mono/packages/kbot/cpp/orchestrator/presets.js

/**
 * orchestrator/presets.js — defaults for IPC integration tests (extend here as suites grow).
 *
 * Llama local runner (llama-basics.test.ts): OpenAI-compatible API at http://localhost:8888/v1,
 * router `ollama` + `base_url` override, model `default` (server picks loaded GGUF).
 */

import { dirname, resolve } from 'node:path';
import { fileURLToPath } from 'node:url';
import { spawn } from 'node:child_process';
import { existsSync } from 'node:fs';

import { probeTcpPort } from './test-commons.js';

const __dirname = dirname(fileURLToPath(import.meta.url));

export const platform = {
  isWin: process.platform === 'win32',
};

/** kbot/cpp root (parent of orchestrator/). */
export const paths = {
  orchestratorDir: __dirname,
  cppRoot: resolve(__dirname, '..'),
  /** Same as packages/kbot/cpp/scripts/run-7b.sh — llama-server on :8888 */
  run7bScript: resolve(__dirname, '../scripts/run-7b.sh'),
};

/** Dist binary name for the current OS. */
export function exeName() {
  return platform.isWin ? 'kbot.exe' : 'kbot';
}

/** Absolute path to kbot binary given orchestrator/ directory (where test-ipc.mjs lives). */
export function distExePath(orchestratorDir) {
  return resolve(orchestratorDir, '..', 'dist', exeName());
}

/** UDS / TCP listen argument passed to `kbot worker --uds <arg>`. */
export const uds = {
  tcpPort: 4001,
  unixPath: '/tmp/kbot-test-ipc.sock',
  /** Value for `--uds` on this OS (Windows: port string; Unix: socket path). */
  workerArg() {
    return platform.isWin ? String(this.tcpPort) : this.unixPath;
  },
  /** Options for `net.connect` to reach the worker. */
  connectOpts(cppUdsArg) {
    return platform.isWin
      ? { port: this.tcpPort, host: '127.0.0.1' }
      : cppUdsArg;
  },
};

/** Millisecond timeouts — tune per step in new tests. */
export const timeouts = {
  ipcDefault: 5000,
  kbotAi: 180_000,
  /** Llama local arithmetic (same order of magnitude as kbot-ai). */
  llamaKbotAi: 180_000,
  /** Max wait for :8888 after spawning run-7b.sh (model load can be slow). */
  llamaServerStart: Number(process.env.KBOT_LLAMA_START_TIMEOUT_MS || 600_000),
  connectAttempts: 15,
  connectRetryMs: 400,
  postShutdownMs: 200,
};

export const router = {
  default: 'openrouter',
  fromEnv() {
    return process.env.KBOT_ROUTER || this.default;
  },
};

/**
 * Local llama.cpp HTTP server — mirrors tests/unit/llama-basics.test.ts (LLAMA_OPTS).
 * Uses router `ollama` so api_key resolves to dummy `ollama`; `base_url` points at :8888/v1.
 */
export const llama = {
  get port() {
    return Number(process.env.KBOT_LLAMA_PORT || 8888);
  },
  get host() {
    return process.env.KBOT_LLAMA_HOST || '127.0.0.1';
  },
  get baseURL() {
    return process.env.KBOT_LLAMA_BASE_URL || `http://localhost:${this.port}/v1`;
  },
  router: 'ollama',
  get model() {
    return process.env.KBOT_LLAMA_MODEL || 'default';
  },
  prompts: {
    /** Same idea as llama-basics completion tests. */
    add5_3: 'What is 5 + 3? Reply with just the number, nothing else.',
  },
};

/**
 * IPC payload for kbot-ai → local llama-server (OpenAI-compatible).
 * Pass `base_url` so LLMClient uses port 8888 instead of default ollama :11434.
 */
export function kbotAiPayloadLlamaLocal(overrides = {}) {
  const merged = {
    prompt: llama.prompts.add5_3,
    router: llama.router,
    model: llama.model,
    base_url: llama.baseURL,
    ...overrides,
  };
  merged.base_url = merged.base_url ?? merged.baseURL ?? llama.baseURL;
  delete merged.baseURL;
  return merged;
}

/** Stock prompts and assertions helpers for LLM smoke tests. */
export const prompts = {
  germanyCapital: 'What is the capital of Germany? Reply in one short sentence.',
};

/** Build `kbot-ai` IPC payload from env + presets (OpenRouter-friendly defaults). */
export function kbotAiPayloadFromEnv() {
  const payload = {
    prompt: process.env.KBOT_IPC_PROMPT || prompts.germanyCapital,
    router: router.fromEnv(),
  };
  if (process.env.KBOT_IPC_MODEL) {
    payload.model = process.env.KBOT_IPC_MODEL;
  }
  return payload;
}

/** True when using the default Germany prompt (for optional Berlin assertion). */
export function usingDefaultGermanyPrompt() {
  return !process.env.KBOT_IPC_PROMPT;
}

/**
 * If nothing listens on llama.port, optionally spawn `scripts/run-7b.sh` (requires `sh` on PATH, e.g. Git Bash on Windows).
 *
 * @param {{ autostart?: boolean, startTimeoutMs?: number }} [opts]
 * @returns {Promise<{ ok: boolean, alreadyRunning: boolean, started?: boolean, pid?: number }>}
 */
export async function ensureLlamaLocalServer(opts = {}) {
  const autostart = opts.autostart ?? true;
  const startTimeoutMs = opts.startTimeoutMs ?? timeouts.llamaServerStart;
  const host = llama.host;
  const port = llama.port;
  const scriptPath = paths.run7bScript;

  if (await probeTcpPort(host, port, 1500)) {
    return { ok: true, alreadyRunning: true };
  }

  if (!autostart) {
    throw new Error(
      `[llama] Nothing listening on ${host}:${port}. Start the server (e.g. sh scripts/run-7b.sh), or remove KBOT_IPC_LLAMA_AUTOSTART=0 to allow autostart`
    );
  }

  if (!existsSync(scriptPath)) {
    throw new Error(`[llama] Script missing: ${scriptPath}`);
  }

  console.log(`[llama] Port ${port} closed — starting ${scriptPath} (timeout ${startTimeoutMs}ms) …`);

  const child = spawn('sh', [scriptPath], {
    detached: true,
    stdio: 'ignore',
    cwd: dirname(scriptPath),
    env: { ...process.env },
  });
  child.unref();

  const deadline = Date.now() + startTimeoutMs;
  while (Date.now() < deadline) {
    if (await probeTcpPort(host, port, 1500)) {
      return { ok: true, alreadyRunning: false, started: true, pid: child.pid };
    }
    await new Promise((r) => setTimeout(r, 1500));
  }

  throw new Error(
    `[llama] Server did not open ${host}:${port} within ${startTimeoutMs}ms — check llama-server / GPU / model path`
  );
}