openmausbot 0.1.77 → 0.1.79

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

@@ -22,9 +22,9 @@
 //                                          never move bots or sections
 //   request_credential(id, reason?)       → show a secure, allowlisted key card
 //   list_routines()                       → inspect this bot's scheduled work
-//   propose_routine(...)                  → show a confirmation card for a new routine
-//   propose_routine_action(...)           → show a confirmation card for a routine change
-//   propose_profile(...)                  → show a confirmation card for a profile change
+//   propose_routine(...)                  → apply or request confirmation for a new routine
+//   propose_routine_action(...)           → apply or request confirmation for a routine change
+//   propose_profile(...)                  → apply or request confirmation for a profile change
 //
 // Speaks raw JSON-RPC 2.0 over stdio (no MCP SDK — house style, matches
 // computer-proxy / permission-proxy). All state comes from env, injected by
@@ -35,6 +35,7 @@
 //   OMB_TURN_DEPTH   this turn's comms depth (the harness refuses recursion)
 import readline from "node:readline";
 import { CREDENTIAL_TARGETS, isCredentialTargetId } from "../../shared/credential-request.js";
+import { normalizeCronSchedule } from "../../shared/routine-schedule.js";
 import { agentToolAnnotations } from "../agent-tool-policy.js";
 import { peerName } from "../peer-roster.js";
 const HARNESS = process.env.OMB_HARNESS_URL ?? "http://127.0.0.1:8799";
@@ -86,12 +87,22 @@ const WEEKDAYS = [
 const ROUTINE_SCHEDULE_SCHEMA = {
     type: "object",
     additionalProperties: false,
-    description: 'Either {"type":"once","at":RFC3339} for one future run, {"type":"weekly","time":"HH:MM","weekdays":[...]} for chosen days, {"type":"daily","time":"HH:MM"} for every day, or {"type":"interval","every_minutes":15} to repeat. Intervals can optionally be limited with weekdays, window_start + window_end, and ends_at.',
+    description: 'Use {"type":"cron","expression":"0 9 1 * *","timeZone":"Asia/Kolkata"} for 09:00 on the first of each month, once with at for one future run, weekly with time + weekdays, daily with time, or interval with every_minutes for elapsed-time repetition. Intervals can optionally be limited with weekdays, window_start + window_end, and ends_at.',
     properties: {
         type: {
             type: "string",
-            enum: ["once", "weekly", "daily", "interval"],
-            description: "once = a single future run; weekly = chosen weekdays; daily = every day; interval = every N minutes.",
+            enum: ["once", "weekly", "daily", "interval", "cron"],
+            description: "once = a single future run; weekly = chosen weekdays; daily = every day; interval = every N elapsed minutes; cron = a calendar rule in an explicit timezone.",
+        },
+        expression: {
+            type: "string",
+            maxLength: 256,
+            description: "Only for cron: five fields, minute hour day-of-month month weekday. Examples: 0 9 1 * * = monthly on day 1 at 09:00; 0 9 L * * = last day of each month; 0 9 * * MON#2 = second Monday of each month. Lists, ranges and steps are supported. No seconds, year, or @ macros. Never substitute daily AI date checks for a calendar rule.",
+        },
+        timeZone: {
+            type: "string",
+            maxLength: 128,
+            description: "Required for cron: explicit IANA timezone, for example Asia/Kolkata, America/New_York, or UTC. Use the user's requested zone; resolve ambiguity before proposing. Do not send a numeric offset or local-time abbreviation.",
         },
         at: {
             type: "string",
@@ -114,7 +125,7 @@ const ROUTINE_SCHEDULE_SCHEMA = {
         },
         starts_at: {
             type: "string",
-            description: "Optional for type interval: RFC3339 date-time with an explicit timezone offset that anchors the cadence. Omit to start one interval after confirmation.",
+            description: "Optional for type interval: RFC3339 date-time with an explicit timezone offset that anchors the cadence. Omit to start one interval after the routine is applied (immediately with granted Full Access, otherwise after confirmation).",
         },
         window_start: {
             type: "string",
@@ -157,7 +168,8 @@ const SHORT_WEEKDAYS = {
 };
 const SUPPORTED_SCHEDULES = 'Supported schedules: {"type":"once","at":"2026-09-01T09:00:00+05:30"} (future RFC3339 with explicit offset), ' +
     '{"type":"weekly","time":"09:00","weekdays":["monday","friday"]}, {"type":"daily","time":"09:00"}, ' +
-    'or {"type":"interval","every_minutes":15,"weekdays":["monday","friday"],"window_start":"09:00","window_end":"17:00"}.';
+    '{"type":"interval","every_minutes":15,"weekdays":["monday","friday"],"window_start":"09:00","window_end":"17:00"}, ' +
+    'or {"type":"cron","expression":"0 9 1 * *","timeZone":"Asia/Kolkata"} (monthly at 09:00 on day 1; five fields and an explicit IANA timezone).';
 /** A schedule as the harness accepts it, or a message telling the model
  * exactly what to send instead. Coercion first, error second: models
  * routinely stringify nested objects, say "daily", or shorten weekday
@@ -182,7 +194,9 @@ function normalizeScheduleInput(args) {
             ? ["type", "time", "weekdays"]
             : type === "interval"
                 ? ["type", "every_minutes", "everyMinutes", "starts_at", "anchorAt", "weekdays", "every_day", "window_start", "window_end", "window", "all_day", "ends_at", "endsAt", "never_ends"]
-                : null;
+                : type === "cron"
+                    ? ["type", "expression", "timeZone"]
+                    : null;
     // Provider conversions may send unused optional fields as null. Ignore
     // those, but never silently discard an actual scheduling constraint (for
     // example timezone or a misspelled starts_at) and approve different work.
@@ -190,6 +204,14 @@ function normalizeScheduleInput(args) {
     if (unsupported) {
         return { error: `Unsupported ${type} schedule field "${unsupported}". Weekly and daily times use the computer's timezone from list_routines. ${SUPPORTED_SCHEDULES}` };
     }
+    if (type === "cron") {
+        try {
+            return { schedule: { ...normalizeCronSchedule({ type, expression: raw.expression, timeZone: raw.timeZone }) } };
+        }
+        catch (error) {
+            return { error: `${error instanceof Error ? error.message : "Invalid cron schedule"}. ${SUPPORTED_SCHEDULES}` };
+        }
+    }
     if (type === "once") {
         if (typeof raw.at !== "string" || !raw.at.trim()) {
             return { error: `A once schedule needs "at": a future RFC3339 date-time with an explicit offset, for example 2026-09-01T09:00:00+05:30.` };
@@ -311,7 +333,7 @@ function normalizeScheduleInput(args) {
             },
         };
     }
-    if (type === "cron" || type === "hourly" || type === "minutes") {
+    if (type === "hourly" || type === "minutes") {
         return { error: `Use an interval schedule for every-N-minutes work. ${SUPPORTED_SCHEDULES}` };
     }
     return { error: `Unknown schedule type "${type || "(missing)"}". ${SUPPORTED_SCHEDULES}` };
@@ -342,9 +364,10 @@ const ROUTINE_FIELDS_SCHEMA = {
     },
     continuity: {
         type: "boolean",
-        description: "Opt in to using the latest completed run's bounded report as historical context. Defaults to false; set false in an update to start fresh again. Shown on the confirmation card.",
+        description: "Opt in to using the latest completed run's bounded report as historical context. Defaults to false; set false in an update to start fresh again. Included in the applied result or pending confirmation.",
     },
 };
+const PROPOSAL_OUTCOME = " Read the result: granted Full Access may apply the change immediately. If applied, continue the requested work without another confirmation. Only a pending result requires ending the turn and waiting for the in-app decision. Never claim success from the permission mode alone; report failed or cancelled results honestly. This does not elevate another bot's execution permissions.";
 const TOOLS = [
     {
         name: "list_shared_computers",
@@ -367,13 +390,14 @@ const TOOLS = [
     },
     {
         name: "coordinate_bots",
-        description: "Ask existing OpenMausBot teammates for advice or assign concrete work. From normal chat each assignment gets a separate recipient conversation; from a room it defaults to this room. Use group_id from list_room_targets for a specific room. Name 1-4 bot_ids: they receive only your brief and use their own model, tools and permissions. Busy bots queue. They can consult their specialists; all results return here and resume you automatically. Include exact file paths, constraints and what must be verified. After sending all assignments, END your turn; do not poll or wait. On return, resolve tradeoffs, verify the requested outcome and request concrete corrections if necessary before giving one final answer. Do not send acknowledgements as new work.",
+        description: "Ask existing OpenMausBot teammates for advice or assign concrete work. From normal chat every assignment you send a teammate continues your one standing conversation with that teammate, so they keep the context of what you asked before; from a room it defaults to this room. Use group_id from list_room_targets for a specific room. Name 1-4 bot_ids: they receive only your brief and use their own model, tools and permissions. Busy bots queue. They can consult their specialists; all results return here and resume you automatically. Include exact file paths, constraints and what must be verified. After sending all assignments, END your turn; do not poll or wait. On return, resolve tradeoffs, verify the requested outcome and request concrete corrections if necessary before giving one final answer. Do not send acknowledgements as new work.",
         inputSchema: { type: "object", additionalProperties: false, properties: {
-                group_id: { type: "string", description: "Optional destination room. Omit for this room, or separate recipient tasks when chatting directly." },
+                group_id: { type: "string", description: "Optional destination room. Omit for this room, or your standing conversation with each teammate when chatting directly." },
                 bot_ids: { type: "array", items: { type: "string" }, minItems: 1, maxItems: 4, uniqueItems: true },
                 message: { type: "string", minLength: 1, maxLength: 4000, description: "Self-contained question or task for these teammates. Send separate requests when responsibilities differ." },
                 request_key: { type: "string", description: "A short unique assignment key. Reuse for an identical retry." },
                 rework: { type: "boolean", description: "True only for concrete additional work from someone who already completed a request." },
+                label: { type: "string", description: "Optional short name (one line, at most 60 characters) for this job. Used only when the teammate is still working on your previous assignment and this one therefore runs in its own thread beside your standing conversation." },
             }, required: ["bot_ids", "message", "request_key"] },
     },
     {
@@ -497,7 +521,7 @@ const TOOLS = [
     },
     {
         name: "propose_team_setup",
-        description: "Chief of Staff only: propose all requested specialist creation, profile/model configuration, and authorized team moves in ONE combined review card. Nothing changes until the user applies it. Use exact catalog engine/model IDs from list_team_setup. Combine all fields for each bot; use the same create key or botId to coalesce repeated entries. New teams must be named explicitly in newTeams and have a specialist in this plan; the card also asks to authorize your access to just those new teams. Existing unauthorized teams cannot be included. Models change bot defaults for groups/new threads; existing threads and execution permissions stay unchanged. After proposing, end your turn. The decision and structured result automatically resume you once; do not ask again, poll, or repeat the proposal.",
+        description: "Chief of Staff only: submit all requested specialist creation, profile/model configuration, and authorized team moves in ONE combined plan. Use exact catalog engine/model IDs from list_team_setup. Combine all fields for each bot; use the same create key or botId to coalesce repeated entries. New teams must be named explicitly in newTeams and have a specialist in this plan; access is granted only to those new teams. Existing unauthorized teams cannot be included. Models change bot defaults for groups/new threads; existing threads and execution permissions stay unchanged. If review is pending, the decision and structured result automatically resume you once; do not ask again, poll, or repeat the proposal." + PROPOSAL_OUTCOME,
         inputSchema: {
             type: "object", additionalProperties: false,
             properties: {
@@ -524,14 +548,14 @@ const TOOLS = [
     },
     {
         name: "propose_bot_deletion",
-        description: "Chief of Staff only: when the user explicitly asks to delete a named teammate, create a separate confirmation card for that exact bot. Deletion removes its conversations, memory, instructions and skills; generated project files remain. Running work and owned computers can block deletion. Never delete yourself, substitute an archive, or put deletion into a setup batch. End your turn after proposing; the decision and result resume you once.",
+        description: "Chief of Staff only: when the user explicitly asks to delete a named teammate, submit a separate deletion request for that exact bot. Deletion removes its conversations, memory, instructions and skills; generated project files remain. Running work and owned computers can block deletion. Never delete yourself, substitute an archive, or put deletion into a setup batch. If review is pending, the decision and result resume you once." + PROPOSAL_OUTCOME,
         inputSchema: { type: "object", additionalProperties: false, properties: {
                 bot_id: { type: "string", minLength: 1 }, reason: { type: "string", minLength: 1, maxLength: 500 },
             }, required: ["bot_id", "reason"] },
     },
     {
         name: "create_room",
-        description: "Create a room in your own section when the user asks for one (maximum four per turn). Chiefs only. Choose active peers from list_bots; you are included automatically as the default responder. This creates no turns or messages. Section moves stay with the user. If peer approval is enabled, ask the user to make the room change instead.",
+        description: "Create a room in your own section when the user asks for one (maximum four per turn). Chiefs only. Choose active peers from list_bots; you are included automatically as the default responder. This creates no turns or messages. Section moves stay with the user. Follow the tool result under the effective access level; if permission is refused, ask the user to make the room change instead, without trying another route.",
         inputSchema: {
             type: "object",
             additionalProperties: false,
@@ -555,7 +579,7 @@ const TOOLS = [
     },
     {
         name: "manage_room",
-        description: "Manage a room from list_rooms: rename it, change its bulletin, or add/remove/set members. Chiefs only, within your own section and allowed peers; keep yourself as a member. Busy rooms, pending approvals and team-goal leads are protected. You cannot move rooms or bots between sections. If peer approval is enabled or the change is refused, ask the user to make the change instead.",
+        description: "Manage a room from list_rooms: rename it, change its bulletin, or add/remove/set members. Chiefs only, within your own section and allowed peers; keep yourself as a member. Busy rooms, pending approvals and team-goal leads are protected. You cannot move rooms or bots between sections. Follow the tool result under the effective access level; if the change is refused, report the blocker and ask the user to make the change instead, without trying another route.",
         inputSchema: {
             type: "object",
             additionalProperties: false,
@@ -663,7 +687,7 @@ const TOOLS = [
     },
     {
         name: "propose_routine",
-        description: "Prepare a new routine after the user explicitly asks to schedule recurring or future work. Call list_routines first for relative dates or times so you use its authoritative current time and timezone. This only creates a durable confirmation card; it does NOT enable the routine. Resolve ambiguous dates, times, timezone, destination, or instructions with the user first, and always give one-time schedules an explicit RFC3339 offset. After calling it, end the turn and do not claim the routine exists until the user confirms the card. If the user asks for the routine to run as ANOTHER bot in your section, call list_bots and pass that bot's id as for_bot_id.",
+        description: "Prepare a new routine after the user explicitly asks to schedule recurring or future work. Call list_routines first for relative dates or times so you use its authoritative current time and timezone. Convert calendar requests (monthly dates, last days, nth weekdays) into a validated five-field cron schedule with an explicit IANA timeZone; keep elapsed every-N-minutes work as interval. Never approximate unsupported requests with a different weekly schedule or an AI date-check routine; explain the limitation instead. Resolve ambiguous dates, times, timezone, destination, or instructions with the user first, and always give one-time schedules an explicit RFC3339 offset. If the user asks for the routine to run as ANOTHER bot in your section, call list_bots and pass that bot's id as for_bot_id; each run retains that bot's own permissions." + PROPOSAL_OUTCOME,
         inputSchema: {
             type: "object",
             additionalProperties: false,
@@ -679,7 +703,7 @@ const TOOLS = [
     },
     {
         name: "propose_routine_action",
-        description: "Prepare a user-requested change to one of this bot's existing routines. This only creates a durable confirmation card; it does NOT apply the change. Use list_routines first to get the routine id. After calling it, end the turn and do not claim the action completed until the user confirms the card.",
+        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,
         inputSchema: {
             type: "object",
             additionalProperties: false,
@@ -702,7 +726,7 @@ const TOOLS = [
     },
     {
         name: "propose_profile",
-        description: "Propose changes to your own name, title, description, standing instructions (SOUL.md), or working folder (cwd). This only creates a confirmation card; nothing changes until the user approves it. After calling it, end the turn and do not claim the change is applied. Keep SOUL.md short — who you are and the rules you never break; put step-by-step procedure into a skill instead. A Chief of Staff may pass for_bot_id (from list_bots) to propose a change for another bot in its section.",
+        description: "Submit user-requested changes to your own name, title, description, standing instructions (SOUL.md), or working folder (cwd). Keep SOUL.md short — who you are and the rules you never break; put step-by-step procedure into a skill instead. A Chief of Staff may pass for_bot_id (from list_bots) for a requested change to another bot in its section." + PROPOSAL_OUTCOME,
         inputSchema: {
             type: "object",
             additionalProperties: false,
@@ -732,7 +756,7 @@ const TOOLS = [
     },
     {
         name: "skill_manage",
-        description: "Stage a new or updated reusable SKILL.md for the user to review. Create stays inactive until approval; update leaves the current version unchanged until approval. Never update unless the user explicitly asked to revise that named skill. After calling this, end the turn and wait for the in-app decision.",
+        description: "Submit a new or updated reusable SKILL.md. Never update unless the user explicitly asked to revise that named skill. While review is pending, a create stays inactive and an update leaves the current version unchanged." + PROPOSAL_OUTCOME,
         inputSchema: {
             type: "object",
             additionalProperties: false,
@@ -752,7 +776,7 @@ const TOOLS = [
                 },
                 gist: {
                     type: "string",
-                    description: "Optional one-line summary shown on the user's confirmation card.",
+                    description: "Optional one-line summary of the skill change, included in its applied result or pending review.",
                 },
                 source: {
                     type: "string",
@@ -877,7 +901,26 @@ function routineFields(args) {
         fields.continuity = args.continuity;
     return { fields };
 }
+/** Full Access is decided by the harness, not inferred from a model claim or
+ * local environment flag. Missing state preserves older pending responses. */
+function completedProposalResult(r, subject) {
+    const state = r.state;
+    const result = jsonRecord(r.result) ? r.result : undefined;
+    const error = typeof r.error === "string" ? r.error : typeof result?.error === "string" ? result.error : undefined;
+    const attention = error ?? (r.settlementPending && typeof r.message === "string" ? r.message : undefined);
+    if ((!state || state === "pending") && !error)
+        return undefined;
+    const summary = typeof r.summary === "string" && r.summary.trim() ? `\n\n${r.summary.trim()}` : "";
+    const details = result ? `\n\nResult: ${JSON.stringify(result)}` : "";
+    if (state !== "applied" || (result?.state !== undefined && result.state !== "applied")) {
+        return { text: `The request for ${subject} did not complete successfully.${error ? ` ${error}` : ""}${summary}${details}\n\nDo not claim it was applied. Address the reported blocker rather than repeating the request or asking for a duplicate confirmation.`, isError: true };
+    }
+    return { text: `Applied ${subject}.${summary}${details}${attention ? `\n\nNeeds attention: ${attention}` : ""}\n\nNo additional confirmation is needed. Continue the requested work; do not wait for a review card or ask the user to approve this change again.` };
+}
 function confirmationResult(r, fallback, noun = "routine") {
+    const completed = completedProposalResult(r, fallback);
+    if (completed)
+        return completed;
     const summary = typeof r.summary === "string" && r.summary.trim() ? `\n\n${r.summary.trim()}` : "";
     return {
         text: `A confirmation card is now visible to the user for ${fallback}.${summary}\n\nThis change has not been applied yet. End this turn and wait for the user to confirm or deny the card; do not claim the ${noun} was created or changed before confirmation.`,
@@ -903,7 +946,7 @@ async function callTool(name, args) {
     if (name === "coordinate_bots") {
         const r = await api("/api/internal/coordinate-bots", { method: "POST", body: JSON.stringify({
                 groupId: args.group_id, botIds: args.bot_ids, message: args.message,
-                requestKey: args.request_key, rework: args.rework,
+                requestKey: args.request_key, rework: args.rework, label: args.label,
             }) });
         return { text: JSON.stringify(r), ...(r.error ? { isError: true } : {}) };
     }
@@ -1177,6 +1220,9 @@ async function callTool(name, args) {
                 ...(deleting ? { targetBotId: args.bot_id, reason: args.reason } : { plan: args }),
             }),
         });
+        const completed = completedProposalResult(result, deleting ? "the requested bot deletion" : "the requested team setup");
+        if (completed)
+            return completed;
         return { text: `One review card is visible: ${String(result.title)}. Nothing has been applied. End this turn; the decision and structured result resume you automatically once. Do not ask again, poll, or repeat this proposal.` };
     }
     if (name === "create_bot") {
@@ -1493,7 +1539,7 @@ async function callTool(name, args) {
         const skills = Array.isArray(r.skills) ? r.skills : [];
         const staged = Array.isArray(r.staged) ? r.staged : [];
         if (!skills.length && !staged.length) {
-            return { text: "This bot has no imported skills and nothing staged. Use skill_manage action=\"create\" to stage one for the user to confirm." };
+            return { text: "This bot has no imported skills and nothing staged. Use skill_manage action=\"create\" for a user-requested skill, then follow its applied or pending result." };
         }
         const live = skills.length
             ? skills.map((skill) => {
@@ -1546,7 +1592,10 @@ async function callTool(name, args) {
             }),
         });
         const nameLabel = typeof r.name === "string" ? r.name : "the skill";
-        const warningText = Array.isArray(r.warnings) && r.warnings.length ? `\n\nScan warnings (shown to the user):\n- ${r.warnings.join("\n- ")}` : "";
+        const warningText = Array.isArray(r.warnings) && r.warnings.length ? `\n\nScan warnings:\n- ${r.warnings.join("\n- ")}` : "";
+        const completed = completedProposalResult(r, args.action === "update" ? `the update to skill “${nameLabel}”` : `the new skill “${nameLabel}”`);
+        if (completed)
+            return { ...completed, text: completed.text + warningText };
         const status = args.action === "update"
             ? "The current version remains unchanged until the user reviews and applies the update."
             : "The skill is staged and inactive until the user reviews and enables it.";

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

@@ -1,7 +1,8 @@
 import { newEventId, newId } from "../contracts.js";
 import { appendNative } from "./native.js";
 const DRIVER_KIND = "boxAgent";
-const BOX_API = "https://ascii.dev/api/box/v1";
+// 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 MODELS = {
     default: "claude-fable-5",
     options: [
@@ -10,7 +11,28 @@ const MODELS = {
         { id: "gpt-5.4", label: "GPT-5.4 (Codex) · on the box" },
     ],
 };
-const providerFor = (model) => (model.startsWith("gpt") ? "codex" : "claude-code");
+/** The box runs every harness ascii.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\/?$/, "");
+    catalog = await fetch(`${root}/api/provider-models`, { signal: AbortSignal.timeout(15_000) })
+        .then((res) => (res.ok ? res.json() : null))
+        .catch(() => null);
+}
+const providerFor = (model) => {
+    const slash = model.indexOf("/");
+    if (slash > 0 && catalog?.[model.slice(0, slash)])
+        return { provider: model.slice(0, slash), model: model.slice(slash + 1) };
+    for (const [provider, harness] of Object.entries(catalog ?? {})) {
+        if (harness.models?.some((m) => m?.id === model))
+            return { provider, model };
+    }
+    return { provider: model.startsWith("gpt") ? "codex" : "claude-code", model };
+};
 function decodeConfig(raw) {
     const o = (raw ?? {});
     return { pollMs: typeof o.pollMs === "number" ? o.pollMs : 2500 };
@@ -70,9 +92,11 @@ export const BoxAgentDriver = {
             ]
                 .filter((s) => s !== undefined)
                 .join("\n");
+            if (!catalog)
+                await loadCatalog();
             const started = await api(`/boxes/${boxId}/prompt`, {
                 method: "POST",
-                body: JSON.stringify({ provider: providerFor(model), model, prompt }),
+                body: JSON.stringify({ ...providerFor(model), prompt }),
             });
             appendNative(threadId, { dir: "out", source: "box.prompt", msg: { model, prompt, response: started } });
             // real shape (2026-08): {type:"prompt.queued", promptId, promptRun:{id,…},
@@ -95,6 +119,8 @@ export const BoxAgentDriver = {
                 const startedAt = Date.now();
                 let lastText = "";
                 let pendingText = "";
+                /** Why the box could not answer (login expired, model refused, …). */
+                let problem = null;
                 /** Emit unflushed deltas as assistant_text and reset pendingText. */
                 const flushAssistantText = () => {
                     const text = pendingText;
@@ -120,6 +146,11 @@ export const BoxAgentDriver = {
                         const events = await api(`/boxes/${boxId}/events`).catch(() => null);
                         const list = events?.events ?? events?.items ?? [];
                         for (const ev of list) {
+                            // The stream is the whole conversation from its start; `taskId`
+                            // names the prompt run an event belongs to. Earlier turns are
+                            // history, not this answer.
+                            if (promptId && ev.taskId && String(ev.taskId) !== promptId)
+                                continue;
                             const id = String(ev.id ?? ev.eventId ?? JSON.stringify(ev).slice(0, 120));
                             if (seen.has(id))
                                 continue;
@@ -135,6 +166,11 @@ export const BoxAgentDriver = {
                             if (/assistant|message|output|response/i.test(kind) && typeof text === "string" && text.trim()) {
                                 ingest(text);
                             }
+                            else if (/usage_limit|error|fail/i.test(kind)) {
+                                const why = ev.data?.summary ?? ev.data?.message ?? ev.data?.error ?? ev.message;
+                                if (typeof why === "string" && why.trim())
+                                    problem = why.trim();
+                            }
                             else if (/tool|command|exec|browse/i.test(kind)) {
                                 flushAssistantText();
                                 emit({
@@ -168,14 +204,22 @@ export const BoxAgentDriver = {
                                 if (typeof result === "string" && result.trim() && result !== lastText) {
                                     ingest(result);
                                 }
-                                if (!pendingText.trim() && !lastText.trim())
+                                if (!pendingText.trim() && !lastText.trim()) {
+                                    // a run that ends with nothing said and a recorded problem
+                                    // (login expired, …) is a failure the person must see
+                                    if (problem)
+                                        throw new Error(problem);
                                     pendingText = "(finished)";
+                                }
                                 flushAssistantText();
                                 active.delete(threadId);
                                 emit({ ...base(threadId, turnId), type: "turn.completed", ok: true, stopReason: null, cost: null });
                                 return;
                             }
                             if (/failed|error|cancelled|interrupted/i.test(state)) {
+                                const runError = [run?.error, run?.failureReason, run?.message].find((v) => typeof v === "string" && v.trim());
+                                if (problem || runError || /failed|error/i.test(state))
+                                    throw new Error(problem ?? runError ?? `the box run ${state}`);
                                 flushAssistantText();
                                 active.delete(threadId);
                                 emit({ ...base(threadId, turnId), type: "turn.completed", ok: false, stopReason: state, cost: null });