined, "off");
+
+    const model = {
+      api: "openai-completions",
+      provider: "openrouter",
+      id: "deepseek/deepseek-r1",
+    } as Model<"openai-completions">;
+    const context: Context = { messages: [] };
+    void agent.streamFn?.(model, context, {});
+
+    expect(payloads).toHaveLength(1);
+    expect(payloads[0]).not.toHaveProperty("reasoning");
+    expect(payloads[0]).not.toHaveProperty("reasoning_effort");
+  });
+
+  it("injects reasoning.effort when thinkingLevel is non-off for OpenRouter", () => {
+    const payloads: Record<string, unknown>[] = [];
+    const baseStreamFn: StreamFn = (_model, _context, options) => {
+      const payload: Record<string, unknown> = {};
+      options?.onPayload?.(payload);
+      payloads.push(payload);
+      return {} as ReturnType<StreamFn>;
+    };
+    const agent = { streamFn: baseStreamFn };
+
+    applyExtraParamsToAgent(agent, undefined, "openrouter", "openrouter/auto", undefined, "low");
+
+    const model = {
+      api: "openai-completions",
+      provider: "openrouter",
+      id: "openrouter/auto",
+    } as Model<"openai-completions">;
+    const context: Context = { messages: [] };
+    void agent.streamFn?.(model, context, {});
+
+    expect(payloads).toHaveLength(1);
+    expect(payloads[0]?.reasoning).toEqual({ effort: "low" });
+  });
+
   it("adds OpenRouter attribution headers to stream options", () => {
     const { calls, agent } = createOptionsCaptureAgent();
 
diff --git a/src/agents/pi-embedded-runner/extra-params.ts b/src/agents/pi-embedded-runner/extra-params.ts
index 8ebacf6df68..66b077af232 100644
--- a/src/agents/pi-embedded-runner/extra-params.ts
+++ b/src/agents/pi-embedded-runner/extra-params.ts
@@ -435,24 +435,31 @@ function createOpenRouterWrapper(
           // only the nested one is sent.
           delete payloadObj.reasoning_effort;
 
-          const existingReasoning = payloadObj.reasoning;
+          // When thinking is "off", do not inject reasoning at all.
+          // Some models (e.g. deepseek/deepseek-r1) require reasoning and reject
+          // { effort: "none" } with "Reasoning is mandatory for this endpoint and
+          // cannot be disabled." Omitting the field lets each model use its own
+          // default reasoning behavior.
+          if (thinkingLevel !== "off") {
+            const existingReasoning = payloadObj.reasoning;
 
-          // OpenRouter treats reasoning.effort and reasoning.max_tokens as
-          // alternative controls. If max_tokens is already present, do not
-          // inject effort and do not overwrite caller-supplied reasoning.
-          if (
-            existingReasoning &&
-            typeof existingReasoning === "object" &&
-            !Array.isArray(existingReasoning)
-          ) {
-            const reasoningObj = existingReasoning as Record<string, unknown>;
-            if (!("max_tokens" in reasoningObj) && !("effort" in reasoningObj)) {
-              reasoningObj.effort = mapThinkingLevelToOpenRouterReasoningEffort(thinkingLevel);
+            // OpenRouter treats reasoning.effort and reasoning.max_tokens as
+            // alternative controls. If max_tokens is already present, do not
+            // inject effort and do not overwrite caller-supplied reasoning.
+            if (
+              existingReasoning &&
+              typeof existingReasoning === "object" &&
+              !Array.isArray(existingReasoning)
+            ) {
+              const reasoningObj = existingReasoning as Record<string, unknown>;
+              if (!("max_tokens" in reasoningObj) && !("effort" in reasoningObj)) {
+                reasoningObj.effort = mapThinkingLevelToOpenRouterReasoningEffort(thinkingLevel);
+              }
+            } else if (!existingReasoning) {
+              payloadObj.reasoning = {
+                effort: mapThinkingLevelToOpenRouterReasoningEffort(thinkingLevel),
+              };
             }
-          } else if (!existingReasoning) {
-            payloadObj.reasoning = {
-              effort: mapThinkingLevelToOpenRouterReasoningEffort(thinkingLevel),
-            };
           }
         }
         onPayload?.(payload);