/**
 * Lifecycle manager for a standalone llama.cpp router-mode `llama-server` process, used by the
 * VISION_API="llama-server" strand (see VisionAnalyzer.ts). Router mode auto-discovers GGUF
 * models from the local HF cache and loads/unloads them on demand per request, entirely isolated
 * from any agent-model process (LM Studio Bionic, Unsloth Studio) -- live-verified 2026-10-01 that
 * loading/using a vision model through a separate router-mode process never evicts an agent model
 * loaded elsewhere, which is the whole reason this strand exists. See
 * made-for-bionic/planning/made-for-preset-rollout-plan.md for the root-cause writeup of the
 * eviction behavior this works around.
 */
import fs from "fs";
import os from "os";
import path from "path";
import { spawn } from "child_process";
import { getLogsDir } from "../core-bundle.mjs";

const HEALTH_POLL_INTERVAL_MS = 1_000;
const HEALTH_POLL_TIMEOUT_MS = 30_000;
const HEALTH_FETCH_TIMEOUT_MS = 2_000;

let activePort: number | null = null;
let idleStopTimer: NodeJS.Timeout | null = null;

function cancelIdleStop(): void {
  if (idleStopTimer !== null) {
    clearTimeout(idleStopTimer);
    idleStopTimer = null;
  }
}

function normalizePath(value: string): string {
  const trimmed = value.trim();
  if (!trimmed) return "";
  return trimmed.startsWith("~") ? path.join(os.homedir(), trimmed.slice(1)) : trimmed;
}

function pidFilePath(): string {
  return path.join(getLogsDir(), "llama-server-router.pid");
}

function logFilePath(): string {
  return path.join(getLogsDir(), "llama-server-router.log");
}

function presetFilePath(): string {
  return path.join(getLogsDir(), "llama-server-router-preset.ini");
}

function isProcessAlive(pid: number): boolean {
  try {
    process.kill(pid, 0);
    return true;
  } catch {
    return false;
  }
}

function readPid(): number | null {
  try {
    const raw = fs.readFileSync(pidFilePath(), "utf-8").trim();
    const pid = parseInt(raw, 10);
    return Number.isFinite(pid) && pid > 0 ? pid : null;
  } catch {
    return null;
  }
}

function fetchWithTimeout(url: string, timeoutMs: number): Promise<Response> {
  const controller = new AbortController();
  const timer = setTimeout(() => controller.abort(), timeoutMs);
  return fetch(url, { signal: controller.signal }).finally(() => clearTimeout(timer));
}

async function isHealthy(host: string, port: number): Promise<boolean> {
  try {
    const res = await fetchWithTimeout(`http://${host}:${port}/v1/models`, HEALTH_FETCH_TIMEOUT_MS);
    return res.ok;
  } catch {
    return false;
  }
}

async function pollHealth(host: string, port: number): Promise<boolean> {
  const deadline = Date.now() + HEALTH_POLL_TIMEOUT_MS;
  while (Date.now() < deadline) {
    if (await isHealthy(host, port)) return true;
    await new Promise<void>((resolve) => setTimeout(resolve, HEALTH_POLL_INTERVAL_MS));
  }
  return false;
}

export interface LlamaServerRouterConfig {
  host: string;
  port: number;
  binaryPath: string;
  /**
   * --ctx-size applied to every router-spawned child model via a `[*]` preset section. Without
   * this, llama.cpp router mode defaults each auto-loaded model to its own native context length
   * -- live-verified 2026-10-01: Qwen3-VL-8B-Instruct's native 262144 context blew up to ~45GB
   * VRAM with no cap; capped at 8192 it used ~7.3GB for the same model. Configured via
   * src/config.ts's llamaServerCtxSize / mcp/config.ts's LLAMA_SERVER_CTX_SIZE, never hardcoded here.
   */
  ctxSize: number;
  /** --models-max passed to the router server. Configured via src/config.ts's llamaServerModelsMax / mcp/config.ts's LLAMA_SERVER_MODELS_MAX. */
  modelsMax: number;
  /** Minutes of inactivity before an owned router process is stopped again, 0 disables idle shutdown. Configured via src/config.ts's llamaServerIdleTtlMinutes / mcp/config.ts's LLAMA_SERVER_IDLE_TTL_MINUTES. */
  idleTtlMinutes: number;
  /**
   * Registers an absolute-path GGUF model (not an hf-repo id) as a named custom preset section,
   * so the router can serve it without relying on HF-cache auto-discovery. Only set when
   * VISION_API=llama-server's modelKey is an absolute .gguf file path (see ensureLlamaServerVisionModelReady).
   */
  localModel?: {
    name: string;
    modelPath: string;
    mmprojPath: string;
  };
}

function buildPresetIni(ctxSize: number, localModel: LlamaServerRouterConfig["localModel"]): string {
  let ini = `[*]\nctx-size = ${ctxSize}\n`;
  if (localModel) {
    ini += `\n[${localModel.name}]\nmodel = ${localModel.modelPath}\nmmproj = ${localModel.mmprojPath}\n`;
  }
  return ini;
}

async function listRouterModelIds(host: string, port: number): Promise<string[] | null> {
  try {
    const res = await fetchWithTimeout(`http://${host}:${port}/v1/models`, HEALTH_FETCH_TIMEOUT_MS);
    if (!res.ok) return null;
    const data = (await res.json()) as any;
    const models: any[] = Array.isArray(data?.data) ? data.data : [];
    return models.map((m) => String(m?.id ?? "").trim()).filter(Boolean);
  } catch {
    return null;
  }
}

/**
 * Ensures a router-mode llama-server is listening at host:port, spawning it if nothing answers
 * there yet. Idempotent and safe to call before every request -- a already-healthy endpoint
 * (ours or adopted) is trusted as-is, since router mode's own model management is independent of
 * which process started the server. Reschedules this process's own idle-shutdown timer on every
 * call (see releaseIdleLlamaServerRouter()).
 */
export async function ensureLlamaServerRouterRunning(config: LlamaServerRouterConfig): Promise<void> {
  const { host, port, ctxSize, modelsMax, idleTtlMinutes, localModel } = config;
  cancelIdleStop();
  activePort = port;
  const binaryPath = normalizePath(config.binaryPath);
  if (!binaryPath) {
    throw new Error("LLAMA_SERVER_BINARY is not configured.");
  }
  if (!fs.existsSync(binaryPath)) {
    throw new Error(`llama-server binary does not exist: ${binaryPath}`);
  }

  if (await isHealthy(host, port)) {
    if (localModel) {
      const registeredIds = await listRouterModelIds(host, port);
      const alreadyRegistered = registeredIds !== null && registeredIds.includes(localModel.name);
      if (!alreadyRegistered) {
        // The running router's preset doesn't know this local model yet -- router mode has no
        // hot-reload endpoint for --models-preset, so the only way to register a new local model
        // is to rewrite the ini and restart the process we ourselves own.
        await stopLlamaServerRouter(port);
      }
    }
    if (await isHealthy(host, port)) {
      scheduleIdleStop(port, idleTtlMinutes);
      return;
    }
  }

  const stalePid = readPid();
  if (stalePid !== null) {
    if (isProcessAlive(stalePid)) {
      throw new Error(
        `Port ${port} is not responding, but a previously spawned llama-server router (PID ${stalePid}) is still alive. Check logs: ${logFilePath()}`
      );
    }
    try {
      fs.unlinkSync(pidFilePath());
    } catch {}
  }

  fs.mkdirSync(getLogsDir(), { recursive: true });
  fs.writeFileSync(presetFilePath(), buildPresetIni(ctxSize, localModel));

  const args = [
    "--host", host,
    "--port", String(port),
    "--models-max", String(modelsMax),
    "--models-preset", presetFilePath(),
  ];

  const logStream = fs.createWriteStream(logFilePath(), { flags: "a" });
  const child = spawn(binaryPath, args, {
    detached: true,
    stdio: ["ignore", "pipe", "pipe"],
    env: { ...process.env, HOME: os.homedir() },
  });
  child.stdout?.pipe(logStream);
  child.stderr?.pipe(logStream);
  child.once("error", (error) => {
    try {
      fs.appendFileSync(logFilePath(), `[llama-server-router] process error: ${error.message}\n`);
    } catch {}
  });
  child.unref();

  if (!child.pid) {
    throw new Error(`Failed to spawn llama-server router process. Check logs: ${logFilePath()}`);
  }
  fs.writeFileSync(pidFilePath(), String(child.pid));

  const ready = await pollHealth(host, port);
  if (!ready) {
    throw new Error(
      `llama-server router did not become healthy within ${HEALTH_POLL_TIMEOUT_MS / 1000}s on ${host}:${port}. Check logs: ${logFilePath()}`
    );
  }

  scheduleIdleStop(port, idleTtlMinutes);
}

/** 0 disables idle shutdown entirely (process stays until the host process itself exits). */
function scheduleIdleStop(port: number, idleTtlMinutes: number): void {
  cancelIdleStop();
  if (idleTtlMinutes <= 0) return;
  idleStopTimer = setTimeout(() => {
    idleStopTimer = null;
    if (activePort !== port) return;
    void stopLlamaServerRouter(port).catch((error) => {
      try {
        fs.appendFileSync(logFilePath(), `[llama-server-router] failed to stop idle router: ${error instanceof Error ? error.message : String(error)}\n`);
      } catch {}
    });
  }, idleTtlMinutes * 60_000);
  idleStopTimer.unref();
}

/** Stops the router process this plugin itself spawned (tracked via the PID file); a no-op if nothing is running. */
export async function stopLlamaServerRouter(port: number): Promise<void> {
  cancelIdleStop();
  const pid = readPid();
  if (pid === null) return;
  if (isProcessAlive(pid)) {
    try {
      process.kill(pid, "SIGTERM");
    } catch {}
  }
  try {
    fs.unlinkSync(pidFilePath());
  } catch {}
  if (activePort === port) activePort = null;
}

/**
 * Lifecycle manager for a standalone llama.cpp router-mode `llama-server` process, used by the
 * VISION_API="llama-server" strand (see VisionAnalyzer.ts). Router mode auto-discovers GGUF
 * models from the local HF cache and loads/unloads them on demand per request, entirely isolated
 * from any agent-model process (LM Studio Bionic, Unsloth Studio) -- live-verified 2026-10-01 that
 * loading/using a vision model through a separate router-mode process never evicts an agent model
 * loaded elsewhere, which is the whole reason this strand exists. See
 * made-for-bionic/planning/made-for-preset-rollout-plan.md for the root-cause writeup of the
 * eviction behavior this works around.
 */
import fs from "fs";
import os from "os";
import path from "path";
import { spawn } from "child_process";
import { getLogsDir } from "../core-bundle.mjs";

const HEALTH_POLL_INTERVAL_MS = 1_000;
const HEALTH_POLL_TIMEOUT_MS = 30_000;
const HEALTH_FETCH_TIMEOUT_MS = 2_000;

let activePort: number | null = null;
let idleStopTimer: NodeJS.Timeout | null = null;

function cancelIdleStop(): void {
  if (idleStopTimer !== null) {
    clearTimeout(idleStopTimer);
    idleStopTimer = null;
  }
}

function normalizePath(value: string): string {
  const trimmed = value.trim();
  if (!trimmed) return "";
  return trimmed.startsWith("~") ? path.join(os.homedir(), trimmed.slice(1)) : trimmed;
}

function pidFilePath(): string {
  return path.join(getLogsDir(), "llama-server-router.pid");
}

function logFilePath(): string {
  return path.join(getLogsDir(), "llama-server-router.log");
}

function presetFilePath(): string {
  return path.join(getLogsDir(), "llama-server-router-preset.ini");
}

function isProcessAlive(pid: number): boolean {
  try {
    process.kill(pid, 0);
    return true;
  } catch {
    return false;
  }
}

function readPid(): number | null {
  try {
    const raw = fs.readFileSync(pidFilePath(), "utf-8").trim();
    const pid = parseInt(raw, 10);
    return Number.isFinite(pid) && pid > 0 ? pid : null;
  } catch {
    return null;
  }
}

function fetchWithTimeout(url: string, timeoutMs: number): Promise<Response> {
  const controller = new AbortController();
  const timer = setTimeout(() => controller.abort(), timeoutMs);
  return fetch(url, { signal: controller.signal }).finally(() => clearTimeout(timer));
}

async function isHealthy(host: string, port: number): Promise<boolean> {
  try {
    const res = await fetchWithTimeout(`http://${host}:${port}/v1/models`, HEALTH_FETCH_TIMEOUT_MS);
    return res.ok;
  } catch {
    return false;
  }
}

async function pollHealth(host: string, port: number): Promise<boolean> {
  const deadline = Date.now() + HEALTH_POLL_TIMEOUT_MS;
  while (Date.now() < deadline) {
    if (await isHealthy(host, port)) return true;
    await new Promise<void>((resolve) => setTimeout(resolve, HEALTH_POLL_INTERVAL_MS));
  }
  return false;
}

export interface LlamaServerRouterConfig {
  host: string;
  port: number;
  binaryPath: string;
  /**
   * --ctx-size applied to every router-spawned child model via a `[*]` preset section. Without
   * this, llama.cpp router mode defaults each auto-loaded model to its own native context length
   * -- live-verified 2026-10-01: Qwen3-VL-8B-Instruct's native 262144 context blew up to ~45GB
   * VRAM with no cap; capped at 8192 it used ~7.3GB for the same model. Configured via
   * src/config.ts's llamaServerCtxSize / mcp/config.ts's LLAMA_SERVER_CTX_SIZE, never hardcoded here.
   */
  ctxSize: number;
  /** --models-max passed to the router server. Configured via src/config.ts's llamaServerModelsMax / mcp/config.ts's LLAMA_SERVER_MODELS_MAX. */
  modelsMax: number;
  /** Minutes of inactivity before an owned router process is stopped again, 0 disables idle shutdown. Configured via src/config.ts's llamaServerIdleTtlMinutes / mcp/config.ts's LLAMA_SERVER_IDLE_TTL_MINUTES. */
  idleTtlMinutes: number;
  /**
   * Registers an absolute-path GGUF model (not an hf-repo id) as a named custom preset section,
   * so the router can serve it without relying on HF-cache auto-discovery. Only set when
   * VISION_API=llama-server's modelKey is an absolute .gguf file path (see ensureLlamaServerVisionModelReady).
   */
  localModel?: {
    name: string;
    modelPath: string;
    mmprojPath: string;
  };
}

function buildPresetIni(ctxSize: number, localModel: LlamaServerRouterConfig["localModel"]): string {
  let ini = `[*]\nctx-size = ${ctxSize}\n`;
  if (localModel) {
    ini += `\n[${localModel.name}]\nmodel = ${localModel.modelPath}\nmmproj = ${localModel.mmprojPath}\n`;
  }
  return ini;
}

async function listRouterModelIds(host: string, port: number): Promise<string[] | null> {
  try {
    const res = await fetchWithTimeout(`http://${host}:${port}/v1/models`, HEALTH_FETCH_TIMEOUT_MS);
    if (!res.ok) return null;
    const data = (await res.json()) as any;
    const models: any[] = Array.isArray(data?.data) ? data.data : [];
    return models.map((m) => String(m?.id ?? "").trim()).filter(Boolean);
  } catch {
    return null;
  }
}

/**
 * Ensures a router-mode llama-server is listening at host:port, spawning it if nothing answers
 * there yet. Idempotent and safe to call before every request -- a already-healthy endpoint
 * (ours or adopted) is trusted as-is, since router mode's own model management is independent of
 * which process started the server. Reschedules this process's own idle-shutdown timer on every
 * call (see releaseIdleLlamaServerRouter()).
 */
export async function ensureLlamaServerRouterRunning(config: LlamaServerRouterConfig): Promise<void> {
  const { host, port, ctxSize, modelsMax, idleTtlMinutes, localModel } = config;
  cancelIdleStop();
  activePort = port;
  const binaryPath = normalizePath(config.binaryPath);
  if (!binaryPath) {
    throw new Error("LLAMA_SERVER_BINARY is not configured.");
  }
  if (!fs.existsSync(binaryPath)) {
    throw new Error(`llama-server binary does not exist: ${binaryPath}`);
  }

  if (await isHealthy(host, port)) {
    if (localModel) {
      const registeredIds = await listRouterModelIds(host, port);
      const alreadyRegistered = registeredIds !== null && registeredIds.includes(localModel.name);
      if (!alreadyRegistered) {
        // The running router's preset doesn't know this local model yet -- router mode has no
        // hot-reload endpoint for --models-preset, so the only way to register a new local model
        // is to rewrite the ini and restart the process we ourselves own.
        await stopLlamaServerRouter(port);
      }
    }
    if (await isHealthy(host, port)) {
      scheduleIdleStop(port, idleTtlMinutes);
      return;
    }
  }

  const stalePid = readPid();
  if (stalePid !== null) {
    if (isProcessAlive(stalePid)) {
      throw new Error(
        `Port ${port} is not responding, but a previously spawned llama-server router (PID ${stalePid}) is still alive. Check logs: ${logFilePath()}`
      );
    }
    try {
      fs.unlinkSync(pidFilePath());
    } catch {}
  }

  fs.mkdirSync(getLogsDir(), { recursive: true });
  fs.writeFileSync(presetFilePath(), buildPresetIni(ctxSize, localModel));

  const args = [
    "--host", host,
    "--port", String(port),
    "--models-max", String(modelsMax),
    "--models-preset", presetFilePath(),
  ];

  const logStream = fs.createWriteStream(logFilePath(), { flags: "a" });
  const child = spawn(binaryPath, args, {
    detached: true,
    stdio: ["ignore", "pipe", "pipe"],
    env: { ...process.env, HOME: os.homedir() },
  });
  child.stdout?.pipe(logStream);
  child.stderr?.pipe(logStream);
  child.once("error", (error) => {
    try {
      fs.appendFileSync(logFilePath(), `[llama-server-router] process error: ${error.message}\n`);
    } catch {}
  });
  child.unref();

  if (!child.pid) {
    throw new Error(`Failed to spawn llama-server router process. Check logs: ${logFilePath()}`);
  }
  fs.writeFileSync(pidFilePath(), String(child.pid));

  const ready = await pollHealth(host, port);
  if (!ready) {
    throw new Error(
      `llama-server router did not become healthy within ${HEALTH_POLL_TIMEOUT_MS / 1000}s on ${host}:${port}. Check logs: ${logFilePath()}`
    );
  }

  scheduleIdleStop(port, idleTtlMinutes);
}

/** 0 disables idle shutdown entirely (process stays until the host process itself exits). */
function scheduleIdleStop(port: number, idleTtlMinutes: number): void {
  cancelIdleStop();
  if (idleTtlMinutes <= 0) return;
  idleStopTimer = setTimeout(() => {
    idleStopTimer = null;
    if (activePort !== port) return;
    void stopLlamaServerRouter(port).catch((error) => {
      try {
        fs.appendFileSync(logFilePath(), `[llama-server-router] failed to stop idle router: ${error instanceof Error ? error.message : String(error)}\n`);
      } catch {}
    });
  }, idleTtlMinutes * 60_000);
  idleStopTimer.unref();
}

/** Stops the router process this plugin itself spawned (tracked via the PID file); a no-op if nothing is running. */
export async function stopLlamaServerRouter(port: number): Promise<void> {
  cancelIdleStop();
  const pid = readPid();
  if (pid === null) return;
  if (isProcessAlive(pid)) {
    try {
      process.kill(pid, "SIGTERM");
    } catch {}
  }
  try {
    fs.unlinkSync(pidFilePath());
  } catch {}
  if (activePort === port) activePort = null;
}