@addai/node 0.24.1 → 0.26.0

package/dist/gemini-spawn.js

@@ -104,8 +104,8 @@ function writeGeminiWorkspaceSettings(cwd, servers) {
             ...(s.env && Object.keys(s.env).length > 0 ? { env: s.env } : {}),
         };
     }
-    // Start with the host's settings so auth method (security.auth.selectedType
-    // = oauth-personal etc.) and theme carry through. Override just mcpServers.
+    // Start with the host's settings so auth method and theme carry through.
+    // Override just mcpServers — and the auth type, below.
     const hostSettingsPath = path.join(os.homedir(), '.gemini', 'settings.json');
     let baseSettings = {};
     if (fs.existsSync(hostSettingsPath)) {
@@ -115,6 +115,24 @@ function writeGeminiWorkspaceSettings(cwd, servers) {
         catch { /* corrupt — start fresh */ }
     }
     const merged = { ...baseSettings, mcpServers };
+    // Google retired individual Gemini Code Assist OAuth on 18 June 2026: a CLI
+    // set to 'oauth-personal' now answers IneligibleTierError and points at
+    // Antigravity, whatever else is configured. Most hosts still carry that
+    // retired value in ~/.gemini/settings.json, so inheriting it verbatim means
+    // the run cannot start.
+    //
+    // An API key is the supported path that survives (as is vertex-ai for
+    // enterprise), so when one is present the auth type is forced to match it
+    // rather than trusting whatever the host was left set to. childEnv already
+    // passes GEMINI_API_KEY through from the daemon's own environment.
+    if (process.env.GEMINI_API_KEY) {
+        const security = { ...(baseSettings.security ?? {}) };
+        security.auth = {
+            ...(security.auth ?? {}),
+            selectedType: 'gemini-api-key',
+        };
+        merged.security = security;
+    }
     const settingsPath = path.join(dir, 'settings.json');
     fs.writeFileSync(settingsPath, JSON.stringify(merged, null, 2), { mode: 0o600 });
     return settingsPath;
@@ -274,6 +292,40 @@ function spawnGemini(input) {
     if (input.resumeSessionId && isGeminiSessionId(input.resumeSessionId)) {
         args.push('--resume', input.resumeSessionId);
     }
+    // Fail early and legibly when there is nothing to authenticate with. Without
+    // this the run dies inside gemini with "IneligibleTierError: This client is no
+    // longer supported for Gemini Code Assist for individuals", which reads like
+    // an account problem rather than a missing key on this machine.
+    if (!process.env.GEMINI_API_KEY) {
+        const host = path.join(os.homedir(), '.gemini', 'settings.json');
+        let selected = 'oauth-personal';
+        try {
+            const raw = JSON.parse(fs.readFileSync(host, 'utf8'));
+            selected = raw.security?.auth?.selectedType ?? selected;
+        }
+        catch { /* no host settings — the default is the retired one anyway */ }
+        if (selected === 'oauth-personal') {
+            const events = (0, events_1.bufferedEvents)([
+                {
+                    type: 'error',
+                    code: 'gemini_auth_retired',
+                    message: 'gemini cannot start on this machine: Google retired individual ' +
+                        'Gemini Code Assist OAuth on 18 June 2026, and this host is still set ' +
+                        "to 'oauth-personal'. Set GEMINI_API_KEY on the node (or switch the " +
+                        'host to vertex-ai for an enterprise licence), or point this entity at ' +
+                        'another agent.',
+                },
+                { type: 'turn_complete', stopReason: 'failed' },
+            ]);
+            return {
+                pid: undefined,
+                sessionId: null,
+                kill: () => { },
+                onEvent: events.onEvent,
+                done: Promise.resolve(1),
+            };
+        }
+    }
     // No HOME override — gemini's auth + shell-out tools all need the
     // real $HOME. Workspace-scope settings.json (written above) is enough
     // to override the host's mcpServers without poisoning everything else.

package/dist/grok-spawn.js

@@ -74,7 +74,6 @@ const grok_binary_1 = require("./grok-binary");
 const win_1 = require("./win");
 const events_1 = require("./events");
 const think_split_1 = require("./think-split");
-const mcp_headers_1 = require("./mcp-headers");
 /** Real config dir for the logged-in account. GROK_HOME points here so the
  *  entity uses the same account and token refresh works. */
 function realGrokHome() {
@@ -161,44 +160,16 @@ function writeProjectMcpConfig(workingDirectory, servers) {
     }
     const lines = [];
     for (const s of usable) {
-        // Remote MCP servers are BRIDGED through mcp-remote, because grok cannot
-        // speak HTTP MCP itself.
-        //
-        // Its bundled README documents `url` + a `headers` inline table, and this
-        // code used to emit exactly that. Measured against grok 1.0.3, that entry
-        // is SILENTLY IGNORED: in one run grok loaded ten stdio servers and skipped
-        // both url-based ones — ours and a pre-existing third-party entry — with no
-        // error for either. The docs describe a transport the binary does not have.
-        //
-        // That is why entities on grok lost every Connection the moment the
-        // registry moved from `npx` to `http`: the servers were simply not there,
-        // and grok reported it to the model as "(auth required)".
-        //
-        // mcp-remote is the standard stdio<->HTTP shim (grok's own README reaches
-        // for it in its Linear example) and it is verified working here. The cost
-        // is an npx start per remote server on grok runs only — the very cost the
-        // HTTP move removed for every other agent. Revisit when grok ships real
-        // remote-MCP support.
-        //
-        // Folder trust still gates repo-local config.toml servers, so the --trust
-        // launch flag remains load-bearing for these entries to start at all.
-        if ((0, mcp_headers_1.isRemoteMcp)(s.command)) {
-            const url = (0, mcp_headers_1.remoteMcpUrl)(s.args);
-            if (!url) {
-                console.error(`[grok] MCP ${s.slug}: ${s.command} server has no URL — skipped`);
-                continue;
-            }
-            // The bridge itself is shared — see AGENTS_WITHOUT_HTTP_MCP for which
-            // agents need it and the evidence for each.
-            const bridge = (0, mcp_headers_1.remoteMcpBridge)(url, s.env);
-            const bridged = (0, win_1.wrapMcpCommandForPlatform)(bridge.command, bridge.args);
-            lines.push(`[mcp_servers.${tomlStr(s.slug)}]`);
-            lines.push(`command = ${tomlStr(bridged.command)}`);
-            lines.push(`args = [${bridged.args.map(tomlStr).join(', ')}]`);
-            lines.push('enabled = true');
-            // The first run on a machine downloads mcp-remote, which 30s does not cover.
-            lines.push('startup_timeout_sec = 90');
-            lines.push('');
+        // Remote MCP servers (registry command='http'|'sse', args=[url]) have no
+        // verified grok config spelling, and writing the row verbatim is worse
+        // than skipping: grok would try to exec a binary called `http`, the
+        // server would never start, and its tools would vanish with no error —
+        // the exact silent failure that cost an entity its GitHub access after a
+        // provider-ladder hop. Skip it, and say so loudly enough to be findable.
+        if (s.command === 'http' || s.command === 'sse') {
+            console.error(`[grok] MCP ${s.slug}: remote ${s.command} servers are not supported by the grok ` +
+                `adapter — skipping. Its tools will be ABSENT from this run. Use claude (or codex, ` +
+                `which supports url + bearer_token_env_var) for entities that depend on it.`);
             continue;
         }
         // `cmd /c` wrapper for npm-shim commands on native Windows (no-op on POSIX).

package/dist/heartbeat.js

@@ -13,6 +13,8 @@ const supabase_client_1 = require("./supabase-client");
 const store_1 = require("./store");
 const capabilities_1 = require("./capabilities");
 const request_pump_1 = require("./request-pump");
+const pump_1 = require("./flows/pump");
+const manager_1 = require("./desktop/manager");
 const config_1 = require("./config");
 let timer = null;
 let stopped = false;
@@ -59,6 +61,19 @@ async function tick() {
         // whatever it already had (its own default of 4).
         if (res && res.max_concurrent_runs != null)
             (0, request_pump_1.setMaxConcurrent)(res.max_concurrent_runs);
+        // Flows and desktops ride the same heartbeat for the same reason: they are
+        // fields on this node's own row and we are already talking every 30s.
+        //
+        // Each is applied only when the server actually said something. An older
+        // server that has not had the migration says nothing, and nothing must
+        // read as "leave it alone" — never as "turn flows off" (which would
+        // silently strand a queue) or "cap desktops at zero".
+        if (res && typeof res.flows_enabled === 'boolean')
+            (0, pump_1.setFlowsEnabled)(res.flows_enabled);
+        if (res && res.max_concurrent_flows != null)
+            (0, pump_1.setMaxConcurrentFlows)(res.max_concurrent_flows);
+        if (res && res.max_concurrent_desktops != null)
+            (0, manager_1.setMaxConcurrentDesktops)(res.max_concurrent_desktops);
     }
     catch (err) {
         if (err instanceof supabase_client_1.RpcError && err.status === 401) {
@@ -124,6 +139,15 @@ async function runTick() {
             p_token: pairing.daemonToken,
             p_request_ids: (0, request_pump_1.activeRequestIdList)(),
         });
+        // Flow runs are claimed rows too, and the server's reclaim sweep is just
+        // as capable of requeueing a healthy one that happens to be quiet. Sent in
+        // the same tick so the two never drift apart under load; the empty call
+        // matters, because it is what tells the server this build reports flow
+        // runs at all.
+        await (0, supabase_client_1.rpc)('flow_run_heartbeat', {
+            p_token: pairing.daemonToken,
+            p_run_ids: (0, pump_1.activeFlowRunIds)(),
+        });
         lastRunBeatError = '';
     }
     catch (err) {

package/dist/index.js

@@ -56,6 +56,7 @@ const autostart_1 = require("./autostart");
 const sleep_detector_1 = require("./sleep-detector");
 const claude_config_1 = require("./claude-config");
 const request_pump_1 = require("./request-pump");
+const pump_1 = require("./flows/pump");
 const projects_1 = require("./projects");
 const manager_1 = require("./desktop/manager");
 const relay_client_1 = require("./desktop/relay-client");
@@ -210,6 +211,12 @@ async function start(argv = []) {
     (0, projects_1.start)();
     // …and to keep desktops and the containers behind them in agreement.
     (0, manager_1.startDesktopManager)();
+    // The flows pump starts idle and stays that way until the heartbeat reports
+    // this machine has flows switched on. Starting it unconditionally is right:
+    // the alternative is a daemon that only notices the switch on restart, and
+    // "turn it on, then restart the node" is not a setup step anyone should have
+    // to be told about.
+    (0, pump_1.start)();
     // …and dial out to the relay. That socket is this node's live link: the
     // server pushes work down it the instant it exists, so a command no longer
     // waits up to 30s for the next heartbeat to notice it. The heartbeat and
@@ -231,6 +238,7 @@ async function start(argv = []) {
         (0, manager_1.stopDesktopManager)();
         (0, relay_client_1.stopRelayClient)();
         (0, request_pump_1.stop)();
+        (0, pump_1.stop)();
         (0, heartbeat_1.stop)();
         (0, auto_update_1.stop)();
         sleepMonitor.stop();
@@ -241,10 +249,20 @@ async function start(argv = []) {
         // leave paused flow runs stuck `waiting` until server-side reclaim
         // (5 min) requeued or failed them.
         try {
-            const { drained, timedOut } = await (0, request_pump_1.drain)(drainTimeoutMs);
-            lastDrain = { drained, timedOut };
-            if (drained || timedOut) {
-                console.log(`[shutdown] drained ${drained} request(s); ${timedOut} still in-flight (timed out)`);
+            // Both pools drain against the SAME deadline, in parallel. Draining them
+            // one after the other would give a machine holding both up to twice the
+            // timeout to shut down, and the timeout is there because something
+            // upstream is waiting for this process to be gone.
+            const [requests, flows] = await Promise.all([
+                (0, request_pump_1.drain)(drainTimeoutMs),
+                (0, pump_1.drain)(drainTimeoutMs),
+            ]);
+            lastDrain = { drained: requests.drained, timedOut: requests.timedOut };
+            if (requests.drained || requests.timedOut) {
+                console.log(`[shutdown] drained ${requests.drained} request(s); ${requests.timedOut} still in-flight (timed out)`);
+            }
+            if (flows.drained || flows.timedOut) {
+                console.log(`[shutdown] drained ${flows.drained} flow run(s); ${flows.timedOut} still in-flight (timed out)`);
             }
         }
         catch (err) {

package/dist/mcp-headers.d.ts

@@ -20,12 +20,18 @@ export declare function remoteMcpHeaders(env?: Record<string, string> | null): R
  *   claude  YES  — proven in production; entities use these every day.
  *   codex   YES  — `codex mcp list` shows a `url` entry under Url, enabled.
  *   kimi    YES  — `kimi mcp list` prints "<slug> (http): <url>".
- *   gemini  ?    — UNVERIFIED. The CLI on hand cannot authenticate at all
- *                  ("no longer supported for Gemini Code Assist for
- *                  individuals"), so its httpUrl handling was never observed
- *                  connecting. No entity runs gemini today. If one ever does,
- *                  verify before trusting it — this exact assumption is what
- *                  cost us grok.
+ *   gemini  YES  — its shipped bundle treats httpUrl as a network transport
+ *                  (hasNetworkTransport), carries StreamableHTTPClientTransport,
+ *                  and documents "httpUrl (for Streamable HTTP), url (for SSE),
+ *                  and command (for stdio)" with a headers map. Read out of the
+ *                  installed bundle rather than the docs, because docs are what
+ *                  lied about grok. It was never watched connecting, though:
+ *                  gemini-cli cannot authenticate at all on a personal account
+ *                  since Google retired individual Gemini Code Assist OAuth on
+ *                  18 June 2026 (IneligibleTierError -> migrate to Antigravity).
+ *                  API-key and enterprise auth are unaffected. So gemini's
+ *                  blocker is AUTH, not MCP transport — bridging it would fix
+ *                  nothing. No entity runs gemini today.
  *   grok    NO   — grok 1.0.3 SILENTLY IGNORES url-based entries. In one run it
  *                  loaded ten stdio servers and skipped every remote one, ours
  *                  and a third party's, with no error, while its own bundled

package/dist/mcp-headers.js

@@ -57,12 +57,18 @@ function remoteMcpHeaders(env) {
  *   claude  YES  — proven in production; entities use these every day.
  *   codex   YES  — `codex mcp list` shows a `url` entry under Url, enabled.
  *   kimi    YES  — `kimi mcp list` prints "<slug> (http): <url>".
- *   gemini  ?    — UNVERIFIED. The CLI on hand cannot authenticate at all
- *                  ("no longer supported for Gemini Code Assist for
- *                  individuals"), so its httpUrl handling was never observed
- *                  connecting. No entity runs gemini today. If one ever does,
- *                  verify before trusting it — this exact assumption is what
- *                  cost us grok.
+ *   gemini  YES  — its shipped bundle treats httpUrl as a network transport
+ *                  (hasNetworkTransport), carries StreamableHTTPClientTransport,
+ *                  and documents "httpUrl (for Streamable HTTP), url (for SSE),
+ *                  and command (for stdio)" with a headers map. Read out of the
+ *                  installed bundle rather than the docs, because docs are what
+ *                  lied about grok. It was never watched connecting, though:
+ *                  gemini-cli cannot authenticate at all on a personal account
+ *                  since Google retired individual Gemini Code Assist OAuth on
+ *                  18 June 2026 (IneligibleTierError -> migrate to Antigravity).
+ *                  API-key and enterprise auth are unaffected. So gemini's
+ *                  blocker is AUTH, not MCP transport — bridging it would fix
+ *                  nothing. No entity runs gemini today.
  *   grok    NO   — grok 1.0.3 SILENTLY IGNORES url-based entries. In one run it
  *                  loaded ten stdio servers and skipped every remote one, ours
  *                  and a third party's, with no error, while its own bundled

package/dist/session-runner.js

@@ -59,6 +59,7 @@ const events_1 = require("./events");
 const paths_1 = require("./paths");
 const projects_1 = require("./projects");
 const claude_print_1 = require("./claude-print");
+const subagent_usage_1 = require("./subagent-usage");
 const attachments_1 = require("./attachments");
 const codex_spawn_1 = require("./codex-spawn");
 const diskguard_1 = require("./diskguard");
@@ -808,6 +809,26 @@ function isOwnSessionDir(dir) {
     const resolved = path.resolve(dir);
     return resolved.startsWith(root + path.sep);
 }
+/**
+ * Report what this run's subagents cost, before the run goes terminal.
+ *
+ * Awaited rather than fired-and-forgotten: the server recomputes a run's
+ * rollup from its events the moment the status flips, so an event that lands
+ * a beat later would be counted only if something else came along to trigger
+ * a recompute. Failure here must never change the run's outcome — a missing
+ * figure is a worse report, not a worse run.
+ */
+async function emitSubagentUsage(requestId, cwd, startedAtMs) {
+    try {
+        const sub = await (0, subagent_usage_1.readSubagentUsage)(cwd, startedAtMs);
+        if (!sub)
+            return;
+        await emit(requestId, 'subagent_usage', { agents: sub.agents, usage: sub.usage });
+    }
+    catch (err) {
+        console.error(`[session-runner] subagent usage [reqId=${requestId}] failed: ${err.message}`);
+    }
+}
 function eventPayload(e) {
     const { type: _type, ...rest } = e;
     return rest;
@@ -1110,6 +1131,9 @@ async function runClaudeTui(req, cwd, installed) {
             await new Promise(r => setTimeout(r, 250));
         }
     }
+    // Floor for "which subagent transcripts belong to this attempt" — see
+    // subagent-usage.ts.
+    const startedAtMs = Date.now();
     let spawn;
     try {
         spawn = (0, claude_spawn_1.spawnClaudeForRuntime)({
@@ -1509,7 +1533,8 @@ async function runClaudeTui(req, cwd, installed) {
             // is still non-terminal, reopening the server-side reclaim double-spawn
             // window. Resolve only after the chain settles, like the print runners
             // which `await finalizeTerminal`.
-            const done = (0, supabase_client_1.rpc)('runtime_get_request_status', { p_token: token(), p_request_id: req.id })
+            const done = emitSubagentUsage(req.id, cwd, startedAtMs)
+                .then(() => (0, supabase_client_1.rpc)('runtime_get_request_status', { p_token: token(), p_request_id: req.id }))
                 .then(currentStatus => {
                 if (currentStatus === 'canceled')
                     return;
@@ -1547,6 +1572,10 @@ async function runClaudePrint(req, cwd, installed) {
         return;
     }
     const permissionMode = req.agent === 'claude-bypass' ? 'bypassPermissions' : (req.permission_mode ?? undefined);
+    // Floor for "which subagent transcripts belong to this attempt" — see
+    // subagent-usage.ts. Taken before the spawn so nothing this run writes
+    // can fall outside the window.
+    const startedAtMs = Date.now();
     let handle;
     try {
         handle = (0, claude_print_1.spawnClaudePrint)({
@@ -1621,6 +1650,8 @@ async function runClaudePrint(req, cwd, installed) {
     hang.stop();
     if (handle.sessionId)
         await setStatus(req.id, { spawnedSessionId: handle.sessionId });
+    // The agent has exited, so every subagent transcript it wrote is complete.
+    await emitSubagentUsage(req.id, cwd, startedAtMs);
     void emit(req.id, 'session_end', { reason: 'claude_exit', exitCode });
     const currentStatus = await (0, supabase_client_1.rpc)('runtime_get_request_status', { p_token: token(), p_request_id: req.id }).catch(() => null);
     if (currentStatus !== 'canceled') {

package/dist/subagent-usage.d.ts

@@ -0,0 +1,19 @@
+export interface SubagentUsage {
+    /** How many subagent transcripts contributed. */
+    agents: number;
+    /** Anthropic-shaped so the server normalizes it with every other usage
+     *  block instead of needing a special case. */
+    usage: {
+        input_tokens: number;
+        output_tokens: number;
+        cache_read_input_tokens: number;
+        cache_creation_input_tokens: number;
+    };
+}
+/**
+ * Everything this run's subagents burned, or null if it spawned none.
+ *
+ * Null rather than zeros on purpose: "no subagents" and "subagents that cost
+ * nothing" are different claims, and only the first one is ever true.
+ */
+export declare function readSubagentUsage(cwd: string | null | undefined, sinceMs: number): Promise<SubagentUsage | null>;

package/dist/subagent-usage.js

@@ -0,0 +1,173 @@
+"use strict";
+// What the subagents cost.
+//
+// Claude Code streams the MAIN agent loop to stdout and reports that turn's
+// usage on its `result` line — which is what the daemon forwards and what
+// Entity Studio has always displayed. A Task/Agent subagent is invisible to
+// that stream: it runs inside the parent process, and its transcript is
+// written to a SEPARATE `agent-*.jsonl` beside the session file. Its tokens
+// are never mentioned on stdout, and the tool_result the parent receives
+// carries the subagent's text and nothing else.
+//
+// Measured on a real run before this existed: main loop 3,658,620 tokens,
+// twenty subagents 5,479,638 more. The run was reported at 40% of what it
+// actually cost. This reads the other 60% off the disk the agent just wrote.
+var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    var desc = Object.getOwnPropertyDescriptor(m, k);
+    if (!desc || ("get" in desc ? !m.__esModule : desc.writable || desc.configurable)) {
+      desc = { enumerable: true, get: function() { return m[k]; } };
+    }
+    Object.defineProperty(o, k2, desc);
+}) : (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    o[k2] = m[k];
+}));
+var __setModuleDefault = (this && this.__setModuleDefault) || (Object.create ? (function(o, v) {
+    Object.defineProperty(o, "default", { enumerable: true, value: v });
+}) : function(o, v) {
+    o["default"] = v;
+});
+var __importStar = (this && this.__importStar) || (function () {
+    var ownKeys = function(o) {
+        ownKeys = Object.getOwnPropertyNames || function (o) {
+            var ar = [];
+            for (var k in o) if (Object.prototype.hasOwnProperty.call(o, k)) ar[ar.length] = k;
+            return ar;
+        };
+        return ownKeys(o);
+    };
+    return function (mod) {
+        if (mod && mod.__esModule) return mod;
+        var result = {};
+        if (mod != null) for (var k = ownKeys(mod), i = 0; i < k.length; i++) if (k[i] !== "default") __createBinding(result, mod, k[i]);
+        __setModuleDefault(result, mod);
+        return result;
+    };
+})();
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.readSubagentUsage = readSubagentUsage;
+const fs = __importStar(require("fs"));
+const path = __importStar(require("path"));
+const readline = __importStar(require("readline"));
+const paths_1 = require("./paths");
+const claude_binary_1 = require("./claude-binary");
+const isAgentTranscript = (name) => /^agent-.+\.jsonl$/i.test(name);
+/**
+ * Which subagent transcripts belong to this run.
+ *
+ * +Ai gives most runs their own scratch directory (~/.ainode/sessions/<id>),
+ * and Claude Code names the transcript folder after the working directory —
+ * so for those the whole folder is this run's and nothing else's. A run pinned
+ * to a project checkout shares its folder with every other run against that
+ * project, so there we fall back to "written since this attempt started",
+ * which is exact unless two runs are working the same checkout concurrently.
+ */
+function transcriptsFor(cwd, sinceMs) {
+    const dir = path.join(paths_1.CLAUDE_PROJECTS_DIR, (0, claude_binary_1.encodeProjectPath)(cwd));
+    let names;
+    try {
+        names = fs.readdirSync(dir);
+    }
+    catch {
+        return [];
+    }
+    const exclusive = path.resolve(cwd).startsWith(path.resolve(paths_1.RUNTIME_SESSIONS_DIR) + path.sep);
+    const out = [];
+    for (const name of names) {
+        if (!isAgentTranscript(name))
+            continue;
+        const file = path.join(dir, name);
+        if (exclusive) {
+            out.push(file);
+            continue;
+        }
+        try {
+            if (fs.statSync(file).mtimeMs >= sinceMs)
+                out.push(file);
+        }
+        catch { /* vanished */ }
+    }
+    return out;
+}
+const num = (v) => (typeof v === 'number' && Number.isFinite(v) ? v : 0);
+/**
+ * Sum one transcript's usage.
+ *
+ * Claude Code writes the same message id more than once — the text first, the
+ * tool_use block after — and both copies repeat the identical usage. Summing
+ * blind therefore roughly doubles the answer, so this dedupes by message id.
+ */
+async function sumTranscript(file, into) {
+    let sawAny = false;
+    const seen = new Set();
+    let stream;
+    try {
+        stream = fs.createReadStream(file);
+    }
+    catch {
+        return false;
+    }
+    try {
+        for await (const line of readline.createInterface({ input: stream, crlfDelay: Infinity })) {
+            if (!line.trim())
+                continue;
+            let o;
+            try {
+                o = JSON.parse(line);
+            }
+            catch {
+                continue;
+            }
+            if (o.type !== 'assistant')
+                continue;
+            const message = o.message;
+            const u = message?.usage;
+            if (!u)
+                continue;
+            const id = typeof message?.id === 'string' ? message.id : '';
+            if (id) {
+                if (seen.has(id))
+                    continue;
+                seen.add(id);
+            }
+            sawAny = true;
+            into.usage.input_tokens += num(u.input_tokens);
+            into.usage.output_tokens += num(u.output_tokens);
+            into.usage.cache_read_input_tokens += num(u.cache_read_input_tokens);
+            into.usage.cache_creation_input_tokens += num(u.cache_creation_input_tokens);
+        }
+    }
+    catch {
+        // A half-written or unreadable transcript costs us that subagent's
+        // tokens, not the run's terminal status.
+    }
+    return sawAny;
+}
+/**
+ * Everything this run's subagents burned, or null if it spawned none.
+ *
+ * Null rather than zeros on purpose: "no subagents" and "subagents that cost
+ * nothing" are different claims, and only the first one is ever true.
+ */
+async function readSubagentUsage(cwd, sinceMs) {
+    if (!cwd)
+        return null;
+    const files = transcriptsFor(cwd, sinceMs);
+    if (!files.length)
+        return null;
+    const total = {
+        agents: 0,
+        usage: {
+            input_tokens: 0, output_tokens: 0,
+            cache_read_input_tokens: 0, cache_creation_input_tokens: 0,
+        },
+    };
+    for (const file of files) {
+        if (await sumTranscript(file, total))
+            total.agents++;
+    }
+    if (!total.agents)
+        return null;
+    return total;
+}

package/dist/types.d.ts

@@ -25,6 +25,14 @@ export type RuntimeEvent = {
     type: 'turn_complete';
     usage?: Record<string, unknown>;
     stopReason?: string;
+}
+/** What this run's subagents cost, read off their own transcripts once the
+ *  agent has exited. Never reaches stdout, so it is never in turn_complete
+ *  — see subagent-usage.ts. Emitted at most once per attempt. */
+ | {
+    type: 'subagent_usage';
+    agents: number;
+    usage: Record<string, unknown>;
 } | {
     type: 'error';
     code: string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@addai/node",
-  "version": "0.24.1",
+  "version": "0.26.0",
   "description": "Daemon that pairs a machine with your +Ai account and runs Claude / Codex / Kimi / Gemini agents on its behalf. Reachable via Supabase from Vault, Entity Studio, or any other +Ai surface.",
   "license": "MIT",
   "keywords": [
@@ -49,6 +49,7 @@
     "node": ">=18"
   },
   "dependencies": {
+    "@addai/node-flows": "^1.0.0",
     "node-pty": "^1.1.0",
     "ws": "^8.21.3"
   },