openmausbot 0.1.89 → 0.1.90

package/dist-server/server/container-computer.js

@@ -69,6 +69,19 @@ export function perBotLocalVmTarget(botId) {
         label: digest,
     };
 }
+/** A pool-mode seat (issue #1654): one of N desktops shared by every
+ * conversation, addressed by seat index. Identities stay disjoint from the
+ * shared singleton and from per-bot digests, so discovery, labels, and lease
+ * lanes never confuse the three modes. */
+export function poolLocalVmTarget(seat) {
+    return {
+        key: `pool:${seat}`,
+        containerName: `${CONTAINER}-p${seat}`,
+        workspaceDir: join(DATA_DIR, "vm-homes", `pool-${seat}`),
+        viewerPort: null,
+        label: `pool-${seat}`,
+    };
+}
 /** Only provisioning creates this durable directory; idle removal keeps it. */
 export function localVmWorkspaceExists(target) {
     try {

package/dist-server/server/control-client.js

@@ -1,8 +1,8 @@
 // The proxy-side half of computer control. The harness keeps the record of
 // who is driving (server/computer-control.ts); the per-turn computer
 // processes consult it through this client before acting, because the
-// action paths themselves never traverse the harness — a box click goes
-// straight to the box's REST API, and a Local VM / VPS click rides a
+// action paths themselves never traverse the harness — a boat click goes
+// straight to the boat's REST API, and a Local VM / VPS click rides a
 // transparent stdio bridge into Cua Driver.
 //
 // Failure posture: CLOSED once configured. A missing/expired turn capability

package/dist-server/server/delegations.js

@@ -259,8 +259,10 @@ export function pendingDelegationSnapshot() {
     })));
 }
 /** How many handoffs one turn may queue. Small on purpose: this is the only
- * thing standing between a confused bot and a fan-out of real turns. */
-const MAX_QUEUED_PER_THREAD = 4;
+ * thing standing between a confused bot and a fan-out of real turns. Six,
+ * not four: a lead running a weekly check-in over a five-member team is the
+ * ordinary case, and hitting the cap there silently dropped one teammate. */
+const MAX_QUEUED_PER_THREAD = 6;
 /** Validate and enqueue a delegation. Pushes a "Delegated to @B: reason"
  * chip to the source thread so the user can see what was queued. */
 export function queueDelegation(bus, from, item, maxDepth, sourceThreadId = from.threadId) {
@@ -838,7 +840,13 @@ export function buildDelegationFailurePrompt(targetName, reason) {
         "Take over: tell the user what failed in plain terms, then decide the next step — retry with a narrower task, do the work yourself, or propose an alternative. Do not re-delegate the exact same task unchanged.",
     ].join("\n\n");
 }
-export const DELEGATION_WAKE_MAX_PER_WINDOW = 3;
+// Must not be smaller than MAX_QUEUED_PER_THREAD: every delegate reply is
+// one wake, so a fan-out that the queue cap allows has to be resumable
+// within the window. Three failed a five-member weekly check-in whose
+// replies all landed inside four minutes ("Delegation follow-up limit
+// reached"), while the same routine passed a week earlier only because the
+// replies happened to spread over twenty.
+export const DELEGATION_WAKE_MAX_PER_WINDOW = 6;
 export const DELEGATION_WAKE_WINDOW_MS = 5 * 60 * 1000;
 /** Bounded auto-wake budget per source thread. A delegation completion
  * wakes the source; if that source re-delegates and the new completion

package/dist-server/server/drivers/acp/core.js

@@ -43,7 +43,7 @@ export function skipSubscriptionAuthForLocalInject(model) {
 import { newEventId, newId, TurnNotStartedError } from "../../contracts.js";
 import { augmentedPath } from "../../env-path.js";
 import { supportsApprovalMode } from "../../../shared/approval-mode.js";
-import { parseAskQuestions, parseChoices, questionAnswersByQuestion } from "../../../shared/ask-question.js";
+import { MAX_QUESTION_TEXT, parseAskQuestions, parseChoices, questionAnswersByQuestion } from "../../../shared/ask-question.js";
 import { appendNative } from "../native.js";
 import { acpPermissionCommand, permissionLaunchCwd } from "../permission-command.js";
 import { commandSummary, toolDetailPreview } from "../../tool-summary.js";
@@ -275,7 +275,7 @@ export function createAcpDriver(support) {
                 };
                 const allowedCredentials = new Set(support.credentialEnv ?? []);
                 // two lists, one rule: foreign PROVIDER keys must not flip a CLI's
-                // billing off its own login, and WORKSPACE credentials (box token,
+                // billing off its own login, and WORKSPACE credentials (boat token,
                 // voice key, …) are the harness's secrets — riding along in
                 // `...process.env` is not a grant. A driver keeps only what its
                 // credentialEnv allowlist names.
@@ -439,7 +439,7 @@ export function createAcpDriver(support) {
                 }
                 // The bot's computer, mounted exactly like the Claude driver does:
                 // host and sandbox Cua connections expose Cua Driver's own MCP server.
-                // (A cloud box is not mounted here at all: a cloud turn runs ON the box.)
+                // (A cloud boat is not mounted here at all: a cloud turn runs ON the boat.)
                 if (turn.integrations?.localComputer) {
                     const local = turn.integrations.localComputer;
                     servers.push({
@@ -722,7 +722,8 @@ export function createAcpDriver(support) {
                     }
                     const tool = kind === "execute" ? "shell" : kind === "edit" ? "edit" : kind || "tool";
                     const isShellCommand = !isQuestion && kind === "execute" && !/^mcp(?:__|[.:])/i.test(String(toolCall.title ?? ""));
-                    const summary = String(toolCall.rawInput?.command ?? toolCall.title ?? tool).slice(0, 200);
+                    const rawSummary = String(toolCall.rawInput?.command ?? toolCall.title ?? tool);
+                    const summary = rawSummary.slice(0, isQuestion ? MAX_QUESTION_TEXT : 200);
                     // One structured question beside the flat choices: the richer card
                     // renders from it while older clients keep answering through
                     // `choices`. Built once here so the emit and the answer path can

package/dist-server/server/drivers/agents-call.js

@@ -278,7 +278,7 @@ function routineFields(args) {
     const runOn = destination(args.run_on ?? args.runOn);
     const timeoutMinutes = args.timeout_minutes ?? args.timeoutMinutes;
     if (runOn != null && runOn !== "maus" && runOn !== "cloud") {
-        return { fields, error: 'Use run_on="maus" for the bot’s current model and configured computer (including VPS), or run_on="box" only for the Box-hosted agent. Legacy "cloud" also means Box.' };
+        return { fields, error: 'Use run_on="maus" for the bot’s current model and configured computer (including VPS), or run_on="box" only for the Boat-hosted agent. Legacy "cloud" also means Boat.' };
     }
     if (timeoutMinutes != null && (typeof timeoutMinutes !== "number" || !Number.isInteger(timeoutMinutes) || timeoutMinutes < 5 || timeoutMinutes > 240)) {
         return { fields, error: "timeout_minutes must be a whole number from 5 to 240. Use clear_timeout to remove a limit." };
@@ -982,11 +982,14 @@ export async function callTool(name, args, context) {
         if (!routineId || !action) {
             return { text: "propose_routine_action needs a routine_id and supported action.", isError: true };
         }
+        const forBotId = String(args.for_bot_id ?? "").trim();
         const body = {
             fromBotId: BOT_ID,
             fromThreadId: THREAD_ID,
             action,
             routineId,
+            // JSON.stringify drops the key entirely when no target was named
+            ...(forBotId ? { forBotId } : {}),
         };
         if (action === "update") {
             if (!jsonRecord(args.changes)) {
@@ -1098,6 +1101,7 @@ export async function callTool(name, args, context) {
                 action: args.action,
                 text: args.text,
                 oldText: args.old_text,
+                ...(typeof args.until === "string" && args.until.trim() ? { until: args.until.trim() } : {}),
             }),
         });
         if (r.error || r.ok !== true) {

package/dist-server/server/drivers/agents-catalog.js

@@ -128,8 +128,9 @@ const ROUTINE_FIELDS_SCHEMA = {
     schedule: ROUTINE_SCHEDULE_SCHEMA,
     run_on: {
         type: "string",
+        // "box" is Boat's historical run_on destination id (agents wire contract).
         enum: ["maus", "box"],
-        description: "Default maus keeps the bot's selected model and configured computer, INCLUDING a self-hosted VPS. Omit this field for normal schedules. box explicitly switches the agent to the Box-hosted runner; it requires Box setup and is not the generic cloud/VPS option. Legacy cloud values from list_routines mean box, not VPS.",
+        description: "Default maus keeps the bot's selected model and configured computer, INCLUDING a self-hosted VPS. Omit this field for normal schedules. box explicitly switches the agent to the Boat-hosted runner; it requires Boat setup and is not the generic cloud/VPS option. Legacy cloud values from list_routines mean box, not VPS.",
     },
     timeout_minutes: {
         type: "integer",
@@ -293,7 +294,7 @@ const toolDefinitions = (externalRuntime) => [
         description: "Choose where this conversation does computer work. Call with no arguments to inspect actual available choices and the current place. For a task needing computer interaction, select the requested place, or auto to choose a suitable configured computer without asking the user to use menus. OpenMausBot reuses an existing computer first; with a configured provider it can start or provision one when needed. Do not provision for ordinary chat or just to inspect availability. A pending result means end this turn immediately: OpenMausBot updates the conversation selector and resumes the original request with that computer's real tools. Do not use the old tools after requesting a switch, repeat the task, or claim the action is done. This cannot change permissions, override Off, or switch a teammate/routine/channel.",
         inputSchema: { type: "object", additionalProperties: false, properties: {
                 surface: { type: "string", enum: ["auto", "cloud", "vm", "local", "browser"],
-                    description: "auto = suitable configured computer, cloud = remote Box/VPS, vm = isolated Local VM, local = user's own desktop, browser = built-in browser. Omit to list." },
+                    description: "auto = suitable configured computer, cloud = remote Boat/VPS, vm = isolated Local VM, local = user's own desktop, browser = built-in browser. Omit to list." },
             } },
     },
     {
@@ -523,6 +524,7 @@ const toolDefinitions = (externalRuntime) => [
                 action: { type: "string", enum: ["append", "replace", "remove", "supersede"] },
                 text: { type: "string", minLength: 1, description: "Non-blank new text for append, replace, or supersede: the fact itself, without a date or bullet. Omit for remove; use remove to delete a passage." },
                 old_text: { type: "string", minLength: 1, description: "Exact unique existing passage for replace, supersede, or remove. Omit for append." },
+                until: { type: "string", pattern: "^\\d{4}-\\d{2}-\\d{2}$", description: "Optional, for append or supersede: YYYY-MM-DD, the last day a temporary fact holds (an exam this weekend, a trip next week). After that day the entry is hidden from your memory." },
             },
             required: ["action"],
         },
@@ -614,12 +616,16 @@ const toolDefinitions = (externalRuntime) => [
     },
     {
         name: "propose_routine_action",
-        description: "Prepare a user-requested change to one of this bot's existing routines. Use list_routines first to get the routine id." + PROPOSAL_OUTCOME,
+        description: "Prepare a user-requested change to one of this bot's existing routines. Use list_routines first to get the routine id. If the user asks to change ANOTHER bot's routine and that bot is in your section, call list_bots and pass that bot's id as for_bot_id; the routine keeps its owner and every run keeps that bot's engine and permissions." + PROPOSAL_OUTCOME,
         inputSchema: {
             type: "object",
             additionalProperties: false,
             properties: {
                 routine_id: { type: "string", minLength: 1, description: "Routine id from list_routines." },
+                for_bot_id: {
+                    type: "string",
+                    description: "Only when the requested change targets ANOTHER bot's routine and that bot is in your section: that bot's id from list_bots. Learn the routine id from that bot's own routines (it can run list_routines). Omit to change one of your own routines.",
+                },
                 action: {
                     type: "string",
                     enum: ["update", "pause", "resume", "run_now", "delete"],

package/dist-server/server/drivers/{boxagent.js → boatagent.js}

@@ -1,9 +1,10 @@
 import { newEventId, newId } from "../contracts.js";
 import { appendNative } from "./native.js";
 import { OMB_ASK_TOOL, answerWithoutPreamble, askQuestionSummary, capAnswerEcho, ombAskProtocolPrompt, parseOmbAskQuestions, questionChoices, stripOmbAskBlock, } from "../../shared/ask-question.js";
+// "boxAgent" is Boat's historical driver kind (catalog and wire); keep it.
 const DRIVER_KIND = "boxAgent";
 // overridable so tests and a dev backend can be pointed at instead of the live provider
-const BOX_API = process.env.OMB_BOX_API || "https://ascii.dev/api/box/v1";
+const BOAT_API = process.env.OMB_BOX_API || "https://ascii.dev/api/box/v1";
 const MODELS = {
     default: "claude-fable-5",
     options: [
@@ -15,14 +16,14 @@ const MODELS = {
 /** Any fence whose info string names the ask protocol, even when its body
  * does not parse: the marker for "the model tried to ask and failed". */
 const ASK_FENCE_ANY = /(^|\n)[ \t]{0,3}(`{3,}|~{3,})[ \t]*omb-ask\b/;
-/** The box runs every harness boat.dev ships (claude-code, codex, pi, opencode,
+/** The boat runs every harness boat.dev ships (claude-code, codex, pi, opencode,
  * prime-agent, kimi). Which one a model id belongs to comes from the public
  * catalog, `GET /api/provider-models` at the API root: an object keyed by
  * harness, each with its `models`. A bot that arrives here from another engine
  * carries that engine's model id, so this is what lets it keep its model. */
 let catalog = null;
 async function loadCatalog() {
-    const root = BOX_API.replace(/\/api\/box\/v1\/?$/, "");
+    const root = BOAT_API.replace(/\/api\/box\/v1\/?$/, "");
     catalog = await fetch(`${root}/api/provider-models`, { signal: AbortSignal.timeout(15_000) })
         .then((res) => (res.ok ? res.json() : null))
         .catch(() => null);
@@ -44,7 +45,7 @@ function decodeConfig(raw) {
         ...(typeof o.askTimeoutMs === "number" ? { askTimeoutMs: o.askTimeoutMs } : {}),
     };
 }
-export const BoxAgentDriver = {
+export const BoatAgentDriver = {
     driverKind: DRIVER_KIND,
     metadata: { displayName: "Computer", supportsMultipleInstances: false },
     models: MODELS,
@@ -74,7 +75,7 @@ export const BoxAgentDriver = {
             createdAt: new Date().toISOString(),
         });
         const api = async (path, opts = {}) => {
-            const res = await fetch(`${BOX_API}${path}`, {
+            const res = await fetch(`${BOAT_API}${path}`, {
                 ...opts,
                 headers: { authorization: `Bearer ${token}`, "content-type": "application/json", ...opts.headers },
                 signal: opts.signal ?? AbortSignal.timeout(30_000),
@@ -120,7 +121,7 @@ export const BoxAgentDriver = {
                 });
                 appendNative(threadId, { dir: "out", source: "box.prompt", msg: { model, prompt: promptText, response: started } });
                 // real shape (2026-08): {type:"prompt.queued", promptId, promptRun:{id,…},
-                // id:<box id>} — never fall back to the bare id, it's the box's
+                // id:<boat id>} — never fall back to the bare id, it's the boat's
                 return started?.promptRun?.id ?? started?.prompt?.id ?? started?.promptId ?? null;
             };
             const promptId = await postPrompt(prompt);
@@ -149,7 +150,7 @@ export const BoxAgentDriver = {
                 const seen = new Set();
                 let lastText = "";
                 let pendingText = "";
-                /** Why the box could not answer (login expired, model refused, …). */
+                /** Why the boat could not answer (login expired, model refused, …). */
                 let problem = null;
                 /** Emit unflushed deltas as assistant_text and reset pendingText.
                  * The omb-ask block is protocol, not prose: it streamed raw (the
@@ -229,7 +230,7 @@ export const BoxAgentDriver = {
                         return { ok: ok && !cancelled, stopReason: cancelled ? "interrupted" : stopReason };
                     }
                     // The turn is still open, so this prompt continues it: same turnId,
-                    // same accounting, and the answer reaches the box the way the ask
+                    // same accounting, and the answer reaches the boat the way the ask
                     // contract promised — Q:/A: blocks, capped for echo.
                     const continuation = [
                         capAnswerEcho(answerWithoutPreamble(reply)),
@@ -238,7 +239,7 @@ export const BoxAgentDriver = {
                     ].join("\n");
                     const nextPromptId = await postPrompt(continuation);
                     // Stop can land while the continuation POST is in flight: the
-                    // interrupt inside cancel() then hits a box with no active run,
+                    // interrupt inside cancel() then hits a boat with no active run,
                     // and the continuation would start after it. Interrupt the run
                     // that just started before ending the turn.
                     if (cancelled) {
@@ -247,9 +248,9 @@ export const BoxAgentDriver = {
                     }
                     return await settleRun(nextPromptId);
                 };
-                /** Poll one box run to its settle. */
+                /** Poll one boat run to its settle. */
                 const settleRun = async (runPromptId) => {
-                    const startedAt = Date.now(); // one 30-min ceiling per box run, ask chains included
+                    const startedAt = Date.now(); // one 30-min ceiling per boat run, ask chains included
                     lastText = "";
                     pendingText = "";
                     problem = null;
@@ -355,7 +356,7 @@ export const BoxAgentDriver = {
         };
         const snapshot = async () => {
             if (!token) {
-                return { state: "unavailable", reason: 'no Box token — add {"box":{"token":"…"}} to ~/.openmausbot/config.json' };
+                return { state: "unavailable", reason: 'no Boat token — add {"box":{"token":"…"}} to ~/.openmausbot/config.json' };
             }
             try {
                 await api("/me");

package/dist-server/server/drivers/builtIn.js

@@ -1,5 +1,5 @@
 import { AntigravityDriver } from "./antigravity.js";
-import { BoxAgentDriver } from "./boxagent.js";
+import { BoatAgentDriver } from "./boatagent.js";
 import { ClaudeDriver } from "./claude.js";
 import { CodexDriver } from "./codex.js";
 import { GrokDriver } from "./grok.js";
@@ -32,7 +32,7 @@ export const BUILT_IN_DRIVERS = [
     ClaudeDriver,
     CodexDriver,
     AntigravityDriver,
-    BoxAgentDriver,
+    BoatAgentDriver,
     MinimaxDriver,
     MistralDriver,
 ];

package/dist-server/server/drivers/{chat-box-tools.js → chat-boat-tools.js}

@@ -1,4 +1,4 @@
-import { isolatedRemoteCommand, runCommand, screenshotBox } from "../box.js";
+import { isolatedRemoteCommand, runCommand, screenshotBoat } from "../boat.js";
 import { createControlClient, CONTROL_REFUSAL_PLAIN } from "../control-client.js";
 const coordinate = { type: "integer", minimum: 0, maximum: 32767 };
 const tool = (name, description, properties = {}, required = []) => ({ name, description, inputSchema: { type: "object", properties, required, additionalProperties: false } });
@@ -19,7 +19,7 @@ const buttons = { left: 1, middle: 2, right: 3 };
 const scrollButtons = { up: 4, down: 5, left: 6, right: 7 };
 /** Uses only the leased descriptor supplied by the harness. Schema validation
  * and user approval happen in the shared chat tool executor before this call. */
-export class ChatBoxClient {
+export class ChatBoatClient {
     closed = false;
     control;
     descriptor;
@@ -45,7 +45,7 @@ export class ChatBoxClient {
         const config = { box: { token: this.descriptor.token } };
         const id = this.descriptor.boxId;
         if (name === "screenshot") {
-            const shot = await screenshotBox(config, "", id, { signal, nativeSize: true });
+            const shot = await screenshotBoat(config, "", id, { signal, nativeSize: true });
             signal.throwIfAborted();
             return { content: [{ type: "image", mimeType: "image/jpeg", data: shot.png }] };
         }

package/dist-server/server/drivers/chat-mcp-tools.js

@@ -7,7 +7,7 @@ import { stripControlPlaneEnv } from "../config.js";
 import { augmentedPath } from "../env-path.js";
 import { killCliTree, spawnCli } from "../procs.js";
 import { chatImage } from "./chat-images.js";
-import { ChatBoxClient } from "./chat-box-tools.js";
+import { ChatBoatClient } from "./chat-boat-tools.js";
 /** The transport cannot safely continue this turn. A dispatched operation may
  * already have taken effect, so callers must not retry it through a new round. */
 export class ChatToolSessionError extends Error {
@@ -314,7 +314,7 @@ export async function mountChatTools(integrations, signal, computerUse = false)
                 throw aborted();
             // Every mounted MCP server can return images when the caller enables
             // image delivery, including custom servers. Text stays bounded below.
-            const client = "boxId" in descriptor ? new ChatBoxClient(descriptor) : new ChatMcpClient(descriptor, computerUse);
+            const client = "boxId" in descriptor ? new ChatBoatClient(descriptor) : new ChatMcpClient(descriptor, computerUse);
             clients.push(client);
             return { name, client, tools: await client.tools(signal) };
         }));

package/dist-server/server/drivers/claude.js

@@ -7,7 +7,7 @@
 // Integrations become MCP servers on the CLI:
 //   - Composio Sessions (connected apps → tools) over streamable HTTP
 //   - the bot's cloud computer (boat.dev) via server/computer-proxy.ts
-//     — screenshot/exec/open_url, the CUA-on-the-box bridge
+//     — screenshot/exec/open_url, the CUA-on-the-boat bridge
 import { createHash, randomBytes } from "node:crypto";
 import { chmodSync, mkdirSync, mkdtempSync, readFileSync, rmSync, unlinkSync, writeFileSync } from "node:fs";
 import { createServer as createNetServer } from "node:net";
@@ -435,7 +435,7 @@ function askQuestions(ask) {
 function askSummary(ask) {
     const questions = askQuestions(ask);
     if (questions)
-        return askQuestionSummary(questions).slice(0, 300);
+        return askQuestionSummary(questions);
     return askInputSummary(ask.input) ?? ask.tool ?? "tool";
 }
 /** Where the hook helper reads this thread's current turn token. Stable per
@@ -752,6 +752,109 @@ function firstText(content) {
     }
     return "";
 }
+/** A turn's own cost from the CLI's total_cost_usd, which is not a per-turn
+ * figure: it is "cumulative across turns in streaming-input sessions — each
+ * result carries the running total so far" (2.1.282), and a retained process
+ * runs turn after turn. So a turn costs the growth since the total its
+ * process reported for the turn before — or, for a process's first turn,
+ * since the total the CLI restored on --resume (see restoredCostBase). With
+ * no known start (null) the turn keeps its whole figure. A total that went
+ * down is not the same count, so it is taken whole too rather than booked as
+ * a negative cost. Rounding to 1e-10 USD removes only the float noise of the
+ * subtraction. */
+export function turnCostFromRunningTotal(total, previous) {
+    if (total === null)
+        return null;
+    if (previous === null || total < previous)
+        return total;
+    return Number((total - previous).toFixed(10));
+}
+export function claudeCostSnapshot(total, modelUsage) {
+    if (typeof total !== "number" || !Number.isFinite(total))
+        return null;
+    const count = (value) => (typeof value === "number" && Number.isFinite(value) ? value : 0);
+    const models = {};
+    if (modelUsage && typeof modelUsage === "object" && !Array.isArray(modelUsage)) {
+        for (const [model, raw] of Object.entries(modelUsage)) {
+            if (!raw || typeof raw !== "object")
+                continue;
+            const u = raw;
+            models[model] = [count(u.inputTokens), count(u.cacheReadInputTokens), count(u.cacheCreationInputTokens), count(u.outputTokens)];
+        }
+    }
+    return { total, models };
+}
+/** The running total a resumed session already carried before this
+ * process's first turn. On --resume the CLI (2.1.282) restores the session's
+ * cost from an earlier state — not always the latest one this driver saw —
+ * so that turn's total_cost_usd and modelUsage include the earlier turns.
+ * The restored state is the earlier state that sits inside the new counts
+ * and leaves exactly this turn's own usage: in one model (usage leaves out
+ * side calls such as a Haiku title) or summed over all models (a turn split
+ * between two); nothing restored is 0. When no state fits exactly — the CLI
+ * saved work that never reported a result, like an interrupted turn — the
+ * latest state inside the new counts stands, so that work is booked once,
+ * with this turn. Either way the latest state wins, not the highest total:
+ * a resume that went back to an older state leaves later, lower totals. */
+export function restoredCostBase(earlier, current, usage) {
+    const turn = [usage.input, usage.cacheRead, usage.cacheWrite, usage.output];
+    const nothing = { total: 0, models: {} };
+    let exact = null;
+    let inside = 0;
+    // oldest first: the session's states in the order they were recorded
+    for (const state of [nothing, ...earlier]) {
+        const within = Object.entries(state.models).every(([model, counts]) => counts.every((n, i) => n <= (current.models[model]?.[i] ?? 0)));
+        if (!within)
+            continue;
+        inside = state.total;
+        const growth = Object.entries(current.models).map(([model, counts]) => counts.map((n, i) => n - (state.models[model]?.[i] ?? 0)));
+        const isTurn = (counts) => counts.every((n, i) => n === turn[i]);
+        const summed = turn.map((_, i) => growth.reduce((sum, counts) => sum + counts[i], 0));
+        if (growth.some(isTurn) || isTurn(summed))
+            exact = state.total;
+    }
+    return exact ?? inside;
+}
+/** Each Claude session's latest cost states, so the first turn after a
+ * --resume can tell what the CLI restored — after an app restart too. Small
+ * by design: a few states for the most recent sessions. */
+const COST_HISTORY_FILE = join(DATA_DIR, "claude-cost-history.json");
+const COST_HISTORY_SESSIONS = 100;
+const COST_HISTORY_STATES = 8;
+function isCostSnapshot(value) {
+    if (!value || typeof value !== "object")
+        return false;
+    const { total, models } = value;
+    return typeof total === "number" && !!models && typeof models === "object" &&
+        Object.values(models).every((counts) => Array.isArray(counts) && counts.length === 4 && counts.every((n) => typeof n === "number"));
+}
+function readCostHistory() {
+    try {
+        const parsed = JSON.parse(readFileSync(COST_HISTORY_FILE, "utf8"));
+        if (!parsed || typeof parsed !== "object" || Array.isArray(parsed))
+            return {};
+        return Object.fromEntries(Object.entries(parsed).map(([id, states]) => [id, Array.isArray(states) ? states.filter(isCostSnapshot) : []]));
+    }
+    catch {
+        return {};
+    }
+}
+function recordCostState(sessionId, state) {
+    const history = readCostHistory();
+    const states = [...(history[sessionId] ?? []), state].slice(-COST_HISTORY_STATES);
+    // most recent session last, so the oldest ones are dropped first
+    delete history[sessionId];
+    history[sessionId] = states;
+    const ids = Object.keys(history);
+    for (const id of ids.slice(0, Math.max(0, ids.length - COST_HISTORY_SESSIONS)))
+        delete history[id];
+    try {
+        writeFileAtomic(COST_HISTORY_FILE, JSON.stringify(history), { mode: 0o600 });
+    }
+    catch {
+        // a lost state only means a later resume keeps its whole figure
+    }
+}
 /** Claude's stream-json input accepts the same image source blocks as the
  * Anthropic Messages API. Keep the old string form for text-only turns so a
  * CLI update cannot disturb the overwhelmingly common path. */
@@ -1155,7 +1258,7 @@ export const ClaudeDriver = {
             mcpServers.ogb = { command: process.execPath, args: [PERM_PROXY_PATH, socketPath], env: { ...NODE_ENV_FLAG }, alwaysLoad: true };
             allowed.push("mcp__ogb");
             // The MCP config carries credentials — a Composio consumer key in a
-            // header, the box token in the computer proxy's env, the comms token in
+            // header, the boat token in the computer proxy's env, the comms token in
             // the agents proxy's env. On argv every one of those is world-readable
             // through `ps` for the life of the turn, to any local process. The CLI
             // accepts a FILE for this flag, so the secrets go in a 0600 file that
@@ -1414,11 +1517,12 @@ export const ClaudeDriver = {
                 idleTimer: null,
                 closing: false,
                 stderr: "",
+                costTotal: undefined,
             };
             sessions.set(threadId, session);
             // settles the TURN, not the process: the CLI stays for the next
             // message until it has been quiet for SESSION_IDLE_MS
-            const settle = (ok, stopReason, cost = null, usage) => {
+            const settle = (ok, stopReason, total = null, usage) => {
                 const t = session.turn;
                 if (!t || t.settled)
                     return;
@@ -1450,11 +1554,31 @@ export const ClaudeDriver = {
                 // retry resumes on a fresh process; healthy sibling sessions stay warm.
                 if (stopReason === "update_required")
                     closeSession(threadId, "update required");
+                // `total` is the CLI's running total for this process; the harness
+                // books turn.completed.cost as this turn's own spend
+                const cost = turnCostFromRunningTotal(total, session.costTotal ?? null);
+                if (total !== null)
+                    session.costTotal = total;
                 emit({ ...base(threadId, t.turnId), type: "turn.completed", ok, stopReason, cost, ...(usage ? { usage } : {}) });
                 if (session.child.exitCode === null && !session.closing)
                     armIdle(threadId);
             };
             const currentTurnId = () => session.turn?.turnId ?? turnId;
+            // The process's first result with a cost says what --resume restored,
+            // which its turns are measured from; every result is kept for a later
+            // resume. A result without one (an API error) decides nothing yet.
+            const noteCostState = (total, modelUsage, usage) => {
+                const state = claudeCostSnapshot(total, modelUsage);
+                if (!state)
+                    return;
+                if (session.costTotal === undefined) {
+                    session.costTotal = session.sessionId
+                        ? restoredCostBase(readCostHistory()[session.sessionId] ?? [], state, usage)
+                        : null;
+                }
+                if (session.sessionId)
+                    recordCostState(session.sessionId, state);
+            };
             const handleLine = (line) => {
                 if (session.closing)
                     return;
@@ -1575,11 +1699,19 @@ export const ClaudeDriver = {
                         // and deny approvals while that user turn is still running.
                         if (o.origin?.kind === "task-notification")
                             break;
-                        // result.usage is this invocation's total — one process per turn,
-                        // so it is the turn's figure. cache reads count as input: they
-                        // are billed (at the cache rate) and they fill the window — but
-                        // they are reported separately too, so the UI can show how much
-                        // of the figure was context re-read rather than new text.
+                        // result.usage is this turn's own figure, "per-turn in
+                        // streaming-input sessions" (2.1.282) even on a retained process.
+                        // cache reads count as input: they are billed (at the cache rate)
+                        // and they fill the window — but they are reported separately
+                        // too, so the UI can show how much of the figure was context
+                        // re-read rather than new text. total_cost_usd is instead the
+                        // process's running total; settle() books this turn's share.
+                        noteCostState(o.total_cost_usd, o.modelUsage, {
+                            input: o.usage?.input_tokens || 0,
+                            cacheRead: o.usage?.cache_read_input_tokens || 0,
+                            cacheWrite: o.usage?.cache_creation_input_tokens || 0,
+                            output: o.usage?.output_tokens || 0,
+                        });
                         settle(o.is_error !== true, session.turn?.authFailed
                             ? "auth_required"
                             : session.turn?.updateRequired
@@ -1877,10 +2009,11 @@ export const ClaudeDriver = {
          * summaries can contain paths, commands, or secrets, so the generic
          * `claude -p "prompt"` shape is not safe for review. No tools or MCP
          * servers are mounted in this isolated process. */
-        const generateReview = (prompt, signal) => new Promise((resolve, reject) => {
-            const child = spawnCli(config.cli, ["-p", "--model", config.managedModels?.[0] ?? "claude-haiku-4-5", "--output-format", "text"], {
+        const generateReview = (prompt, signal, onUsage) => new Promise((resolve, reject) => {
+            const model = config.managedModels?.[0] ?? "claude-haiku-4-5";
+            const child = spawnCli(config.cli, ["-p", "--model", model, "--output-format", onUsage ? "json" : "text"], {
                 stdio: ["pipe", "pipe", "pipe"],
-                env: environment(config.managedModels?.[0] ?? "claude-haiku-4-5"),
+                env: environment(model),
             });
             let stdout = "";
             let stderr = "";
@@ -1919,6 +2052,37 @@ export const ClaudeDriver = {
             });
             child.on("error", (error) => finish(error));
             child.on("close", (code) => {
+                if (settled)
+                    return;
+                if (onUsage) {
+                    try {
+                        const result = JSON.parse(stdout);
+                        if (!result || result.type !== "result")
+                            throw new Error("Claude text generation returned no result");
+                        const count = (value) => typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : undefined;
+                        const cachedInput = count(result.usage?.cache_read_input_tokens);
+                        const inputs = [count(result.usage?.input_tokens), cachedInput, count(result.usage?.cache_creation_input_tokens)]
+                            .filter((value) => value !== undefined);
+                        const models = result.modelUsage && typeof result.modelUsage === "object" && !Array.isArray(result.modelUsage)
+                            ? Object.keys(result.modelUsage) : [];
+                        onUsage({
+                            model: models.length === 1 ? models[0] : models.find(candidate => candidate === model || candidate.startsWith(`${model}-`)) ?? model,
+                            input: inputs.length ? inputs.reduce((sum, value) => sum + value, 0) : undefined,
+                            output: count(result.usage?.output_tokens),
+                            cachedInput,
+                            costUsd: count(result.total_cost_usd),
+                        });
+                        if (result.is_error === true)
+                            throw new Error(typeof result.result === "string" && result.result.trim() ? result.result : stderr.trim() || "Claude text generation failed");
+                        if (typeof result.result !== "string")
+                            throw new Error("Claude text generation returned no text");
+                        stdout = result.result;
+                    }
+                    catch (error) {
+                        finish(error instanceof Error ? error : new Error(String(error)));
+                        return;
+                    }
+                }
                 if (code === 0)
                     finish();
                 else
@@ -2000,7 +2164,7 @@ export const ClaudeDriver = {
                     return () => listeners.delete(listener);
                 },
             },
-            generateText: (prompt, options) => generateReview(prompt, options?.signal),
+            generateText: (prompt, options) => generateReview(prompt, options?.signal, options?.onUsage),
             reviewPermission: generateReview,
             dispose: async () => {
                 try {

package/dist-server/server/drivers/native.js

@@ -11,7 +11,7 @@ export function appendNative(threadId, entry) {
     const file = join(NATIVE_DIR, `${threadId}.ndjson`);
     try {
         // The session-setup messages carry the credentials the agent is handed —
-        // the box and comms tokens ride inside session/new's mcpServers env, and
+        // the boat and comms tokens ride inside session/new's mcpServers env, and
         // an MCP header can carry a Composio key. These files are ordinary
         // 0644 files people paste into bug reports, so values are masked while
         // the shape stays intact.