Function chatCompletion

bin/smallcode.js:2128–2555

(config, messages)

// Make a chat completion request (non-streaming for tool use, streaming for final response)
async function chatCompletion(config, messages) {
  let target = config.activeModelTarget || getModelTarget(config, 'default');
  let requestConfig = withModelTarget(config, target);
  let baseUrl = target.baseUrl;
  const systemMsg = {
    role: 'system',
    content: buildCompactSystemPrompt(currentTaskType, messages),
  };

  try {
    // Strip ANSI escape codes from all message content before sending to model.
    // Thinking models (Qwen3, etc.) will reproduce ANSI codes they see in context,
    // causing corrupted bash commands like "find ... -\x1b[38;2mtype f".
    function stripAnsiFromMsg(msg) {
      if (!msg || typeof msg.content !== 'string') return msg;
      return { ...msg, content: msg.content.replace(/\x1b\[[0-9;]*[a-zA-Z]/g, '').replace(/\x1b\][^\x07]*\x07/g, '') };
    }
    const processedMessages = messages.map(stripAnsiFromMsg);

    // Cache-split (Feature #14): when SMALLCODE_CACHE_SPLIT=true, dynamic
    // context (memory, knowledge) is moved out of the system prompt and into
    // a [CONTEXT] block prepended to the latest user message. This keeps the
    // system prompt stable across turns so remote APIs with prefix caching
    // (Anthropic, OpenAI) get cache hits on the static portion.
    const dynamicCtx = buildDynamicContext(messages);
    if (dynamicCtx) {
      const lastIdx = processedMessages.reduce((last, m, i) => m.role === 'user' ? i : last, -1);
      if (lastIdx >= 0) {
        const lastMsg = processedMessages[lastIdx];
        if (typeof lastMsg.content === 'string') {
          processedMessages[lastIdx] = {
            ...lastMsg,
            content: dynamicCtx + lastMsg.content,
          };
        }
        // If last user message is multimodal (image array), prepend as first text element
        else if (Array.isArray(lastMsg.content)) {
          const firstText = lastMsg.content.find(c => c.type === 'text');
          if (firstText) {
            processedMessages[lastIdx] = {
              ...lastMsg,
              content: [
                { type: 'text', text: dynamicCtx + firstText.text },
                ...lastMsg.content.filter(c => c !== firstText),
              ],
            };
          }
        }
      }
    }

    // Transform messages with images into multimodal format
    // OPTIMIZATION: Only re-extract images from the LAST user message (the new one).
    // Older messages that had images already had their content consumed; re-reading
    // the image from disk on every call is both wasteful (disk I/O) and causes
    // context overflow (a 1MB PNG = ~330k base64 tokens sent on EVERY API call).
    const { extractImages, formatImagesForAPI, modelSupportsVision } = require('../src/session/images');
    const lastUserIdx = processedMessages.length > 0