feat: 支持 GLM 系列模型（智谱 AI）作为推理后端

src/agent/context.ts

@@ -55,10 +55,15 @@ export function estimateTokens(text: string): number {
  * Truncate a tool result to fit within the size limit.
  * Appends a truncation notice if content was trimmed.
  */
-export function truncateToolResult(result: string, maxSize: number = MAX_TOOL_RESULT_SIZE): string {
+export function truncateToolResult(
+  result: string,
+  maxSize: number = MAX_TOOL_RESULT_SIZE,
+): string {
   if (result.length <= maxSize) return result;
-  return result.slice(0, maxSize) +
-    `\n\n[TRUNCATED: ${result.length - maxSize} characters omitted]`;
+  return (
+    result.slice(0, maxSize) +
+    `\n\n[TRUNCATED: ${result.length - maxSize} characters omitted]`
+  );
 }
 
 /**
@@ -95,9 +100,7 @@ export function buildContextMessages(
 ): ChatMessage[] {
   const budget = options?.budget ?? DEFAULT_TOKEN_BUDGET;
 
-  const messages: ChatMessage[] = [
-    { role: "system", content: systemPrompt },
-  ];
+  const messages: ChatMessage[] = [{ role: "system", content: systemPrompt }];
 
   // Calculate token estimates for all turns
   const turnTokens = recentTurns.map((turn) => ({
@@ -127,7 +130,8 @@ export function buildContextMessages(
 
     // Ensure we always summarize at least something
     if (splitIndex === 0) splitIndex = 1;
-    if (splitIndex >= recentTurns.length) splitIndex = Math.max(1, recentTurns.length - 1);
+    if (splitIndex >= recentTurns.length)
+      splitIndex = Math.max(1, recentTurns.length - 1);
 
     const oldTurns = recentTurns.slice(0, splitIndex);
     turnsToRender = recentTurns.slice(splitIndex);
@@ -185,9 +189,7 @@ export function buildContextMessages(
 
       // Add tool results with truncation
       for (const tc of turn.toolCalls) {
-        const rawContent = tc.error
-          ? `Error: ${tc.error}`
-          : tc.result;
+        const rawContent = tc.error ? `Error: ${tc.error}` : tc.result;
         messages.push({
           role: "tool",
           content: truncateToolResult(rawContent),
@@ -260,14 +262,18 @@ export function formatMemoryBlock(memories: MemoryRetrievalResult): string {
   if (memories.workingMemory.length > 0) {
     sections.push("### Working Memory");
     for (const e of memories.workingMemory) {
-      sections.push(`- [${e.contentType}] (p=${e.priority.toFixed(1)}) ${e.content}`);
+      sections.push(
+        `- [${e.contentType}] (p=${e.priority.toFixed(1)}) ${e.content}`,
+      );
     }
   }
 
   if (memories.episodicMemory.length > 0) {
     sections.push("### Recent History");
     for (const e of memories.episodicMemory) {
-      sections.push(`- [${e.eventType}] ${e.summary} (${e.outcome || "neutral"})`);
+      sections.push(
+        `- [${e.eventType}] ${e.summary} (${e.outcome || "neutral"})`,
+      );
     }
   }
 
@@ -281,14 +287,18 @@ export function formatMemoryBlock(memories: MemoryRetrievalResult): string {
   if (memories.proceduralMemory.length > 0) {
     sections.push("### Known Procedures");
     for (const e of memories.proceduralMemory) {
-      sections.push(`- ${e.name}: ${e.description} (${e.steps.length} steps, ${e.successCount}/${e.successCount + e.failureCount} success)`);
+      sections.push(
+        `- ${e.name}: ${e.description} (${e.steps.length} steps, ${e.successCount}/${e.successCount + e.failureCount} success)`,
+      );
     }
   }
 
   if (memories.relationships.length > 0) {
     sections.push("### Known Entities");
     for (const e of memories.relationships) {
-      sections.push(`- ${e.entityName || e.entityAddress}: ${e.relationshipType} (trust: ${e.trustScore.toFixed(1)})`);
+      sections.push(
+        `- ${e.entityName || e.entityAddress}: ${e.relationshipType} (trust: ${e.trustScore.toFixed(1)})`,
+      );
     }
   }
 
@@ -321,20 +331,23 @@ export async function summarizeTurns(
 
   // For many turns, use inference to create a summary
   try {
-    const response = await inference.chat([
+    const response = await inference.chat(
+      [
+        {
+          role: "system",
+          content:
+            "Summarize the following agent activity log into a concise paragraph. Focus on: what was accomplished, what failed, current goals, and important context for the next turn.",
+        },
+        {
+          role: "user",
+          content: turnSummaries.join("\n"),
+        },
+      ],
       {
-        role: "system",
-        content:
-          "Summarize the following agent activity log into a concise paragraph. Focus on: what was accomplished, what failed, current goals, and important context for the next turn.",
+        maxTokens: 500,
+        temperature: 0.1, // GLM-5 requires temperature > 0
       },
-      {
-        role: "user",
-        content: turnSummaries.join("\n"),
-      },
-    ], {
-      maxTokens: 500,
-      temperature: 0,
-    });
+    );
 
     return `Previous activity summary:\n${response.message.content}`;
   } catch {

src/agent/loop.ts

@@ -110,8 +110,14 @@ export async function runAgentLoop(
   };
 
   // Initialize inference router (Phase 2.3)
+  // Merge: defaults < top-level config fields < nested modelStrategy
   const modelStrategyConfig: ModelStrategyConfig = {
     ...DEFAULT_MODEL_STRATEGY_CONFIG,
+    // Bridge top-level config fields for backward compatibility
+    // This ensures inferenceModel/maxTokensPerTurn work without nested modelStrategy
+    ...(config.inferenceModel ? { inferenceModel: config.inferenceModel } : {}),
+    ...(config.maxTokensPerTurn ? { maxTokensPerTurn: config.maxTokensPerTurn } : {}),
+    // Nested modelStrategy takes highest priority
     ...(config.modelStrategy ?? {}),
   };
   const modelRegistry = new ModelRegistry(db.raw);

src/conway/inference.ts

@@ -25,6 +25,8 @@ interface InferenceClientOptions {
   maxTokens: number;
   lowComputeModel?: string;
   openaiApiKey?: string;
+  /** Custom OpenAI API base URL (e.g., for GLM Coding Plan: https://open.bigmodel.cn/api/coding/paas/v4) */
+  openaiApiBaseUrl?: string;
   anthropicApiKey?: string;
   ollamaBaseUrl?: string;
   /** Optional registry lookup — if provided, used before name heuristics */
@@ -36,7 +38,14 @@ type InferenceBackend = "conway" | "openai" | "anthropic" | "ollama";
 export function createInferenceClient(
   options: InferenceClientOptions,
 ): InferenceClient {
-  const { apiUrl, apiKey, openaiApiKey, anthropicApiKey, ollamaBaseUrl, getModelProvider } = options;
+  const {
+    apiUrl,
+    apiKey,
+    openaiApiKey,
+    anthropicApiKey,
+    ollamaBaseUrl,
+    getModelProvider,
+  } = options;
   const httpClient = new ResilientHttpClient({
     baseTimeout: INFERENCE_TIMEOUT_MS,
     retryableStatuses: [429, 500, 502, 503, 504],
@@ -98,13 +107,17 @@ export function createInferenceClient(
     }
 
     const openAiLikeApiUrl =
-      backend === "openai" ? "https://api.openai.com" :
-      backend === "ollama" ? (ollamaBaseUrl as string).replace(/\/$/, "") :
-      apiUrl;
+      backend === "openai"
+        ? options.openaiApiBaseUrl || "https://api.openai.com"
+        : backend === "ollama"
+          ? (ollamaBaseUrl as string).replace(/\/$/, "")
+          : apiUrl;
     const openAiLikeApiKey =
-      backend === "openai" ? (openaiApiKey as string) :
-      backend === "ollama" ? "ollama" :
-      apiKey;
+      backend === "openai"
+        ? (openaiApiKey as string)
+        : backend === "ollama"
+          ? "ollama"
+          : apiKey;
 
     return chatViaOpenAiCompatible({
       model,
@@ -141,9 +154,7 @@ export function createInferenceClient(
   };
 }
 
-function formatMessage(
-  msg: ChatMessage,
-): Record<string, unknown> {
+function formatMessage(msg: ChatMessage): Record<string, unknown> {
   const formatted: Record<string, unknown> = {
     role: msg.role,
     content: msg.content,
@@ -182,9 +193,14 @@ function resolveInferenceBackend(
 
   // Heuristic fallback (model not in registry yet)
   if (keys.anthropicApiKey && /^claude/i.test(model)) return "anthropic";
-  if (keys.openaiApiKey && /^(gpt-[3-9]|gpt-4|gpt-5|o[1-9][-\s.]|o[1-9]$|chatgpt)/i.test(model)) return "openai";
+  // GLM models (智谱): glm-* - 使用 OpenAI 兼容 API
+  if (keys.openaiApiKey && /^glm/i.test(model)) return "openai";
+  if (
+    keys.openaiApiKey &&
+    /^(gpt-[3-9]|gpt-4|gpt-5|o[1-9][-\s.]|o[1-9]$|chatgpt)/i.test(model)
+  )
+    return "openai";
   return "conway";
-
 }
 
 async function chatViaOpenAiCompatible(params: {
@@ -195,7 +211,14 @@ async function chatViaOpenAiCompatible(params: {
   backend: "conway" | "openai" | "ollama";
   httpClient: ResilientHttpClient;
 }): Promise<InferenceResponse> {
-  const resp = await params.httpClient.request(`${params.apiUrl}/v1/chat/completions`, {
+  // GLM Coding Plan uses /chat/completions instead of /v1/chat/completions
+  const isGlmApi =
+    params.apiUrl.includes("bigmodel.cn") ||
+    params.apiUrl.includes("/v4") ||
+    params.apiUrl.includes("/paas");
+  const endpoint = isGlmApi ? "/chat/completions" : "/v1/chat/completions";
+
+  const resp = await params.httpClient.request(`${params.apiUrl}${endpoint}`, {
     method: "POST",
     headers: {
       "Content-Type": "application/json",
@@ -215,7 +238,7 @@ async function chatViaOpenAiCompatible(params: {
     );
   }
 
-  const data = await resp.json() as any;
+  const data = (await resp.json()) as any;
   const choice = data.choices?.[0];
 
   if (!choice) {
@@ -229,15 +252,16 @@ async function chatViaOpenAiCompatible(params: {
     totalTokens: data.usage?.total_tokens || 0,
   };
 
-  const toolCalls: InferenceToolCall[] | undefined =
-    message.tool_calls?.map((tc: any) => ({
+  const toolCalls: InferenceToolCall[] | undefined = message.tool_calls?.map(
+    (tc: any) => ({
       id: tc.id,
       type: "function" as const,
       function: {
         name: tc.function.name,
         arguments: tc.function.arguments,
       },
-    }));
+    }),
+  );
 
   return {
     id: data.id || "",
@@ -269,7 +293,9 @@ async function chatViaAnthropic(params: {
     messages:
       transformed.messages.length > 0
         ? transformed.messages
-        : (() => { throw new Error("Cannot send empty message array to Anthropic API"); })(),
+        : (() => {
+            throw new Error("Cannot send empty message array to Anthropic API");
+          })(),
   };
 
   if (transformed.system) {
@@ -289,23 +315,26 @@ async function chatViaAnthropic(params: {
     body.tool_choice = { type: "auto" };
   }
 
-  const resp = await params.httpClient.request("https://api.anthropic.com/v1/messages", {
-    method: "POST",
-    headers: {
-      "Content-Type": "application/json",
-      "x-api-key": params.anthropicApiKey,
-      "anthropic-version": "2023-06-01",
+  const resp = await params.httpClient.request(
+    "https://api.anthropic.com/v1/messages",
+    {
+      method: "POST",
+      headers: {
+        "Content-Type": "application/json",
+        "x-api-key": params.anthropicApiKey,
+        "anthropic-version": "2023-06-01",
+      },
+      body: JSON.stringify(body),
+      timeout: INFERENCE_TIMEOUT_MS,
     },
-    body: JSON.stringify(body),
-    timeout: INFERENCE_TIMEOUT_MS,
-  });
+  );
 
   if (!resp.ok) {
     const text = await resp.text();
     throw new Error(`Inference error (anthropic): ${resp.status}: ${text}`);
   }
 
-  const data = await resp.json() as any;
+  const data = (await resp.json()) as any;
   const content = Array.isArray(data.content) ? data.content : [];
   const textBlocks = content.filter((c: any) => c?.type === "text");
   const toolUseBlocks = content.filter((c: any) => c?.type === "tool_use");
@@ -353,9 +382,10 @@ async function chatViaAnthropic(params: {
   };
 }
 
-function transformMessagesForAnthropic(
-  messages: ChatMessage[],
-): { system?: string; messages: Array<Record<string, unknown>> } {
+function transformMessagesForAnthropic(messages: ChatMessage[]): {
+  system?: string;
+  messages: Array<Record<string, unknown>>;
+} {
   const systemParts: string[] = [];
   const transformed: Array<Record<string, unknown>> = [];