fix(agents): normalize SiliconFlow Pro thinking=off payload (#25435)

Land PR #25435 from @Zjianru.
Changelog: add 2026.2.24 fix entry with contributor credit.

Co-authored-by: codez <codezhujr@gmail.com>

CHANGELOG.md

@@ -31,6 +31,7 @@ Docs: https://docs.openclaw.ai
 - WhatsApp/Web reconnect: treat close status `440` as non-retryable (including string-form status values), stop reconnect loops immediately, and emit operator guidance to relink after resolving session conflicts. (#25858) Thanks @markmusson.
 - Onboarding/Telegram: keep core-channel onboarding available when plugin registry population is missing by falling back to built-in adapters and continuing wizard setup with actionable recovery guidance. (#25803) Thanks @Suko.
 - Models/Bedrock auth: normalize additional Bedrock provider aliases (`bedrock`, `aws-bedrock`, `aws_bedrock`, `amazon bedrock`) to canonical `amazon-bedrock`, ensuring auth-mode resolution consistently selects AWS SDK fallback. (#25756) Thanks @fwhite13.
+- Providers/SiliconFlow: normalize `thinking="off"` to `thinking: null` for `Pro/*` model payloads to avoid provider-side 400 loops and misleading compaction retries. (#25435) Thanks @Zjianru.
 - Gateway/Models: honor explicit `agents.defaults.models` allowlist refs even when bundled model catalog data is stale, synthesize missing allowlist entries in `models.list`, and allow `sessions.patch`/`/model` selection for those refs without false `model not allowed` errors. (#20291) Thanks @kensipe, @nikolasdehor, and @vincentkoc.
 - Automation/Subagent/Cron reliability: honor `ANNOUNCE_SKIP` in `sessions_spawn` completion/direct announce flows (no user-visible token leaks), add transient direct-announce retries for channel unavailability (for example WhatsApp listener reconnect windows), and include `cron` in the `coding` tool profile so `/tools/invoke` can execute cron actions when explicitly allowed by gateway policy. (#25800, #25656, #25842, #25813, #25822, #25821) Thanks @astra-fer, @aaajiao, @dwight11232-coder, @kevinWangSheng, @widingmarcus-cyber, and @stakeswky.
 - Discord/Proxy + reactions + model picker: thread channel proxy fetch into inbound media/sticker downloads, use proxy-aware gateway metadata fetch for WSL/corporate proxy setups, wire `messages.statusReactions.{emojis,timing}` into Discord reaction lifecycle control, and compact model-picker `custom_id` keys to stay under Discord's 100-char limit while keeping backward-compatible parsing. (#25232, #25507, #25564, #25695) Thanks @openperf, @chilu18, @Yipsh, @lbo728, and @s1korrrr.

src/agents/pi-embedded-runner-extraparams.test.ts

@@ -310,6 +310,68 @@ describe("applyExtraParamsToAgent", () => {
     expect(payloads[0]).toEqual({ reasoning: { max_tokens: 256 } });
   });
 
+  it("normalizes thinking=off to null for SiliconFlow Pro models", () => {
+    const payloads: Record<string, unknown>[] = [];
+    const baseStreamFn: StreamFn = (_model, _context, options) => {
+      const payload: Record<string, unknown> = { thinking: "off" };
+      options?.onPayload?.(payload);
+      payloads.push(payload);
+      return {} as ReturnType<StreamFn>;
+    };
+    const agent = { streamFn: baseStreamFn };
+
+    applyExtraParamsToAgent(
+      agent,
+      undefined,
+      "siliconflow",
+      "Pro/MiniMaxAI/MiniMax-M2.1",
+      undefined,
+      "off",
+    );
+
+    const model = {
+      api: "openai-completions",
+      provider: "siliconflow",
+      id: "Pro/MiniMaxAI/MiniMax-M2.1",
+    } as Model<"openai-completions">;
+    const context: Context = { messages: [] };
+    void agent.streamFn?.(model, context, {});
+
+    expect(payloads).toHaveLength(1);
+    expect(payloads[0]?.thinking).toBeNull();
+  });
+
+  it("keeps thinking=off unchanged for non-Pro SiliconFlow model IDs", () => {
+    const payloads: Record<string, unknown>[] = [];
+    const baseStreamFn: StreamFn = (_model, _context, options) => {
+      const payload: Record<string, unknown> = { thinking: "off" };
+      options?.onPayload?.(payload);
+      payloads.push(payload);
+      return {} as ReturnType<StreamFn>;
+    };
+    const agent = { streamFn: baseStreamFn };
+
+    applyExtraParamsToAgent(
+      agent,
+      undefined,
+      "siliconflow",
+      "deepseek-ai/DeepSeek-V3.2",
+      undefined,
+      "off",
+    );
+
+    const model = {
+      api: "openai-completions",
+      provider: "siliconflow",
+      id: "deepseek-ai/DeepSeek-V3.2",
+    } as Model<"openai-completions">;
+    const context: Context = { messages: [] };
+    void agent.streamFn?.(model, context, {});
+
+    expect(payloads).toHaveLength(1);
+    expect(payloads[0]?.thinking).toBe("off");
+  });
+
   it("adds OpenRouter attribution headers to stream options", () => {
     const { calls, agent } = createOptionsCaptureAgent();
 

src/agents/pi-embedded-runner/extra-params.ts

@@ -408,6 +408,42 @@ function mapThinkingLevelToOpenRouterReasoningEffort(
   return thinkingLevel;
 }
 
+function shouldApplySiliconFlowThinkingOffCompat(params: {
+  provider: string;
+  modelId: string;
+  thinkingLevel?: ThinkLevel;
+}): boolean {
+  return (
+    params.provider === "siliconflow" &&
+    params.thinkingLevel === "off" &&
+    params.modelId.startsWith("Pro/")
+  );
+}
+
+/**
+ * SiliconFlow's Pro/* models reject string thinking modes (including "off")
+ * with HTTP 400 invalid-parameter errors. Normalize to `thinking: null` to
+ * preserve "thinking disabled" intent without sending an invalid enum value.
+ */
+function createSiliconFlowThinkingWrapper(baseStreamFn: StreamFn | undefined): 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") {
+          const payloadObj = payload as Record<string, unknown>;
+          if (payloadObj.thinking === "off") {
+            payloadObj.thinking = null;
+          }
+        }
+        originalOnPayload?.(payload);
+      },
+    });
+  };
+}
+
 /**
  * Create a streamFn wrapper that adds OpenRouter app attribution headers
  * and injects reasoning.effort based on the configured thinking level.
@@ -544,6 +580,13 @@ export function applyExtraParamsToAgent(
     agent.streamFn = createAnthropicBetaHeadersWrapper(agent.streamFn, anthropicBetas);
   }
 
+  if (shouldApplySiliconFlowThinkingOffCompat({ provider, modelId, thinkingLevel })) {
+    log.debug(
+      `normalizing thinking=off to thinking=null for SiliconFlow compatibility (${provider}/${modelId})`,
+    );
+    agent.streamFn = createSiliconFlowThinkingWrapper(agent.streamFn);
+  }
+
   if (provider === "openrouter") {
     log.debug(`applying OpenRouter app attribution headers for ${provider}/${modelId}`);
     // "auto" is a dynamic routing model — we don't know which underlying model