thatgfsj-code 2.2.9 → 3.0.0

package/src/llm/index.ts

@@ -1,200 +1,265 @@
-/**
- * LLM Service - Factory for creating providers
- * Supports all providers + custom relay stations (中转站)
- */
-
-import chalk from 'chalk';
-import type { ChatMessage, ChatResponse, ChatOptions, ToolCall } from '../types.js';
-import type { Tool } from '../tools/types.js';
-import type { LLMProvider, StreamChunk } from './provider.js';
-import type { AIConfig, ProviderName } from '../config/types.js';
-import { PROVIDERS } from '../config/providers.js';
-import { OpenAIProvider } from './openai.js';
-import { AnthropicProvider } from './anthropic.js';
-import { GeminiProvider } from './gemini.js';
-
-export class LLMService {
-  private provider: LLMProvider;
-  private tools: Map<string, Tool> = new Map();
-  private apiKey: string;
-
-  constructor(provider: LLMProvider, apiKey: string) {
-    this.provider = provider;
-    this.apiKey = apiKey;
-  }
-
-  static fromConfig(config: AIConfig): LLMService {
-    const providerName = config.provider || 'siliconflow';
-    const providerConfig = PROVIDERS[providerName];
-
-    const providerCfg = {
-      apiKey: config.apiKey || '',
-      model: config.model || providerConfig.defaultModel,
-      baseUrl: config.baseUrl || providerConfig.baseUrl,
-      temperature: config.temperature ?? 0.7,
-      maxTokens: config.maxTokens ?? 4096,
-    };
-
-    const format = providerConfig.format;
-    let provider: LLMProvider;
-
-    switch (format) {
-      case 'anthropic':
-        provider = new AnthropicProvider(providerCfg);
-        break;
-      case 'gemini':
-        provider = new GeminiProvider(providerCfg);
-        break;
-      default:
-        provider = new OpenAIProvider(providerCfg);
-    }
-
-    return new LLMService(provider, providerCfg.apiKey);
-  }
-
-  registerTools(tools: Tool[]): void {
-    for (const tool of tools) {
-      this.tools.set(tool.name, tool);
-    }
-  }
-
-  getProviderName(): string { return this.provider.name; }
-  hasApiKey(): boolean { return !!this.apiKey; }
-
-  async chat(messages: ChatMessage[], options?: ChatOptions): Promise<ChatResponse> {
-    if (!this.hasApiKey()) throw new Error(this.getNoKeyMessage());
-    const toolsArray = [...this.tools.values()];
-    return this.provider.chat(messages, options, toolsArray.length > 0 ? toolsArray : undefined);
-  }
-
-  /**
-   * Streaming chat with agent loop (tool call support)
-   * - Streams text to the user
-   * - Detects structured tool calls from the API
-   * - Executes tools and loops back for more
-   */
-  async *chatStream(
-    messages: ChatMessage[],
-    options?: ChatOptions & { maxIterations?: number }
-  ): AsyncGenerator<string, ChatResponse> {
-    if (!this.hasApiKey()) throw new Error(this.getNoKeyMessage());
-
-    const maxIterations = options?.maxIterations ?? 10;
-    let currentMessages = [...messages];
-    let iterations = 0;
-
-    while (iterations < maxIterations) {
-      iterations++;
-      const toolsArray = [...this.tools.values()];
-      const hasTools = toolsArray.length > 0;
-
-      let fullContent = '';
-      let detectedToolCalls: ToolCall[] | undefined;
-
-      const stream = this.provider.chatStream(currentMessages, options, hasTools ? toolsArray : undefined);
-
-      for await (const chunk of stream) {
-        if (chunk.type === 'text' && chunk.content) {
-          fullContent += chunk.content;
-          yield chunk.content;
-        } else if (chunk.type === 'tool_calls' && chunk.toolCalls) {
-          detectedToolCalls = chunk.toolCalls;
-        }
-      }
-
-      // If we got tool calls, execute them and loop
-      if (detectedToolCalls && detectedToolCalls.length > 0) {
-        // Add assistant message with tool calls
-        currentMessages.push({
-          role: 'assistant',
-          content: fullContent || '',
-          tool_calls: detectedToolCalls,
-        });
-
-        // Execute each tool
-        for (const toolCall of detectedToolCalls) {
-          const tool = this.tools.get(toolCall.function.name);
-
-          // Yield structured tool call as JSON line
-          yield `\n@@TOOL@@${JSON.stringify({ action: 'call', name: toolCall.function.name, args: toolCall.function.arguments })}\n`;
-
-          if (!tool) {
-            const errMsg = `Tool "${toolCall.function.name}" not found`;
-            currentMessages.push({
-              role: 'tool',
-              content: errMsg,
-              tool_call_id: toolCall.id,
-              name: toolCall.function.name,
-            });
-            yield `@@TOOL@@${JSON.stringify({ action: 'result', error: errMsg })}\n`;
-            continue;
-          }
-
-          try {
-            const params = JSON.parse(toolCall.function.arguments || '{}');
-            const result = await tool.execute(params);
-            const output = result.success ? (result.output || JSON.stringify(result.data)) : (result.error || 'Tool failed');
-
-            currentMessages.push({
-              role: 'tool',
-              content: output,
-              tool_call_id: toolCall.id,
-              name: toolCall.function.name,
-            });
-
-            yield `@@TOOL@@${JSON.stringify({ action: 'result', output })}\n`;
-          } catch (error: any) {
-            const errMsg = `Error: ${error.message}`;
-            currentMessages.push({
-              role: 'tool',
-              content: errMsg,
-              tool_call_id: toolCall.id,
-              name: toolCall.function.name,
-            });
-            yield `@@TOOL@@${JSON.stringify({ action: 'result', error: errMsg })}\n`;
-          }
-        }
-
-        yield '\n';
-        continue;
-      }
-
-      // No tool calls - done
-      return { content: fullContent, role: 'assistant' };
-    }
-
-    return { content: '[Agent loop exceeded maximum iterations]', role: 'assistant' };
-  }
-
-  private truncateArgs(args: string): string {
-    try {
-      const obj = JSON.parse(args || '{}');
-      const entries = Object.entries(obj);
-      if (entries.length === 0) return '';
-      return entries.map(([k, v]) => {
-        const val = typeof v === 'string' && v.length > 50 ? v.slice(0, 50) + '...' : v;
-        return `${k}: ${JSON.stringify(val)}`;
-      }).join(', ');
-    } catch {
-      return args.length > 80 ? args.slice(0, 80) + '...' : args;
-    }
-  }
-
-  private getNoKeyMessage(): string {
-    return [
-      '❌ 未配置 API Key，无法调用 AI。',
-      '',
-      '请先运行: gfcode init',
-      '',
-      '或设置环境变量:',
-      '  export SILICONFLOW_API_KEY="sk-..."',
-      '  export OPENAI_API_KEY="sk-..."',
-      '  export DEEPSEEK_API_KEY="sk-..."',
-    ].join('\n');
-  }
-}
-
-export type { LLMProvider, ProviderConfig, StreamChunk } from './provider.js';
-export { OpenAIProvider } from './openai.js';
-export { AnthropicProvider } from './anthropic.js';
-export { GeminiProvider } from './gemini.js';
+/**
+ * LLM Service - Factory for creating providers
+ * Supports all providers + custom relay stations (中转站)
+ *
+ * v3.0.0+: streaming protocol migrated to structured StreamChunk
+ *   - chatStream yields { type: 'text' | 'tool_calls' | 'thinking' | 'usage' }
+ *   - Tool call execution is now part of the agent loop; consumers see
+ *     { type: 'tool_calls', toolCalls } for dispatch and { type: 'text', content }
+ *     for model output. The legacy @@TOOL@@ sentinel-string protocol is GONE.
+ *   - Usage / cache stats are emitted as the final { type: 'usage' } chunk,
+ *     which the TUI cache stats store consumes.
+ */
+
+import chalk from 'chalk';
+import type { ChatMessage, ChatResponse, ChatOptions, ToolCall, StreamChunk } from '../types.js';
+import type { Tool } from '../tools/types.js';
+import type { LLMProvider } from './provider.js';
+import type { AIConfig, Config, ProviderName } from '../config/types.js';
+import { PROVIDERS } from '../config/providers.js';
+import { OpenAIProvider } from './openai.js';
+import { AnthropicProvider } from './anthropic.js';
+import { GeminiProvider } from './gemini.js';
+
+export class LLMService {
+  private provider: LLMProvider;
+  private tools: Map<string, Tool> = new Map();
+  private apiKey: string;
+
+  constructor(provider: LLMProvider, apiKey: string) {
+    this.provider = provider;
+    this.apiKey = apiKey;
+  }
+
+  static fromConfig(config: AIConfig & { cache?: Config['cache'] }): LLMService {
+    const providerName = config.provider || 'siliconflow';
+    const providerConfig = PROVIDERS[providerName];
+
+    const providerCfg = {
+      apiKey: config.apiKey || '',
+      model: config.model || providerConfig.defaultModel,
+      baseUrl: config.baseUrl || providerConfig.baseUrl,
+      temperature: config.temperature ?? 0.7,
+      maxTokens: config.maxTokens ?? 4096,
+      // v3.0.0: forward cache policy. Anthropic reads this to decide
+      // whether to attach cache_control markers; other providers ignore it.
+      cache: config.cache ?? { enabled: true, ttl: '5m', strategy: 'auto' as const },
+    };
+
+    const format = providerConfig.format;
+    let provider: LLMProvider;
+
+    switch (format) {
+      case 'anthropic':
+        provider = new AnthropicProvider(providerCfg);
+        break;
+      case 'gemini':
+        provider = new GeminiProvider(providerCfg);
+        break;
+      default:
+        provider = new OpenAIProvider(providerCfg);
+    }
+
+    return new LLMService(provider, providerCfg.apiKey);
+  }
+
+  registerTools(tools: Tool[]): void {
+    for (const tool of tools) {
+      this.tools.set(tool.name, tool);
+    }
+  }
+
+  getProviderName(): string { return this.provider.name; }
+  hasApiKey(): boolean { return !!this.apiKey; }
+
+  async chat(messages: ChatMessage[], options?: ChatOptions): Promise<ChatResponse> {
+    if (!this.hasApiKey()) throw new Error(this.getNoKeyMessage());
+    const toolsArray = [...this.tools.values()];
+    return this.provider.chat(messages, options, toolsArray.length > 0 ? toolsArray : undefined);
+  }
+
+  /**
+   * Streaming chat with agent loop (tool call support).
+   *
+   * Structured chunk protocol (replaces the @@TOOL@@ sentinel strings):
+   *   { type: 'text',       content: string }      model output, accumulate + render
+   *   { type: 'tool_calls', toolCalls: ToolCall[] } dispatch tools, results will
+   *                                                be folded into the next request
+   *   { type: 'thinking',   content: string }      reasoning text (consumers may
+   *                                                display in debug mode)
+   *   { type: 'usage',      usage: Usage }         cache hit/miss + token counts;
+   *                                                emitted on the final round
+   *
+   * Returns the final ChatResponse of the agent loop when done.
+   */
+  async *chatStream(
+    messages: ChatMessage[],
+    options?: ChatOptions & { maxIterations?: number }
+  ): AsyncGenerator<StreamChunk, ChatResponse> {
+    if (!this.hasApiKey()) throw new Error(this.getNoKeyMessage());
+
+    const maxIterations = options?.maxIterations ?? 10;
+    let currentMessages = [...messages];
+    let iterations = 0;
+    let lastUsage: ChatResponse['usage'] | undefined;
+
+    while (iterations < maxIterations) {
+      iterations++;
+      const toolsArray = [...this.tools.values()];
+      const hasTools = toolsArray.length > 0;
+
+      let fullContent = '';
+      let detectedToolCalls: ToolCall[] | undefined;
+
+      // Forward stream chunks from the provider. We collect text internally for
+      // tool-call persistence but always re-emit the original chunks unchanged.
+      const stream = this.provider.chatStream(currentMessages, options, hasTools ? toolsArray : undefined);
+
+      for await (const chunk of stream) {
+        if (chunk.type === 'text' && chunk.content) {
+          fullContent += chunk.content;
+          yield chunk;
+        } else if (chunk.type === 'tool_calls' && chunk.toolCalls) {
+          detectedToolCalls = chunk.toolCalls;
+          // Don't yield the raw tool_calls chunk here — we emit one combined
+          // chunk after persisting the assistant message so consumers don't
+          // double-render.
+        } else if (chunk.type === 'thinking') {
+          yield chunk;
+        } else if (chunk.type === 'usage') {
+          lastUsage = chunk.usage;
+          yield chunk;
+        }
+      }
+
+      // If we got tool calls, execute them and loop
+      if (detectedToolCalls && detectedToolCalls.length > 0) {
+        // Add assistant message with tool calls (append-only, preserves prefix cache)
+        currentMessages.push({
+          role: 'assistant',
+          content: fullContent || '',
+          tool_calls: detectedToolCalls,
+        });
+
+        // Execute each tool and emit a single structured tool_calls chunk
+        // describing the dispatch plan. Per-tool results are appended to
+        // currentMessages but NOT yielded as additional chunks — consumers
+        // that want to display result text can introspect tool_calls[*].result
+        // on a synthesized combined chunk, OR we re-emit one tool_calls chunk
+        // at the end with results attached. We choose the simpler approach:
+        // emit ONE tool_calls chunk per iteration with all the calls; results
+        // are surfaced through the next assistant turn's text content.
+        //
+        // v3.0.0 Tool-call Repair (Reasonix P2): on tool failure we APPEND
+        // a `[TOOL_REPAIR]` system message rather than mutating the existing
+        // tool_call message. This preserves the upstream cache prefix —
+        // re-writing an earlier message would shift the prefix by N bytes
+        // and bust the cache for every subsequent round.
+        for (const toolCall of detectedToolCalls) {
+          const tool = this.tools.get(toolCall.function.name);
+
+          if (!tool) {
+            const errMsg = `Tool "${toolCall.function.name}" not found`;
+            // Repair message — explain the failure and tell the model to
+            // try a different tool. Cache-safe because we are APPENDING,
+            // never modifying existing messages.
+            currentMessages.push({
+              role: 'system',
+              content: `[TOOL_REPAIR] Previous tool_call "${toolCall.function.name}" (id=${toolCall.id}) failed: ${errMsg}. Available tools: ${[...this.tools.keys()].join(', ')}.`,
+            });
+            currentMessages.push({
+              role: 'tool',
+              content: errMsg,
+              tool_call_id: toolCall.id,
+              name: toolCall.function.name,
+            });
+            continue;
+          }
+
+          try {
+            const params = JSON.parse(toolCall.function.arguments || '{}');
+            const result = await tool.execute(params);
+            const output = result.success ? (result.output || JSON.stringify(result.data)) : (result.error || 'Tool failed');
+
+            currentMessages.push({
+              role: 'tool',
+              content: output,
+              tool_call_id: toolCall.id,
+              name: toolCall.function.name,
+            });
+
+            if (!result.success) {
+              // Soft failure: tool returned success=false. Same repair pattern.
+              currentMessages.push({
+                role: 'system',
+                content: `[TOOL_REPAIR] Tool "${toolCall.function.name}" returned success=false: ${output}. Consider correcting the arguments and retrying.`,
+              });
+            }
+          } catch (error: any) {
+            const errMsg = `Error: ${error.message}`;
+            currentMessages.push({
+              role: 'tool',
+              content: errMsg,
+              tool_call_id: toolCall.id,
+              name: toolCall.function.name,
+            });
+            // Hard failure: tool.execute threw. Repair message so the model
+            // can see the failure next round and adjust (e.g. fix a path
+            // typo, retry without the optional arg).
+            currentMessages.push({
+              role: 'system',
+              content: `[TOOL_REPAIR] Tool "${toolCall.function.name}" threw an exception: ${errMsg}. Inspect the arguments and retry with a corrected call.`,
+            });
+          }
+        }
+
+        // Emit one tool_calls chunk for this iteration. TUI renders this as a
+        // "tool invoked" entry. Per-tool outputs are summarized on the next
+        // assistant turn.
+        yield { type: 'tool_calls', toolCalls: detectedToolCalls };
+        continue;
+      }
+
+      // No tool calls - done. Return final response (with usage if we have it).
+      return {
+        content: fullContent,
+        role: 'assistant',
+        usage: lastUsage,
+      };
+    }
+
+    return { content: '[Agent loop exceeded maximum iterations]', role: 'assistant' };
+  }
+
+  private truncateArgs(args: string): string {
+    try {
+      const obj = JSON.parse(args || '{}');
+      const entries = Object.entries(obj);
+      if (entries.length === 0) return '';
+      return entries.map(([k, v]) => {
+        const val = typeof v === 'string' && v.length > 50 ? v.slice(0, 50) + '...' : v;
+        return `${k}: ${JSON.stringify(val)}`;
+      }).join(', ');
+    } catch {
+      return args.length > 80 ? args.slice(0, 80) + '...' : args;
+    }
+  }
+
+  private getNoKeyMessage(): string {
+    return [
+      '❌ 未配置 API Key，无法调用 AI。',
+      '',
+      '请先运行: gfcode init',
+      '',
+      '或设置环境变量:',
+      '  export SILICONFLOW_API_KEY="sk-..."',
+      '  export OPENAI_API_KEY="sk-..."',
+      '  export DEEPSEEK_API_KEY="sk-..."',
+    ].join('\n');
+  }
+}
+
+export type { LLMProvider, ProviderConfig, StreamChunk } from './provider.js';
+export { OpenAIProvider } from './openai.js';
+export { AnthropicProvider } from './anthropic.js';
+export { GeminiProvider } from './gemini.js';