agent-termo 0.0.0-stage → 1.0.1

package/src/llm/client.ts

@@ -0,0 +1,192 @@
+// =============================================================================
+// Nvidia NIM LLM Client — Stepfun + OpenAI-compatible models
+// =============================================================================
+
+import OpenAI from 'openai';
+import type { ToolCall, ConversationMessage, ToolDef } from '../types/index.js';
+
+type NIMTool = OpenAI.Chat.ChatCompletionTool;
+
+export class NimClient {
+  private client: OpenAI;
+  private model: string;
+
+  constructor(_projectRoot?: string) {
+    const apiKey = process.env.NIM_API_KEY ?? '';
+    const baseUrl = process.env.NIM_BASE_URL ?? 'https://integrate.api.nvidia.com/v1';
+    this.model = process.env.NIM_MODEL ?? 'stepfun-ai/step-3.7-flash';
+    const timeout = parseInt(process.env.LLM_TIMEOUT ?? '120000', 10);
+
+    this.client = new OpenAI({
+      apiKey,
+      baseURL: baseUrl,
+      timeout,
+      defaultHeaders: { 'User-Agent': 'agent-termo/1.0' },
+    });
+  }
+
+  buildToolDefs(toolSchemas: ToolDef[]): NIMTool[] {
+    return toolSchemas.map((t) => ({
+      type: 'function' as const,
+      function: {
+        name: t.name,
+        description: t.description,
+        parameters: t.parameters as unknown as NIMTool['function']['parameters'],
+      },
+    })) as NIMTool[];
+  }
+
+  getModel(): string {
+    return this.model;
+  }
+
+  /**
+   * Streaming chat completion.
+   * Handles Stepfun's quirks: filters "null" reasoning chunks, strips reasoning field.
+   * Yields { content, toolCalls } at the end with clean text.
+   */
+  async *streamComplete(
+    messages: ConversationMessage[],
+    tools?: NIMTool[],
+    onChunk?: (text: string) => void,
+  ): AsyncGenerator<{ content: string; toolCalls: ToolCall[] }, void, unknown> {
+    const rawChunks: string[] = [];
+    const toolCallMap: Map<number, { id: string; name: string; arguments: string }> = new Map();
+
+    const stream = await this.client.chat.completions.create({
+      model: this.model,
+      messages: messages.map((m) => {
+        if (m.role === 'tool') {
+          return {
+            role: 'tool' as const,
+            content: m.content,
+            tool_call_id: m.id,
+          };
+        }
+        return {
+          role: m.role as 'system' | 'user' | 'assistant',
+          content: m.content,
+        };
+      }),
+      tools: tools,
+      temperature: 0.3,
+      max_tokens: 8192,
+      stream: true,
+    });
+
+    for await (const chunk of stream) {
+      const delta = chunk.choices[0]?.delta;
+
+      if (!delta) continue;
+
+      // Text delta — filter out "null" chunks (Stepfun's reasoning output)
+      const rawContent = delta.content ?? '';
+      // Check: if it's the literal string "null", it's reasoning noise — skip it
+      const isNoise = rawContent === 'null' || rawContent === 'undefined';
+      if (!isNoise && rawContent) {
+        rawChunks.push(rawContent);
+        onChunk?.(rawContent);
+      }
+
+      // Tool call deltas (accumulate across chunks via index)
+      if (delta.tool_calls) {
+        for (const tc of delta.tool_calls) {
+          const idx = tc.index ?? 0;
+          const prev = toolCallMap.get(idx);
+          if (!prev) {
+            toolCallMap.set(idx, {
+              id: tc.id ?? crypto.randomUUID(),
+              name: tc.function?.name ?? '',
+              arguments: tc.function?.arguments ?? '',
+            });
+          } else {
+            prev.arguments += tc.function?.arguments ?? '';
+            if (tc.id && !prev.id) prev.id = tc.id;
+          }
+        }
+      }
+    }
+
+    // For streaming: content is accumulated in rawChunks (null chunks already filtered)
+    // For non-streaming (complete() method): reasoning stripping happens in that method
+    const finalContent = rawChunks.join('');
+
+    // Extract tool calls
+    const toolCalls: ToolCall[] = Array.from(toolCallMap.values()).map((tc) => {
+      let args: Record<string, unknown> = {};
+      try {
+        args = JSON.parse(tc.arguments);
+      } catch {
+        args = { _raw: tc.arguments };
+      }
+      return { id: tc.id, name: tc.name, arguments: args };
+    });
+
+    yield { content: finalContent, toolCalls };
+  }
+
+  /**
+   * Non-streaming completion — for when streaming fails or for retries.
+   */
+  async complete(
+    messages: ConversationMessage[],
+    tools?: NIMTool[],
+  ): Promise<{ content: string; toolCalls: ToolCall[] }> {
+    const response = await this.client.chat.completions.create({
+      model: this.model,
+      messages: messages.map((m) => {
+        if (m.role === 'tool') {
+          return { role: 'tool' as const, content: m.content, tool_call_id: m.id };
+        }
+        return { role: m.role as 'system' | 'user' | 'assistant', content: m.content };
+      }),
+      tools: tools,
+      temperature: 0.3,
+      max_tokens: 8192,
+      stream: false,
+    });
+
+    const msg = response.choices[0]?.message;
+    // Strip reasoning noise
+    let content =
+      msg?.content?.replace(
+        /\[ Reasoning[^\]]*\][^\[]*\[\/ Reasoning\][^\n]*/gi,
+        '',
+      ).trim() ?? '';
+
+    const toolCalls: ToolCall[] = (msg?.tool_calls ?? []).map((tc) => {
+      let args: Record<string, unknown> = {};
+      try {
+        args = JSON.parse(tc.function.arguments ?? '{}');
+      } catch {
+        args = { _raw: tc.function.arguments };
+      }
+      return { id: tc.id, name: tc.function.name, arguments: args };
+    });
+
+    // If no tool calls but content looks like a tool call request, parse it
+    if (toolCalls.length === 0 && content.includes('"tool_calls"')) {
+      try {
+        const parsed = JSON.parse(content);
+        if (parsed.tool_calls?.length) {
+          toolCalls.push(...parsed.tool_calls.map((tc: { id?: string; name?: string; function?: { name?: string; arguments?: unknown } }) => ({
+            id: tc.id ?? crypto.randomUUID(),
+            name: tc.function?.name ?? tc.name ?? '',
+            arguments: typeof tc.function?.arguments === 'string'
+              ? JSON.parse(tc.function.arguments)
+              : tc.function?.arguments ?? {},
+          })));
+          // Remove JSON from content
+          const jsonStart = content.indexOf('{"tool_calls"');
+          if (jsonStart !== -1) content = content.slice(0, jsonStart).trim();
+        }
+      } catch {
+        // Not valid JSON — leave as-is
+      }
+    }
+
+    return { content, toolCalls };
+  }
+}
+
+export { type ToolDef } from '../types/index.js';

package/src/llm/tools.ts

@@ -0,0 +1,206 @@
+// =============================================================================
+// Tool Definitions (OpenAI function calling schema)
+// These are matched to the actual tools implemented in src/tools/
+// =============================================================================
+
+import type { ToolDef } from '../types/index.js';
+
+export const TOOL_DEFINITIONS: ToolDef[] = [
+  // ── File Operations ──────────────────────────────────────────────────────────
+  {
+    name: 'file:read',
+    description: 'Read the complete contents of a file. Use for examining existing code, config, or documentation.',
+    parameters: {
+      type: 'object',
+      properties: {
+        path: {
+          type: 'string',
+          description: 'File path (absolute or relative to project root)',
+        },
+        limit: {
+          type: 'number',
+          description: 'Maximum number of lines to read (0 = all)',
+          default: 0,
+        },
+        offset: {
+          type: 'number',
+          description: 'Line offset to start reading from (0 = start)',
+          default: 0,
+        },
+      },
+      required: ['path'],
+    },
+  },
+  {
+    name: 'file:write',
+    description: 'Write or overwrite a complete file. Use this to create new files or replace existing ones entirely.',
+    parameters: {
+      type: 'object',
+      properties: {
+        path: {
+          type: 'string',
+          description: 'File path to write (relative to project root)',
+        },
+        content: {
+          type: 'string',
+          description: 'Complete file content to write',
+        },
+      },
+      required: ['path', 'content'],
+    },
+  },
+  {
+    name: 'file:edit',
+    description: 'Edit an existing file by replacing an exact section. The old_string must appear verbatim in the file.',
+    parameters: {
+      type: 'object',
+      properties: {
+        path: {
+          type: 'string',
+          description: 'File to edit',
+        },
+        old_string: {
+          type: 'string',
+          description: 'Exact text to find and replace (must be unique in file)',
+        },
+        new_string: {
+          type: 'string',
+          description: 'Replacement text',
+        },
+      },
+      required: ['path', 'old_string', 'new_string'],
+    },
+  },
+  {
+    name: 'file:glob',
+    description: 'Find all files matching a glob pattern. Use to discover project structure.',
+    parameters: {
+      type: 'object',
+      properties: {
+        pattern: {
+          type: 'string',
+          description: 'Glob pattern (e.g. "src/**/*.ts", "**/*.json")',
+        },
+        cwd: {
+          type: 'string',
+          description: 'Root directory to search in (defaults to project root)',
+        },
+      },
+      required: ['pattern'],
+    },
+  },
+  {
+    name: 'grep',
+    description: 'Search for text or regex patterns inside files. Use to find function definitions, imports, etc.',
+    parameters: {
+      type: 'object',
+      properties: {
+        pattern: {
+          type: 'string',
+          description: 'Text or regex pattern to search for',
+        },
+        path: {
+          type: 'string',
+          description: 'File or directory to search in',
+        },
+        glob: {
+          type: 'string',
+          description: 'Only search files matching this glob pattern',
+        },
+        case_sensitive: {
+          type: 'boolean',
+          description: 'Case sensitive search (default true)',
+          default: true,
+        },
+      },
+      required: ['pattern', 'path'],
+    },
+  },
+  // ── Bash ───────────────────────────────────────────────────────────────────
+  {
+    name: 'bash:execute',
+    description: 'Execute a shell command and return its stdout/stderr output. Destructive commands (rm, dd) require user confirmation.',
+    parameters: {
+      type: 'object',
+      properties: {
+        command: {
+          type: 'string',
+          description: 'Shell command to execute',
+        },
+        cwd: {
+          type: 'string',
+          description: 'Working directory (defaults to project root)',
+        },
+        timeout: {
+          type: 'number',
+          description: 'Timeout in milliseconds (default 300000)',
+          default: 300000,
+        },
+      },
+      required: ['command'],
+    },
+  },
+  // ── Git ────────────────────────────────────────────────────────────────────
+  {
+    name: 'git:status',
+    description: 'Show the git working tree status (modified, staged, untracked files).',
+    parameters: {
+      type: 'object',
+      properties: {
+        cwd: {
+          type: 'string',
+          description: 'Repository root (default: project root)',
+        },
+      },
+    },
+  },
+  {
+    name: 'git:commit',
+    description: 'Stage and commit all changes with a conventional commit message.',
+    parameters: {
+      type: 'object',
+      properties: {
+        message: {
+          type: 'string',
+          description: 'Conventional commit message (e.g. "feat: add login page")',
+        },
+        cwd: {
+          type: 'string',
+          description: 'Repository root (default: project root)',
+        },
+      },
+      required: ['message'],
+    },
+  },
+  {
+    name: 'git:push',
+    description: 'Push the current branch to the remote repository.',
+    parameters: {
+      type: 'object',
+      properties: {
+        cwd: {
+          type: 'string',
+          description: 'Repository root (default: project root)',
+        },
+      },
+    },
+  },
+  {
+    name: 'git:log',
+    description: 'Show recent git commit history.',
+    parameters: {
+      type: 'object',
+      properties: {
+        cwd: {
+          type: 'string',
+          description: 'Repository root (default: project root)',
+        },
+        limit: {
+          type: 'number',
+          description: 'Number of commits to show (default: 20)',
+          default: 20,
+        },
+      },
+    },
+  },
+];