ramwisp 0.1.6 → 0.1.8

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ramwisp",
-  "version": "0.1.6",
+  "version": "0.1.8",
   "description": "ramwisp: Claude Code and Codex subagents on their own cloud machines (AWS Nitro Enclaves), with your own subscription. MCP + CLI.",
   "type": "module",
   "bin": {
@@ -19,7 +19,7 @@
     "node": ">=20"
   },
   "scripts": {
-    "test": "node --test test/"
+    "test": "node --test"
   },
   "license": "Apache-2.0",
   "keywords": [
@@ -32,7 +32,7 @@
   "homepage": "https://ramwisp.com",
   "repository": {
     "type": "git",
-    "url": "https://github.com/spectre-systems/wisp-saas",
+    "url": "https://github.com/spectre-systems/ramwisp",
     "directory": "mcp"
   }
 }

package/src/account.js

@@ -41,6 +41,7 @@ export function logout() {
 let client = "cli";
 /** Quem está usando o MCP (vem do clientInfo do initialize: "claude-code", "codex-mcp-client"…). */
 export function setClient(name) { if (name) client = String(name).slice(0, 60); }
+export const getClient = () => client;
 
 export async function call(method, path, body, { token = getToken(), timeoutMs = 30_000 } = {}) {
   const r = await fetch(API + path, {

package/src/client.js

@@ -11,10 +11,16 @@ import { engineCredential } from "./creds.js";
 export const FINAL = ["done", "failed", "killed", "expired"];
 const sealers = new Map();          // id -> Promise (selagem em andamento neste processo)
 const sealErrors = new Map();
+const collecting = new Map();
+const watchers = new Map();        // id -> Promise (recolhe sozinho quando termina)       // id -> Promise (abertura em andamento: duas chamadas juntas não disputam a chave)
 
 export class LoginRequired extends Error {}
 
 const keyFile = (id) => join(ensureDir("jobs"), `${id}.json`);
+// Cópia local do resultado já aberto: uma espera em segundo plano pode recolher o resultado e a resposta dela
+// nunca chegar ao agente; com a cópia, a próxima chamada devolve o mesmo resultado em vez de "already collected".
+export const RESULT_TTL_MS = 7 * 86400_000;
+const resultFile = (id) => join(ensureDir("results"), `${id}.json`);
 const MAX_WORKSPACE = 15 * 1024 * 1024;       // compactado
 const SKIP_DIRS = new Set(["node_modules", ".git", ".venv", "venv", "__pycache__", "dist", "build", ".next", "target", ".cache"]);
 
@@ -64,7 +70,7 @@ export async function spawnAgent(o) {
   const job = await call("POST", "/api/jobs", { engine, ram_gb: o.ram_gb ?? 2, timeout_s: timeout,
     nonce: nonce.toString("base64"), label: o.label });
   writeSecret(keyFile(job.id), JSON.stringify({ id: job.id, priv: exportKey(priv), nonce: nonce.toString("base64"), engine,
-    workspace: ws?.root }));
+    workspace: ws?.root, pid: process.pid }));
   const payload = Buffer.from(JSON.stringify({ engine, model: o.model, mission: o.mission, turns: o.max_turns ?? 20,
     timeout, auth: { kind: cred.kind, value: cred.value }, ...(ws ? { workspace_tgz: ws.tgz.toString("base64") } : {}) }));
   const p = sealWhenReady(job.id, priv, nonce, payload).finally(() => { payload.fill(0); sealers.delete(job.id); });
@@ -126,22 +132,53 @@ function parseOutput(stdout) {
   }
 }
 
-/** Resultado final (e apaga a chave local e a saída cifrada no servidor), ou o status se ainda roda. */
+function cachedResult(id) {
+  try {
+    const f = resultFile(id);
+    if (Date.now() - statSync(f).mtimeMs > RESULT_TTL_MS) { rmSync(f, { force: true }); return null; }
+    return JSON.parse(readFileSync(f, "utf8"));
+  } catch { return null; }
+}
+
+/** Apaga cópias locais com mais de 7 dias (melhor esforço). */
+export function pruneResults() {
+  try {
+    const d = ensureDir("results");
+    for (const n of readdirSync(d)) {
+      const f = join(d, n);
+      if (Date.now() - statSync(f).mtimeMs > RESULT_TTL_MS) rmSync(f, { force: true });
+    }
+  } catch { /* sem pasta ainda */ }
+}
+
+/**
+ * Resultado final, ou o status se ainda roda. Ao abrir, guarda uma cópia local (0600, 7 dias), avisa o servidor
+ * (que apaga a saída cifrada) e apaga a chave. Chamadas seguintes devolvem a cópia, na mesma sessão ou em outra.
+ */
 export async function result(id) {
+  const cached = cachedResult(id);
+  if (cached) return { ...cached, from_local_copy: true };
+  if (collecting.has(id)) return collecting.get(id);
+  const p = fetchResult(id).finally(() => collecting.delete(id));
+  collecting.set(id, p);
+  return p;
+}
+
+async function fetchResult(id) {
   const j = await call("GET", `/api/jobs/${id}`);
   const meta = { id, status: j.status, ram_gb: j.ram_gb, peak_mem_mib: j.peak_mem_mib, cost_usd: j.cost_cents != null ? +(j.cost_cents / 100).toFixed(4) : null };
   if (sealErrors.has(id)) return { ...meta, status: "failed", error: `refused for security: ${sealErrors.get(id)}` };
   if (!FINAL.includes(j.status)) return { ...meta, mem_used_mib: j.mem_used_mib };
   if (j.status !== "done" || !j.output) {
     rmSync(keyFile(id), { force: true });
-    return { ...meta, error: j.error ?? (j.collected_at ? "result was already collected" : j.status) };
+    return { ...meta, error: j.error ?? (j.collected_at
+      ? `result was already collected on another machine (it is kept only where it was opened, in ${ensureDir("results")})`
+      : j.status) };
   }
   const f = keyFile(id);
   if (!existsSync(f)) return { ...meta, error: "the key to open this result is not on this machine" };
   const k = JSON.parse(readFileSync(f, "utf8"));
   const out = JSON.parse(openOutput(importKey(k.priv), Buffer.from(k.enclave_pub, "base64"), Buffer.from(k.nonce, "base64"), j.output).toString());
-  await call("POST", `/api/jobs/${id}/collected`).catch(() => {});
-  rmSync(f, { force: true });
   const parsed = parseOutput(out.stdout ?? "");
   const res = { ...parsed, ...meta, exit_code: out.exit_code, duration_s: out.duration_s };
   if (out.exit_code === 124) res.error = "timeout";
@@ -158,9 +195,71 @@ export async function result(id) {
     res.patch_stat = out.patch_stat;
   }
   if (out.exit_code !== 0 && out.stderr_tail) res.stderr_tail = out.stderr_tail.slice(-1500);
+  // a cópia vem ANTES de avisar o servidor e apagar a chave: se algo cair no meio, o resultado não se perde
+  writeSecret(resultFile(id), JSON.stringify(res));
+  await call("POST", `/api/jobs/${id}/collected`).catch(() => {});
+  rmSync(f, { force: true });
+  pruneResults();
   return res;
 }
 
+const nap = (ms) => new Promise((r) => setTimeout(r, ms).unref());
+
+/**
+ * Recolhe o resultado sozinho quando o subagente termina (guarda a cópia local), sem ninguém bloqueado esperando.
+ * Os timers não seguram o processo: no CLI, sair continua saindo.
+ */
+export function watch(id) {
+  if (watchers.has(id)) return;
+  const p = (async () => {
+    const until = Date.now() + 3 * 3600_000;
+    while (Date.now() < until) {
+      await nap(Number(process.env.WISP_WATCH_MS ?? 10_000));
+      const r = await result(id).catch(() => null);
+      if (r && (FINAL.includes(r.status) || r.error)) return;
+    }
+  })().finally(() => watchers.delete(id));
+  watchers.set(id, p);
+}
+
+/**
+ * Ao subir o MCP: volta a vigiar os subagentes desta máquina que ainda têm chave (sessão anterior caiu, etc.).
+ * Um job que nunca foi selado e cujo processo selador morreu (a sessão acabou antes da máquina subir) não tem como
+ * rodar: a missão só existia na memória daquele processo. Esse é derrubado na hora, para não ficar cobrando à toa.
+ * Se o selador está vivo (outra sessão, ou o ajudante do Codex, que sobe o próprio MCP), só vigia.
+ */
+export function resumeWatches() {
+  let names = [];
+  try { names = readdirSync(ensureDir("jobs")).filter((n) => n.endsWith(".json")); } catch { return; }
+  for (const n of names) {
+    const id = n.slice(0, -5);
+    if (sealers.has(id) || watchers.has(id)) continue;
+    let k;
+    try { k = JSON.parse(readFileSync(join(ensureDir("jobs"), n), "utf8")); } catch { continue; }
+    if (k.enclave_pub || alive(k.pid)) watch(id);
+    else if (getToken()) {
+      killAgent(id).then(() => process.stderr.write(`ramwisp: ${id} was never delivered (the session ended first); stopped it\n`),
+        () => {});
+    }
+  }
+}
+
+function alive(pid) {
+  if (!pid) return false;
+  try { process.kill(pid, 0); return true; } catch (e) { return e.code === "EPERM"; }
+}
+
+/** Espera vários de uma vez: volta quando todos terminarem (ou no limite), com o resultado de cada um. */
+export async function waitAgents(ids, maxWaitS = 900) {
+  const until = Date.now() + maxWaitS * 1000;
+  for (;;) {
+    const agents = await Promise.all(ids.map((id) => result(id).catch((e) => ({ id, error: e.message }))));
+    const done = agents.every((r) => FINAL.includes(r.status) || r.error);
+    if (done || Date.now() > until) return { all_done: done, agents };
+    await nap(3000);
+  }
+}
+
 export async function waitAgent(id, maxWaitS = 900) {
   const until = Date.now() + maxWaitS * 1000;
   for (;;) {

package/src/mcp.js

@@ -1,10 +1,10 @@
 // Servidor MCP (stdio, JSON-RPC por linha). Sem dependências.
 import { createInterface } from "node:readline";
-import { API, getToken, setClient, startLogin } from "./account.js";
-import { LoginRequired, killAgent, listAgents, result, spawnAgent, waitAgent } from "./client.js";
+import { API, getClient, getToken, setClient, startLogin } from "./account.js";
+import { LoginRequired, killAgent, listAgents, result, resumeWatches, spawnAgent, waitAgent, waitAgents, watch } from "./client.js";
 
 const PROTOCOLS = ["2025-06-18", "2025-03-26", "2024-11-05"];
-const VERSION = "0.1.6";
+const VERSION = "0.1.8";
 
 const INSTRUCTIONS = `ramwisp runs Claude Code or Codex subagents on ephemeral cloud machines with the RAM you ask for,
 without loading this machine. Each subagent starts inside an isolated enclave (AWS Nitro): before sending anything,
@@ -15,17 +15,30 @@ When to use: the user asks for remote subagents / "run it on ramwisp" / "spin up
 independent, heavy tasks (builds, test suites, long research) that can run in parallel elsewhere.
 
 How to use it well:
-- Parallel: call spawn_agent for every task first, then wait_agent for each id.
+- Parallel: call spawn_agent for every task first, then wait for all of them with ONE wait_agent call (ids: [...]).
+- Don't block the main conversation while subagents run (minutes to an hour). This MCP collects each result by itself
+  when it finishes and keeps it on this machine for 7 days, so nothing is lost if nobody is waiting. By default,
+  hand the waiting to a small background helper and keep working with the user:
+  - Claude Code: launch a background Agent with model "sonnet" whose only job is to call wait_agent with the ids
+    (repeat while it returns running) and report the full result text back. It must not spawn or kill anything.
+  - Codex: spawn a sub-agent with model "gpt-6-luna" for that same job, then keep working and check on it later.
+  - Without sub-agents: carry on with other work and call agent_result (instant) now and then.
+  Only wait in the main thread yourself if the user explicitly asks to wait, or if this session is about to end
+  (one-shot runs like \`claude -p\` or \`codex exec\`): the task is sent from this process once the machine is up,
+  so ending the session before status is "running" stops the subagent.
 - To work on the user's code, pass workspace (e.g. the project directory): an encrypted copy is sent,
   the subagent works on it and the changes come back as a patch (apply it with the "apply" command after reviewing).
   Without workspace the machine starts empty: put all context in the mission. It has internet (HTTPS) but no git/SSH access of the user.
-- A machine takes ~1-3 min to start; wait_agent waits up to 15 min per call (call it again if it returns running).
+- A machine takes ~1-3 min to start; wait_agent waits up to max_wait_s per call (call it again if it returns running).
 - Always collect with wait_agent or agent_result: the answer can only be decrypted on this machine.
+  Once opened, the result stays readable here for 7 days: calling wait_agent/agent_result again returns the same answer,
+  so a wait that ran in the background never loses it. While running, only status, RAM and cost are visible.
 - Each subagent uses ramwisp credit (the machine) and the user's subscription/key (the model). Don't launch dozens.
 - If a response says the user needs to sign in, show them the link.`;
 
 const TOOLS = [
-  { name: "spawn_agent", description: "Launch an ephemeral subagent on a machine with the requested RAM and return its id right away.",
+  { name: "spawn_agent", annotations: { title: "Launch a subagent", readOnlyHint: false, destructiveHint: false, openWorldHint: true },
+    description: "Launch an ephemeral subagent on a machine with the requested RAM and return its id right away.",
     inputSchema: { type: "object", required: ["mission"], properties: {
       mission: { type: "string", description: "Complete, self-contained task with all the context it needs." },
       engine: { type: "string", enum: ["claude", "codex"], default: "claude" },
@@ -37,14 +50,20 @@ const TOOLS = [
         description: "login = the user's subscription on this machine; key = API key from env; auto = key if present." },
       workspace: { type: "string", description: "Path to a directory/repo ON THIS MACHINE to send along. An encrypted copy is sent (in git: tracked files + new non-ignored files; never anything in .gitignore). The subagent works in ~/work and changes come back as a patch (patch_file + apply command). Max 15 MB compressed." },
       label: { type: "string", description: "Short label VISIBLE in the dashboard (don't put anything sensitive)." } } } },
-  { name: "wait_agent", description: "Wait for the subagent to finish and return the result (field result = the answer).",
-    inputSchema: { type: "object", required: ["id"], properties: {
-      id: { type: "string" }, max_wait_s: { type: "integer", default: 900, maximum: 1800 } } } },
-  { name: "agent_result", description: "Result without waiting: returns the answer or the current status (running, RAM in use).",
+  { name: "wait_agent", annotations: { title: "Wait for a subagent's result", readOnlyHint: true, openWorldHint: false },
+    description: "Wait for the subagent to finish and return the result (field result = the answer). Safe to call again: a result already opened on this machine is returned again.",
+    inputSchema: { type: "object", properties: {
+      id: { type: "string", description: "One subagent id." },
+      ids: { type: "array", items: { type: "string" }, description: "Several ids: returns when all of them finish (or at max_wait_s), with each result." },
+      max_wait_s: { type: "integer", default: 900, maximum: 1800 } } } },
+  { name: "agent_result", annotations: { title: "Read a subagent's result", readOnlyHint: true, openWorldHint: false },
+    description: "Result without waiting: returns the answer or the current status (running, RAM in use).",
     inputSchema: { type: "object", required: ["id"], properties: { id: { type: "string" } } } },
-  { name: "kill_agent", description: "Kill the subagent and destroy its machine immediately.",
+  { name: "kill_agent", annotations: { title: "Kill a subagent", readOnlyHint: false, destructiveHint: true, idempotentHint: true },
+    description: "Kill the subagent and destroy its machine immediately.",
     inputSchema: { type: "object", required: ["id"], properties: { id: { type: "string" } } } },
-  { name: "list_agents", description: "Balance, available RAM sizes and subagents that are running or have results to collect.",
+  { name: "list_agents", annotations: { title: "List subagents and balance", readOnlyHint: true, openWorldHint: false },
+    description: "Balance, available RAM sizes and subagents that are running or have results to collect.",
     inputSchema: { type: "object", properties: {} } },
   { name: "wisp_login", description: "Connect this MCP to the user's ramwisp account (opens the browser).",
     inputSchema: { type: "object", properties: {} } },
@@ -61,6 +80,19 @@ async function loginMessage() {
     `No account yet? Create one there, with free credit. Then repeat the request.`;
 }
 
+/** O que fazer depois do spawn, no idioma de cada cliente: esperar num ajudante pequeno, não na conversa principal. */
+export function nextStep(id, client = getClient()) {
+  const how = /codex/i.test(client)
+    ? `use collaboration.spawn_agent to start a sub-agent with model "gpt-6-luna" whose only job is to call the ramwisp wait_agent tool`
+    : /claude/i.test(client)
+      ? `launch a background Agent with model "sonnet" whose only job is to call mcp__ramwisp__wait_agent`
+      : `use a small background helper to call wait_agent`;
+  return `Running. Don't block the user: ${how} with ids ["${id}"] (plus any other ids you just spawned), ` +
+    `repeating while it returns running, and report the full result. If you can't, call agent_result later. ` +
+    `Keep this session open at least until status is "running": the task is still being sent from here, ` +
+    `and if the session ends first the subagent is stopped. After that the result is collected and kept on this machine automatically.`;
+}
+
 async function runTool(name, a) {
   if (name === "wisp_login") {
     if (getToken()) return { ok: true, msg: `already connected to ${API}` };
@@ -69,8 +101,11 @@ async function runTool(name, a) {
   }
   const attempt = () => {
     switch (name) {
-      case "spawn_agent": return spawnAgent(a);
-      case "wait_agent": return waitAgent(a.id, a.max_wait_s ?? 900);
+      case "spawn_agent": return spawnAgent(a).then((r) => { watch(r.id); return { ...r, next: nextStep(r.id) }; });
+      case "wait_agent":
+        if (Array.isArray(a.ids) && a.ids.length) return waitAgents(a.ids.map(String), a.max_wait_s ?? 900);
+        if (!a.id) throw new Error("pass id or ids");
+        return waitAgent(a.id, a.max_wait_s ?? 900);
       case "agent_result": return result(a.id);
       case "kill_agent": return killAgent(a.id);
       case "list_agents": return listAgents();
@@ -89,6 +124,7 @@ async function runTool(name, a) {
 }
 
 export function serve() {
+  resumeWatches();
   const send = (m) => process.stdout.write(JSON.stringify(m) + "\n");
   const rl = createInterface({ input: process.stdin });
   rl.on("line", async (line) => {