cascade-ai 0.53.0 → 0.55.0

package/dist/index.js

@@ -41,6 +41,12 @@ import cron from 'node-cron';
 // Cascade AI — Multi-tier AI Orchestration System
 var __defProp = Object.defineProperty;
 var __getOwnPropNames = Object.getOwnPropertyNames;
+var __require = /* @__PURE__ */ ((x) => typeof require !== "undefined" ? require : typeof Proxy !== "undefined" ? new Proxy(x, {
+  get: (a, b) => (typeof require !== "undefined" ? require : a)[b]
+}) : x)(function(x) {
+  if (typeof require !== "undefined") return require.apply(this, arguments);
+  throw Error('Dynamic require of "' + x + '" is not supported');
+});
 var __esm = (fn, res) => function __init() {
   return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
 };
@@ -1845,7 +1851,7 @@ function reconcilePrice(model, live) {
 }
 
 // src/constants.ts
-var CASCADE_VERSION = "0.53.0";
+var CASCADE_VERSION = "0.55.0";
 var CASCADE_CONFIG_DIR = ".cascade";
 var CASCADE_MD_FILE = "CASCADE.md";
 var CASCADE_IGNORE_FILE = ".cascadeignore";
@@ -3658,7 +3664,21 @@ var ModelSelector = class {
     this.availableProviders = availableProviders;
     this.availableModels = new Map(Object.entries(MODELS));
   }
+  /**
+   * Verdicts from earlier runs about ids that 404 on this account. Consulted
+   * BEFORE a model enters the pool, which is the only placement that helps: a
+   * T3 wave runs concurrently, so filtering after the first failure still lets
+   * every other worker in the wave call the dead id.
+   */
+  deadModels;
+  setDeadModelStore(store) {
+    this.deadModels = store;
+    for (const [id, m] of [...this.availableModels]) {
+      if (this.deadModels.isDead(m.provider, m.id)) this.availableModels.delete(id);
+    }
+  }
   addDynamicModel(model) {
+    if (this.deadModels?.isDead(model.provider, model.id)) return;
     this.availableModels.set(model.id, model);
     if (!(model.id in MODELS)) this.discovered.add(normalizeModelId(model.id));
   }
@@ -4065,6 +4085,121 @@ var LocalRequestQueue = class {
   }
 };
 
+// src/core/router/dead-models.ts
+var DEAD_MODEL_TTL_MS = 7 * 24 * 60 * 60 * 1e3;
+function deadModelKey(provider, modelId) {
+  return `${provider}:${modelId}`;
+}
+var DeadModelStore = class {
+  constructor(persistence, opts = {}) {
+    this.persistence = persistence;
+    this.ttlMs = opts.ttlMs ?? DEAD_MODEL_TTL_MS;
+    this.now = opts.now ?? (() => Date.now());
+    this.restore();
+  }
+  persistence;
+  records = /* @__PURE__ */ new Map();
+  ttlMs;
+  now;
+  restore() {
+    if (!this.persistence) return;
+    try {
+      for (const r of this.persistence.load()) {
+        if (!this.isExpired(r)) this.records.set(r.key, r);
+      }
+    } catch {
+    }
+  }
+  isExpired(r) {
+    return this.now() - r.lastSeenAt > this.ttlMs;
+  }
+  flush() {
+    if (!this.persistence) return;
+    try {
+      this.persistence.save([...this.records.values()]);
+    } catch {
+    }
+  }
+  /**
+   * Record that a model was found dead. Returns true when this is the first
+   * time (so the caller can log it once rather than once per concurrent
+   * worker in the wave that discovered it).
+   */
+  record(provider, modelId, reason) {
+    const key = deadModelKey(provider, modelId);
+    const now = this.now();
+    const existing = this.records.get(key);
+    if (existing && !this.isExpired(existing)) {
+      existing.lastSeenAt = now;
+      existing.hits += 1;
+      this.flush();
+      return false;
+    }
+    this.records.set(key, {
+      key,
+      provider,
+      modelId,
+      reason,
+      firstSeenAt: now,
+      lastSeenAt: now,
+      hits: 1
+    });
+    this.flush();
+    return true;
+  }
+  /** True when this model is known dead and the verdict has not expired. */
+  isDead(provider, modelId) {
+    const r = this.records.get(deadModelKey(provider, modelId));
+    if (!r) return false;
+    if (this.isExpired(r)) {
+      this.records.delete(r.key);
+      this.flush();
+      return false;
+    }
+    return true;
+  }
+  /** Why a model is considered dead, for /why and the models command. */
+  reasonFor(provider, modelId) {
+    const r = this.records.get(deadModelKey(provider, modelId));
+    return r && !this.isExpired(r) ? r.reason : null;
+  }
+  /** Live records, expired ones filtered out. */
+  list() {
+    const live = [...this.records.values()].filter((r) => !this.isExpired(r));
+    return live.sort((a, b) => b.lastSeenAt - a.lastSeenAt);
+  }
+  /**
+   * Forget a verdict — used when the user fixes the underlying problem (adds
+   * the model to their project, gets quota) and wants it retried now rather
+   * than after the TTL.
+   */
+  forget(provider, modelId) {
+    const removed = this.records.delete(deadModelKey(provider, modelId));
+    if (removed) this.flush();
+    return removed;
+  }
+  clear() {
+    this.records.clear();
+    this.flush();
+  }
+};
+function fileDeadModelPersistence(filePath) {
+  return {
+    load() {
+      const fs25 = __require("fs");
+      if (!fs25.existsSync(filePath)) return [];
+      const raw = JSON.parse(fs25.readFileSync(filePath, "utf-8"));
+      return Array.isArray(raw) ? raw : [];
+    },
+    save(records) {
+      const fs25 = __require("fs");
+      const path26 = __require("path");
+      fs25.mkdirSync(path26.dirname(filePath), { recursive: true });
+      fs25.writeFileSync(filePath, JSON.stringify(records, null, 2));
+    }
+  };
+}
+
 // src/utils/retry.ts
 var CascadeCancelledError = class extends Error {
   constructor(reason) {
@@ -4654,10 +4789,47 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter {
    */
   budgetState = "ok";
   budgetExceededReason;
+  /**
+   * Whether the run/session budget is permanently tripped, and why.
+   *
+   * A tier or worker that catches `BudgetExceededError` internally (T3Worker
+   * does, so one dead-end subtask doesn't crash the whole run) absorbs the
+   * exception into a normal FAILED/PARTIAL result — the tier returns as if
+   * nothing unusual happened. That silently skips `Cascade.run()`'s dedicated
+   * budget-exceeded handling (the `run:budget-exceeded` event and
+   * `lastInterruptedRun` for `/continue`). This lets a caller check the
+   * router's own state directly after a tier returns, regardless of how that
+   * tier chose to report the error internally.
+   */
+  budgetExceededInfo() {
+    if (this.runBudgetExceeded) return { reason: this.runBudgetExceededReason ?? "Per-task budget exceeded." };
+    if (this.budgetState === "exceeded") return { reason: this.budgetExceededReason ?? "Session budget exceeded." };
+    return null;
+  }
   tpmLimiter;
   localQueue;
   taskAnalyzer;
   pendingFeedbackSource;
+  /** Models known dead for this account (404 / retired / not enabled). */
+  deadModels = new DeadModelStore();
+  /**
+   * Back the dead-model memory with durable storage (a JSON file on desktop, a
+   * per-user table in the cloud). Without one the verdicts still work, but only
+   * for the life of the process — so tomorrow's run pays the same wave of 404s
+   * to learn the same fact.
+   */
+  setDeadModelStore(store) {
+    this.deadModels = store;
+    this.selector.setDeadModelStore(store);
+  }
+  /** Models known dead for this account, for `cascade models` and /why. */
+  listDeadModels() {
+    return this.deadModels.list();
+  }
+  /** Retry a model the user has since fixed, without waiting out the TTL. */
+  forgetDeadModel(provider, modelId) {
+    return this.deadModels.forget(provider, modelId);
+  }
   worldStateDB;
   privacyPaths;
   guidanceQueue;
@@ -5022,6 +5194,8 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter {
       }
       if (isModelNotFoundError(errMsg)) {
         this.selector.removeModel(model.id);
+        const firstSighting = this.deadModels.record(model.provider, model.id, errMsg);
+        if (firstSighting) this.emit("model:dead", { provider: model.provider, modelId: model.id, reason: errMsg });
         const depth = (options._notFoundDepth ?? 0) + 1;
         const next = depth <= 3 ? this.selector.selectForTier(tier) : null;
         if (next && next.id !== model.id) {
@@ -6289,6 +6463,13 @@ Now execute your subtask using this context where relevant.`
         this.peerBus?.publish(this.id, assignment.subtaskId, finalOutput2, "FAILED");
         return this.buildResult("ESCALATED", finalOutput2, { checksRun, passed, failed }, issues, correctionAttempts);
       }
+      if (err instanceof Error && err.name === "BudgetExceededError") {
+        issues.push(errMsg);
+        const stopped = output || errMsg;
+        this.setStatus("FAILED", stopped);
+        this.peerBus?.publish(this.id, assignment.subtaskId, stopped, "FAILED");
+        return this.buildResult("FAILED", stopped, { checksRun, passed, failed }, issues, correctionAttempts);
+      }
       issues.push(`Execution error: ${errMsg}`);
       const finalOutput = output || errMsg;
       this.setStatus("FAILED", finalOutput);
@@ -7310,6 +7491,26 @@ var RedactionLayer = class {
   }
 };
 
+// src/core/tiers/escalation-policy.ts
+function sectionNeedsDecision(results) {
+  return results.some((r) => r.status === "ESCALATED");
+}
+function settledEscalationStatus(results) {
+  const hasWork = results.some((r) => r.status === "COMPLETED" || !!r.output);
+  return hasWork ? "PARTIAL" : "FAILED";
+}
+function composeRootSectionOutput(t2Result, completedOutputs) {
+  let output = `${t2Result.sectionSummary}
+
+${completedOutputs.join("\n\n")}`;
+  if (t2Result.status === "FAILED" && t2Result.issues?.length) {
+    output += `
+
+_(incomplete: ${t2Result.issues.join("; ")})_`;
+  }
+  return output;
+}
+
 // src/core/tiers/t2-manager.ts
 function buildT2SystemPrompt(hasPeerMessage) {
   return [
@@ -7335,6 +7536,19 @@ var T2Manager = class extends BaseTier {
   t2PeerBus;
   permissionEscalator;
   toolCreator;
+  /**
+   * Asked when a section ends ESCALATED — i.e. a worker hit something it could
+   * not decide alone. Without this the escalation was a dead end: the status
+   * said "needs a decision" and no decision was ever requested, so an MCP run
+   * that legitimately needed input just stopped there.
+   */
+  escalationCallback;
+  /**
+   * The one allowed retry has been spent. Retrying is bounded because an
+   * escalation loop that can re-ask forever burns a budget on a question that
+   * is not resolving; a second escalation is terminal instead.
+   */
+  escalationRetryUsed = false;
   /** Optional boardroom gate (Moderate / root-T2 runs) — pauses after decomposition. */
   planApprovalCallback;
   /** AbortController for the current T3 wave — aborted on cancel-and-respawn */
@@ -7383,6 +7597,10 @@ var T2Manager = class extends BaseTier {
   setToolCreator(creator) {
     this.toolCreator = creator;
   }
+  /** Ask the user what to do when a section escalates. */
+  setEscalationCallback(cb) {
+    this.escalationCallback = cb;
+  }
   /** Boardroom gate for Moderate (root-T2) runs: pause after decomposition. */
   setPlanApprovalCallback(cb) {
     this.planApprovalCallback = cb;
@@ -7487,9 +7705,57 @@ Guidance (must be followed): ${decision.note}`
         currentAction: "Aggregating T3 results",
         status: "IN_PROGRESS"
       });
-      const summary = await this.aggregateResults(assignment, t3Results);
+      let summary = await this.aggregateResults(assignment, t3Results);
       const issues = t3Results.filter((r) => r.status !== "COMPLETED").flatMap((r) => r.issues);
-      const overallStatus = this.determineStatus(t3Results);
+      let overallStatus = this.determineStatus(t3Results);
+      let userSkipped = false;
+      const hasEscalated = sectionNeedsDecision(t3Results);
+      const settleEscalated = async (reason) => {
+        if (reason) issues.push(reason);
+        summary = await this.aggregateResults(assignment, t3Results, { includeEscalated: true });
+        overallStatus = settledEscalationStatus(t3Results);
+      };
+      if (hasEscalated && this.escalationCallback && this.escalationRetryUsed) {
+        await settleEscalated("Escalated again after the retry \u2014 no further attempts were made.");
+      } else if (hasEscalated && !this.escalationCallback) {
+        await settleEscalated();
+      } else if (hasEscalated && this.escalationCallback) {
+        this.sendStatusUpdate({
+          progressPct: 95,
+          currentAction: "Escalated \u2014 waiting for your decision",
+          status: "ESCALATING"
+        });
+        const decision = await this.escalationCallback({
+          sectionId: assignment.sectionId,
+          sectionTitle: assignment.sectionTitle,
+          issues,
+          summary
+        });
+        this.log(`Escalation decision for "${assignment.sectionTitle}": ${decision.action}`);
+        if (decision.action === "skip") {
+          summary = await this.aggregateResults(assignment, t3Results, { includeEscalated: true });
+          overallStatus = settledEscalationStatus(t3Results);
+          userSkipped = decision.automatic !== true;
+        } else if (decision.action === "retry" || decision.action === "guidance") {
+          this.escalationRetryUsed = true;
+          const guided = decision.action === "guidance" && decision.note?.trim() ? {
+            ...assignment,
+            description: `${assignment.description}
+
+Guidance (must be followed): ${decision.note.trim()}`,
+            // Drop T1's preplanned subtasks so the retry re-decomposes with
+            // the guidance in hand. The workers execute from their subtask
+            // slices, not from this description — leaving the slices in
+            // place gave "Retry with guidance" byte-identical worker
+            // prompts to "Retry as-is", making the note do nothing at all.
+            t3Subtasks: []
+          } : assignment;
+          return await this.execute(guided, taskId, signal);
+        } else {
+          issues.push("Escalated, but no decision was received in time.");
+          overallStatus = "FAILED";
+        }
+      }
       const isOk = overallStatus === "COMPLETED" || overallStatus === "PARTIAL";
       this.setStatus(isOk ? "COMPLETED" : "FAILED", summary);
       const sectionAction = overallStatus === "COMPLETED" ? "Section complete" : overallStatus === "PARTIAL" ? "Section partially complete \u2014 some subtasks failed" : overallStatus === "ESCALATED" ? "Section escalated \u2014 needs a decision" : "Section failed \u2014 no subtask completed";
@@ -7500,7 +7766,8 @@ Guidance (must be followed): ${decision.note}`
         status: overallStatus,
         t3Results,
         sectionSummary: summary,
-        issues
+        issues,
+        ...userSkipped ? { userSkipped: true } : {}
       };
       this.publishSectionOutput(result);
       return result;
@@ -7934,8 +8201,10 @@ HIERARCHY CONTEXT: ${this.hierarchyContext}` : ""),
       result.status === "COMPLETED" ? "COMPLETED" : "FAILED"
     );
   }
-  async aggregateResults(assignment, results) {
-    const completed = results.filter((r) => r.status === "COMPLETED");
+  async aggregateResults(assignment, results, opts = {}) {
+    const completed = results.filter(
+      (r) => r.status === "COMPLETED" || opts.includeEscalated && r.status === "ESCALATED" && r.output
+    );
     if (!completed.length) return `Section ${assignment.sectionTitle} failed \u2014 no T3 workers completed.`;
     const peerOutputs = this.peerSyncBuffer.filter((p) => p.content?.type === "T2_SECTION_OUTPUT").map((p) => `[Peer ${p.fromId} Output]: ${p.content.output}`).join("\n\n");
     const peerContext = peerOutputs ? `
@@ -8141,6 +8410,7 @@ var T1Administrator = class extends BaseTier {
   escalations = [];
   store;
   runBreaker;
+  escalationCallback;
   t2PeerBus = new PeerBus();
   permissionEscalator;
   toolCreator;
@@ -8155,6 +8425,10 @@ var T1Administrator = class extends BaseTier {
     this.toolRegistry = toolRegistry;
     this.config = config;
   }
+  /** Ask the user what to do when any section escalates. */
+  setEscalationCallback(cb) {
+    this.escalationCallback = cb;
+  }
   /** Share the run-wide circuit breaker with every section manager. */
   setRunBreaker(breaker) {
     this.runBreaker = breaker;
@@ -8314,8 +8588,11 @@ Create a CORRECTION PLAN that contains only the new sections needed to fix the i
         reason: `Some T2 managers failed entirely: ${failedSections.map((s) => s.sectionTitle).join(", ")}. Errors: ${failedSections.flatMap((s) => s.issues).join("; ")}`
       };
     }
-    const sectionsText = t2Results.map((r) => `**${r.sectionTitle}**
-${r.sectionSummary}`).join("\n\n");
+    const sectionsText = t2Results.map((r) => {
+      const skipNote = r.userSkipped ? "\n[The user was asked to retry or skip this section and explicitly chose to SKIP it. This is an intentional decision, not a failure \u2014 do not cite it as a reason to reject.]" : "";
+      return `**${r.sectionTitle}**
+${r.sectionSummary}${skipNote}`;
+    }).join("\n\n");
     const prompt = `You are a strict QA Reviewer for the Cascade AI system.
 Review the following execution outputs against the original user prompt.
 
@@ -8325,6 +8602,7 @@ T2 Manager Summaries:
 ${sectionsText}
 
 Does the current state of the workspace and the outputs fully satisfy the user's request?
+A section marked as user-skipped above is accepted as-is by the user's own choice \u2014 treat it as satisfied, not as a gap.
 If yes, reply with exactly: "APPROVED".
 If no, reply with "REJECTED: [Detailed reason explaining exactly what is missing or incorrect]".`;
     try {
@@ -8522,6 +8800,7 @@ SPEC RULES \u2014 each subtask is a self-contained spec slice (workers execute f
     const managers = sections.map((section) => {
       const manager = new T2Manager(this.router, this.toolRegistry, this.id);
       if (this.runBreaker) manager.setRunBreaker(this.runBreaker);
+      if (this.escalationCallback) manager.setEscalationCallback(this.escalationCallback);
       manager.setHierarchyContext(`You are a T2 Manager for the section "${section.sectionTitle}". You are part of a COMPLEX task overseen by T1 Administrator.`);
       if (this.store) {
         manager.setStore(this.store);
@@ -8731,7 +9010,10 @@ SPEC RULES \u2014 each subtask is a self-contained spec slice (workers execute f
   async compileFinalOutput(originalPrompt, plan, t2Results) {
     const completedSections = t2Results.filter((r) => r.status !== "FAILED");
     if (!completedSections.length) {
-      const allIssues = t2Results.flatMap((r) => r.t3Results.flatMap((t) => t.issues));
+      const allIssues = [
+        ...t2Results.flatMap((r) => r.issues ?? []),
+        ...t2Results.flatMap((r) => r.t3Results.flatMap((t) => t.issues))
+      ].filter(Boolean);
       const critical = allIssues.find((i) => i.includes("[CRITICAL_TOOL_ERROR]"));
       const stalled = allIssues.find((i) => /^Stalled:/.test(i));
       const topReason = critical ?? stalled ?? allIssues[0] ?? "no specific reason recorded";
@@ -8817,9 +9099,16 @@ var CAPABILITIES = [
     api: "openai-images"
   },
   {
+    // Imagen, not `gemini-3-pro-image`. That id appears on Google's PRICING
+    // page but 404s on `:predict` ("not found for API version v1beta, or is not
+    // supported for predict") — the Gemini image models are called through
+    // generateContent with an image response modality, while `:predict` is the
+    // Imagen shape this executor implements. A billing line item is not proof
+    // of a callable endpoint, which is the assumption that put the wrong id
+    // here in the first place.
     modality: "image",
     provider: "gemini",
-    modelId: "gemini-3-pro-image",
+    modelId: "imagen-4.0-generate-001",
     api: "gemini-predict"
   },
   // ── Speech (text → audio) ──
@@ -9533,6 +9822,102 @@ var RunBreaker = class {
     return out;
   }
 };
+var MCP_TOOL_PREFIX = "mcp__";
+var MAX_SEGMENT_LENGTH = 24;
+var SEGMENT_HASH_LENGTH = 8;
+function boundedSegment(raw) {
+  const clean = sanitizeToolNameSegment(raw);
+  if (clean.length <= MAX_SEGMENT_LENGTH) return clean;
+  const hash = createHash("sha256").update(raw, "utf8").digest("hex").slice(0, SEGMENT_HASH_LENGTH);
+  return `${clean.slice(0, MAX_SEGMENT_LENGTH - SEGMENT_HASH_LENGTH - 1)}_${hash}`;
+}
+function isStableSegment(raw) {
+  return sanitizeToolNameSegment(raw) === raw && raw.length <= MAX_SEGMENT_LENGTH;
+}
+function sanitizeToolNameSegment(segment) {
+  const cleaned = segment.replace(/[^a-zA-Z0-9-]+/g, "_");
+  let start = 0;
+  let end = cleaned.length;
+  while (start < end && cleaned[start] === "_") start++;
+  while (end > start && cleaned[end - 1] === "_") end--;
+  return cleaned.slice(start, end) || "_";
+}
+function mcpToolName(serverName, toolName) {
+  return `${MCP_TOOL_PREFIX}${boundedSegment(serverName)}__${boundedSegment(toolName)}`;
+}
+function isMcpToolName(name) {
+  return name.startsWith(MCP_TOOL_PREFIX);
+}
+function uniqueMcpServerName(desired, existingNames) {
+  const taken = new Set(existingNames.map(mcpServerPrefix));
+  if (!taken.has(mcpServerPrefix(desired))) return desired;
+  for (let n = 2; n < 1e3; n++) {
+    const candidate = `${desired} (${n})`;
+    if (!taken.has(mcpServerPrefix(candidate))) return candidate;
+  }
+  return `${desired} (${randomUUID().slice(0, 8)})`;
+}
+function disambiguateMcpServerNames(servers) {
+  const seenNames = [];
+  const renames = [];
+  const result = servers.map((s) => {
+    const name = uniqueMcpServerName(s.name, seenNames);
+    seenNames.push(name);
+    if (name === s.name) return s;
+    renames.push({ from: s.name, to: name });
+    return { ...s, name };
+  });
+  return { servers: renames.length ? result : servers, renames };
+}
+function assignMcpToolNames(tools) {
+  const names = new Array(tools.length);
+  const bases = tools.map((t) => mcpToolName(t.server, t.tool));
+  const groups = /* @__PURE__ */ new Map();
+  bases.forEach((base, i) => {
+    const g = groups.get(base);
+    if (g) g.push(i);
+    else groups.set(base, [i]);
+  });
+  const used = new Set(bases);
+  for (const [base, idx] of groups) {
+    if (idx.length === 1) {
+      names[idx[0]] = base;
+      continue;
+    }
+    const raw = (i) => `${tools[i].server}::${tools[i].tool}`;
+    const isClean = (i) => isStableSegment(tools[i].server) && isStableSegment(tools[i].tool);
+    const ordered = [...idx].sort((a, b) => {
+      const la = isClean(a) ? 0 : 1;
+      const lb = isClean(b) ? 0 : 1;
+      if (la !== lb) return la - lb;
+      const ra = raw(a);
+      const rb = raw(b);
+      return ra < rb ? -1 : ra > rb ? 1 : 0;
+    });
+    names[ordered[0]] = base;
+    let n = 2;
+    for (let rank3 = 1; rank3 < ordered.length; rank3++) {
+      let candidate = `${base}_${n}`;
+      while (used.has(candidate)) {
+        n++;
+        candidate = `${base}_${n}`;
+      }
+      names[ordered[rank3]] = candidate;
+      used.add(candidate);
+      n++;
+    }
+  }
+  return names;
+}
+function mcpServerPrefix(serverName) {
+  return `${MCP_TOOL_PREFIX}${boundedSegment(serverName)}__`;
+}
+function removeMcpServerDenials(disabledTools, serverName) {
+  if (!disabledTools?.length) return disabledTools;
+  const prefix = mcpServerPrefix(serverName);
+  const kept = disabledTools.filter((t) => !t.startsWith(prefix));
+  return kept.length === disabledTools.length ? disabledTools : kept;
+}
 var execAsync = promisify(exec);
 var ShellTool = class extends BaseTool {
   name = "shell";
@@ -10948,20 +11333,6 @@ ${text}`;
   }
 };
 
-// src/tools/tool-name.ts
-var MCP_TOOL_PREFIX = "mcp__";
-function sanitizeToolNameSegment(segment) {
-  const cleaned = segment.replace(/[^a-zA-Z0-9-]+/g, "_");
-  let start = 0;
-  let end = cleaned.length;
-  while (start < end && cleaned[start] === "_") start++;
-  while (end > start && cleaned[end - 1] === "_") end--;
-  return cleaned.slice(start, end) || "_";
-}
-function mcpToolName(serverName, toolName) {
-  return `${MCP_TOOL_PREFIX}${sanitizeToolNameSegment(serverName)}__${sanitizeToolNameSegment(toolName)}`;
-}
-
 // src/tools/mcp.ts
 var McpToolWrapper = class extends BaseTool {
   name;
@@ -10970,12 +11341,12 @@ var McpToolWrapper = class extends BaseTool {
   mcpClient;
   serverName;
   toolName;
-  constructor(mcpClient, serverName, toolName, description, inputSchema) {
+  constructor(mcpClient, serverName, toolName, description, inputSchema, registeredName) {
     super();
     this.mcpClient = mcpClient;
     this.serverName = serverName;
     this.toolName = toolName;
-    this.name = mcpToolName(serverName, toolName);
+    this.name = registeredName ?? mcpToolName(serverName, toolName);
     this.description = `[MCP:${serverName}] ${description}`;
     this.inputSchema = inputSchema;
   }
@@ -11046,22 +11417,35 @@ var ToolRegistry = class extends EventEmitter {
   getRegisteredPlugins() {
     return Array.from(this.plugins.keys());
   }
-  /** Registers all tools from an MCP client */
-  registerMcpTools(mcpClient) {
-    const definitions = mcpClient.getToolDefinitions();
-    for (const def of definitions) {
-      const [, serverName, toolName] = def.name.split("::");
-      if (!serverName || !toolName) continue;
+  /**
+   * Registers all tools from an MCP client.
+   *
+   * `isAllowed` receives the FINAL registered name (`mcp__server__tool`) — the
+   * same string the user sees and the same one that reaches the provider — so
+   * a per-tool opt-out in Settings can be expressed as a plain name list.
+   * Filtering here rather than at call time means a deselected tool is genuinely
+   * absent from the definitions handed to the model, so it is never proposed,
+   * never counted against the tool budget, and never has to be refused.
+   */
+  registerMcpTools(mcpClient, isAllowed) {
+    const parsed = mcpClient.getToolDefinitions().flatMap((def) => {
+      const [, server, tool] = def.name.split("::");
+      return server && tool ? [{ def, server, tool }] : [];
+    });
+    const names = assignMcpToolNames(parsed);
+    parsed.forEach(({ def, server, tool }, i) => {
       const wrapper = new McpToolWrapper(
         mcpClient,
-        serverName,
-        toolName,
-        def.description.replace(`[MCP:${serverName}] `, ""),
-        def.inputSchema
+        server,
+        tool,
+        def.description.replace(`[MCP:${server}] `, ""),
+        def.inputSchema,
+        names[i]
       );
+      if (isAllowed && !isAllowed(wrapper.name)) return;
       wrapper.setWorkspaceRoot(this.workspaceRoot);
       this.register(wrapper);
-    }
+    });
   }
   setIgnoredPaths(patterns) {
     this.ignoreMatcher = ignore().add(patterns);
@@ -11398,6 +11782,17 @@ function fileOAuthProvider(storePath) {
     }
   });
 }
+var oauthStoreLocks = /* @__PURE__ */ new Map();
+function withOAuthStoreLock(storePath, fn) {
+  const prev = oauthStoreLocks.get(storePath) ?? Promise.resolve();
+  const run = prev.then(fn, fn);
+  const tail = run.then(() => void 0, () => void 0);
+  oauthStoreLocks.set(storePath, tail);
+  void tail.then(() => {
+    if (oauthStoreLocks.get(storePath) === tail) oauthStoreLocks.delete(storePath);
+  });
+  return run;
+}
 function beginMcpOAuth(provider, serverUrl) {
   return auth(provider, { serverUrl });
 }
@@ -14336,7 +14731,50 @@ var GuidanceQueue = class {
   }
 };
 
+// src/tools/current-page.ts
+var MAX_PAGE_TEXT_CHARS = 4e4;
+var CurrentPageTool = class extends BaseTool {
+  name = "read_current_page";
+  description = 'Read the web page the user currently has open in the built-in browser \u2014 its URL, title, and visible text. Use this whenever the user refers to "this page", "what I am looking at", or asks about something on screen. Prefer it over web_fetch for the open page: it sees content behind a login and content rendered by JavaScript, which web_fetch cannot.';
+  inputSchema = {
+    type: "object",
+    properties: {},
+    required: []
+  };
+  provider;
+  constructor(provider) {
+    super();
+    this.provider = provider;
+  }
+  // Reading what is already on the user's screen changes nothing and reaches
+  // nothing they have not already opened themselves.
+  isDangerous() {
+    return false;
+  }
+  async execute(_input, _options) {
+    let page2;
+    try {
+      page2 = await this.provider();
+    } catch (err) {
+      return `Error: could not read the open page \u2014 ${err instanceof Error ? err.message : String(err)}`;
+    }
+    if (!page2) {
+      return "No page is open in the built-in browser. Ask the user to open one, or use web_fetch for a public URL.";
+    }
+    const text = page2.text.length > MAX_PAGE_TEXT_CHARS ? `${page2.text.slice(0, MAX_PAGE_TEXT_CHARS)}
+
+[truncated \u2014 the page is longer than ${MAX_PAGE_TEXT_CHARS} characters]` : page2.text;
+    return [
+      `URL: ${page2.url}`,
+      `Title: ${page2.title}`,
+      "",
+      text.trim() || "[the page rendered no readable text]"
+    ].join("\n");
+  }
+};
+
 // src/core/cascade.ts
+var ESCALATION_DECISION_TIMEOUT_MS = 5 * 6e4;
 function buildContextualPrompt(prompt, history = []) {
   const recent = history.slice(-6);
   if (recent.length === 0) return prompt;
@@ -14391,6 +14829,13 @@ var Cascade = class _Cascade extends EventEmitter {
       }
     });
     this.toolRegistry = new ToolRegistry(this.config.tools, workspacePath);
+    try {
+      const nodePath = __require("path");
+      this.router.setDeadModelStore(new DeadModelStore(
+        fileDeadModelPersistence(nodePath.join(workspacePath, ".cascade", "dead-models.json"))
+      ));
+    } catch {
+    }
     this.multimodal = new MultimodalRegistry(
       (this.config.providers ?? []).map((p) => p.type)
     );
@@ -14552,6 +14997,94 @@ var Cascade = class _Cascade extends EventEmitter {
       });
     });
   }
+  // ── Escalation decisions ────────────────────────────────────────────
+  //
+  // A section that ends ESCALATED used to be a dead end: the Cockpit showed
+  // "Section escalated — needs a decision" and no decision was ever requested,
+  // so the run stopped there having spent a full orchestration. This is the
+  // same round-trip shape as plan approval, with one deliberate difference —
+  // an unanswered escalation FAILS the section instead of proceeding.
+  //
+  // That direction is the safe one here. Plan approval defaults to "yes" on
+  // timeout because the plan is already the model's considered proposal;
+  // an escalation exists precisely BECAUSE the worker was not confident, so
+  // acting unattended is the option most likely to be wrong. And in cloud a
+  // waiting run holds server resources, so hanging is not free either.
+  // Keyed by request id, NOT a single slot. A Complex run dispatches each wave
+  // of ready sections with `Promise.all` (t1-administrator.ts), so two sections
+  // can be waiting on a decision at the same moment. A single field made that
+  // case fail hard rather than merely confusingly: the second escalation
+  // overwrote the first's resolver, the user's answer resolved only the second,
+  // and then the FIRST section's timer found the slot already empty and
+  // returned without resolving anything — parking the run forever, past its own
+  // timeout. Same class of bug as the twelve-failover dead-model race: a wave
+  // is concurrent, so per-run singletons are never safe inside it.
+  pendingEscalations = /* @__PURE__ */ new Map();
+  async requestEscalationDecision(ctx, taskId, signal) {
+    if (this.config.autonomy === "auto") return { action: "skip", automatic: true };
+    if (this.listenerCount("escalation:decision-required") === 0) return { action: "skip", automatic: true };
+    if (signal?.aborted) return { action: "skip", automatic: true };
+    const requestId = randomUUID();
+    return await new Promise((resolve2) => {
+      let timeout;
+      const settle = (decision) => {
+        if (!this.pendingEscalations.delete(requestId)) return;
+        if (timeout) clearTimeout(timeout);
+        signal?.removeEventListener("abort", onAbort);
+        resolve2(decision);
+      };
+      const onAbort = () => settle({ action: "skip", automatic: true });
+      signal?.addEventListener("abort", onAbort, { once: true });
+      timeout = setTimeout(() => {
+        this.emit("escalation:timeout", { taskId, requestId, sectionId: ctx.sectionId });
+        settle({ action: "timeout" });
+      }, ESCALATION_DECISION_TIMEOUT_MS);
+      this.pendingEscalations.set(requestId, settle);
+      this.emit("escalation:decision-required", {
+        taskId,
+        requestId,
+        sectionId: ctx.sectionId,
+        sectionTitle: ctx.sectionTitle,
+        issues: ctx.issues,
+        summary: ctx.summary,
+        timeoutMs: ESCALATION_DECISION_TIMEOUT_MS
+      });
+    });
+  }
+  /**
+   * Resolve a pending escalation from a REPL / dashboard / cloud listener.
+   *
+   * `requestId` comes back from the `escalation:decision-required` event and is
+   * what makes an answer land on the section that asked. It is optional so a
+   * host with only one escalation in flight (and any caller that predates the
+   * id) still works: without it the oldest outstanding request is answered,
+   * which is correct whenever there is exactly one.
+   *
+   * `automatic` defaults to false — a call into this method is normally a real
+   * person answering the prompt. Pass `true` from a host that is itself
+   * settling the gate on nobody's behalf (a REST caller who never had a
+   * chance to connect, a session halt with no per-section review behind it),
+   * so T2 doesn't set `userSkipped` for a decision no one actually reviewed —
+   * see EscalationDecision.automatic.
+   */
+  resolveEscalation(action, note, requestId, automatic) {
+    const decision = { action, ...note ? { note } : {}, ...automatic ? { automatic: true } : {} };
+    if (requestId) {
+      this.pendingEscalations.get(requestId)?.(decision);
+      return;
+    }
+    const oldest = this.pendingEscalations.keys().next();
+    if (!oldest.done) this.pendingEscalations.get(oldest.value)?.(decision);
+  }
+  /**
+   * Release every parked escalation — run teardown, so none outlive the run.
+   * `automatic: true`: teardown resolving a section nobody answered is the
+   * same system-decided case as the early-return skips above, not a person
+   * reviewing and accepting that section's output.
+   */
+  releasePendingEscalations(action = "skip") {
+    for (const settle of Array.from(this.pendingEscalations.values())) settle({ action, automatic: true });
+  }
   /**
    * Resolve a pending boardroom plan approval from a REPL / dashboard listener.
    * An optional `note` re-plans and re-asks; an optional `editedPlan` is applied
@@ -14685,6 +15218,20 @@ ${input}` }],
     this.mediaSink = sink;
     this.registerMediaTools(this.workspacePath);
   }
+  /**
+   * Let the agent read the page the user has open in the host's browser.
+   *
+   * Host-supplied because only a host WITH a browser can answer it — so the
+   * tool is simply absent in the CLI and in hosted runs, rather than present
+   * and always failing. Registered outside `enabledTools` for the same reason
+   * media tools are: that allowlist is a blast-radius control for tools that
+   * touch the machine, and reading a page the user already opened touches
+   * nothing. `disabledTools` still removes it.
+   */
+  setCurrentPageProvider(provider) {
+    if ((this.config.tools?.disabledTools ?? []).includes("read_current_page")) return;
+    this.toolRegistry.register(new CurrentPageTool(provider));
+  }
   registerMediaTools(workspacePath) {
     const denied = new Set(this.config.tools?.disabledTools ?? []);
     const permitted = (name) => !denied.has(name);
@@ -14817,8 +15364,12 @@ ${last.partialOutput}` : "");
         for (const server of this.config.tools.mcpServers) {
           try {
             const authProvider = server.oauthStore ? fileOAuthProvider(server.oauthStore) : void 0;
-            await this.mcpClient.connect(server, authProvider ? { authProvider } : {});
-            this.toolRegistry.registerMcpTools(this.mcpClient);
+            const connect = () => this.mcpClient.connect(server, authProvider ? { authProvider } : {});
+            await (server.oauthStore ? withOAuthStoreLock(server.oauthStore, connect) : connect());
+            this.toolRegistry.registerMcpTools(
+              this.mcpClient,
+              (name) => !(this.config.tools?.disabledTools ?? []).includes(name)
+            );
           } catch (err) {
             console.error(`Failed to connect to MCP server "${server.name}":`, err);
           }
@@ -15283,6 +15834,7 @@ ${prompt}` : prompt;
       } else if (complexity === "Moderate") {
         const t2 = new T2Manager(this.router, this.toolRegistry, "root");
         t2.setRunBreaker(runBreaker);
+        t2.setEscalationCallback((ctx) => this.requestEscalationDecision(ctx, taskId, options.signal));
         t2.setPresenter(true);
         t2.setHierarchyContext("You are the ROOT Manager for this task. There is no T1 Administrator involved in this run. You are responsible for decomposing the task and managing T3 workers directly.");
         if (identityPrompt) {
@@ -15327,11 +15879,17 @@ ${prompt}` : prompt;
         t2Results = [t2Result];
         const completed = t2Result.t3Results.filter((r) => r.status === "COMPLETED");
         if (completed.length > 0) {
-          finalOutput = t2Result.sectionSummary + "\n\n" + completed.map((r) => r.output).join("\n\n");
+          finalOutput = composeRootSectionOutput(
+            t2Result,
+            completed.map((r) => typeof r.output === "string" ? r.output : JSON.stringify(r.output))
+          );
         } else {
           const partial = t2Result.t3Results.map((r) => typeof r.output === "string" ? r.output : "").filter((o) => o.trim()).sort((a, b) => b.length - a.length)[0];
           const reasons = Array.from(
-            new Set(t2Result.t3Results.flatMap((r) => r.issues ?? []).filter(Boolean))
+            new Set([
+              ...t2Result.issues ?? [],
+              ...t2Result.t3Results.flatMap((r) => r.issues ?? [])
+            ].filter(Boolean))
           );
           if (partial) {
             finalOutput = reasons.length ? `${partial}
@@ -15344,6 +15902,7 @@ _(incomplete: ${reasons.join("; ")})_` : partial;
       } else {
         const t1 = new T1Administrator(this.router, this.toolRegistry, this.config);
         t1.setRunBreaker(runBreaker);
+        t1.setEscalationCallback((ctx) => this.requestEscalationDecision(ctx, taskId, options.signal));
         t1.setPresenter(true);
         t1.setHierarchyContext("You are the top-level Administrator. You are responsible for the overall plan and supervising multiple T2 Managers.");
         if (identityPrompt) {
@@ -15368,6 +15927,8 @@ _(incomplete: ${reasons.join("; ")})_` : partial;
         finalOutput = result.output;
         t2Results = result.t2Results;
       }
+      const budgetInfo = this.router.budgetExceededInfo();
+      if (budgetInfo) throw new CascadeRouter.BudgetExceededError(budgetInfo.reason);
       const trip = runBreaker.reason();
       if (trip) {
         finalOutput = [
@@ -15408,6 +15969,10 @@ ${finalOutput}` : ""
         escalator.cancelAllPending();
       } catch {
       }
+      try {
+        this.releasePendingEscalations();
+      } catch {
+      }
       this.router.restoreTierModels();
       this.router.setRunSignal(void 0);
       if (this.taskAnalyzer) {
@@ -16724,6 +17289,13 @@ var ConfigManager = class {
   }
   async load() {
     this.config = await this.loadConfig();
+    const servers = this.config.tools?.mcpServers;
+    const { servers: disambiguated, renames } = servers?.length ? disambiguateMcpServerNames(servers) : { servers, renames: [] };
+    const mcpNamesChanged = renames.length > 0;
+    if (mcpNamesChanged && this.config.tools) {
+      this.config.tools.mcpServers = disambiguated;
+      this.renameMcpServerReferences(renames);
+    }
     this.ignore = new CascadeIgnore();
     await this.ignore.load(this.workspacePath);
     this.cascadeMd = await loadCascadeMd(this.workspacePath);
@@ -16732,6 +17304,7 @@ var ConfigManager = class {
     await this.injectEnvKeys();
     this.config.providers = mergeGlobalCredentials(this.config.providers, loadGlobalCredentials(this.globalDir));
     await this.ensureDefaultIdentity();
+    if (mcpNamesChanged) await this.save();
   }
   getConfig() {
     return this.config;
@@ -16815,6 +17388,66 @@ var ConfigManager = class {
       this.config.providers.push({ type: "ollama" });
     }
   }
+  /**
+   * Follow a disambiguation rename into `tools.mcpTrusted`.
+   *
+   * `mcpTrusted` is matched by EXACT name (`McpClient.connect()`). Two shapes
+   * of rename reach here, and they need OPPOSITE treatment:
+   *
+   * - Distinct raw names that only collide via sanitizing (`foo bar` /
+   *   `foo@bar`): each had its OWN trust entry, and the old string now belongs
+   *   to nothing — the server that used it moved away, so the entry has to
+   *   move with it. A plain rewrite is correct here.
+   * - Literally identical names (a hand-edited or duplicated config: two rows
+   *   both named `foo`): there is only ONE trust entry for both, since
+   *   `mcpTrusted` itself is deduplicated. `this.config.tools.mcpServers` has
+   *   already been updated to the disambiguated list by the time this runs
+   *   (see `load()`), so the untouched survivor is STILL named `foo` — a plain
+   *   rewrite would move the one entry entirely onto the renamed row and
+   *   leave the survivor untrusted. The entry must be kept for the survivor
+   *   AND granted to the renamed identity, not moved.
+   *
+   * Renames are grouped by `from` before either treatment is applied, and
+   * each group is settled in one step. Two DIFFERENT rows can start out with
+   * the exact same raw name and BOTH get renamed away (e.g. three servers
+   * named `foo bar`, `foo@bar`, `foo@bar` — the two `foo@bar` rows collide
+   * with `foo bar` and with each other, so both are renamed to distinct
+   * identities while `foo bar` survives untouched). Processing renames one at
+   * a time in that shape drops coverage: the first rename's REPLACE branch
+   * would erase `from` from `trusted` entirely, so the guard on the second
+   * rename with the same `from` sees it already gone and silently skips —
+   * granting trust to only one of the two renamed identities. Grouping first
+   * means every renamed identity that shares a `from` is added or substituted
+   * together, in the one pass that still sees the original entry.
+   *
+   * `tools.disabledTools` deliberately gets no equivalent treatment. It's
+   * matched by sanitized PREFIX, and a prefix collision is exactly what made
+   * an existing entry ambiguous between the two servers in the first place —
+   * there is no way to tell, from the stored string alone, which of the two
+   * a denial was meant for. Leaving those entries untouched resolves that
+   * ambiguity for free: the survivor keeps the original (now unique) prefix,
+   * so an old entry keeps applying to it, while the renamed server starts
+   * clean under its new prefix. Rewriting the entry to follow the rename
+   * would do the opposite — move a denial that already worked for the
+   * survivor onto a server it may never have meant to cover.
+   */
+  renameMcpServerReferences(renames) {
+    const tools = this.config.tools;
+    if (!tools?.mcpTrusted?.length || !renames.length) return;
+    const currentNames = new Set((tools.mcpServers ?? []).map((s) => s.name));
+    const byFrom = /* @__PURE__ */ new Map();
+    for (const { from, to } of renames) {
+      const tos = byFrom.get(from);
+      if (tos) tos.push(to);
+      else byFrom.set(from, [to]);
+    }
+    let trusted = tools.mcpTrusted;
+    for (const [from, tos] of byFrom) {
+      if (!trusted.includes(from)) continue;
+      trusted = currentNames.has(from) ? [...trusted, ...tos] : [...trusted.filter((n) => n !== from), ...tos];
+    }
+    tools.mcpTrusted = Array.from(new Set(trusted));
+  }
   async ensureDefaultIdentity() {
     const existing = this.store.getDefaultIdentity();
     if (existing) return;
@@ -16937,6 +17570,14 @@ var DashboardSocket = class {
     this.secret = options.secret;
     this.setupHandlers();
   }
+  /** How many sockets are currently in a room — 0 means nobody can answer. */
+  roomSize(room) {
+    return this.io.sockets.adapter.rooms.get(room)?.size ?? 0;
+  }
+  /** Is this exact connection still live? A reconnect issues a NEW socket id. */
+  hasSocket(socketId) {
+    return this.io.sockets.sockets.has(socketId);
+  }
   broadcastToRoom(room, event, data) {
     this.io.to(room).emit(event, data);
   }
@@ -17008,6 +17649,22 @@ var DashboardSocket = class {
       });
     });
   }
+  /**
+   * A client subscribed to a session's room.
+   *
+   * The host uses this to replay a gate the run is already parked on. Socket.IO
+   * does not buffer room emissions, so a prompt broadcast while the renderer was
+   * reconnecting is simply gone; without a replay the section waits out its full
+   * timeout with the user sitting in front of a screen that never asked.
+   */
+  onJoinSession(callback) {
+    this.io.on("connection", (socket) => {
+      socket.on("join:session", (payload) => {
+        const { sessionId } = normalizeSessionSubscriptionPayload(payload);
+        if (sessionId) callback(sessionId, socket.id);
+      });
+    });
+  }
   onSessionRate(callback) {
     this.io.on("connection", (socket) => {
       socket.on("session:rate", (payload) => {
@@ -17045,6 +17702,31 @@ var DashboardSocket = class {
       });
     });
   }
+  /**
+   * A section escalated and the run is parked waiting for an answer. The
+   * desktop shows the escalation modal on `escalation:decision-required` and
+   * answers here.
+   *
+   * Unlike a plan decision, silence is NOT taken as consent — the SDK fails
+   * the section on timeout — so this handler only ever forwards a real choice.
+   */
+  onEscalationDecision(callback) {
+    this.io.on("connection", (socket) => {
+      socket.on("escalation:decide", (payload) => {
+        const action = payload?.action;
+        if (typeof payload?.sessionId !== "string") return;
+        if (action !== "retry" && action !== "skip" && action !== "guidance") return;
+        callback({
+          sessionId: payload.sessionId,
+          // Identifies WHICH parked section this answers — sections in a wave
+          // run concurrently, so more than one can be waiting.
+          requestId: typeof payload.requestId === "string" ? payload.requestId : void 0,
+          action,
+          note: typeof payload.note === "string" && payload.note.trim() ? payload.note.trim() : void 0
+        });
+      });
+    });
+  }
   onSessionSteer(callback) {
     this.io.on("connection", (socket) => {
       socket.on("session:steer", (payload) => {
@@ -17203,6 +17885,7 @@ function aggregateCostStats(sessions, opts = {}) {
 
 // src/dashboard/server.ts
 var __dirname$1 = path24.dirname(fileURLToPath(import.meta.url));
+var ORPHANED_ESCALATION_GRACE_MS = 3e4;
 var DashboardServer = class {
   app;
   httpServer;
@@ -17213,6 +17896,13 @@ var DashboardServer = class {
   globalStore = null;
   broadcastTimer = null;
   activeSessions = /* @__PURE__ */ new Map();
+  /**
+   * Supplied by the desktop shell so a run can read the page the user has
+   * open in the built-in browser. Absent in the CLI, where there is no
+   * browser — so the tool is simply not registered rather than present and
+   * always failing.
+   */
+  currentPageProvider;
   activeControllers = /* @__PURE__ */ new Map();
   /**
    * Run taskIds per chat session — file snapshots are keyed by the run's
@@ -17227,6 +17917,17 @@ var DashboardServer = class {
    * map is how that answer reaches the run that's blocked on it.
    */
   pendingApprovals = /* @__PURE__ */ new Map();
+  /**
+   * The escalation prompt a session is currently parked on, kept so it can be
+   * delivered to a client that was not reachable when it was raised.
+   *
+   * Socket.IO buffers nothing: a prompt emitted while the renderer is mid
+   * reconnect is gone, and the section then waits out its full five minutes
+   * with the user in front of a screen that never asked them anything. Holding
+   * the payload lets `join:session` replay it, and lets a run with no reachable
+   * client be settled instead of parked (see ORPHANED_ESCALATION_GRACE_MS).
+   */
+  pendingEscalations = /* @__PURE__ */ new Map();
   /**
    * The orchestration decision trail ("why") of each session's most recent
    * run — captured when the run ends so the desktop's Why panel can show it
@@ -17304,20 +18005,29 @@ var DashboardServer = class {
       if (forceTier) cfg = { ...cfg, routing: { ...cfg.routing, forceTier } };
       const cascade = new Cascade(cfg, this.workspacePath, this.store);
       this.activeSessions.set(sessionId, cascade);
+      if (this.currentPageProvider) cascade.setCurrentPageProvider(this.currentPageProvider);
       cascade.on("stream:token", (e) => {
-        this.socket.emitToSocket(socketId, "stream:token", { sessionId, tierId: e.tierId, text: e.text, primary: e.primary });
+        this.emitToSessionClients(socketId, sessionId, "stream:token", { sessionId, tierId: e.tierId, text: e.text, primary: e.primary });
       });
       cascade.on("tier:status", (e) => {
-        this.socket.emitToSocket(socketId, "tier:status", { sessionId, ...e });
+        this.emitToSessionClients(socketId, sessionId, "tier:status", { sessionId, ...e });
       });
       cascade.on("permission:user-required", (e) => {
-        this.socket.emitToSocket(socketId, "permission:user-required", { sessionId, ...e });
+        this.emitToSessionClients(socketId, sessionId, "permission:user-required", { sessionId, ...e });
       });
       cascade.on("peer:message", (e) => {
         this.socket.emitPeerMessage(e);
       });
       cascade.on("plan:approval-required", (e) => {
-        this.socket.emitToSocket(socketId, "plan:approval-required", { sessionId, ...e });
+        this.emitToSessionClients(socketId, sessionId, "plan:approval-required", { sessionId, ...e });
+      });
+      cascade.on("escalation:decision-required", (e) => {
+        this.raiseEscalation(sessionId, socketId, e);
+      });
+      cascade.on("escalation:timeout", (e) => {
+        const ev = e;
+        if (ev.requestId) this.clearPendingEscalation(ev.requestId);
+        this.emitToSessionClients(socketId, sessionId, "escalation:timeout", { sessionId, ...e });
       });
       try {
         const result = await cascade.run({
@@ -17328,7 +18038,7 @@ var DashboardServer = class {
         this.recordSessionTask(sessionId, result.taskId);
         this.persistRunEnd(sessionId, title, prompt, result.output, "COMPLETED", result);
         this.captureWhy(sessionId, cascade, result);
-        this.socket.emitToSocket(socketId, "session:complete", { sessionId, result });
+        this.emitToSessionClients(socketId, sessionId, "session:complete", { sessionId, result });
         this.socket.broadcast("cost:update", {
           sessionId,
           totalTokens: result.usage.totalTokens,
@@ -17341,7 +18051,7 @@ var DashboardServer = class {
       } catch (err) {
         this.persistRunEnd(sessionId, title, prompt, void 0, "FAILED");
         this.captureWhy(sessionId, cascade);
-        this.socket.emitToSocket(socketId, "session:error", {
+        this.emitToSessionClients(socketId, sessionId, "session:error", {
           sessionId,
           error: err instanceof Error ? err.message : String(err)
         });
@@ -17349,6 +18059,7 @@ var DashboardServer = class {
         this.activeSessions.delete(sessionId);
         this.activeControllers.delete(sessionId);
         this.denyPendingApprovals(sessionId);
+        this.clearSessionEscalations(sessionId);
       }
     });
     this.socket.onPlanDecision(({ sessionId, approved, note, editedPlan }) => {
@@ -17358,10 +18069,20 @@ var DashboardServer = class {
         editedPlan
       );
     });
+    this.socket.onEscalationDecision(({ sessionId, requestId, action, note }) => {
+      if (requestId) this.clearPendingEscalation(requestId);
+      else this.clearOldestSessionEscalation(sessionId);
+      this.activeSessions.get(sessionId)?.resolveEscalation(action, note, requestId);
+    });
+    this.socket.onJoinSession((sessionId, socketId) => {
+      this.replayEscalations(sessionId, socketId);
+    });
     this.socket.onSessionHalt((sessionId) => {
       this.activeControllers.get(sessionId)?.abort();
       this.denyPendingApprovals(sessionId);
       this.activeSessions.get(sessionId)?.resolvePlanApproval(false);
+      this.clearSessionEscalations(sessionId);
+      this.activeSessions.get(sessionId)?.resolveEscalation("skip", void 0, void 0, true);
     });
     this.socket.onApprovalResponse(({ requestId, approved, always }) => {
       const pending = this.pendingApprovals.get(requestId);
@@ -17376,6 +18097,10 @@ var DashboardServer = class {
       }
     });
   }
+  /** Wire the host's browser view in (desktop only). */
+  setCurrentPageProvider(provider) {
+    this.currentPageProvider = provider;
+  }
   async start() {
     const isLoopback = this.host === "127.0.0.1" || this.host === "::1" || this.host === "localhost";
     if (!isLoopback) {
@@ -17664,6 +18389,116 @@ var DashboardServer = class {
       this.activeSessions.delete(sessionId);
     }
   }
+  /**
+   * Is anyone able to answer for this session right now?
+   *
+   * The originating socket first — that connection is definitely watching this
+   * run — then the session room, which is where a client lands after a
+   * reconnect issues it a new socket id.
+   */
+  hasResponder(sessionId, socketId) {
+    if (socketId && this.socket.hasSocket(socketId)) return true;
+    return this.socket.roomSize(`session:${sessionId}`) > 0;
+  }
+  /**
+   * Send a run event to whoever is watching this session.
+   *
+   * A captured socket id goes stale the moment Socket.IO reconnects — the
+   * client is back, but under a new id, and `emitToSocket` on the old one is a
+   * silent no-op. Falling back to the session room keeps the run addressable
+   * across a reconnect; emitting to one OR the other (never both) keeps a
+   * client that is both the originator and a room member from seeing every
+   * token twice.
+   */
+  emitToSessionClients(socketId, sessionId, event, data) {
+    if (socketId && this.socket.hasSocket(socketId)) this.socket.emitToSocket(socketId, event, data);
+    else this.socket.broadcastToRoom(`session:${sessionId}`, event, data);
+  }
+  /**
+   * Raise an escalation prompt, and make sure it can still be answered.
+   *
+   * Held in `pendingEscalations` — keyed by requestId, because a Complex wave
+   * dispatches sections concurrently and two can be parked at once — so a
+   * client that joins late (or rejoins after a reconnect) gets every waiting
+   * prompt replayed. Keying by session would drop the first of a pair: its
+   * grace timer would be cleared by the second, leaving it parked for the full
+   * timeout with nobody able to answer it. If nobody is reachable at all, the gate is
+   * settled as `skip` after a short grace period rather than parked for the
+   * full timeout: that matches the SDK's policy when no listener exists, and
+   * five minutes of a held run is a real cost for a REST caller who has already
+   * hung up. The grace period exists so an ordinary two-second reconnect is not
+   * mistaken for an absent user.
+   */
+  raiseEscalation(sessionId, socketId, event) {
+    const requestId = typeof event.requestId === "string" && event.requestId ? event.requestId : randomUUID();
+    const payload = { sessionId, ...event };
+    const entry = { payload, sessionId, requestId, raisedAt: Date.now() };
+    this.pendingEscalations.set(requestId, entry);
+    if (this.hasResponder(sessionId, socketId)) {
+      this.emitToSessionClients(socketId, sessionId, "escalation:decision-required", payload);
+      return;
+    }
+    entry.graceTimer = setTimeout(() => {
+      if (this.pendingEscalations.get(requestId) !== entry) return;
+      if (this.hasResponder(sessionId, socketId)) return;
+      this.pendingEscalations.delete(requestId);
+      this.activeSessions.get(sessionId)?.resolveEscalation("skip", "No client was connected to answer.", requestId, true);
+    }, ORPHANED_ESCALATION_GRACE_MS);
+  }
+  /**
+   * Replay EVERY parked prompt to a client that just subscribed to the session.
+   *
+   * Plural because a Complex run dispatches sections concurrently, so a wave can
+   * leave more than one waiting at the same time — the same reason the SDK keys
+   * its gates and the clients queue theirs.
+   *
+   * `timeoutMs` is recomputed as the time actually LEFT, because the SDK's timer
+   * has been running since the prompt was raised. Sent verbatim, a client
+   * reconnecting four minutes in would show a fresh five-minute countdown and
+   * then watch the section fail with four of those minutes still on the clock.
+   * A remaining duration is used rather than an absolute deadline so the two
+   * sides never have to agree on a wall clock.
+   */
+  replayEscalations(sessionId, socketId) {
+    for (const entry of this.pendingEscalations.values()) {
+      if (entry.sessionId !== sessionId) continue;
+      if (entry.graceTimer) {
+        clearTimeout(entry.graceTimer);
+        entry.graceTimer = void 0;
+      }
+      const originalTimeout = typeof entry.payload.timeoutMs === "number" ? entry.payload.timeoutMs : 0;
+      const remaining = Math.max(0, originalTimeout - (Date.now() - entry.raisedAt));
+      this.socket.emitToSocket(socketId, "escalation:decision-required", { ...entry.payload, timeoutMs: remaining });
+    }
+  }
+  /** Forget ONE parked prompt — the one that was answered or timed out. */
+  clearPendingEscalation(requestId) {
+    const entry = this.pendingEscalations.get(requestId);
+    if (!entry) return;
+    if (entry.graceTimer) clearTimeout(entry.graceTimer);
+    this.pendingEscalations.delete(requestId);
+  }
+  /**
+   * Forget the session's OLDEST prompt — for a client that answered without a
+   * requestId, which is exactly the gate `resolveEscalation` will settle.
+   * Insertion order makes `values()` oldest-first.
+   */
+  clearOldestSessionEscalation(sessionId) {
+    for (const [id, entry] of this.pendingEscalations) {
+      if (entry.sessionId !== sessionId) continue;
+      if (entry.graceTimer) clearTimeout(entry.graceTimer);
+      this.pendingEscalations.delete(id);
+      return;
+    }
+  }
+  /** Forget every prompt a session was parked on (run ended, or halted). */
+  clearSessionEscalations(sessionId) {
+    for (const [id, entry] of this.pendingEscalations) {
+      if (entry.sessionId !== sessionId) continue;
+      if (entry.graceTimer) clearTimeout(entry.graceTimer);
+      this.pendingEscalations.delete(id);
+    }
+  }
   /** Deny + clear any approvals still pending for a session (run end / abort). */
   denyPendingApprovals(sessionId) {
     for (const [id, pending] of this.pendingApprovals) {
@@ -18164,6 +18999,7 @@ ${prompt}`;
       void (async () => {
         const cascade = new Cascade(this.config, this.workspacePath, this.store);
         this.activeSessions.set(sessionId, cascade);
+        if (this.currentPageProvider) cascade.setCurrentPageProvider(this.currentPageProvider);
         cascade.on("stream:token", (e) => {
           this.socket.broadcastToRoom(`session:${sessionId}`, "stream:token", { sessionId, tierId: e.tierId, text: e.text, primary: e.primary });
         });
@@ -18179,6 +19015,14 @@ ${prompt}`;
         cascade.on("plan:approval-required", (e) => {
           this.socket.broadcastToRoom(`session:${sessionId}`, "plan:approval-required", { sessionId, ...e });
         });
+        cascade.on("escalation:decision-required", (e) => {
+          this.raiseEscalation(sessionId, void 0, e);
+        });
+        cascade.on("escalation:timeout", (e) => {
+          const ev = e;
+          if (ev.requestId) this.clearPendingEscalation(ev.requestId);
+          this.socket.broadcastToRoom(`session:${sessionId}`, "escalation:timeout", { sessionId, ...e });
+        });
         try {
           const result = await cascade.run({
             prompt: runPrompt,
@@ -18206,6 +19050,7 @@ ${prompt}`;
         } finally {
           this.activeSessions.delete(sessionId);
           this.denyPendingApprovals(sessionId);
+          this.clearSessionEscalations(sessionId);
         }
       })();
     });
@@ -18538,6 +19383,41 @@ var HooksRunner = class {
   }
 };
 
+// src/mcp/discover.ts
+async function discoverMcpTools(servers) {
+  const results = [];
+  for (const server of servers) {
+    const client = new McpClient({ trustedServers: [server.name] });
+    try {
+      const authProvider = server.oauthStore ? fileOAuthProvider(server.oauthStore) : void 0;
+      const connect = () => client.connect(server, authProvider ? { authProvider } : {});
+      await (server.oauthStore ? withOAuthStoreLock(server.oauthStore, connect) : connect());
+      const parsed = client.getToolDefinitions().flatMap((def) => {
+        const [, serverName, toolName] = def.name.split("::");
+        return serverName && toolName ? [{ def, server: serverName, tool: toolName }] : [];
+      });
+      const names = assignMcpToolNames(parsed);
+      const tools = parsed.map(({ def, server: s, tool }, i) => ({
+        server: s,
+        tool,
+        name: names[i],
+        description: def.description.replace(`[MCP:${s}] `, "")
+      }));
+      results.push({ server: server.name, tools });
+    } catch (err) {
+      results.push({
+        server: server.name,
+        tools: [],
+        error: err instanceof Error ? err.message : String(err)
+      });
+    } finally {
+      await client.disconnectAll().catch(() => {
+      });
+    }
+  }
+  return results;
+}
+
 // src/retrieval/chunk.ts
 var DEFAULTS2 = { targetChars: 2e3, overlapChars: 200, minChars: 80 };
 function splitLongBlock(block, max) {
@@ -19076,11 +19956,55 @@ function mergeByKey(local, incoming, key) {
   for (const i of incoming) map.set(key(i), i);
   return [...map.values()];
 }
+function uniqueServerName(desired, takenPrefixes) {
+  if (!takenPrefixes.has(mcpServerPrefix(desired))) return desired;
+  let n = 2;
+  while (takenPrefixes.has(mcpServerPrefix(`${desired} ${n}`))) n++;
+  return `${desired} ${n}`;
+}
+function mergeMcpServers(local, incoming) {
+  const rewrite = /* @__PURE__ */ new Map();
+  if (!incoming?.length) return { servers: local, synced: [], rewrite };
+  const byName = /* @__PURE__ */ new Map();
+  for (const l of local ?? []) byName.set(l.name, l);
+  const prefixes = new Set([...byName.values()].map((m) => mcpServerPrefix(m.name)));
+  const synced = [];
+  for (const inc of incoming) {
+    if (byName.has(inc.name)) {
+      byName.set(inc.name, inc);
+      synced.push(inc);
+      continue;
+    }
+    const name = uniqueServerName(inc.name, prefixes);
+    if (name !== inc.name) rewrite.set(mcpServerPrefix(inc.name), mcpServerPrefix(name));
+    const server = name === inc.name ? inc : { ...inc, name };
+    byName.set(name, server);
+    synced.push(server);
+    prefixes.add(mcpServerPrefix(name));
+  }
+  return { servers: [...byName.values()], synced, rewrite };
+}
+function rewriteDisabledTools(tools, rewrite) {
+  if (!tools?.length || rewrite.size === 0) return tools;
+  return tools.map((t) => {
+    for (const [from, to] of rewrite) if (t.startsWith(from)) return to + t.slice(from.length);
+    return t;
+  });
+}
+function mergeDisabledTools(local, incoming, syncedServers) {
+  const prefixes = (syncedServers ?? []).map((m) => mcpServerPrefix(m.name));
+  const covered = (tool) => prefixes.some((p) => tool.startsWith(p));
+  const kept = (local ?? []).filter((t) => !covered(t));
+  const merged = [.../* @__PURE__ */ new Set([...kept, ...incoming ?? []])];
+  return merged.length ? merged : void 0;
+}
 function gatherSyncBundle(config) {
-  const bundle = { v: 1 };
+  const bundle = { v: 2 };
   if (config.providers?.length) bundle.providers = config.providers;
   if (config.tools?.webSearch) bundle.webSearch = config.tools.webSearch;
   if (config.tools?.mcpServers?.length) bundle.mcpServers = config.tools.mcpServers;
+  const syncedDisabledTools = config.tools?.disabledTools?.filter(isMcpToolName);
+  if (syncedDisabledTools?.length) bundle.disabledTools = syncedDisabledTools;
   if (config.models) bundle.models = config.models;
   if (config.budget) bundle.budget = config.budget;
   const prefs = {};
@@ -19095,7 +20019,16 @@ function applySyncBundle(bundle, config) {
   const next = { ...config, tools: { ...config.tools } };
   if (bundle.providers) next.providers = mergeByKey(config.providers, bundle.providers, providerSig) ?? [];
   if (bundle.webSearch) next.tools.webSearch = { ...config.tools?.webSearch, ...bundle.webSearch };
-  if (bundle.mcpServers) next.tools.mcpServers = mergeByKey(config.tools?.mcpServers, bundle.mcpServers, (m) => m.name);
+  const mcp = mergeMcpServers(config.tools?.mcpServers, bundle.mcpServers);
+  if (bundle.mcpServers) next.tools.mcpServers = mcp.servers;
+  const carriesSelections = bundle.v >= 2 || bundle.disabledTools !== void 0;
+  if (carriesSelections && (bundle.mcpServers || bundle.disabledTools)) {
+    next.tools.disabledTools = mergeDisabledTools(
+      config.tools?.disabledTools,
+      rewriteDisabledTools(bundle.disabledTools, mcp.rewrite),
+      bundle.mcpServers ? mcp.synced : void 0
+    );
+  }
   if (bundle.models) next.models = { ...config.models, ...bundle.models };
   if (bundle.budget) next.budget = { ...config.budget, ...bundle.budget };
   if (bundle.prefs) {
@@ -19108,6 +20041,6 @@ function applySyncBundle(bundle, config) {
   return next;
 }
 
-export { AZURE_BASE_MODELS, AZURE_BASE_URL_TEMPLATE, AuditLogger, CASCADE_AUDIT_FILE, CASCADE_CONFIG_DIR, CASCADE_CONFIG_FILE, CASCADE_DASHBOARD_SECRET_FILE, CASCADE_DB_FILE, CASCADE_IGNORE_FILE, CASCADE_KEYSTORE_FILE, CASCADE_MD_FILE, CASCADE_VERSION, CHARS_PER_TOKEN, COMPLEXITY_T2_COUNT, Cascade, CascadeCancelledError, CascadeIgnore, CascadeRouter, CascadeToolError, CloudClient, CodeSearchTool, ConfigManager, DEFAULT_API_PORT, DEFAULT_APPROVAL_REQUIRED, DEFAULT_AUTO_SUMMARIZE_AT, DEFAULT_CLOUD_URL, DEFAULT_CONTEXT_LIMIT, DEFAULT_DASHBOARD_PORT, DEFAULT_MAX_SESSION_MESSAGES, DEFAULT_RETENTION_DAYS, DEFAULT_THEME, DashboardServer, FileMcpOAuthStore, GLOBAL_CONFIG_DIR, GLOBAL_CREDENTIALS_FILE, GLOBAL_DB_FILE, GLOBAL_KEYSTORE_FILE, GLOBAL_RUNTIME_DB_FILE, GraphRetriever, GraphSearchTool, HooksRunner, Keystore, LLMReranker, LM_STUDIO_BASE_URL, MODELS, McpClient, McpOAuthProvider, MemoryStore, OLLAMA_BASE_URL, OpenAICompatibleEmbedder, PROVIDER_DISPLAY_NAMES, Retriever, SqliteVectorStore, T1Administrator, T1_MODEL_PRIORITY, T2Manager, T2_MODEL_PRIORITY, T3Worker, T3_MODEL_PRIORITY, THEME_NAMES, TOOL_NAMES, TaskScheduler, Telemetry, ToolRegistry, VISION_MODEL_PRIORITY, WorkspaceIndex, applySyncBundle, azureModelForDeployment, beginMcpOAuth, buildManifest, buildSessionTranscript, cagCharBudget, chatCompleterFromProviders, chunkCode, chunkText2 as chunkText, completeMcpOAuth, connectMcpWithLoopbackOAuth, createCascade, decryptJSON as decryptSyncBlob, diffManifest, discoverMcpAuthServer, distillSessionFacts, embedderFromProviders, encryptJSON as encryptSyncBlob, fileOAuthProvider, gatherSyncBundle, hashContent, heuristicCodeChunker, inferAzureBaseModel, nodeHttpFetch, parseRankOrder, planRetrieval, preferIpv4Host, reciprocalRankFusion, refreshMcpToken, runCascade, sessionWorthRemembering, streamCascade };
+export { AZURE_BASE_MODELS, AZURE_BASE_URL_TEMPLATE, AuditLogger, CASCADE_AUDIT_FILE, CASCADE_CONFIG_DIR, CASCADE_CONFIG_FILE, CASCADE_DASHBOARD_SECRET_FILE, CASCADE_DB_FILE, CASCADE_IGNORE_FILE, CASCADE_KEYSTORE_FILE, CASCADE_MD_FILE, CASCADE_VERSION, CHARS_PER_TOKEN, COMPLEXITY_T2_COUNT, Cascade, CascadeCancelledError, CascadeIgnore, CascadeRouter, CascadeToolError, CloudClient, CodeSearchTool, ConfigManager, CurrentPageTool, DEFAULT_API_PORT, DEFAULT_APPROVAL_REQUIRED, DEFAULT_AUTO_SUMMARIZE_AT, DEFAULT_CLOUD_URL, DEFAULT_CONTEXT_LIMIT, DEFAULT_DASHBOARD_PORT, DEFAULT_MAX_SESSION_MESSAGES, DEFAULT_RETENTION_DAYS, DEFAULT_THEME, DashboardServer, FileMcpOAuthStore, GLOBAL_CONFIG_DIR, GLOBAL_CREDENTIALS_FILE, GLOBAL_DB_FILE, GLOBAL_KEYSTORE_FILE, GLOBAL_RUNTIME_DB_FILE, GraphRetriever, GraphSearchTool, HooksRunner, Keystore, LLMReranker, LM_STUDIO_BASE_URL, MAX_PAGE_TEXT_CHARS, MCP_TOOL_PREFIX, MODELS, McpClient, McpOAuthProvider, MemoryStore, OLLAMA_BASE_URL, OpenAICompatibleEmbedder, PROVIDER_DISPLAY_NAMES, Retriever, SqliteVectorStore, T1Administrator, T1_MODEL_PRIORITY, T2Manager, T2_MODEL_PRIORITY, T3Worker, T3_MODEL_PRIORITY, THEME_NAMES, TOOL_NAMES, TaskScheduler, Telemetry, ToolRegistry, VISION_MODEL_PRIORITY, WorkspaceIndex, applySyncBundle, azureModelForDeployment, beginMcpOAuth, buildManifest, buildSessionTranscript, cagCharBudget, chatCompleterFromProviders, chunkCode, chunkText2 as chunkText, completeMcpOAuth, connectMcpWithLoopbackOAuth, createCascade, decryptJSON as decryptSyncBlob, diffManifest, disambiguateMcpServerNames, discoverMcpAuthServer, discoverMcpTools, distillSessionFacts, embedderFromProviders, encryptJSON as encryptSyncBlob, fileOAuthProvider, gatherSyncBundle, hashContent, heuristicCodeChunker, inferAzureBaseModel, isMcpToolName, mcpServerPrefix, mcpToolName, nodeHttpFetch, parseRankOrder, planRetrieval, preferIpv4Host, reciprocalRankFusion, refreshMcpToken, removeMcpServerDenials, runCascade, sessionWorthRemembering, streamCascade, uniqueMcpServerName };
 //# sourceMappingURL=index.js.map
 //# sourceMappingURL=index.js.map