Function chatCompletion

bin/model_client.js:17–93

* Make a chat completion request (non-streaming, for tool use). * @param {object} ctx - Shared context { config, conversationHistory, memoryStore, skillManager, pluginLoader, currentTaskType, tokenTracker, sessionStore, getAllTools, _fullscreenRef }

(ctx)

 * @param {object} ctx - Shared context { config, conversationHistory, memoryStore, skillManager, pluginLoader, currentTaskType, tokenTracker, sessionStore, getAllTools, _fullscreenRef }
 */
async function chatCompletion(ctx) {
  const { config, conversationHistory, tokenTracker, sessionStore } = ctx;
  const target = config.activeModelTarget || getModelTarget(config, 'default');
  const requestConfig = withModelTarget(config, target);
  const baseUrl = target.baseUrl;

  const systemMsg = {
    role: 'system',
    content: buildSystemPrompt(ctx),
  };

  try {
    const { extractImages, formatImagesForAPI, modelSupportsVision } = require('../src/session/images');
    const processedMessages = conversationHistory.map(msg => {
      if (msg.role !== 'user' || typeof msg.content !== 'string') return msg;
      const images = extractImages(msg.content, process.cwd());
      if (images.length === 0 || !modelSupportsVision(target.model)) return msg;
      return { ...msg, content: [{ type: 'text', text: msg.content }, ...formatImagesForAPI(images)] };
    });

    const _tools = ctx.getAllTools(config);
    // Collapse any mid-conversation system messages into a single leading one
    // so strict chat templates (Qwen3/Qwen3.5 under llama.cpp --jinja) don't
    // 400 with "System message must be at the beginning." See issue #62.
    const { consolidateSystemMessages } = require('../src/session/message_normalizer');
    const body = {
      model: target.model,
      messages: consolidateSystemMessages([systemMsg, ...processedMessages]),
      temperature: 0.1,
      max_tokens: 4096,
    };
    // Only include tools when there are tools to send — some endpoints (OpenWebUI)
    // error on an empty tools array rather than treating it as "no tools".
    if (_tools && _tools.length > 0) {
      body.tools = _tools;
    }

    const headers = buildAuthHeaders(requestConfig);

    const response = await fetch(`${baseUrl}/chat/completions`, {
      method: 'POST',
      headers,
      body: JSON.stringify(body),
    });

    if (!response.ok) {
      const err = await response.text();
      if (response.status >= 400 && response.status < 500) {
        await new Promise(r => setTimeout(r, 2000));
        try {
          const retry = await fetch(`${baseUrl}/chat/completions`, { method: 'POST', headers, body: JSON.stringify(body) });
          if (retry.ok) return await retry.json();
        } catch {}
      }
      // Redact the error response — providers sometimes echo the request
      // back, including the Authorization header value, when responding
      // with 401/403. Never print raw provider errors verbatim.
      console.log(`  \x1b[31m✗ API error ${response.status}: ${redactString(err.slice(0, 200))}\x1b[0m`);