@addai/node 0.25.0 → 0.26.0

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": "0.25.0",
+  "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": [