feat(mcplocal+mcpd): wake-recipe config + wake-task execution (v2 Stage 1)

First half of v2 — mcplocal can now declare a hibernating backend and
respond to a `wake` task by running a configured recipe. v2 Stage 2
will wire mcpd to dispatch the wake task before relaying inference.

Config (LlmProviderFileEntry):
- New \`wake\` block on a published provider:
    wake:
      type: http        # or: command
      url: ...           # http only
      method: POST       # http only, default POST
      headers: {...}     # http only
      body: ...          # http only
      command: ...       # command only
      args: [...]        # command only
      maxWaitSeconds: 60 # how long to poll isAvailable() after wake fires

Registrar (mcplocal):
- At publish time, providers with a wake recipe whose isAvailable()
  returns false report initialStatus=hibernating to mcpd. Without a
  wake recipe (legacy v1) or when already up, status stays active.
- handleWakeTask: runs the recipe (HTTP request OR child-process
  spawn), then polls isAvailable() up to maxWaitSeconds, sending a
  heartbeat each loop so mcpd's GC sweep doesn't time us out
  mid-boot. Reports { ok, ms } on success or { error } on
  timeout/recipe failure via the existing _provider-task/:id/result.
- Replaces the v1 stub that rejected wake tasks with "not implemented".

mcpd VirtualLlmService:
- RegisterProviderInput gains optional initialStatus ('active' |
  'hibernating'). The register/upsert path uses it for both new and
  reconnecting rows. Defaults to 'active' so v1 publishers still
  work unchanged.
- Provider-register route's coercer accepts the new field.

Tests: 3 new in registrar.test.ts cover initialStatus selection
(hibernating when wake configured + unavailable, active otherwise,
active when no wake even if unavailable). 8/8 registrar tests, 833/833
mcpd unchanged.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/mcpd/src/routes/virtual-llms.ts

@@ -150,6 +150,7 @@ function coerceProviderInput(raw: unknown): {
   tier?: string;
   description?: string;
   extraConfig?: Record<string, unknown>;
+  initialStatus?: 'active' | 'hibernating';
 } {
   if (raw === null || typeof raw !== 'object') {
     throw Object.assign(new Error('provider entry must be an object'), { statusCode: 400 });
@@ -170,5 +171,11 @@ function coerceProviderInput(raw: unknown): {
   if (o['extraConfig'] !== null && typeof o['extraConfig'] === 'object') {
     out.extraConfig = o['extraConfig'] as Record<string, unknown>;
   }
+  // Only accept the two values v2 actually defines. Anything else falls
+  // through to the service default (active) — matches v1 publishers that
+  // don't know about this field.
+  if (o['initialStatus'] === 'active' || o['initialStatus'] === 'hibernating') {
+    out.initialStatus = o['initialStatus'];
+  }
   return out;
 }

src/mcpd/src/services/virtual-llm.service.ts

@@ -37,6 +37,15 @@ export interface RegisterProviderInput {
   tier?: string;
   description?: string;
   extraConfig?: Record<string, unknown>;
+  /**
+   * Optional. Lets the publisher hint that the underlying backend is
+   * asleep — mcpd records the row as `hibernating` and will dispatch a
+   * `wake` task before any inference. Defaults to `active` (today's
+   * behavior). v2 publishers (mcplocal with a configured wake recipe)
+   * pass 'hibernating' when `LlmProvider.isAvailable()` returns false at
+   * publish time.
+   */
+  initialStatus?: 'active' | 'hibernating';
 }
 
 export interface RegisterResult {
@@ -112,6 +121,7 @@ export class VirtualLlmService implements IVirtualLlmService {
     const llms: Llm[] = [];
 
     for (const p of input.providers) {
+      const initialStatus = p.initialStatus ?? 'active';
       const existing = await this.repo.findByName(p.name);
       if (existing === null) {
         const created = await this.repo.create({
@@ -123,7 +133,7 @@ export class VirtualLlmService implements IVirtualLlmService {
           ...(p.extraConfig !== undefined ? { extraConfig: p.extraConfig } : {}),
           kind: 'virtual',
           providerSessionId: sessionId,
-          status: 'active',
+          status: initialStatus,
           lastHeartbeatAt: now,
           inactiveSince: null,
         });
@@ -156,7 +166,7 @@ export class VirtualLlmService implements IVirtualLlmService {
         ...(p.extraConfig !== undefined ? { extraConfig: p.extraConfig } : {}),
         kind: 'virtual',
         providerSessionId: sessionId,
-        status: 'active',
+        status: initialStatus,
         lastHeartbeatAt: now,
         inactiveSince: null,
       });

src/mcplocal/src/http/config.ts

@@ -80,8 +80,25 @@ export interface LlmProviderFileEntry {
    * Default: false — existing setups don't change behavior.
    */
   publish?: boolean;
+  /**
+   * Optional wake recipe for hibernating backends. When set, a provider
+   * whose `isAvailable()` returns false at registrar start time is
+   * published as `status=hibernating`. The next inference request that
+   * lands on the row triggers this recipe; mcplocal polls
+   * `isAvailable()` until it returns true (or times out) and then flips
+   * the row to active so mcpd can dispatch the queued inference.
+   *
+   * Two recipe types:
+   *   - `http`: POST to a URL (e.g. an external sleep/wake controller)
+   *   - `command`: spawn a shell command (e.g. `systemctl --user start vllm`)
+   */
+  wake?: WakeRecipe;
 }
 
+export type WakeRecipe =
+  | { type: 'http'; url: string; method?: 'GET' | 'POST'; headers?: Record<string, string>; body?: string; maxWaitSeconds?: number }
+  | { type: 'command'; command: string; args?: string[]; maxWaitSeconds?: number };
+
 export interface ProjectLlmOverride {
   model?: string;
   provider?: string;

src/mcplocal/src/main.ts

@@ -215,6 +215,7 @@ async function maybeStartVirtualLlmRegistrar(
       model: entry.model ?? entry.name,
     };
     if (entry.tier !== undefined) item.tier = entry.tier;
+    if (entry.wake !== undefined) item.wake = entry.wake;
     published.push(item);
   }
   if (published.length === 0) return null;

src/mcplocal/src/providers/registrar.ts

@@ -27,7 +27,9 @@ import http from 'node:http';
 import https from 'node:https';
 import { promises as fs } from 'node:fs';
 import { dirname } from 'node:path';
+import { spawn } from 'node:child_process';
 import type { LlmProvider, CompletionOptions } from './types.js';
+import type { WakeRecipe } from '../http/config.js';
 
 export interface RegistrarLogger {
   info: (msg: string) => void;
@@ -45,6 +47,13 @@ export interface RegistrarPublishedProvider {
   tier?: 'fast' | 'heavy';
   /** Optional human-readable description for `mcpctl get llm`. */
   description?: string;
+  /**
+   * Optional wake recipe for backends that hibernate. When provided AND
+   * `provider.isAvailable()` returns false at registrar start, the row is
+   * published with status=hibernating; on the first incoming `wake` task
+   * the registrar runs this recipe and waits for the backend to come up.
+   */
+  wake?: WakeRecipe;
 }
 
 export interface RegistrarOptions {
@@ -140,15 +149,28 @@ export class VirtualLlmRegistrar {
   }
 
   private async register(): Promise<void> {
-    const body: Record<string, unknown> = {
-      providers: this.opts.publishedProviders.map((p) => ({
+    // Decide initial status per provider. A provider with a wake recipe
+    // that's NOT currently available comes up as hibernating; otherwise
+    // active (today's behavior). isAvailable() is forgiving — any
+    // unexpected throw is treated as "not available" so a transient
+    // network blip during boot doesn't crash the registrar.
+    const providers = await Promise.all(this.opts.publishedProviders.map(async (p) => {
+      let initialStatus: 'active' | 'hibernating' = 'active';
+      if (p.wake !== undefined) {
+        let alive = false;
+        try { alive = await p.provider.isAvailable(); } catch { alive = false; }
+        if (!alive) initialStatus = 'hibernating';
+      }
+      return {
         name: p.provider.name,
         type: p.type,
         model: p.model,
         ...(p.tier !== undefined ? { tier: p.tier } : {}),
         ...(p.description !== undefined ? { description: p.description } : {}),
-      })),
+        initialStatus,
       };
+    }));
+    const body: Record<string, unknown> = { providers };
     if (this.sessionId !== null) body['providerSessionId'] = this.sessionId;
 
     const res = await postJson(
@@ -276,9 +298,51 @@ export class VirtualLlmRegistrar {
       void this.handleInferTask(task);
       return;
     }
-    // Wake tasks are reserved for v2 — acknowledge with an error so mcpd
-    // surfaces a clean failure rather than waiting forever.
-    void this.postResult(task.taskId, { error: 'wake task type not implemented in this client (v2)' });
+    if (task.kind === 'wake') {
+      void this.handleWakeTask(task);
+      return;
+    }
+  }
+
+  /**
+   * Run the configured wake recipe and poll the provider until it comes
+   * up. Sends a `{ status: 200, body: { ok: true } }` result on success;
+   * `{ error }` on timeout or recipe failure. While waiting, also bumps
+   * the heartbeat so mcpd's GC sweep doesn't decide we're stale mid-wake.
+   */
+  private async handleWakeTask(task: { kind: 'wake'; taskId: string; llmName: string }): Promise<void> {
+    const published = this.opts.publishedProviders.find((p) => p.provider.name === task.llmName);
+    if (published === undefined) {
+      await this.postResult(task.taskId, { error: `provider '${task.llmName}' not registered locally` });
+      return;
+    }
+    if (published.wake === undefined) {
+      await this.postResult(task.taskId, { error: `provider '${task.llmName}' has no wake recipe configured` });
+      return;
+    }
+
+    try {
+      await runWakeRecipe(published.wake);
+      // Poll isAvailable() until it comes up (or timeout). Heartbeat
+      // every poll tick so mcpd doesn't time us out while we're waiting
+      // on a slow boot.
+      const maxWaitMs = (published.wake.maxWaitSeconds ?? 60) * 1000;
+      const started = Date.now();
+      while (Date.now() - started < maxWaitMs) {
+        let alive = false;
+        try { alive = await published.provider.isAvailable(); } catch { alive = false; }
+        if (alive) {
+          await this.heartbeatOnce();
+          await this.postResult(task.taskId, { status: 200, body: { ok: true, ms: Date.now() - started } });
+          return;
+        }
+        await this.heartbeatOnce();
+        await new Promise((r) => setTimeout(r, 1500));
+      }
+      await this.postResult(task.taskId, { error: `provider '${task.llmName}' did not come up within ${String(maxWaitMs)}ms` });
+    } catch (err) {
+      await this.postResult(task.taskId, { error: `wake recipe failed: ${(err as Error).message}` });
+    }
   }
 
   private async handleInferTask(task: InferTask): Promise<void> {
@@ -373,6 +437,68 @@ function openAiStreamChunk(
   };
 }
 
+/**
+ * Execute a wake recipe. Returns when the recipe completes; throws if it
+ * fails. Doesn't itself poll for provider readiness — that's the caller's
+ * job (handleWakeTask polls isAvailable() with its own timeout).
+ *
+ * `http`: fires the configured request and considers any 2xx a success.
+ *   The remote service is expected to be a "wake controller" that returns
+ *   quickly; if the underlying boot takes minutes, the controller should
+ *   return 202 and the readiness poll catches up.
+ *
+ * `command`: spawns the binary with args, waits for exit. Non-zero exit
+ *   is treated as failure. stdout/stderr are discarded — the recipe's job
+ *   is to *trigger* a wake, not to produce output.
+ */
+async function runWakeRecipe(recipe: WakeRecipe): Promise<void> {
+  if (recipe.type === 'http') {
+    const u = new URL(recipe.url);
+    const driver = u.protocol === 'https:' ? https : http;
+    const method = recipe.method ?? 'POST';
+    const headers: Record<string, string> = { ...(recipe.headers ?? {}) };
+    const body = recipe.body;
+    if (body !== undefined) {
+      headers['Content-Length'] = String(Buffer.byteLength(body));
+    }
+    await new Promise<void>((resolve, reject) => {
+      const req = driver.request({
+        hostname: u.hostname,
+        port: u.port || (u.protocol === 'https:' ? 443 : 80),
+        path: u.pathname + u.search,
+        method,
+        headers,
+        timeout: 30_000,
+      }, (res) => {
+        const status = res.statusCode ?? 0;
+        // Drain so the socket can be reused/freed.
+        res.resume();
+        if (status >= 200 && status < 300) resolve();
+        else reject(new Error(`wake HTTP returned ${String(status)}`));
+      });
+      req.on('error', reject);
+      req.on('timeout', () => { req.destroy(); reject(new Error('wake HTTP timed out')); });
+      if (body !== undefined) req.write(body);
+      req.end();
+    });
+    return;
+  }
+  if (recipe.type === 'command') {
+    await new Promise<void>((resolve, reject) => {
+      const child = spawn(recipe.command, recipe.args ?? [], {
+        stdio: 'ignore',
+      });
+      child.on('error', reject);
+      child.on('exit', (code) => {
+        if (code === 0) resolve();
+        else reject(new Error(`wake command exited with code ${String(code)}`));
+      });
+    });
+    return;
+  }
+  throw new Error(`unknown wake recipe type`);
+}
+
 interface PostResponse { statusCode: number; body: string }
 
 /** Tiny JSON POST helper used by all of the registrar's mcpd calls. */

src/mcplocal/tests/registrar.test.ts

@@ -142,13 +142,17 @@ describe('VirtualLlmRegistrar', () => {
       const registerCall = fake.calls.find((c) => c.path === '/api/v1/llms/_provider-register');
       expect(registerCall).toBeDefined();
       expect(registerCall!.method).toBe('POST');
-      const body = JSON.parse(registerCall!.body) as { providers: Array<{ name: string; type: string; model: string; tier: string }> };
-      expect(body.providers).toEqual([{
+      const body = JSON.parse(registerCall!.body) as { providers: Array<Record<string, unknown>> };
+      expect(body.providers).toHaveLength(1);
+      expect(body.providers[0]).toMatchObject({
         name: 'vllm-local',
         type: 'openai',
         model: 'qwen',
         tier: 'fast',
-      }]);
+        // v2 always sends initialStatus; defaults to 'active' when no
+        // wake recipe is configured.
+        initialStatus: 'active',
+      });
       expect(registerCall!.headers['authorization']).toBe('Bearer tok-abc');
 
       // Sticky session id persisted.
@@ -219,6 +223,107 @@ describe('VirtualLlmRegistrar', () => {
     }
   });
 
+  // ── v2: hibernating + wake recipe ──
+
+  it('publishes initialStatus=hibernating when provider is unavailable AND wake is configured', async () => {
+    const fake = await startFakeServer();
+    try {
+      const sleeping: LlmProvider = {
+        name: 'vllm-local',
+        async complete() { throw new Error('not running'); },
+        async listModels() { return []; },
+        async isAvailable() { return false; },
+      };
+      const registrar = new VirtualLlmRegistrar({
+        mcpdUrl: fake.url,
+        token: 't',
+        publishedProviders: [{
+          provider: sleeping,
+          type: 'openai',
+          model: 'm',
+          wake: { type: 'http', url: 'http://localhost:9999/wake', maxWaitSeconds: 1 },
+        }],
+        sessionFilePath: join(tempDir, 'provider-session'),
+        log: silentLog(),
+        heartbeatIntervalMs: 60_000,
+      });
+      await registrar.start();
+      await new Promise((r) => setTimeout(r, 20));
+
+      const registerCall = fake.calls.find((c) => c.path === '/api/v1/llms/_provider-register');
+      const body = JSON.parse(registerCall!.body) as { providers: Array<{ initialStatus?: string }> };
+      expect(body.providers[0]!.initialStatus).toBe('hibernating');
+      registrar.stop();
+    } finally {
+      await fake.close();
+    }
+  });
+
+  it('publishes initialStatus=active when provider is available even with a wake recipe', async () => {
+    const fake = await startFakeServer();
+    try {
+      const awake: LlmProvider = {
+        name: 'vllm-local',
+        async complete() { throw new Error('not used'); },
+        async listModels() { return []; },
+        async isAvailable() { return true; },
+      };
+      const registrar = new VirtualLlmRegistrar({
+        mcpdUrl: fake.url,
+        token: 't',
+        publishedProviders: [{
+          provider: awake,
+          type: 'openai',
+          model: 'm',
+          wake: { type: 'http', url: 'http://localhost:9999/wake' },
+        }],
+        sessionFilePath: join(tempDir, 'provider-session'),
+        log: silentLog(),
+        heartbeatIntervalMs: 60_000,
+      });
+      await registrar.start();
+      await new Promise((r) => setTimeout(r, 20));
+
+      const registerCall = fake.calls.find((c) => c.path === '/api/v1/llms/_provider-register');
+      const body = JSON.parse(registerCall!.body) as { providers: Array<{ initialStatus?: string }> };
+      expect(body.providers[0]!.initialStatus).toBe('active');
+      registrar.stop();
+    } finally {
+      await fake.close();
+    }
+  });
+
+  it('publishes initialStatus=active when no wake recipe is configured (legacy path)', async () => {
+    const fake = await startFakeServer();
+    try {
+      // Provider intentionally returns false but has no wake recipe →
+      // legacy v1 publishers don't get hibernation behavior.
+      const sleeping: LlmProvider = {
+        name: 'vllm-local',
+        async complete() { return { content: '', toolCalls: [], usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 }, finishReason: 'stop' }; },
+        async listModels() { return []; },
+        async isAvailable() { return false; },
+      };
+      const registrar = new VirtualLlmRegistrar({
+        mcpdUrl: fake.url,
+        token: 't',
+        publishedProviders: [{ provider: sleeping, type: 'openai', model: 'm' }],
+        sessionFilePath: join(tempDir, 'provider-session'),
+        log: silentLog(),
+        heartbeatIntervalMs: 60_000,
+      });
+      await registrar.start();
+      await new Promise((r) => setTimeout(r, 20));
+
+      const registerCall = fake.calls.find((c) => c.path === '/api/v1/llms/_provider-register');
+      const body = JSON.parse(registerCall!.body) as { providers: Array<{ initialStatus?: string }> };
+      expect(body.providers[0]!.initialStatus).toBe('active');
+      registrar.stop();
+    } finally {
+      await fake.close();
+    }
+  });
+
   it('throws when mcpd returns non-201 from /_provider-register', async () => {
     const fake = await startFakeServer();
     fake.handler = (_req, res, _body) => {