clawdie-ai/setup/llama-cpp.ts

/**
 * setup/llama-cpp.ts — Provision the llama-cpp jail for local inference.
 *
 * Creates the jail if missing, installs llama-cpp, and prepares
 * directories for models. Does not start llama-server by default.
 */
import { execSync, spawnSync } from 'child_process';

import {
  FEATURE_LLAMA_CPP,
  LLAMA_CPP_INTERNAL_DOMAIN,
  LOCAL_LLM_PROVIDER,
  LLAMA_CPP_JAIL_IP,
  RUNTIME_ID,
  SUBNET_BASE,
} from '../src/config.js';
import { logger } from '../src/logger.js';
import { loadPackageList, mountPkgCacheInJail } from './packages.js';
import { getPlatform } from './platform.js';
import { emitStatus } from './status.js';
import { maybeEnableTailscaleInJail } from './tailscale.js';
import {
  bastille,
  jailExists,
  detectFreeBSDRelease,
} from './bastille-helpers.js';

const LOG = 'logs/setup.log';
const MODELS_DIR = '/var/db/llm-models';

export async function run(_args: string[]): Promise<void> {
  const explicitJailName = (process.env.LLAMA_CPP_JAIL_NAME || '').trim();
  const hostname = LLAMA_CPP_INTERNAL_DOMAIN;
  const runBastille = (args: string[]) => bastille(...args);

  if (!FEATURE_LLAMA_CPP && LOCAL_LLM_PROVIDER !== 'llama_cpp') {
    emitStatus('SETUP_LLAMA_CPP', {
      STATUS: 'skipped',
      REASON: 'feature_disabled',
      LOG,
    });
    logger.info(
      'llama-cpp skipped — FEATURE_LLAMA_CPP disabled and provider not llama_cpp',
    );
    return;
  }

  if (getPlatform() !== 'freebsd') {
    emitStatus('SETUP_LLAMA_CPP', {
      STATUS: 'failed',
      ERROR: 'unsupported_platform',
      LOG,
    });
    process.exit(1);
  }

  try {
    const jailName = explicitJailName || `${RUNTIME_ID}-llamacpp`;

    const exists = jailExists(jailName);

    if (!exists) {
      const release = detectFreeBSDRelease();
      const gateway = process.env.WARDEN_GATEWAY || `${SUBNET_BASE}.1`;
      const bridge = process.env.WARDEN_BRIDGE || 'warden0';

      logger.info(
        { jailName, ip: LLAMA_CPP_JAIL_IP, release },
        'Creating llama-cpp jail',
      );
      const create = bastille(
        'create',
        // thin jail (no -T): keep model state in mounted storage, not a copied base.
        '-B',
        '-g',
        gateway,
        jailName,
        release,
        `${LLAMA_CPP_JAIL_IP}/24`,
        bridge,
      );
      if (!create.ok) {
        throw new Error(`bastille create failed: ${create.output}`);
      }
      bastille('config', jailName, 'set', 'host.hostname', hostname);
      bastille('restart', jailName);
    } else {
      logger.info(
        { jailName },
        'llama-cpp jail already exists, skipping creation',
      );
    }

    mountPkgCacheInJail(jailName);
    const packages = loadPackageList('llama-cpp-jail.txt');
    const pkg = bastille('pkg', jailName, 'install', '-y', ...packages);
    if (!pkg.ok) {
      logger.warn(
        { output: pkg.output },
        'llama-cpp package install had warnings',
      );
    }

    maybeEnableTailscaleInJail(runBastille, jailName, jailName);

    bastille('cmd', jailName, 'install', '-d', '-m', '755', MODELS_DIR);

    emitStatus('SETUP_LLAMA_CPP', {
      STATUS: 'success',
      JAIL_NAME: jailName,
      JAIL_IP: LLAMA_CPP_JAIL_IP,
      MODELS_DIR,
      NOTE: 'models_required',
      LOG,
    });
  } catch (error) {
    const message = error instanceof Error ? error.message : String(error);
    emitStatus('SETUP_LLAMA_CPP', {
      STATUS: 'failed',
      ERROR: message,
      LOG,
    });
    throw error;
  }
}