From b6febba2bf365d9a1d4dc6f3dc4dd2dbf87e025f Mon Sep 17 00:00:00 2001
From: "Siren.W" <sirenexcelsior@gmail.com>
Date: Tue, 19 May 2026 21:45:59 +0800
Subject: [PATCH] Fix standalone DeepSeek L1 extraction

---
 src/adapters/standalone/llm-runner.ts | 36 +++++++++++++++++++--------
 src/gateway/config.ts                 | 12 +++++++++
 2 files changed, 38 insertions(+), 10 deletions(-)

diff --git a/src/adapters/standalone/llm-runner.ts b/src/adapters/standalone/llm-runner.ts
index 2f7c9e1..8724c30 100644
--- a/src/adapters/standalone/llm-runner.ts
+++ b/src/adapters/standalone/llm-runner.ts
@@ -49,6 +49,8 @@ export interface StandaloneLLMConfig {
   maxTokens?: number;
   /** Request timeout in milliseconds (default: 120_000). */
   timeoutMs?: number;
+  /** Disable thinking/reasoning mode (DeepSeek: sets thinking.type=disabled in request body). */
+  disableThinking?: boolean;
 }
 
 // ============================
@@ -149,12 +151,6 @@ function createSandboxedTools(workspaceDir: string, logger?: Logger) {
   };
 }
 
-/** Read-only tool subset — used when enableTools=false to avoid empty tools rejection. */
-function createReadOnlyTools(workspaceDir: string, logger?: Logger) {
-  const all = createSandboxedTools(workspaceDir, logger);
-  return { read_file: all.read_file };
-}
-
 // ============================
 // StandaloneLLMRunner
 // ============================
@@ -185,29 +181,49 @@ export class StandaloneLLMRunner implements LLMRunner {
 
     this.logger?.debug?.(
       `${TAG} run() start: taskId=${params.taskId}, model=${this.model}, ` +
-      `tools=${this.enableTools}, timeout=${timeoutMs}ms`,
+      `tools=${this.enableTools}, disableThinking=${this.config.disableThinking === true}, timeout=${timeoutMs}ms`,
     );
 
     // Create OpenAI-compatible provider via AI SDK
     // Use "compatible" mode to call /chat/completions (not Responses API),
     // which works with all OpenAI-compatible backends (DeepSeek, Qwen, etc.)
+    const disableThinking = this.config.disableThinking === true;
     const provider = createOpenAI({
       baseURL: this.config.baseUrl,
       apiKey: this.config.apiKey,
       compatibility: "compatible",
+      ...(disableThinking
+        ? {
+            fetch: async (url, init) => {
+              if (typeof init?.body === "string") {
+                try {
+                  const body = JSON.parse(init.body);
+                  body.thinking = { type: "disabled" };
+                  init = { ...init, body: JSON.stringify(body) };
+                } catch (error) {
+                  const errMsg = error instanceof Error ? error.message : String(error);
+                  this.logger?.warn?.(`${TAG} failed to inject thinking=disabled: ${errMsg}`);
+                }
+              }
+              return fetch(url, init);
+            },
+          }
+        : {}),
     });
 
-    // Select tools based on mode
+    // For pure text tasks like L1 extraction, avoid exposing any tools.
+    // DeepSeek may opportunistically issue tool calls even when the prompt
+    // asks for plain JSON, which can lead to empty/non-JSON output.
     const tools = this.enableTools
       ? createSandboxedTools(workspaceDir, this.logger)
-      : createReadOnlyTools(workspaceDir, this.logger);
+      : undefined;
 
     try {
       const result = await generateText({
         model: provider.chat(this.model),
         system: params.systemPrompt,
         prompt: params.prompt,
-        tools,
+        ...(tools ? { tools } : {}),
         stopWhen: stepCountIs(this.enableTools ? MAX_TOOL_ITERATIONS : 1),
         maxOutputTokens: maxTokens,
         abortSignal: AbortSignal.timeout(timeoutMs),
diff --git a/src/gateway/config.ts b/src/gateway/config.ts
index 81c55ad..1a91321 100644
--- a/src/gateway/config.ts
+++ b/src/gateway/config.ts
@@ -92,6 +92,7 @@ export function loadGatewayConfig(overrides?: Partial<GatewayConfig>): GatewayCo
     model: env("TDAI_LLM_MODEL") ?? str(llmConfig, "model") ?? "gpt-4o",
     maxTokens: envInt("TDAI_LLM_MAX_TOKENS") ?? num(llmConfig, "maxTokens") ?? 4096,
     timeoutMs: envInt("TDAI_LLM_TIMEOUT_MS") ?? num(llmConfig, "timeoutMs") ?? 120_000,
+    disableThinking: envBool("TDAI_LLM_DISABLE_THINKING") ?? bool(llmConfig, "disableThinking") ?? false,
   };
 
   // Memory config (reuse the plugin's parseConfig for full compatibility)
@@ -184,6 +185,17 @@ function envInt(key: string): number | undefined {
   return Number.isFinite(n) ? n : undefined;
 }
 
+function envBool(key: string): boolean | undefined {
+  const v = env(key);
+  if (!v) return undefined;
+  return v === "1" || v.toLowerCase() === "true";
+}
+
+function bool(src: Record<string, unknown>, key: string): boolean | undefined {
+  const v = src[key];
+  return typeof v === "boolean" ? v : undefined;
+}
+
 function obj(c: Record<string, unknown>, key: string): Record<string, unknown> {
   const v = c[key];
   return v && typeof v === "object" && !Array.isArray(v) ? v as Record<string, unknown> : {};