Function buildCompactSystemPrompt

bin/smallcode.js:1925–1981

(taskType, messages)

// loop. A stable system prompt lets llama.cpp reuse its KV cache across turns.
// Set SMALLCODE_CACHE_SPLIT=false to revert to the old behaviour.
function buildCompactSystemPrompt(taskType, messages) {
  const cacheSplit = process.env.SMALLCODE_CACHE_SPLIT !== 'false'; // default: true
  const os = process.platform === 'win32' ? 'Windows' : process.platform === 'darwin' ? 'macOS' : 'Linux';
  const osHint = process.platform === 'win32' ? '\nUse "dir" not "ls", "type" not "cat". No bash-only commands.' : '';

  // Bootstrap detection (Feature 11): compact project summary at the top so
  // the model knows its runtime, build/test commands, and entry point without
  // burning tool calls on discovery.
  let bootstrapLine = '';
  try {
    if (_bootstrapDetector) {
      const raw = _bootstrapDetector.formatForPrompt();
      // Prefix as a workspace hint so the model treats it as tool context,
      // not as a project description to regurgitate when users ask questions.
      // Without this framing models answer "tell me about the project" from
      // the bootstrap summary instead of reading README.md.
      if (raw) bootstrapLine = raw.replace('\n\nProject:', '\n\nWorkspace context:');
    }
  } catch {}

  let prompt = `You are SmallCode, a coding agent. Working directory: ${process.cwd()}
OS: ${os}${osHint}${bootstrapLine}

Rules: Use patch for edits (not full rewrites). Prefer compound tools. Be concise. ACT immediately — do not ask for confirmation unless the task is genuinely ambiguous. If asked to read a file, read it. If asked to create something, create it. If asked about the project, read README.md or relevant files — do not answer from the workspace context line above.

CRITICAL — large file rule: write_file calls are limited to 60 lines / ~8KB. llama.cpp's JSON parser crashes on larger tool calls. For any file over 60 lines: (1) write_file with just the skeleton (imports + empty stubs), then (2) use multiple patch calls to fill in each function/section. Never put more than 60 lines in a single write_file content field.`;

  // Only add tool-use instructions for tasks that need tools
  if (taskType !== 'explanation') {
    prompt += `\nUse graph_search/explain_symbol for "how does X work" questions. Use list_projects for workspace overview.`;
  }

  // Only add BoneScript for backend tasks
  if (taskType === 'backend') {
    prompt += `\n\nFor Node.js backends: write a .bone file → bone_check → bone_compile. Don't hand-write routes.`;
  }

  // Web research tools — only advertise when web browsing is enabled. Small
  // models trust this prose over the raw `tools` array and otherwise refuse
  // research tasks ("my tools are for code files only") even though the web
  // tools are available (issue #58).
  if (taskType !== 'explanation' && String(process.env.SMALLCODE_WEB_BROWSE).toLowerCase() === 'true') {
    prompt += `\n\nWEB RESEARCH — you have live internet access: web_search (query → results with URLs) and web_fetch (URL → page text). Use them for research / look-up / current-info tasks; report a short summary with source URL(s). Do NOT write scripts to fetch the web.`;
  }

  if (cacheSplit) {
    // Dynamic context goes into a separate [CONTEXT] user message — see
    // buildDynamicContext(). Plan + plugins stay here (system role = authoritative).
    prompt += getPluginPrompts() + getActivePlanContext() + getTestRunnerContext();
    return prompt;
  }

  // Legacy behavior: everything in the system prompt
  prompt += getMemoryContext(messages) + getSkillContext(messages) + getPluginPrompts() + getKnowledgeContext(messages) + getRagContext(messages) + getActivePlanContext() + getTestRunnerContext();

  return prompt;
}