@bolt-foundry/gambit 0.8.5-rc.6 → 0.8.5

package/script/gambit/simulator-ui/dist/bundle.js

@@ -1013,17 +1013,17 @@ var require_react_jsx_runtime_development = __commonJS({
       function isValidElement(object) {
         return "object" === typeof object && null !== object && object.$$typeof === REACT_ELEMENT_TYPE;
       }
-      var React40 = require__(), REACT_ELEMENT_TYPE = Symbol.for("react.transitional.element"), REACT_PORTAL_TYPE = Symbol.for("react.portal"), REACT_FRAGMENT_TYPE = Symbol.for("react.fragment"), REACT_STRICT_MODE_TYPE = Symbol.for("react.strict_mode"), REACT_PROFILER_TYPE = Symbol.for("react.profiler"), REACT_CONSUMER_TYPE = Symbol.for("react.consumer"), REACT_CONTEXT_TYPE = Symbol.for("react.context"), REACT_FORWARD_REF_TYPE = Symbol.for("react.forward_ref"), REACT_SUSPENSE_TYPE = Symbol.for("react.suspense"), REACT_SUSPENSE_LIST_TYPE = Symbol.for("react.suspense_list"), REACT_MEMO_TYPE = Symbol.for("react.memo"), REACT_LAZY_TYPE = Symbol.for("react.lazy"), REACT_ACTIVITY_TYPE = Symbol.for("react.activity"), REACT_CLIENT_REFERENCE = Symbol.for("react.client.reference"), ReactSharedInternals = React40.__CLIENT_INTERNALS_DO_NOT_USE_OR_WARN_USERS_THEY_CANNOT_UPGRADE, hasOwnProperty = Object.prototype.hasOwnProperty, isArrayImpl = Array.isArray, createTask = console.createTask ? console.createTask : function() {
+      var React41 = require__(), REACT_ELEMENT_TYPE = Symbol.for("react.transitional.element"), REACT_PORTAL_TYPE = Symbol.for("react.portal"), REACT_FRAGMENT_TYPE = Symbol.for("react.fragment"), REACT_STRICT_MODE_TYPE = Symbol.for("react.strict_mode"), REACT_PROFILER_TYPE = Symbol.for("react.profiler"), REACT_CONSUMER_TYPE = Symbol.for("react.consumer"), REACT_CONTEXT_TYPE = Symbol.for("react.context"), REACT_FORWARD_REF_TYPE = Symbol.for("react.forward_ref"), REACT_SUSPENSE_TYPE = Symbol.for("react.suspense"), REACT_SUSPENSE_LIST_TYPE = Symbol.for("react.suspense_list"), REACT_MEMO_TYPE = Symbol.for("react.memo"), REACT_LAZY_TYPE = Symbol.for("react.lazy"), REACT_ACTIVITY_TYPE = Symbol.for("react.activity"), REACT_CLIENT_REFERENCE = Symbol.for("react.client.reference"), ReactSharedInternals = React41.__CLIENT_INTERNALS_DO_NOT_USE_OR_WARN_USERS_THEY_CANNOT_UPGRADE, hasOwnProperty = Object.prototype.hasOwnProperty, isArrayImpl = Array.isArray, createTask = console.createTask ? console.createTask : function() {
         return null;
       };
-      React40 = {
+      React41 = {
         react_stack_bottom_frame: function(callStackForError) {
           return callStackForError();
         }
       };
       var specialPropKeyWarningShown;
       var didWarnAboutElementRef = {};
-      var unknownOwnerDebugStack = React40.react_stack_bottom_frame.bind(React40, UnknownOwner)();
+      var unknownOwnerDebugStack = React41.react_stack_bottom_frame.bind(React41, UnknownOwner)();
       var unknownOwnerDebugTask = createTask(getTaskName(UnknownOwner));
       var didWarnAboutKeySpread = {};
       exports.Fragment = REACT_FRAGMENT_TYPE;
@@ -1347,7 +1347,7 @@ var require_react_dom_development = __commonJS({
         return dispatcher;
       }
       "undefined" !== typeof __REACT_DEVTOOLS_GLOBAL_HOOK__ && "function" === typeof __REACT_DEVTOOLS_GLOBAL_HOOK__.registerInternalModuleStart && __REACT_DEVTOOLS_GLOBAL_HOOK__.registerInternalModuleStart(Error());
-      var React40 = require__(), Internals = {
+      var React41 = require__(), Internals = {
         d: {
           f: noop,
           r: function() {
@@ -1363,7 +1363,7 @@ var require_react_dom_development = __commonJS({
         },
         p: 0,
         findDOMNode: null
-      }, REACT_PORTAL_TYPE = Symbol.for("react.portal"), ReactSharedInternals = React40.__CLIENT_INTERNALS_DO_NOT_USE_OR_WARN_USERS_THEY_CANNOT_UPGRADE;
+      }, REACT_PORTAL_TYPE = Symbol.for("react.portal"), ReactSharedInternals = React41.__CLIENT_INTERNALS_DO_NOT_USE_OR_WARN_USERS_THEY_CANNOT_UPGRADE;
       "function" === typeof Map && null != Map.prototype && "function" === typeof Map.prototype.forEach && "function" === typeof Set && null != Set.prototype && "function" === typeof Set.prototype.clear && "function" === typeof Set.prototype.forEach || console.error("React depends on Map and Set built-in types. Make sure that you load a polyfill in older browsers. https://reactjs.org/link/react-polyfills");
       exports.__DOM_INTERNALS_DO_NOT_USE_OR_WARN_USERS_THEY_CANNOT_UPGRADE = Internals;
       exports.createPortal = function(children, container) {
@@ -2718,7 +2718,7 @@ var require_react_dom_client_development = __commonJS({
         "number" === type && getActiveElement(node.ownerDocument) === node || node.defaultValue === "" + value || (node.defaultValue = "" + value);
       }
       function validateOptionProps(element, props) {
-        null == props.value && ("object" === typeof props.children && null !== props.children ? React40.Children.forEach(props.children, function(child) {
+        null == props.value && ("object" === typeof props.children && null !== props.children ? React41.Children.forEach(props.children, function(child) {
           null == child || "string" === typeof child || "number" === typeof child || "bigint" === typeof child || didWarnInvalidChild || (didWarnInvalidChild = true, console.error("Cannot infer the option value of complex children. Pass a `value` prop or use a plain string as children to <option>."));
         }) : null == props.dangerouslySetInnerHTML || didWarnInvalidInnerHTML || (didWarnInvalidInnerHTML = true, console.error("Pass a `value` prop if you set dangerouslyInnerHTML so React knows which value should be selected.")));
         null == props.selected || didWarnSelectedSetOnOption || (console.error("Use the `defaultValue` or `value` props on <select> instead of setting `selected` on <option>."), didWarnSelectedSetOnOption = true);
@@ -13575,14 +13575,14 @@ var require_react_dom_client_development = __commonJS({
         container[internalContainerInstanceKey] && (container._reactRootContainer ? console.error("You are calling ReactDOMClient.createRoot() on a container that was previously passed to ReactDOM.render(). This is not supported.") : console.error("You are calling ReactDOMClient.createRoot() on a container that has already been passed to createRoot() before. Instead, call root.render() on the existing root instead if you want to update it."));
       }
       "undefined" !== typeof __REACT_DEVTOOLS_GLOBAL_HOOK__ && "function" === typeof __REACT_DEVTOOLS_GLOBAL_HOOK__.registerInternalModuleStart && __REACT_DEVTOOLS_GLOBAL_HOOK__.registerInternalModuleStart(Error());
-      var Scheduler = require__2(), React40 = require__(), ReactDOM = require__3(), assign = Object.assign, REACT_LEGACY_ELEMENT_TYPE = Symbol.for("react.element"), REACT_ELEMENT_TYPE = Symbol.for("react.transitional.element"), REACT_PORTAL_TYPE = Symbol.for("react.portal"), REACT_FRAGMENT_TYPE = Symbol.for("react.fragment"), REACT_STRICT_MODE_TYPE = Symbol.for("react.strict_mode"), REACT_PROFILER_TYPE = Symbol.for("react.profiler"), REACT_CONSUMER_TYPE = Symbol.for("react.consumer"), REACT_CONTEXT_TYPE = Symbol.for("react.context"), REACT_FORWARD_REF_TYPE = Symbol.for("react.forward_ref"), REACT_SUSPENSE_TYPE = Symbol.for("react.suspense"), REACT_SUSPENSE_LIST_TYPE = Symbol.for("react.suspense_list"), REACT_MEMO_TYPE = Symbol.for("react.memo"), REACT_LAZY_TYPE = Symbol.for("react.lazy");
+      var Scheduler = require__2(), React41 = require__(), ReactDOM = require__3(), assign = Object.assign, REACT_LEGACY_ELEMENT_TYPE = Symbol.for("react.element"), REACT_ELEMENT_TYPE = Symbol.for("react.transitional.element"), REACT_PORTAL_TYPE = Symbol.for("react.portal"), REACT_FRAGMENT_TYPE = Symbol.for("react.fragment"), REACT_STRICT_MODE_TYPE = Symbol.for("react.strict_mode"), REACT_PROFILER_TYPE = Symbol.for("react.profiler"), REACT_CONSUMER_TYPE = Symbol.for("react.consumer"), REACT_CONTEXT_TYPE = Symbol.for("react.context"), REACT_FORWARD_REF_TYPE = Symbol.for("react.forward_ref"), REACT_SUSPENSE_TYPE = Symbol.for("react.suspense"), REACT_SUSPENSE_LIST_TYPE = Symbol.for("react.suspense_list"), REACT_MEMO_TYPE = Symbol.for("react.memo"), REACT_LAZY_TYPE = Symbol.for("react.lazy");
       Symbol.for("react.scope");
       var REACT_ACTIVITY_TYPE = Symbol.for("react.activity");
       Symbol.for("react.legacy_hidden");
       Symbol.for("react.tracing_marker");
       var REACT_MEMO_CACHE_SENTINEL = Symbol.for("react.memo_cache_sentinel");
       Symbol.for("react.view_transition");
-      var MAYBE_ITERATOR_SYMBOL = Symbol.iterator, REACT_CLIENT_REFERENCE = Symbol.for("react.client.reference"), isArrayImpl = Array.isArray, ReactSharedInternals = React40.__CLIENT_INTERNALS_DO_NOT_USE_OR_WARN_USERS_THEY_CANNOT_UPGRADE, ReactDOMSharedInternals = ReactDOM.__DOM_INTERNALS_DO_NOT_USE_OR_WARN_USERS_THEY_CANNOT_UPGRADE, NotPending = Object.freeze({
+      var MAYBE_ITERATOR_SYMBOL = Symbol.iterator, REACT_CLIENT_REFERENCE = Symbol.for("react.client.reference"), isArrayImpl = Array.isArray, ReactSharedInternals = React41.__CLIENT_INTERNALS_DO_NOT_USE_OR_WARN_USERS_THEY_CANNOT_UPGRADE, ReactDOMSharedInternals = ReactDOM.__DOM_INTERNALS_DO_NOT_USE_OR_WARN_USERS_THEY_CANNOT_UPGRADE, NotPending = Object.freeze({
         pending: false,
         data: null,
         method: null,
@@ -16555,7 +16555,7 @@ var require_react_dom_client_development = __commonJS({
         }
       };
       (function() {
-        var isomorphicReactPackageVersion = React40.version;
+        var isomorphicReactPackageVersion = React41.version;
         if ("19.2.4" !== isomorphicReactPackageVersion) throw Error('Incompatible React versions: The "react" and "react-dom" packages must have the exact same version. Instead got:\n  - react:      ' + (isomorphicReactPackageVersion + "\n  - react-dom:  19.2.4\nLearn more: https://react.dev/warnings/version-mismatch"));
       })();
       "function" === typeof Map && null != Map.prototype && "function" === typeof Map.prototype.forEach && "function" === typeof Set && null != Set.prototype && "function" === typeof Set.prototype.clear && "function" === typeof Set.prototype.forEach || console.error("React depends on Map and Set built-in types. Make sure that you load a polyfill in older browsers. https://react.dev/link/react-polyfills");
@@ -16651,8 +16651,8 @@ var require_client = __commonJS({
 });
 
 // simulator-ui/src/main.tsx
-var import_jsx_runtime43 = __toESM(require_jsx_runtime());
-var import_react40 = __toESM(require__());
+var import_jsx_runtime44 = __toESM(require_jsx_runtime());
+var import_react41 = __toESM(require__());
 var import_client = __toESM(require_client());
 
 // simulator-ui/src/DocsPage.tsx
@@ -16675,7 +16675,7 @@ var buildWorkspacePath = (tab, workspaceId, opts) => {
   return `${base}/${encodeURIComponent(runId)}`;
 };
 var parseWorkspaceRoute = (pathname) => {
-  const match = pathname.match(/^\/workspaces\/([^/]+)\/(debug|build|test|grade)(?:\/([^/]+))?$/);
+  const match = pathname.match(/^\/workspaces\/([^/]+)\/(debug|build|test|grade|verify)(?:\/([^/]+))?$/);
   if (!match) return null;
   const rawId = decodeURIComponent(match[1]);
   const tab = match[2];
@@ -16714,16 +16714,19 @@ var DEFAULT_WORKSPACE_DEBUG_PATH = buildWorkspacePath("debug");
 var DEFAULT_TEST_PATH = buildWorkspacePath("test");
 var DEFAULT_BUILD_PATH = buildWorkspacePath("build");
 var DEFAULT_GRADE_PATH = buildWorkspacePath("grade");
+var DEFAULT_VERIFY_PATH = buildWorkspacePath("verify");
 var buildGradePath = (workspaceId, gradeRunId) => buildWorkspacePath("grade", workspaceId, {
   runId: gradeRunId
 });
 var buildTestPath = (workspaceId, runId) => buildWorkspacePath("test", workspaceId, {
   runId
 });
+var buildVerifyPath = (workspaceId) => buildWorkspacePath("verify", workspaceId);
 var DURABLE_STREAM_PREFIX = "/api/durable-streams/stream/";
 var SIMULATOR_STREAM_ID = "gambit-simulator";
 var WORKSPACE_STREAM_ID = "gambit-workspace";
 var buildTabEnabled = Boolean(window.__GAMBIT_BUILD_TAB_ENABLED__);
+var verifyTabEnabled = Boolean(window.__GAMBIT_VERIFY_TAB_ENABLED__);
 var workspaceOnboardingEnabled = Boolean(window.__GAMBIT_WORKSPACE_ONBOARDING__);
 var workspaceIdFromWindow = window.__GAMBIT_WORKSPACE_ID__ ?? null;
 var chatAccordionEnabled = Boolean(window.__GAMBIT_CHAT_ACCORDION_ENABLED__);
@@ -17484,6 +17487,27 @@ ${normalizedText}`;
       });
       continue;
     }
+    if (itemType.includes(":")) {
+      const extensionData = Object.hasOwn(item, "data") ? item.data : item;
+      const extensionText = (() => {
+        try {
+          return JSON.stringify(extensionData);
+        } catch {
+          return String(extensionData);
+        }
+      })();
+      if (!extensionText || extensionText === "{}") continue;
+      const outputIndex = typeof payload.output_index === "number" ? String(payload.output_index) : "";
+      const actionScope2 = asString(record.actionCallId) || asString(record.runId);
+      const baseExtensionId = asString(item.id) || asString(payload.item_id) || (outputIndex ? `extension-${outputIndex}` : itemType);
+      upsertReasoning({
+        reasoningId: scopedId(actionScope2, `extension:${baseExtensionId}`),
+        text: `${itemType}: ${extensionText}`,
+        raw: item,
+        mode: "replace"
+      });
+      continue;
+    }
     if (itemType !== "reasoning") continue;
     let text = "";
     const summary = item.summary;
@@ -17620,7 +17644,7 @@ function deriveReasoningByAssistant(traces) {
   }
   return buckets;
 }
-var RESPOND_TOOL_NAME = "gambit_respond";
+var LEGACY_RESPOND_TOOL_NAME = "gambit_respond";
 var stringifyMessageContent = (value) => {
   if (value === null || value === void 0) return "";
   if (typeof value === "string") return value;
@@ -17641,7 +17665,7 @@ var safeParseJson = (text) => {
 var summarizeRespondMessage = (message) => {
   if (!message || message.role !== "tool") return null;
   const name = typeof message.name === "string" ? message.name : void 0;
-  if (name !== RESPOND_TOOL_NAME) return null;
+  if (name !== LEGACY_RESPOND_TOOL_NAME) return null;
   const parsed = safeParseJson(typeof message.content === "string" ? message.content : "");
   const payload = parsed && typeof parsed === "object" ? "payload" in parsed ? parsed.payload : parsed : void 0;
   const status = typeof parsed?.status === "number" ? parsed.status : void 0;
@@ -17683,7 +17707,7 @@ function buildConversationEntries(state) {
         message: {
           role: "assistant",
           content: respondSummary.displayText,
-          name: RESPOND_TOOL_NAME
+          name: LEGACY_RESPOND_TOOL_NAME
         },
         feedback: ref ? feedbackByRef.get(ref.id) : void 0,
         respond: {
@@ -17736,6 +17760,12 @@ function normalizeAppPath(input) {
     }
     return DEFAULT_GRADE_PATH;
   }
+  if (trimmed === "/verify") {
+    if (window.location.pathname !== DEFAULT_VERIFY_PATH) {
+      window.history.replaceState({}, "", DEFAULT_VERIFY_PATH);
+    }
+    return DEFAULT_VERIFY_PATH;
+  }
   if (trimmed === "/build") {
     if (window.location.pathname !== DEFAULT_BUILD_PATH) {
       window.history.replaceState({}, "", DEFAULT_BUILD_PATH);
@@ -17748,7 +17778,7 @@ function normalizeAppPath(input) {
     }
     return DEFAULT_WORKSPACE_DEBUG_PATH;
   }
-  if (/^\/workspaces\/[^/]+\/(debug|build)$/.test(trimmed) || /^\/workspaces\/[^/]+\/(test|grade)(?:\/[^/]+)?$/.test(trimmed)) {
+  if (/^\/workspaces\/[^/]+\/(debug|build|verify)$/.test(trimmed) || /^\/workspaces\/[^/]+\/(test|grade)(?:\/[^/]+)?$/.test(trimmed)) {
     return trimmed;
   }
   if (trimmed.startsWith("/debug/workspaces/")) {
@@ -17758,7 +17788,7 @@ function normalizeAppPath(input) {
     window.history.replaceState({}, "", next);
     return next;
   }
-  if (trimmed.startsWith("/workspaces/") && !trimmed.includes("/debug") && !trimmed.includes("/test") && !trimmed.includes("/grade") && !trimmed.includes("/build") && trimmed !== DEFAULT_WORKSPACE_DEBUG_PATH) {
+  if (trimmed.startsWith("/workspaces/") && !trimmed.includes("/debug") && !trimmed.includes("/test") && !trimmed.includes("/grade") && !trimmed.includes("/build") && !trimmed.includes("/verify") && trimmed !== DEFAULT_WORKSPACE_DEBUG_PATH) {
     const remainder = trimmed.slice("/workspaces/".length);
     if (remainder && remainder !== "new") {
       const decoded = decodeURIComponent(remainder);
@@ -18374,6 +18404,11 @@ code:not(pre *) {
   width: 100%;
   min-height: 0;
 }
+.verify-shell .verify-layout {
+  grid-template-columns: 320px minmax(0, 1fr);
+  width: 100%;
+  min-height: 0;
+}
 @media (max-width: 1100px) {
   .calibrate-shell .calibrate-layout {
     grid-template-columns: minmax(0, 1fr);
@@ -18382,6 +18417,9 @@ code:not(pre *) {
     position: static;
     max-height: none;
   }
+  .verify-shell .verify-layout {
+    grid-template-columns: minmax(0, 1fr);
+  }
 }
 .calibrate-shell .calibrate-runner {
   flex: 0 0 auto;
@@ -18391,6 +18429,227 @@ code:not(pre *) {
   min-height: 0;
   overflow-y: auto;
 }
+.verify-controls {
+  display: flex;
+  flex-direction: column;
+  gap: 12px;
+}
+.verify-controls-header {
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+}
+.verify-number-grid {
+  display: grid;
+  grid-template-columns: repeat(2, minmax(0, 1fr));
+  gap: 8px;
+}
+.verify-number-field {
+  display: flex;
+  flex-direction: column;
+  gap: 6px;
+  font-size: 12px;
+  color: var(--color-text-muted);
+  font-weight: 600;
+}
+.verify-number-field input {
+  border: 1px solid var(--color-border-strong);
+  border-radius: calc(10px * var(--corner-radius-scale, 1));
+  corner-shape: squircle;
+  padding: 6px 8px;
+  background: var(--color-surface);
+  color: var(--color-text);
+  font-size: 13px;
+}
+.verify-results {
+  display: flex;
+  flex-direction: column;
+  gap: 12px;
+  min-height: 0;
+  overflow-y: auto;
+}
+.verify-status-row {
+  display: flex;
+  flex-wrap: wrap;
+  align-items: flex-start;
+  justify-content: flex-start;
+  gap: 10px;
+}
+.verify-status-main {
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+}
+.verify-status-meta {
+  font-size: 12px;
+  color: var(--color-text-muted);
+}
+.verify-progress-row {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 8px;
+  font-size: 12px;
+  color: var(--color-text-muted);
+}
+.verify-verdict-badge {
+  margin-left: auto;
+  border-radius: calc(10px * var(--corner-radius-scale, 1));
+  corner-shape: squircle;
+  border: 1px solid var(--color-border);
+  padding: 4px 8px;
+  font-size: 11px;
+  font-weight: 700;
+  letter-spacing: 0.02em;
+}
+.verify-verdict-badge--pass {
+  background: var(--color-success-soft);
+  border-color: var(--color-success-strong);
+  color: var(--color-success-strong);
+}
+.verify-verdict-badge--warn {
+  background: var(--color-warning-soft);
+  border-color: var(--color-warning-border);
+  color: var(--color-warning);
+}
+.verify-verdict-badge--fail {
+  background: var(--color-danger-soft);
+  border-color: var(--color-danger-strong);
+  color: var(--color-danger-strong);
+}
+.verify-metric-grid {
+  display: grid;
+  grid-template-columns: repeat(2, minmax(0, 1fr));
+  gap: 8px;
+}
+@media (max-width: 800px) {
+  .verify-metric-grid {
+    grid-template-columns: minmax(0, 1fr);
+  }
+}
+.verify-metric-card {
+  border: 1px solid var(--color-border);
+  border-radius: calc(12px * var(--corner-radius-scale, 1));
+  corner-shape: squircle;
+  padding: 8px;
+  background: var(--color-surface-muted);
+}
+.verify-metric-label {
+  font-size: 11px;
+  text-transform: uppercase;
+  letter-spacing: 0.02em;
+  color: var(--color-text-muted);
+}
+.verify-metric-value {
+  margin-top: 4px;
+  font-size: 20px;
+  font-weight: 700;
+  color: var(--color-text);
+}
+.verify-sample-size-row {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 8px;
+  flex-wrap: wrap;
+}
+.verify-sample-size-copy {
+  display: flex;
+  flex-direction: column;
+}
+.verify-sample-size-row .verify-metric-value {
+  margin-top: 4px;
+}
+.verify-sample-scope-select {
+  min-width: 170px;
+}
+.verify-sample-scope-select .gds-listbox-field-label {
+  display: none;
+}
+.verify-metric-value--compact {
+  font-size: 14px;
+}
+.verify-section {
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+}
+.verify-section-header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 8px;
+  flex-wrap: wrap;
+}
+.verify-section-controls {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  flex-wrap: wrap;
+}
+.verify-section-sort {
+  min-width: 150px;
+}
+.verify-section-sort .gds-listbox-field-label {
+  display: none;
+}
+.verify-outlier-list {
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+}
+.verify-outlier-card {
+  border: 1px solid var(--color-border);
+  border-radius: calc(12px * var(--corner-radius-scale, 1));
+  corner-shape: squircle;
+  padding: 8px;
+  background: var(--color-surface-muted);
+  display: flex;
+  flex-direction: column;
+  gap: 6px;
+}
+.verify-outlier-header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 8px;
+}
+.verify-outlier-meta {
+  font-size: 12px;
+  color: var(--color-text-muted);
+}
+.verify-outlier-links {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 8px;
+  font-size: 12px;
+}
+.verify-request-list {
+  list-style: none;
+  margin: 0;
+  padding: 0;
+  display: flex;
+  flex-direction: column;
+  gap: 6px;
+}
+.verify-request-row {
+  border: 1px solid var(--color-border);
+  border-radius: calc(10px * var(--corner-radius-scale, 1));
+  corner-shape: squircle;
+  padding: 6px 8px;
+  background: var(--color-surface-muted);
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  flex-wrap: wrap;
+  font-size: 12px;
+}
+.verify-request-index {
+  color: var(--color-text-muted);
+  min-width: 28px;
+}
+.verify-request-error {
+  color: var(--color-danger-strong);
+}
 .calibrate-run-card {
   border: 1px solid var(--color-border);
   border-radius: calc(14px * var(--corner-radius-scale, 1));
@@ -20110,6 +20369,67 @@ code:not(pre *) {
 .workbench-chat-current.is-history {
   transform: translateX(85%);
 }
+.workbench-chat-readonly-overlay {
+  position: absolute;
+  inset: 0;
+  z-index: 2;
+  background: rgba(248, 250, 252, 0.88);
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  padding: 16px;
+}
+.workbench-chat-readonly-card {
+  background: var(--color-surface);
+  border: 1px solid var(--color-border);
+  border-radius: calc(16px * var(--corner-radius-scale, 1));
+  corner-shape: squircle;
+  box-shadow: 0 12px 30px rgba(15, 23, 42, 0.08);
+  width: 100%;
+  max-width: 640px;
+  padding: 16px;
+  display: flex;
+  flex-direction: column;
+  gap: 10px;
+  text-align: left;
+}
+.workbench-chat-readonly-title {
+  margin: 0;
+  font-size: 16px;
+}
+.workbench-chat-readonly-copy {
+  margin: 0;
+  font-size: 13px;
+  color: var(--color-text-muted);
+}
+.workbench-chat-readonly-actions {
+  display: flex;
+  flex-wrap: wrap;
+  align-items: center;
+  gap: 8px;
+}
+.workbench-chat-command-row {
+  display: flex;
+  align-items: flex-start;
+  gap: 8px;
+}
+.workbench-chat-command-code {
+  margin: 0;
+  flex: 1;
+  max-width: 100%;
+  padding: 10px 12px;
+  border: 1px solid var(--color-border);
+  border-radius: calc(10px * var(--corner-radius-scale, 1));
+  corner-shape: squircle;
+  background: var(--color-surface-muted);
+  overflow-x: auto;
+}
+.workbench-chat-command-code code {
+  font-size: 12px;
+  line-height: 1.4;
+  white-space: pre-wrap;
+  word-break: break-all;
+}
 .gds-accordion .gds-accordion-open-only {
   display: none;
 }
@@ -20334,6 +20654,23 @@ code:not(pre *) {
   opacity: 0.6;
   cursor: not-allowed;
 }
+.gds-listbox--size-small .gds-listbox-field-label {
+  margin-bottom: 4px;
+  font-size: 12px;
+}
+.gds-listbox--size-small .gds-listbox-trigger {
+  padding: 4px 28px 4px 10px;
+  gap: 0;
+}
+.gds-listbox--size-small .gds-listbox-label {
+  font-size: 13px;
+}
+.gds-listbox--size-small .gds-listbox-meta {
+  display: none;
+}
+.gds-listbox--size-small .gds-listbox-caret {
+  right: 9px;
+}
 .gds-listbox-label {
   font-weight: 600;
   font-size: 14px;
@@ -21655,26 +21992,42 @@ function formatScoreLabel(score) {
   return score > 0 ? `+${score}` : String(score);
 }
 function formatContextLabel(context) {
-  if (context.source === "scenario_run_error") return "Error";
+  if (context.source === "scenario_run_error") return "Scenario error";
+  if (context.source === "grader_run_error") return "Grader error";
   if (context.source === "message_rating") {
     return formatScoreLabel(context.score);
   }
+  if (context.source === "verify_outlier") return "Verify";
   return "Flag";
 }
 function formatContextTooltip(context) {
-  if (context.source === "scenario_run_error") return context.error;
-  if (context.source === "message_rating") {
-    return context.reason?.trim() || `Rating ${formatScoreLabel(context.score)}`;
+  switch (context.source) {
+    case "scenario_run_error":
+    case "grader_run_error":
+      return context.error;
+    case "message_rating":
+      return context.reason?.trim() || `Rating ${formatScoreLabel(context.score)}`;
+    case "grading_flag":
+    case "verify_outlier":
+      return context.message;
+  }
+}
+function getVerifyOutlierClass(context) {
+  if (typeof context.instability === "boolean") {
+    return getScoreClass(context.instability ? -1 : 1);
   }
-  return context.message;
+  if (typeof context.score === "number" && Number.isFinite(context.score)) {
+    return getScoreClass(context.score);
+  }
+  return getScoreClass(0);
 }
 function WorkbenchComposerChip(props) {
   const { className, context, enabled, onEnabledChange, onRemove, testId, ...rest } = props;
   const label = formatContextLabel(context);
   const tooltip = formatContextTooltip(context);
   const score = "score" in context ? context.score : void 0;
-  const badgeClassName = classNames("workbench-context-chip", context.source === "scenario_run_error" && "workbench-context-chip--error", context.source === "grading_flag" && "workbench-context-chip--flag", context.source === "grading_flag" && typeof score === "number" && getScoreClass(score), context.source === "message_rating" && "workbench-context-chip--rating", context.source === "message_rating" && typeof score === "number" && getScoreClass(score));
-  const testIds = context.source === "scenario_run_error" ? {
+  const badgeClassName = classNames("workbench-context-chip", (context.source === "scenario_run_error" || context.source === "grader_run_error") && "workbench-context-chip--error", context.source === "grading_flag" && "workbench-context-chip--flag", context.source === "grading_flag" && typeof score === "number" && getScoreClass(score), context.source === "verify_outlier" && "workbench-context-chip--flag", context.source === "verify_outlier" && getVerifyOutlierClass(context), context.source === "message_rating" && "workbench-context-chip--rating", context.source === "message_rating" && typeof score === "number" && getScoreClass(score));
+  const testIds = context.source === "scenario_run_error" || context.source === "grader_run_error" ? {
     badge: "workbench-error-chip",
     toggle: "workbench-error-chip-toggle",
     remove: "workbench-error-chip-remove"
@@ -21685,7 +22038,7 @@ function WorkbenchComposerChip(props) {
   const content = context.source === "grading_flag" ? /* @__PURE__ */ (0, import_jsx_runtime21.jsx)(Icon, {
     name: "flag",
     size: 10
-  }) : label;
+  }) : context.source === "verify_outlier" ? "Verify" : label;
   if (isPassive) {
     return /* @__PURE__ */ (0, import_jsx_runtime21.jsx)(Tooltip, {
       content: tooltip,
@@ -23306,6 +23659,12 @@ function WorkspaceProvider(props) {
           ];
         });
       }
+      if (!data.run) {
+        throw new Error(data.error ?? "Calibration response missing run payload");
+      }
+      if (data.run.status !== "completed") {
+        throw new Error(data.run.error ?? `Calibration run ended with status ${data.run.status}`);
+      }
       setGradeError(null);
       return data;
     } catch (err) {
@@ -23500,6 +23859,9 @@ function useBuildChat() {
 }
 
 // simulator-ui/src/Chat.tsx
+function isWorkbenchErrorChip(chip) {
+  return chip.source === "scenario_run_error" || chip.source === "grader_run_error";
+}
 var ERROR_CONTEXT_START_MARKER = "[gambit:error-context/v1]";
 var ERROR_CONTEXT_END_MARKER = "[/gambit:error-context/v1]";
 var WORKBENCH_CONTEXT_START_MARKER = "[gambit:workbench-context/v2]";
@@ -23514,12 +23876,12 @@ function parseWorkbenchContext(value) {
   }
   const workspaceId = typeof record.workspaceId === "string" ? record.workspaceId : void 0;
   const runId = typeof record.runId === "string" ? record.runId : void 0;
-  if (record.source === "scenario_run_error") {
+  if (record.source === "scenario_run_error" || record.source === "grader_run_error") {
     if (typeof record.error !== "string" || record.error.trim().length === 0) {
       return null;
     }
     return {
-      source: "scenario_run_error",
+      source: record.source,
       workspaceId,
       runId,
       capturedAt: record.capturedAt,
@@ -23557,6 +23919,21 @@ function parseWorkbenchContext(value) {
       message: record.message
     };
   }
+  if (record.source === "verify_outlier") {
+    if (typeof record.outlierKey !== "string" || record.outlierKey.trim().length === 0 || typeof record.message !== "string" || record.message.trim().length === 0) {
+      return null;
+    }
+    return {
+      source: "verify_outlier",
+      workspaceId,
+      runId,
+      capturedAt: record.capturedAt,
+      outlierKey: record.outlierKey,
+      instability: typeof record.instability === "boolean" ? record.instability : void 0,
+      score: typeof record.score === "number" && Number.isFinite(record.score) ? record.score : void 0,
+      message: record.message
+    };
+  }
   return null;
 }
 function decodeLegacyWorkbenchErrorContext(content) {
@@ -23648,7 +24025,7 @@ function UserMessageContent(props) {
         children: decoded.contexts.map((context, index) => /* @__PURE__ */ (0, import_jsx_runtime27.jsx)(WorkbenchComposerChip, {
           className: "workbench-context-chip--transcript",
           context,
-          testId: context.source === "scenario_run_error" ? "workbench-transcript-error-chip" : void 0
+          testId: context.source === "scenario_run_error" ? "workbench-transcript-error-chip" : context.source === "grader_run_error" ? "workbench-transcript-error-chip" : void 0
         }, `${context.source}-${index}`))
       }),
       showBody && /* @__PURE__ */ (0, import_jsx_runtime27.jsx)("div", {
@@ -23811,13 +24188,13 @@ function ChatView(props) {
       return;
     }
     if (!onScenarioErrorChipChange) return;
-    const errorChip = next.find((chip) => chip.source === "scenario_run_error");
+    const errorChip = next.find(isWorkbenchErrorChip);
     if (!errorChip) {
       onScenarioErrorChipChange(null);
       return;
     }
     onScenarioErrorChipChange({
-      source: "scenario_run_error",
+      source: errorChip.source,
       workspaceId: errorChip.workspaceId,
       runId: errorChip.runId,
       capturedAt: errorChip.capturedAt,
@@ -23835,18 +24212,20 @@ function ChatView(props) {
     const message = chatDraft.trim();
     const activeChips = resolvedComposerChips.filter((chip) => chip.enabled);
     const activeChipIds = new Set(activeChips.map((chip) => chip.chipId));
-    const activeContexts = activeChips.map((chip) => {
-      if (chip.source === "scenario_run_error") {
-        return {
-          source: "scenario_run_error",
+    const activeContexts = [];
+    for (const chip of activeChips) {
+      if (isWorkbenchErrorChip(chip)) {
+        activeContexts.push({
+          source: chip.source,
           workspaceId: chip.workspaceId,
           runId: chip.runId,
           capturedAt: chip.capturedAt,
           error: chip.error
-        };
+        });
+        continue;
       }
       if (chip.source === "message_rating") {
-        return {
+        activeContexts.push({
           source: "message_rating",
           workspaceId: chip.workspaceId,
           runId: chip.runId,
@@ -23856,19 +24235,35 @@ function ChatView(props) {
           statePointer: chip.statePointer,
           score: chip.score,
           reason: chip.reason
-        };
+        });
+        continue;
       }
-      return {
-        source: "grading_flag",
-        workspaceId: chip.workspaceId,
-        runId: chip.runId,
-        capturedAt: chip.capturedAt,
-        flagId: chip.flagId,
-        refId: chip.refId,
-        score: chip.score,
-        message: chip.message
-      };
-    });
+      if (chip.source === "grading_flag") {
+        activeContexts.push({
+          source: "grading_flag",
+          workspaceId: chip.workspaceId,
+          runId: chip.runId,
+          capturedAt: chip.capturedAt,
+          flagId: chip.flagId,
+          refId: chip.refId,
+          score: chip.score,
+          message: chip.message
+        });
+        continue;
+      }
+      if (chip.source === "verify_outlier") {
+        activeContexts.push({
+          source: "verify_outlier",
+          workspaceId: chip.workspaceId,
+          runId: chip.runId,
+          capturedAt: chip.capturedAt,
+          outlierKey: chip.outlierKey,
+          instability: chip.instability,
+          score: chip.score,
+          message: chip.message
+        });
+      }
+    }
     if (!message && activeContexts.length === 0) return;
     const outboundMessage = activeContexts.length > 0 ? encodeWorkbenchMessageWithContext(message, activeContexts) : message;
     setOptimisticUser({
@@ -24201,9 +24596,21 @@ function WorkbenchDrawer(props) {
   const [chatHistoryLoading, setChatHistoryLoading] = (0, import_react28.useState)(false);
   const [chatHistoryError, setChatHistoryError] = (0, import_react28.useState)(null);
   const [copiedStatePath, setCopiedStatePath] = (0, import_react28.useState)(false);
+  const [copiedCodexLoginCommand, setCopiedCodexLoginCommand] = (0, import_react28.useState)(false);
+  const [codexTrustPending, setCodexTrustPending] = (0, import_react28.useState)(false);
+  const [codexTrustError, setCodexTrustError] = (0, import_react28.useState)(null);
+  const [codexTrustSuccess, setCodexTrustSuccess] = (0, import_react28.useState)(null);
+  const [codexWorkspaceWriteEnabled, setCodexWorkspaceWriteEnabled] = (0, import_react28.useState)(null);
+  const [codexWorkspaceLoggedIn, setCodexWorkspaceLoggedIn] = (0, import_react28.useState)(null);
+  const [codexLoginStatusText, setCodexLoginStatusText] = (0, import_react28.useState)(null);
+  const [codexTrustedPath, setCodexTrustedPath] = (0, import_react28.useState)(null);
+  const [codexTrustOverlayDismissed, setCodexTrustOverlayDismissed] = (0, import_react28.useState)(false);
   const initializedChipTrackingRef = (0, import_react28.useRef)(false);
   const seenRatingChipIdsRef = (0, import_react28.useRef)(/* @__PURE__ */ new Set());
   const seenFlagChipIdsRef = (0, import_react28.useRef)(/* @__PURE__ */ new Set());
+  const showCodexTrustOverlay = (codexWorkspaceWriteEnabled === false || codexWorkspaceLoggedIn === false) && !codexTrustOverlayDismissed || Boolean(codexTrustError);
+  const workspaceIdForTrust = (sessionId ?? run.id) || void 0;
+  const codexLoginCommand = codexTrustedPath ? `CODEX_HOME="${codexTrustedPath}/.codex" codex login` : 'CODEX_HOME="<workspace>/.codex" codex login';
   const resolvedStatePath = (0, import_react28.useMemo)(() => {
     if (statePath) return statePath;
     const meta = sessionDetail?.meta;
@@ -24421,9 +24828,49 @@ function WorkbenchDrawer(props) {
     initializedChipTrackingRef.current = false;
     seenRatingChipIdsRef.current.clear();
     seenFlagChipIdsRef.current.clear();
+    setCodexTrustPending(false);
+    setCodexTrustError(null);
+    setCodexTrustSuccess(null);
+    setCodexWorkspaceWriteEnabled(null);
+    setCodexWorkspaceLoggedIn(null);
+    setCodexLoginStatusText(null);
+    setCodexTrustedPath(null);
+    setCopiedCodexLoginCommand(false);
+    setCodexTrustOverlayDismissed(false);
   }, [
     sessionId
   ]);
+  (0, import_react28.useEffect)(() => {
+    if (!open) return;
+    if (!workspaceIdForTrust) return;
+    let canceled = false;
+    setCodexTrustError(null);
+    fetch(`/api/codex/trust-workspace?workspaceId=${encodeURIComponent(workspaceIdForTrust)}`).then(async (response) => {
+      const payload = await response.json();
+      if (!response.ok || payload.ok === false) {
+        throw new Error(payload.error || response.statusText);
+      }
+      if (canceled) return;
+      setCodexWorkspaceWriteEnabled(payload.writeEnabled === true);
+      setCodexWorkspaceLoggedIn(payload.codexLoggedIn === true);
+      setCodexLoginStatusText(typeof payload.codexLoginStatus === "string" ? payload.codexLoginStatus : null);
+      setCodexTrustedPath(typeof payload.trustedPath === "string" ? payload.trustedPath : null);
+    }).catch((err) => {
+      if (canceled) return;
+      setCodexWorkspaceWriteEnabled(null);
+      setCodexWorkspaceLoggedIn(null);
+      setCodexLoginStatusText(null);
+      setCodexTrustError(err instanceof Error ? err.message : String(err));
+    }).finally(() => {
+      if (canceled) return;
+    });
+    return () => {
+      canceled = true;
+    };
+  }, [
+    open,
+    workspaceIdForTrust
+  ]);
   (0, import_react28.useEffect)(() => {
     if (loading) return;
     const currentRatingChipIds = new Set(resolvedFeedbackItems.map(({ entry }) => `rating:${entry.messageRefId}:${entry.id}`));
@@ -24460,18 +24907,20 @@ function WorkbenchDrawer(props) {
       let didChange = false;
       const syncedChips = composerChips.filter((chip) => {
         if (chip.source === "message_rating") {
+          if (!chip.chipId.startsWith("rating:")) return true;
           const stillExists = ratingByChipId.has(chip.chipId);
           if (!stillExists) didChange = true;
           return stillExists;
         }
         if (chip.source === "grading_flag") {
+          if (!chip.chipId.startsWith("flag:")) return true;
           const stillExists = flagByChipId.has(chip.chipId);
           if (!stillExists) didChange = true;
           return stillExists;
         }
         return true;
       }).map((chip) => {
-        const latest = chip.source === "message_rating" ? ratingByChipId.get(chip.chipId) : chip.source === "grading_flag" ? flagByChipId.get(chip.chipId) : void 0;
+        const latest = chip.source === "message_rating" && chip.chipId.startsWith("rating:") ? ratingByChipId.get(chip.chipId) : chip.source === "grading_flag" && chip.chipId.startsWith("flag:") ? flagByChipId.get(chip.chipId) : void 0;
         if (!latest) return chip;
         const next = {
           ...chip,
@@ -24539,6 +24988,13 @@ function WorkbenchDrawer(props) {
   }, [
     resolvedStatePath
   ]);
+  const handleCopyCodexLoginCommand = (0, import_react28.useCallback)(() => {
+    navigator.clipboard?.writeText(codexLoginCommand);
+    setCopiedCodexLoginCommand(true);
+    window.setTimeout(() => setCopiedCodexLoginCommand(false), 1200);
+  }, [
+    codexLoginCommand
+  ]);
   (0, import_react28.useEffect)(() => {
     if (!open) return;
     if (!onClose) return;
@@ -24553,6 +25009,55 @@ function WorkbenchDrawer(props) {
     onClose,
     open
   ]);
+  const trustWorkspaceInCodex = (0, import_react28.useCallback)(async () => {
+    setCodexTrustPending(true);
+    setCodexTrustError(null);
+    setCodexTrustSuccess(null);
+    try {
+      const statusResponse = await fetch(`/api/codex/trust-workspace?workspaceId=${encodeURIComponent(workspaceIdForTrust ?? "")}`);
+      const statusPayload = await statusResponse.json();
+      if (!statusResponse.ok || statusPayload.ok === false) {
+        throw new Error(statusPayload.error || statusResponse.statusText);
+      }
+      if (statusPayload.writeEnabled === true && statusPayload.codexLoggedIn === true) {
+        setCodexWorkspaceWriteEnabled(true);
+        setCodexWorkspaceLoggedIn(true);
+        setCodexTrustSuccess("Workspace is already configured for Codex writes.");
+        setCodexTrustOverlayDismissed(true);
+        return;
+      }
+      setCodexWorkspaceWriteEnabled(statusPayload.writeEnabled === true);
+      setCodexWorkspaceLoggedIn(statusPayload.codexLoggedIn === true);
+      setCodexLoginStatusText(typeof statusPayload.codexLoginStatus === "string" ? statusPayload.codexLoginStatus : null);
+      setCodexTrustedPath(typeof statusPayload.trustedPath === "string" ? statusPayload.trustedPath : null);
+      const response = await fetch("/api/codex/trust-workspace", {
+        method: "POST",
+        headers: {
+          "content-type": "application/json"
+        },
+        body: JSON.stringify({
+          workspaceId: workspaceIdForTrust
+        })
+      });
+      const payload = await response.json();
+      if (!response.ok || payload.ok === false) {
+        throw new Error(payload.error || response.statusText);
+      }
+      const trustedPath = typeof payload.trustedPath === "string" ? payload.trustedPath : "workspace";
+      setCodexTrustSuccess(`Codex write enabled for: ${trustedPath}`);
+      setCodexWorkspaceWriteEnabled(payload.writeEnabled === true);
+      setCodexWorkspaceLoggedIn(payload.codexLoggedIn === true);
+      setCodexLoginStatusText(typeof payload.codexLoginStatus === "string" ? payload.codexLoginStatus : null);
+      setCodexTrustedPath(typeof payload.trustedPath === "string" ? payload.trustedPath : null);
+      setCodexTrustOverlayDismissed(payload.codexLoggedIn === true);
+    } catch (err) {
+      setCodexTrustError(err instanceof Error ? err.message : String(err));
+    } finally {
+      setCodexTrustPending(false);
+    }
+  }, [
+    workspaceIdForTrust
+  ]);
   if (!open) return null;
   return /* @__PURE__ */ (0, import_jsx_runtime31.jsx)("aside", {
     className: "workbench-drawer-docked",
@@ -24653,12 +25158,87 @@ function WorkbenchDrawer(props) {
                       })
                     ]
                   }),
-                  /* @__PURE__ */ (0, import_jsx_runtime31.jsx)("div", {
+                  /* @__PURE__ */ (0, import_jsx_runtime31.jsxs)("div", {
                     className: `workbench-chat-current${chatHistoryOpen ? " is-history" : ""}`,
-                    children: /* @__PURE__ */ (0, import_jsx_runtime31.jsx)(Chat, {
-                      composerChips,
-                      onComposerChipsChange
-                    })
+                    children: [
+                      showCodexTrustOverlay && /* @__PURE__ */ (0, import_jsx_runtime31.jsx)("div", {
+                        className: "workbench-chat-readonly-overlay",
+                        children: /* @__PURE__ */ (0, import_jsx_runtime31.jsxs)("div", {
+                          className: "workbench-chat-readonly-card",
+                          children: [
+                            /* @__PURE__ */ (0, import_jsx_runtime31.jsx)("h3", {
+                              className: "workbench-chat-readonly-title",
+                              children: "Codex setup required"
+                            }),
+                            codexWorkspaceWriteEnabled === false && /* @__PURE__ */ (0, import_jsx_runtime31.jsx)("p", {
+                              className: "workbench-chat-readonly-copy",
+                              children: "Codex write access is disabled for this workspace. Trust this workspace to enable file edits."
+                            }),
+                            /* @__PURE__ */ (0, import_jsx_runtime31.jsx)("div", {
+                              className: "workbench-chat-readonly-actions",
+                              children: codexWorkspaceWriteEnabled === false && /* @__PURE__ */ (0, import_jsx_runtime31.jsx)(Button, {
+                                variant: "primary",
+                                onClick: () => trustWorkspaceInCodex(),
+                                disabled: codexTrustPending,
+                                children: codexTrustPending ? "Trusting..." : "Trust workspace"
+                              })
+                            }),
+                            codexWorkspaceLoggedIn === false && /* @__PURE__ */ (0, import_jsx_runtime31.jsxs)(import_jsx_runtime31.Fragment, {
+                              children: [
+                                /* @__PURE__ */ (0, import_jsx_runtime31.jsx)("p", {
+                                  className: "workbench-chat-readonly-copy",
+                                  children: "Codex login is required for this workspace."
+                                }),
+                                /* @__PURE__ */ (0, import_jsx_runtime31.jsx)("p", {
+                                  className: "workbench-chat-readonly-copy",
+                                  children: "Run this in this workspace to authenticate Codex, then restart Gambit."
+                                }),
+                                /* @__PURE__ */ (0, import_jsx_runtime31.jsxs)("div", {
+                                  className: "workbench-chat-command-row",
+                                  children: [
+                                    /* @__PURE__ */ (0, import_jsx_runtime31.jsx)("pre", {
+                                      className: "workbench-chat-command-code",
+                                      children: /* @__PURE__ */ (0, import_jsx_runtime31.jsx)("code", {
+                                        children: codexLoginCommand
+                                      })
+                                    }),
+                                    /* @__PURE__ */ (0, import_jsx_runtime31.jsxs)(Button, {
+                                      variant: "secondary",
+                                      size: "small",
+                                      onClick: handleCopyCodexLoginCommand,
+                                      children: [
+                                        /* @__PURE__ */ (0, import_jsx_runtime31.jsx)(Icon, {
+                                          name: copiedCodexLoginCommand ? "copied" : "copy",
+                                          size: 14
+                                        }),
+                                        copiedCodexLoginCommand ? "Copied" : "Copy"
+                                      ]
+                                    })
+                                  ]
+                                })
+                              ]
+                            }),
+                            codexLoginStatusText && !/^not logged in$/i.test(codexLoginStatusText.trim()) && /* @__PURE__ */ (0, import_jsx_runtime31.jsx)(Callout, {
+                              children: codexLoginStatusText
+                            }),
+                            codexTrustError && /* @__PURE__ */ (0, import_jsx_runtime31.jsx)("div", {
+                              className: "error",
+                              children: codexTrustError
+                            }),
+                            /* @__PURE__ */ (0, import_jsx_runtime31.jsx)(Button, {
+                              variant: "secondary",
+                              onClick: () => setCodexTrustOverlayDismissed(true),
+                              disabled: codexTrustPending,
+                              children: "Dismiss"
+                            })
+                          ]
+                        })
+                      }),
+                      /* @__PURE__ */ (0, import_jsx_runtime31.jsx)(Chat, {
+                        composerChips,
+                        onComposerChipsChange
+                      })
+                    ]
                   })
                 ]
               })
@@ -24672,6 +25252,9 @@ function WorkbenchDrawer(props) {
           content: /* @__PURE__ */ (0, import_jsx_runtime31.jsxs)("div", {
             className: "workbench-ratings",
             children: [
+              codexTrustSuccess && /* @__PURE__ */ (0, import_jsx_runtime31.jsx)(Callout, {
+                children: codexTrustSuccess
+              }),
               showCopyStatePath && handleCopyStatePath && /* @__PURE__ */ (0, import_jsx_runtime31.jsxs)(import_jsx_runtime31.Fragment, {
                 children: [
                   /* @__PURE__ */ (0, import_jsx_runtime31.jsxs)(Button, {
@@ -25140,7 +25723,7 @@ function MessageBubble(props) {
               className: "respond-meta",
               children: [
                 /* @__PURE__ */ (0, import_jsx_runtime35.jsx)(Badge, {
-                  children: "gambit_respond"
+                  children: "legacy respond envelope"
                 }),
                 typeof entry.respond?.status === "number" && /* @__PURE__ */ (0, import_jsx_runtime35.jsxs)(Badge, {
                   variant: "ghost",
@@ -25253,16 +25836,20 @@ function FeedbackControls(props) {
   const [errorMessage, setErrorMessage] = (0, import_react32.useState)(null);
   const requestSeqRef = (0, import_react32.useRef)(0);
   (0, import_react32.useEffect)(() => {
-    setReason(feedback?.reason ?? "");
-    if (typeof feedback?.score === "number" || feedback?.reason !== void 0) {
-      setStatus("saved");
-    } else {
-      setStatus("idle");
+    const isEditing = status === "unsaved" || status === "saving";
+    if (!isEditing) {
+      setReason(feedback?.reason ?? "");
+      if (typeof feedback?.score === "number" || feedback?.reason !== void 0) {
+        setStatus("saved");
+      } else {
+        setStatus("idle");
+      }
+      setErrorMessage(null);
     }
-    setErrorMessage(null);
   }, [
     feedback?.reason,
-    feedback?.score
+    feedback?.score,
+    status
   ]);
   (0, import_react32.useEffect)(() => {
     if (typeof feedback?.score === "number") {
@@ -25963,7 +26550,7 @@ var import_jsx_runtime36 = __toESM(require_jsx_runtime());
 var import_react33 = __toESM(require__());
 var import_react_dom2 = __toESM(require__3());
 function Listbox(props) {
-  const { value, options, placeholder = "Select", disabled = false, onChange, label, labelClassName, id } = props;
+  const { value, options, placeholder = "Select", disabled = false, onChange, label, labelClassName, id, popoverMatchTriggerWidth = true, popoverMinWidth, popoverAlign = "left", size = "default" } = props;
   const [open, setOpen] = (0, import_react33.useState)(false);
   const rootRef = (0, import_react33.useRef)(null);
   const triggerRef = (0, import_react33.useRef)(null);
@@ -25983,17 +26570,40 @@ function Listbox(props) {
     options,
     value
   ]);
+  const selectedTriggerMeta = (0, import_react33.useMemo)(() => {
+    if (!selected) return null;
+    if ("triggerMeta" in selected) return selected.triggerMeta ?? null;
+    return selected.meta ?? null;
+  }, [
+    selected
+  ]);
   const updatePopover = (0, import_react33.useCallback)(() => {
     const trigger = triggerRef.current;
     if (!trigger) return;
     const rect = trigger.getBoundingClientRect();
-    setPopoverStyle({
+    const style = {
       position: "fixed",
-      top: rect.bottom + 6,
-      left: rect.left,
-      width: rect.width
-    });
-  }, []);
+      top: rect.bottom + 6
+    };
+    if (popoverMatchTriggerWidth) {
+      style.width = rect.width;
+    } else if (typeof popoverMinWidth === "number" && Number.isFinite(popoverMinWidth)) {
+      style.minWidth = popoverMinWidth;
+    }
+    if (popoverAlign === "right") {
+      style.left = popoverMatchTriggerWidth ? rect.right - rect.width : rect.right;
+      if (!popoverMatchTriggerWidth) {
+        style.transform = "translateX(-100%)";
+      }
+    } else {
+      style.left = rect.left;
+    }
+    setPopoverStyle(style);
+  }, [
+    popoverAlign,
+    popoverMatchTriggerWidth,
+    popoverMinWidth
+  ]);
   (0, import_react33.useLayoutEffect)(() => {
     if (!open) return;
     updatePopover();
@@ -26032,8 +26642,9 @@ function Listbox(props) {
     open,
     updatePopover
   ]);
+  const rootClassName = size === "small" ? "gds-listbox gds-listbox--size-small" : "gds-listbox";
   return /* @__PURE__ */ (0, import_jsx_runtime36.jsxs)("div", {
-    className: "gds-listbox",
+    className: rootClassName,
     ref: rootRef,
     children: [
       label && /* @__PURE__ */ (0, import_jsx_runtime36.jsx)("label", {
@@ -26054,11 +26665,11 @@ function Listbox(props) {
         ref: triggerRef,
         children: [
           /* @__PURE__ */ (0, import_jsx_runtime36.jsx)(ScrollingText, {
-            text: selected?.label ?? placeholder,
+            text: selected?.triggerLabel ?? selected?.label ?? placeholder,
             className: "gds-listbox-label"
           }),
-          selected?.meta && /* @__PURE__ */ (0, import_jsx_runtime36.jsx)(ScrollingText, {
-            text: selected.meta,
+          selectedTriggerMeta && /* @__PURE__ */ (0, import_jsx_runtime36.jsx)(ScrollingText, {
+            text: selectedTriggerMeta,
             className: "gds-listbox-meta"
           }),
           /* @__PURE__ */ (0, import_jsx_runtime36.jsx)("span", {
@@ -26176,7 +26787,7 @@ var scenarioRunIdFromCalibrationRun = (run) => {
   const scenarioRunId = meta.scenarioRunId;
   return typeof scenarioRunId === "string" && scenarioRunId.trim().length > 0 ? scenarioRunId : null;
 };
-function GradePage({ setNavActions, onAppPathChange, activeWorkspaceId, onFlagsUpdate, onOptimisticToggleFlag, onOptimisticFlagReason, requestedGradeRunId }) {
+function GradePage({ setNavActions, onAppPathChange, activeWorkspaceId, onFlagsUpdate, onOptimisticToggleFlag, onOptimisticFlagReason, onAddErrorToWorkbench, requestedGradeRunId }) {
   const workspaceGrade = useWorkspaceGrade();
   const { loading, error, running, graders, sessions, sessionDetail, loadData, loadSessionDetail, runGrader: runGrade, toggleFlag: toggleGradeFlag, updateFlagReason: updateGradeFlagReason } = workspaceGrade;
   const workspaceRouting = useWorkspaceRouting();
@@ -26757,8 +27368,23 @@ function GradePage({ setNavActions, onAppPathChange, activeWorkspaceId, onFlagsU
             gap: 12
           },
           children: [
-            error && /* @__PURE__ */ (0, import_jsx_runtime39.jsx)("div", {
-              className: "error",
+            error && /* @__PURE__ */ (0, import_jsx_runtime39.jsx)(Callout, {
+              variant: "danger",
+              title: "Grader run failed",
+              actions: /* @__PURE__ */ (0, import_jsx_runtime39.jsx)(Button, {
+                variant: "secondary",
+                size: "small",
+                onClick: () => onAddErrorToWorkbench?.({
+                  source: "grader_run_error",
+                  workspaceId: selectedSessionId ?? activeWorkspaceId ?? void 0,
+                  runId: routeGradeRunId ?? expandedRunId ?? void 0,
+                  error
+                }),
+                disabled: !onAddErrorToWorkbench,
+                "data-testid": "grade-add-error-to-chat",
+                children: "Add to chat"
+              }),
+              "data-testid": "grade-error-callout",
               children: error
             }),
             loading && /* @__PURE__ */ (0, import_jsx_runtime39.jsx)("div", {
@@ -26769,11 +27395,21 @@ function GradePage({ setNavActions, onAppPathChange, activeWorkspaceId, onFlagsU
               children: [
                 /* @__PURE__ */ (0, import_jsx_runtime39.jsx)("div", {
                   className: "flex-column gap-4",
-                  children: /* @__PURE__ */ (0, import_jsx_runtime39.jsx)("div", {
+                  children: /* @__PURE__ */ (0, import_jsx_runtime39.jsxs)("div", {
                     className: "flex-row items-center gap-8",
-                    children: /* @__PURE__ */ (0, import_jsx_runtime39.jsx)("strong", {
-                      children: "Grader runs"
-                    })
+                    children: [
+                      /* @__PURE__ */ (0, import_jsx_runtime39.jsx)("strong", {
+                        children: "Grader runs"
+                      }),
+                      /* @__PURE__ */ (0, import_jsx_runtime39.jsxs)("span", {
+                        className: "secondary-note",
+                        children: [
+                          "(",
+                          filteredSessionRuns.length,
+                          ")"
+                        ]
+                      })
+                    ]
                   })
                 }),
                 runItems.length === 0 && /* @__PURE__ */ (0, import_jsx_runtime39.jsx)(Callout, {
@@ -27052,8 +27688,23 @@ function GradePage({ setNavActions, onAppPathChange, activeWorkspaceId, onFlagsU
                                   ]
                                 })
                               }),
-                              item.error && /* @__PURE__ */ (0, import_jsx_runtime39.jsx)("div", {
-                                className: "error",
+                              item.error && /* @__PURE__ */ (0, import_jsx_runtime39.jsx)(Callout, {
+                                variant: "danger",
+                                title: "Grade run failed",
+                                actions: /* @__PURE__ */ (0, import_jsx_runtime39.jsx)(Button, {
+                                  variant: "secondary",
+                                  size: "small",
+                                  onClick: () => onAddErrorToWorkbench?.({
+                                    source: "grader_run_error",
+                                    workspaceId: selectedSessionId ?? activeWorkspaceId ?? void 0,
+                                    runId: item.runId,
+                                    error: item.error
+                                  }),
+                                  disabled: !onAddErrorToWorkbench,
+                                  "data-testid": "grade-run-add-error-to-chat",
+                                  children: "Add to chat"
+                                }),
+                                "data-testid": "grade-run-error-callout",
                                 children: item.error
                               }),
                               isOpen && /* @__PURE__ */ (0, import_jsx_runtime39.jsxs)("div", {
@@ -27125,7 +27776,7 @@ var import_react38 = __toESM(require__());
 var import_jsx_runtime40 = __toESM(require_jsx_runtime());
 var import_react37 = __toESM(require__());
 function TestBotChatPanel(props) {
-  const { run, runWorkspaceId, runStatusLabel, activeWorkspaceId, requestedRunNotFound, canStart, canRunPersona, hasPersonaSelection, botJsonErrorCount, deckJsonErrorCount, missingBotInput, missingDeckInit, lastInitFill, isUserStart, showStartOverlay, canStartAssistant, canSendChat, chatDraft, setChatDraft, chatError, optimisticUser, streamingUser, streamingAssistant, startRun, stopRun, handleNewChat, handleSendChat, handleStartAssistant, onScore, onReasonChange, onAddScenarioErrorToWorkbench } = props;
+  const { run, runWorkspaceId, runStatusLabel, activeWorkspaceId, requestedRunNotFound, canStart, canRunPersona, hasPersonaSelection, botJsonErrorCount, deckJsonErrorCount, missingBotInput, missingDeckInit, lastInitFill, isUserStart, showStartOverlay, canStartAssistant, canSendChat, chatDraft, setChatDraft, chatError, optimisticUser, streamingUser, streamingAssistant, startRun, stopRun, handleNewChat, handleSendChat, handleStartAssistant, onScore, onReasonChange, onAddErrorToWorkbench } = props;
   const transcriptRef = (0, import_react37.useRef)(null);
   const lastRunMessageCountRef = (0, import_react37.useRef)(0);
   (0, import_react37.useEffect)(() => {
@@ -27333,12 +27984,13 @@ function TestBotChatPanel(props) {
         actions: /* @__PURE__ */ (0, import_jsx_runtime40.jsx)(Button, {
           variant: "secondary",
           size: "small",
-          onClick: () => onAddScenarioErrorToWorkbench?.({
+          onClick: () => onAddErrorToWorkbench?.({
+            source: "scenario_run_error",
             workspaceId: runWorkspaceId ?? activeWorkspaceId ?? void 0,
             runId: run.id,
             error: run.error
           }),
-          disabled: !onAddScenarioErrorToWorkbench,
+          disabled: !onAddErrorToWorkbench,
           "data-testid": "testbot-add-error-to-chat",
           children: "Add to chat"
         }),
@@ -27416,7 +28068,7 @@ function TestBotChatPanel(props) {
                               className: "respond-meta",
                               children: [
                                 /* @__PURE__ */ (0, import_jsx_runtime40.jsx)(Badge, {
-                                  children: "gambit_respond"
+                                  children: "legacy respond envelope"
                                 }),
                                 typeof message.respondStatus === "number" && /* @__PURE__ */ (0, import_jsx_runtime40.jsxs)(Badge, {
                                   variant: "ghost",
@@ -27624,7 +28276,7 @@ var getScenarioTitle2 = (summary, opts) => {
   return fromDeckCatalogById ?? fromDeckCatalogByPath ?? fromDeckLabel ?? fromDeckId ?? fromPath ?? summary.runId;
 };
 function TestBotPage(props) {
-  const { onReplaceTestBotSession, onResetTestBotSession, activeWorkspaceId, requestedRunId, resetToken, setNavActions, onFeedbackPersisted, onAddScenarioErrorToWorkbench } = props;
+  const { onReplaceTestBotSession, onResetTestBotSession, activeWorkspaceId, requestedRunId, resetToken, setNavActions, onFeedbackPersisted, onAddErrorToWorkbench } = props;
   const deckStorageKey = "gambit:test:selected-deck";
   const [testDecks, setTestDecks] = (0, import_react38.useState)([]);
   const [selectedDeckId, setSelectedDeckId] = (0, import_react38.useState)(null);
@@ -28746,7 +29398,7 @@ function TestBotPage(props) {
           handleStartAssistant,
           onScore: handleTestBotScore,
           onReasonChange: handleTestBotReason,
-          onAddScenarioErrorToWorkbench
+          onAddErrorToWorkbench
         })
       ]
     })
@@ -29147,103 +29799,1421 @@ function BuildPage(props) {
   });
 }
 
-// simulator-ui/src/main.tsx
-var globalStyleEl = document.createElement("style");
-globalStyleEl.textContent = globalStyles;
-document.head.appendChild(globalStyleEl);
-function useSimulator() {
-  const [connectionStatus, setConnectionStatus] = (0, import_react40.useState)("connecting");
-  const [savedState, setSavedState] = (0, import_react40.useState)(null);
-  const [traceEvents, setTraceEvents] = (0, import_react40.useState)([]);
-  const [errors, setErrors] = (0, import_react40.useState)([]);
-  const [streamText, setStreamText] = (0, import_react40.useState)("");
-  const [isRunning, setIsRunning] = (0, import_react40.useState)(false);
-  const [connectSeq, setConnectSeq] = (0, import_react40.useState)(0);
-  (0, import_react40.useEffect)(() => {
-    const streamId = SIMULATOR_STREAM_ID;
-    const streamUrl = buildDurableStreamUrl(streamId, getDurableStreamOffset(streamId));
-    const source = new EventSource(streamUrl);
-    setConnectionStatus("connecting");
-    source.onopen = () => {
-      setConnectionStatus("connected");
-      setErrors([]);
+// simulator-ui/src/VerifyPage.tsx
+var import_jsx_runtime43 = __toESM(require_jsx_runtime());
+var import_react40 = __toESM(require__());
+
+// simulator-ui/src/verify_metrics.ts
+var VERIFY_CONSISTENCY_THRESHOLDS = {
+  minSampleSize: 6,
+  instabilityScoreDelta: 1.5,
+  pass: {
+    agreementMin: 0.9,
+    maxSpread: 1,
+    maxInstabilityCount: 0
+  },
+  warn: {
+    agreementMin: 0.75,
+    maxSpread: 2,
+    maxInstabilityCount: 2
+  }
+};
+var pickPayload = (result) => {
+  if (!result || typeof result !== "object") return {};
+  const record = result;
+  if (record.payload && typeof record.payload === "object" && !Array.isArray(record.payload)) {
+    return record.payload;
+  }
+  return record;
+};
+var extractScoreReasonPass = (result) => {
+  const payload = pickPayload(result);
+  const score = typeof payload.score === "number" && Number.isFinite(payload.score) ? payload.score : void 0;
+  const reason = typeof payload.reason === "string" ? payload.reason : void 0;
+  if (typeof payload.pass === "boolean") {
+    return {
+      score,
+      reason,
+      pass: payload.pass
     };
-    source.onerror = () => {
-      setConnectionStatus("error");
-      setErrors((prev) => prev.includes("Stream connection error") ? prev : [
-        ...prev,
-        "Stream connection error"
-      ]);
-      setIsRunning(false);
-      setStreamText("");
+  }
+  if (typeof payload.passed === "boolean") {
+    return {
+      score,
+      reason,
+      pass: payload.passed
     };
-    const handleMessage = (event) => {
-      let msg = null;
-      try {
-        msg = JSON.parse(event.data);
-      } catch (err) {
-        console.error("[sim] failed to parse stream event payload", err);
-        return;
-      }
-      const parsedOffset = Number(event.lastEventId);
-      if (Number.isFinite(parsedOffset)) {
-        setDurableStreamOffset(streamId, parsedOffset + 1);
-      }
-      if (!msg || typeof msg !== "object") return;
-      if (msg.type === "state") {
-        setSavedState(msg.state);
-        if (Array.isArray(msg.state.traces)) {
-          setTraceEvents(msg.state.traces);
-        }
-      } else if (msg.type === "stream") {
-        if (typeof msg.chunk === "string" && msg.chunk.length > 0) {
-          setStreamText((prev) => prev + msg.chunk);
-        }
-      } else if (msg.type === "result") {
-        setIsRunning(false);
-        setStreamText("");
-      } else if (msg.type === "trace" && msg.event) {
-        setTraceEvents((prev) => [
-          ...prev,
-          msg.event
-        ].slice(-200));
-      } else if (msg.type === "error") {
-        const message = msg.message ?? "Unknown error";
-        setErrors((prev) => [
-          ...prev,
-          message
-        ]);
-        if (msg.runId || message !== "Run already in progress") {
-          setIsRunning(false);
-        }
-        setStreamText("");
-      }
+  }
+  if (typeof payload.verdict === "string") {
+    const verdict = payload.verdict.trim().toLowerCase();
+    if (verdict === "pass") return {
+      score,
+      reason,
+      pass: true
     };
-    const simulatorEventTypes = [
-      "ready",
-      "pong",
-      "stream",
-      "result",
-      "trace",
-      "state",
-      "error"
-    ];
-    for (const type of simulatorEventTypes) {
-      source.addEventListener(type, handleMessage);
+    if (verdict === "fail") return {
+      score,
+      reason,
+      pass: false
+    };
+  }
+  if (typeof score === "number") {
+    return {
+      score,
+      reason,
+      pass: score >= 0
+    };
+  }
+  return {
+    score,
+    reason
+  };
+};
+var flattenRunExamples = (run) => {
+  if (!run.result || typeof run.result !== "object") return [];
+  const record = run.result;
+  if (record.mode === "turns" && Array.isArray(record.turns)) {
+    const buckets = [];
+    record.turns.forEach((turn, fallbackIndex) => {
+      if (!turn || typeof turn !== "object") return;
+      const turnRecord = turn;
+      const index = typeof turnRecord.index === "number" && Number.isFinite(turnRecord.index) ? Math.max(0, Math.round(turnRecord.index)) : fallbackIndex;
+      const messageRefId = typeof turnRecord.messageRefId === "string" && turnRecord.messageRefId.trim().length > 0 ? turnRecord.messageRefId : void 0;
+      const key = messageRefId ? `ref:${messageRefId}` : `turn:${index}`;
+      const label = `Assistant turn ${fallbackIndex + 1}`;
+      const parsed2 = extractScoreReasonPass(turnRecord.result);
+      buckets.push({
+        key,
+        label,
+        points: [
+          {
+            runId: run.id,
+            runAt: run.runAt,
+            score: parsed2.score,
+            pass: parsed2.pass,
+            reason: parsed2.reason,
+            turnIndex: index,
+            messageRefId
+          }
+        ]
+      });
+    });
+    return buckets;
+  }
+  const parsed = extractScoreReasonPass(run.result);
+  return [
+    {
+      key: "conversation",
+      label: "Conversation score",
+      points: [
+        {
+          runId: run.id,
+          runAt: run.runAt,
+          score: parsed.score,
+          pass: parsed.pass,
+          reason: parsed.reason
+        }
+      ]
     }
-    return () => {
-      for (const type of simulatorEventTypes) {
-        source.removeEventListener(type, handleMessage);
-      }
-      source.close();
-      setConnectionStatus("closed");
-      setIsRunning(false);
-      setStreamText("");
+  ];
+};
+var median = (values) => {
+  if (!values.length) return null;
+  const sorted = [
+    ...values
+  ].sort((a, b) => a - b);
+  const mid = Math.floor(sorted.length / 2);
+  if (sorted.length % 2 === 1) return sorted[mid];
+  return (sorted[mid - 1] + sorted[mid]) / 2;
+};
+var round2 = (value) => Math.round(value * 100) / 100;
+var resolveVerdict = (input) => {
+  const t = VERIFY_CONSISTENCY_THRESHOLDS;
+  if (input.sampleSize < t.minSampleSize) {
+    return {
+      verdict: "WARN",
+      reason: `Need at least ${t.minSampleSize} samples before issuing a firm verdict.`
     };
-  }, [
+  }
+  if (input.agreementRate === null) {
+    return {
+      verdict: "WARN",
+      reason: "No comparable pass/fail evidence was found in the sampled runs."
+    };
+  }
+  const spreadMax = input.spreadMax ?? 0;
+  if (input.agreementRate >= t.pass.agreementMin && spreadMax <= t.pass.maxSpread && input.instabilityCount <= t.pass.maxInstabilityCount) {
+    return {
+      verdict: "PASS",
+      reason: "Agreement, spread, and instability all meet PASS thresholds."
+    };
+  }
+  if (input.agreementRate >= t.warn.agreementMin && spreadMax <= t.warn.maxSpread && input.instabilityCount <= t.warn.maxInstabilityCount) {
+    return {
+      verdict: "WARN",
+      reason: "Some variation was detected, but results remain within WARN thresholds."
+    };
+  }
+  return {
+    verdict: "FAIL",
+    reason: "Agreement/spread instability exceeds WARN thresholds."
+  };
+};
+function buildVerifyConsistencyReport(runs) {
+  const completedRuns = runs.filter((run) => run.status === "completed");
+  const sampleSize = completedRuns.length;
+  const bucketsByKey = /* @__PURE__ */ new Map();
+  completedRuns.forEach((run) => {
+    flattenRunExamples(run).forEach((entry) => {
+      const existing = bucketsByKey.get(entry.key);
+      if (!existing) {
+        bucketsByKey.set(entry.key, {
+          key: entry.key,
+          label: entry.label,
+          points: [
+            ...entry.points
+          ]
+        });
+        return;
+      }
+      existing.points.push(...entry.points);
+    });
+  });
+  const outliers = [];
+  let agreementVotes = 0;
+  let agreementTotal = 0;
+  const scoreDeltas = [];
+  bucketsByKey.forEach((bucket) => {
+    const scores = bucket.points.map((point) => point.score).filter((score) => typeof score === "number" && Number.isFinite(score));
+    const minScore = scores.length ? Math.min(...scores) : null;
+    const maxScore = scores.length ? Math.max(...scores) : null;
+    const scoreDelta = minScore !== null && maxScore !== null ? round2(maxScore - minScore) : null;
+    const passVotes = bucket.points.map((point) => point.pass).filter((pass) => typeof pass === "boolean");
+    const passCount = passVotes.filter((value) => value).length;
+    const failCount = passVotes.length - passCount;
+    const agreementRate2 = passVotes.length > 0 ? round2(Math.max(passCount, failCount) / passVotes.length) : null;
+    if (passVotes.length > 0) {
+      agreementVotes += Math.max(passCount, failCount);
+      agreementTotal += passVotes.length;
+    }
+    if (scoreDelta !== null) {
+      scoreDeltas.push(scoreDelta);
+    }
+    const passFlip = passCount > 0 && failCount > 0;
+    const instability = passFlip || scoreDelta !== null && scoreDelta > VERIFY_CONSISTENCY_THRESHOLDS.instabilityScoreDelta;
+    const minPoint = minScore === null ? void 0 : bucket.points.find((point) => point.score === minScore);
+    const maxPoint = maxScore === null ? void 0 : bucket.points.find((point) => point.score === maxScore);
+    outliers.push({
+      key: bucket.key,
+      label: bucket.label,
+      sampleSize: bucket.points.length,
+      agreementRate: agreementRate2,
+      scoreDelta,
+      passFlip,
+      instability,
+      minScore,
+      maxScore,
+      minRunId: minPoint?.runId,
+      maxRunId: maxPoint?.runId,
+      turnIndex: maxPoint?.turnIndex ?? minPoint?.turnIndex,
+      messageRefId: maxPoint?.messageRefId ?? minPoint?.messageRefId
+    });
+  });
+  outliers.sort((a, b) => {
+    if (a.instability !== b.instability) return a.instability ? -1 : 1;
+    if (a.passFlip !== b.passFlip) return a.passFlip ? -1 : 1;
+    const aDelta = a.scoreDelta ?? -1;
+    const bDelta = b.scoreDelta ?? -1;
+    if (aDelta !== bDelta) return bDelta - aDelta;
+    if (a.sampleSize !== b.sampleSize) return b.sampleSize - a.sampleSize;
+    return a.label.localeCompare(b.label);
+  });
+  const agreementRate = agreementTotal > 0 ? round2(agreementVotes / agreementTotal) : null;
+  const scoreSpreadMin = scoreDeltas.length ? Math.min(...scoreDeltas) : null;
+  const scoreSpreadMax = scoreDeltas.length ? Math.max(...scoreDeltas) : null;
+  const scoreSpreadMedian = median(scoreDeltas);
+  const instabilityCount = outliers.filter((entry) => entry.instability).length;
+  const verdict = resolveVerdict({
+    sampleSize,
+    agreementRate,
+    spreadMax: scoreSpreadMax,
+    instabilityCount
+  });
+  return {
+    sampleSize,
+    comparableExampleCount: outliers.length,
+    agreementRate,
+    scoreSpreadMin: scoreSpreadMin === null ? null : round2(scoreSpreadMin),
+    scoreSpreadMedian: scoreSpreadMedian === null ? null : round2(scoreSpreadMedian),
+    scoreSpreadMax: scoreSpreadMax === null ? null : round2(scoreSpreadMax),
+    instabilityCount,
+    verdict: verdict.verdict,
+    verdictReason: verdict.reason,
+    outliers
+  };
+}
+
+// simulator-ui/src/VerifyPage.tsx
+var MAX_BATCH_SIZE = 24;
+var MAX_BATCH_CONCURRENCY = 6;
+var DEFAULT_BATCH_SIZE = 8;
+var DEFAULT_BATCH_CONCURRENCY = 3;
+var NO_SCENARIO_RUN_VALUE = "__workspace_context__";
+var parseScenarioRunSummary2 = (value) => {
+  if (!value || typeof value !== "object") return null;
+  const summary = value;
+  const scenarioRunId = typeof summary.scenarioRunId === "string" ? summary.scenarioRunId : null;
+  if (!scenarioRunId) return null;
+  return {
+    scenarioRunId,
+    lastEventSeq: typeof summary.lastEventSeq === "number" && Number.isFinite(summary.lastEventSeq) ? summary.lastEventSeq : void 0,
+    updatedAt: typeof summary.updatedAt === "string" ? summary.updatedAt : void 0,
+    selectedScenarioDeckId: typeof summary.selectedScenarioDeckId === "string" ? summary.selectedScenarioDeckId : void 0,
+    selectedScenarioDeckLabel: typeof summary.selectedScenarioDeckLabel === "string" ? summary.selectedScenarioDeckLabel : void 0,
+    scenarioConfigPath: typeof summary.scenarioConfigPath === "string" ? summary.scenarioConfigPath : void 0
+  };
+};
+var getScenarioTitle3 = (summary) => {
+  const fromDeckLabel = typeof summary.selectedScenarioDeckLabel === "string" && summary.selectedScenarioDeckLabel.trim().length > 0 ? summary.selectedScenarioDeckLabel : null;
+  const fromDeckId = typeof summary.selectedScenarioDeckId === "string" && summary.selectedScenarioDeckId.trim().length > 0 ? scenarioNameFromValue(summary.selectedScenarioDeckId) ?? summary.selectedScenarioDeckId : null;
+  const fromPath = scenarioNameFromValue(summary.scenarioConfigPath ?? null) ?? botFilename(summary.scenarioConfigPath ?? null);
+  return fromDeckLabel ?? fromDeckId ?? fromPath ?? summary.scenarioRunId;
+};
+var scenarioRunIdFromCalibrationRun2 = (run) => {
+  if (!run.input || typeof run.input !== "object") return null;
+  const input = run.input;
+  const session = input.session;
+  if (!session || typeof session !== "object") return null;
+  const meta = session.meta;
+  if (!meta || typeof meta !== "object") return null;
+  const scenarioRunId = meta.scenarioRunId;
+  return typeof scenarioRunId === "string" && scenarioRunId.trim().length > 0 ? scenarioRunId : null;
+};
+var parseRunAt = (run) => {
+  const parsed = Date.parse(run.runAt ?? "");
+  return Number.isFinite(parsed) ? parsed : -1;
+};
+var clampInt = (value, min, max) => {
+  const rounded = Number.isFinite(value) ? Math.round(value) : min;
+  return Math.max(min, Math.min(max, rounded));
+};
+var formatSignedScore = (value) => {
+  if (typeof value !== "number" || !Number.isFinite(value)) return "\u2014";
+  return `${value > 0 ? "+" : ""}${value}`;
+};
+var scoreBadgeVariant = (value) => {
+  if (typeof value !== "number" || !Number.isFinite(value)) return "ghost";
+  if (value < 0) return "error";
+  if (value > 0) return "completed";
+  return "idle";
+};
+function VerifyPage({ setNavActions, onAppPathChange, activeWorkspaceId, composerChips, onComposerChipsChange }) {
+  const { loading, error, graders, sessions, sessionDetail, loadData, loadSessionDetail } = useWorkspaceGrade();
+  const workspaceRouting = useWorkspaceRouting();
+  const [selectedSessionId, setSelectedSessionId] = (0, import_react40.useState)(activeWorkspaceId ?? null);
+  const [selectedGraderId, setSelectedGraderId] = (0, import_react40.useState)(null);
+  const [selectedScenarioRunId, setSelectedScenarioRunId] = (0, import_react40.useState)(null);
+  const [batchSize, setBatchSize] = (0, import_react40.useState)(DEFAULT_BATCH_SIZE);
+  const [batchConcurrency, setBatchConcurrency] = (0, import_react40.useState)(DEFAULT_BATCH_CONCURRENCY);
+  const [batchState, setBatchState] = (0, import_react40.useState)({
+    batchId: 0,
+    status: "idle",
+    requested: 0,
+    concurrency: 0,
+    completed: 0,
+    failed: 0,
+    active: 0,
+    initialRunIds: [],
+    requests: []
+  });
+  const [reportScope, setReportScope] = (0, import_react40.useState)("all_matching");
+  const [inconsistentOnly, setInconsistentOnly] = (0, import_react40.useState)(false);
+  const [exampleSort, setExampleSort] = (0, import_react40.useState)("default");
+  const batchSeqRef = (0, import_react40.useRef)(0);
+  const updateVerifyPath = (0, import_react40.useCallback)((sessionId) => {
+    const targetPath = buildVerifyPath(sessionId);
+    if (window.location.pathname === targetPath) return;
+    window.history.replaceState({}, "", targetPath);
+    onAppPathChange?.(targetPath);
+  }, [
+    onAppPathChange
+  ]);
+  const navigateToAppPath = (0, import_react40.useCallback)((nextPath) => {
+    if (window.location.pathname === nextPath) return;
+    window.history.pushState({}, "", nextPath);
+    onAppPathChange?.(nextPath);
+  }, [
+    onAppPathChange
+  ]);
+  const handleInternalLinkClick = (0, import_react40.useCallback)((event, href) => {
+    if (event.metaKey || event.ctrlKey || event.shiftKey || event.altKey || event.button !== 0) {
+      return;
+    }
+    event.preventDefault();
+    navigateToAppPath(href);
+  }, [
+    navigateToAppPath
+  ]);
+  const loadVerifyData = (0, import_react40.useCallback)(async () => {
+    await loadData({
+      workspaceId: activeWorkspaceId ?? null
+    });
+  }, [
+    activeWorkspaceId,
+    loadData
+  ]);
+  (0, import_react40.useEffect)(() => {
+    loadVerifyData();
+  }, [
+    loadVerifyData
+  ]);
+  (0, import_react40.useEffect)(() => {
+    setSelectedSessionId((prev) => {
+      if (activeWorkspaceId) return activeWorkspaceId;
+      if (prev && sessions.some((session) => session.id === prev)) return prev;
+      return sessions[0]?.id ?? null;
+    });
+  }, [
+    activeWorkspaceId,
+    sessions
+  ]);
+  (0, import_react40.useEffect)(() => {
+    setSelectedGraderId((prev) => {
+      if (prev && graders.some((grader) => grader.id === prev)) return prev;
+      return graders[0]?.id ?? null;
+    });
+  }, [
+    graders
+  ]);
+  (0, import_react40.useEffect)(() => {
+    if (!selectedSessionId) return;
+    updateVerifyPath(selectedSessionId);
+  }, [
+    selectedSessionId,
+    updateVerifyPath
+  ]);
+  (0, import_react40.useEffect)(() => {
+    loadSessionDetail(selectedSessionId).catch((err) => {
+      console.error(err);
+    });
+  }, [
+    loadSessionDetail,
+    selectedSessionId
+  ]);
+  const selectedSession = (0, import_react40.useMemo)(() => sessions.find((session) => session.id === selectedSessionId) ?? null, [
+    sessions,
+    selectedSessionId
+  ]);
+  const selectedGrader = (0, import_react40.useMemo)(() => graders.find((grader) => grader.id === selectedGraderId) ?? null, [
+    graders,
+    selectedGraderId
+  ]);
+  const scenarioRunOptions = (0, import_react40.useMemo)(() => {
+    const meta = sessionDetail?.meta && typeof sessionDetail.meta === "object" ? sessionDetail.meta : {};
+    const fromList = Array.isArray(meta.scenarioRunSummaries) ? meta.scenarioRunSummaries.map((entry) => parseScenarioRunSummary2(entry)) : [];
+    const fromCurrent = parseScenarioRunSummary2(meta.scenarioRunSummary);
+    const all = [
+      ...fromList,
+      fromCurrent
+    ].filter((entry) => Boolean(entry));
+    const deduped = /* @__PURE__ */ new Map();
+    all.forEach((entry) => {
+      const existing = deduped.get(entry.scenarioRunId);
+      if (!existing) {
+        deduped.set(entry.scenarioRunId, entry);
+        return;
+      }
+      const existingSeq = existing.lastEventSeq ?? -1;
+      const nextSeq = entry.lastEventSeq ?? -1;
+      if (nextSeq > existingSeq) {
+        deduped.set(entry.scenarioRunId, entry);
+        return;
+      }
+      if (nextSeq === existingSeq) {
+        const existingStamp = existing.updatedAt ?? "";
+        const nextStamp = entry.updatedAt ?? "";
+        if (nextStamp.localeCompare(existingStamp) > 0) {
+          deduped.set(entry.scenarioRunId, entry);
+        }
+      }
+    });
+    return [
+      ...deduped.values()
+    ].sort((a, b) => {
+      const aTime = Date.parse(a.updatedAt ?? "");
+      const bTime = Date.parse(b.updatedAt ?? "");
+      const aValidTime = Number.isFinite(aTime) ? aTime : -1;
+      const bValidTime = Number.isFinite(bTime) ? bTime : -1;
+      if (aValidTime !== bValidTime) return bValidTime - aValidTime;
+      const aSeq = a.lastEventSeq ?? -1;
+      const bSeq = b.lastEventSeq ?? -1;
+      if (aSeq !== bSeq) return bSeq - aSeq;
+      return b.scenarioRunId.localeCompare(a.scenarioRunId);
+    });
+  }, [
+    sessionDetail?.meta
+  ]);
+  (0, import_react40.useEffect)(() => {
+    const hasOption = (runId) => Boolean(runId && scenarioRunOptions.some((entry) => entry.scenarioRunId === runId));
+    const meta = sessionDetail?.meta && typeof sessionDetail.meta === "object" ? sessionDetail.meta : {};
+    const currentScenarioRunId = typeof meta.scenarioRunId === "string" && meta.scenarioRunId.trim().length > 0 ? meta.scenarioRunId : null;
+    const nextRunId = hasOption(workspaceRouting.testRunId) ? workspaceRouting.testRunId : hasOption(selectedScenarioRunId) ? selectedScenarioRunId : hasOption(currentScenarioRunId) ? currentScenarioRunId : scenarioRunOptions[0]?.scenarioRunId ?? null;
+    if (selectedScenarioRunId !== nextRunId) {
+      setSelectedScenarioRunId(nextRunId);
+    }
+    if (workspaceRouting.testRunId !== nextRunId) {
+      workspaceRouting.setTestRunId(nextRunId);
+    }
+  }, [
+    scenarioRunOptions,
+    selectedScenarioRunId,
+    sessionDetail?.meta,
+    workspaceRouting
+  ]);
+  const sessionRuns = (0, import_react40.useMemo)(() => {
+    if (!selectedSession?.gradingRuns) return [];
+    return [
+      ...selectedSession.gradingRuns
+    ].sort((a, b) => parseRunAt(b) - parseRunAt(a));
+  }, [
+    selectedSession?.gradingRuns
+  ]);
+  const filteredRuns = (0, import_react40.useMemo)(() => {
+    const latestScenarioRunIdFromRuns = sessionRuns.map((run) => scenarioRunIdFromCalibrationRun2(run)).find((runId) => Boolean(runId));
+    const hasOption = (runId) => Boolean(runId && scenarioRunOptions.some((entry) => entry.scenarioRunId === runId));
+    const meta = sessionDetail?.meta && typeof sessionDetail.meta === "object" ? sessionDetail.meta : {};
+    const currentScenarioRunId = typeof meta.scenarioRunId === "string" && meta.scenarioRunId.trim().length > 0 ? meta.scenarioRunId : null;
+    const activeScenarioRunFilterId = hasOption(workspaceRouting.testRunId) ? workspaceRouting.testRunId : hasOption(selectedScenarioRunId) ? selectedScenarioRunId : hasOption(currentScenarioRunId) ? currentScenarioRunId : scenarioRunOptions[0]?.scenarioRunId ?? latestScenarioRunIdFromRuns ?? null;
+    return sessionRuns.filter((run) => {
+      if (selectedGraderId && run.graderId !== selectedGraderId) return false;
+      if (!activeScenarioRunFilterId) return true;
+      return scenarioRunIdFromCalibrationRun2(run) === activeScenarioRunFilterId;
+    });
+  }, [
+    scenarioRunOptions,
+    selectedGraderId,
+    selectedScenarioRunId,
+    sessionDetail?.meta,
+    sessionRuns,
+    workspaceRouting.testRunId
+  ]);
+  const runConsistencySample = (0, import_react40.useCallback)(async (payload) => {
+    const res = await fetch("/api/calibrate/run", {
+      method: "POST",
+      headers: {
+        "content-type": "application/json"
+      },
+      body: JSON.stringify(payload)
+    });
+    const body = await res.json().catch(() => ({}));
+    if (!res.ok) {
+      throw new Error(body.error ?? res.statusText);
+    }
+    if (!body.run) {
+      throw new Error(body.error ?? "Calibration run response missing run payload");
+    }
+    return {
+      run: body.run,
+      error: body.error
+    };
+  }, []);
+  const updateBatchRequest = (0, import_react40.useCallback)((batchId, index, patch) => {
+    setBatchState((prev) => {
+      if (prev.batchId !== batchId) return prev;
+      if (index < 0 || index >= prev.requests.length) return prev;
+      const nextRequests = prev.requests.map((request, requestIndex) => requestIndex === index ? {
+        ...request,
+        ...patch
+      } : request);
+      const completed = nextRequests.filter((request) => request.status === "completed").length;
+      const failed = nextRequests.filter((request) => request.status === "error").length;
+      const active = nextRequests.filter((request) => request.status === "running").length;
+      const terminal = completed + failed === prev.requested && active === 0;
+      const nextStatus = terminal ? failed > 0 ? "error" : "completed" : "running";
+      return {
+        ...prev,
+        requests: nextRequests,
+        completed,
+        failed,
+        active,
+        status: nextStatus,
+        finishedAt: terminal ? (/* @__PURE__ */ new Date()).toISOString() : prev.finishedAt
+      };
+    });
+  }, []);
+  const runConsistencyBatch = (0, import_react40.useCallback)(async () => {
+    if (!selectedSessionId || !selectedGraderId) return;
+    const nextBatchSize = clampInt(batchSize, 1, MAX_BATCH_SIZE);
+    const nextConcurrency = clampInt(batchConcurrency, 1, Math.min(MAX_BATCH_CONCURRENCY, nextBatchSize));
+    const batchId = batchSeqRef.current + 1;
+    batchSeqRef.current = batchId;
+    const now = (/* @__PURE__ */ new Date()).toISOString();
+    const initialRunIds = filteredRuns.map((run) => run.id);
+    const initialRequests = Array.from({
+      length: nextBatchSize
+    }, (_, index) => ({
+      requestId: `${batchId}:${index + 1}`,
+      status: "queued"
+    }));
+    setBatchState({
+      batchId,
+      status: "running",
+      startedAt: now,
+      finishedAt: void 0,
+      requested: nextBatchSize,
+      concurrency: nextConcurrency,
+      completed: 0,
+      failed: 0,
+      active: 0,
+      initialRunIds,
+      requests: initialRequests
+    });
+    let cursor = 0;
+    const workers = Array.from({
+      length: nextConcurrency
+    }, () => (async () => {
+      while (true) {
+        const nextIndex = cursor;
+        cursor += 1;
+        if (nextIndex >= nextBatchSize) return;
+        if (batchSeqRef.current !== batchId) return;
+        updateBatchRequest(batchId, nextIndex, {
+          status: "running"
+        });
+        try {
+          const response = await runConsistencySample({
+            workspaceId: selectedSessionId,
+            graderId: selectedGraderId,
+            scenarioRunId: selectedScenarioRunId ?? void 0
+          });
+          if (batchSeqRef.current !== batchId) return;
+          if (response.run.status !== "completed") {
+            updateBatchRequest(batchId, nextIndex, {
+              status: "error",
+              runId: response.run.id,
+              error: response.run.error ?? `Calibration run ended with status ${response.run.status}`
+            });
+            continue;
+          }
+          updateBatchRequest(batchId, nextIndex, {
+            status: "completed",
+            runId: response.run.id
+          });
+        } catch (err) {
+          if (batchSeqRef.current !== batchId) return;
+          updateBatchRequest(batchId, nextIndex, {
+            status: "error",
+            error: err instanceof Error ? err.message : String(err)
+          });
+        }
+      }
+    })());
+    await Promise.all(workers);
+    if (batchSeqRef.current !== batchId) return;
+    await loadData({
+      workspaceId: selectedSessionId
+    }).catch(() => {
+    });
+  }, [
+    batchConcurrency,
+    batchSize,
+    filteredRuns,
+    loadData,
+    runConsistencySample,
+    selectedGraderId,
+    selectedScenarioRunId,
+    selectedSessionId,
+    updateBatchRequest
+  ]);
+  const batchInitialRunIdSet = (0, import_react40.useMemo)(() => new Set(batchState.initialRunIds), [
+    batchState.initialRunIds
+  ]);
+  const activeBatchRuns = (0, import_react40.useMemo)(() => {
+    if (!batchState.startedAt) return [];
+    const startedAt = Date.parse(batchState.startedAt);
+    const hasStartedAt = Number.isFinite(startedAt);
+    return filteredRuns.filter((run) => {
+      if (batchInitialRunIdSet.has(run.id)) return false;
+      if (!hasStartedAt) return true;
+      const runAt = parseRunAt(run);
+      return runAt < 0 || runAt >= startedAt - 2e3;
+    });
+  }, [
+    batchInitialRunIdSet,
+    batchState.startedAt,
+    filteredRuns
+  ]);
+  const completedBatchRuns = (0, import_react40.useMemo)(() => activeBatchRuns.filter((run) => run.status === "completed"), [
+    activeBatchRuns
+  ]);
+  const historicalCompletedRuns = (0, import_react40.useMemo)(() => filteredRuns.filter((run) => run.status === "completed"), [
+    filteredRuns
+  ]);
+  const reportRuns = (0, import_react40.useMemo)(() => reportScope === "current_batch" ? completedBatchRuns : historicalCompletedRuns, [
+    completedBatchRuns,
+    historicalCompletedRuns,
+    reportScope
+  ]);
+  const consistencyReport = (0, import_react40.useMemo)(() => buildVerifyConsistencyReport(reportRuns), [
+    reportRuns
+  ]);
+  const queuedCount = (0, import_react40.useMemo)(() => batchState.requests.filter((request) => request.status === "queued").length, [
+    batchState.requests
+  ]);
+  const canRun = Boolean(selectedSessionId && selectedGraderId && batchState.status !== "running");
+  const displayedOutliers = (0, import_react40.useMemo)(() => {
+    const filtered = inconsistentOnly ? consistencyReport.outliers.filter((outlier) => outlier.instability) : consistencyReport.outliers;
+    if (exampleSort === "default") return filtered;
+    const next = [
+      ...filtered
+    ];
+    if (exampleSort === "delta_desc") {
+      next.sort((a, b) => {
+        const aDelta = a.scoreDelta ?? -1;
+        const bDelta = b.scoreDelta ?? -1;
+        if (aDelta !== bDelta) return bDelta - aDelta;
+        return a.label.localeCompare(b.label);
+      });
+      return next;
+    }
+    if (exampleSort === "agreement_asc") {
+      next.sort((a, b) => {
+        const aAgreement = a.agreementRate ?? Number.POSITIVE_INFINITY;
+        const bAgreement = b.agreementRate ?? Number.POSITIVE_INFINITY;
+        if (aAgreement !== bAgreement) return aAgreement - bAgreement;
+        return a.label.localeCompare(b.label);
+      });
+      return next;
+    }
+    if (exampleSort === "samples_desc") {
+      next.sort((a, b) => {
+        if (a.sampleSize !== b.sampleSize) return b.sampleSize - a.sampleSize;
+        return a.label.localeCompare(b.label);
+      });
+      return next;
+    }
+    next.sort((a, b) => a.label.localeCompare(b.label));
+    return next;
+  }, [
+    consistencyReport.outliers,
+    exampleSort,
+    inconsistentOnly
+  ]);
+  const reportScopeLabel = reportScope === "current_batch" ? "current batch" : "all matching runs";
+  const resolvedComposerChips = (0, import_react40.useMemo)(() => composerChips ?? [], [
+    composerChips
+  ]);
+  const composerChipIds = (0, import_react40.useMemo)(() => new Set(resolvedComposerChips.map((chip) => chip.chipId)), [
+    resolvedComposerChips
+  ]);
+  const mergeComposerChip = (0, import_react40.useCallback)((base, chip) => {
+    const next = [
+      ...base
+    ];
+    const existingIndex = next.findIndex((entry) => entry.chipId === chip.chipId);
+    if (existingIndex >= 0) {
+      next[existingIndex] = {
+        ...next[existingIndex],
+        ...chip,
+        enabled: true
+      };
+      return next;
+    }
+    next.push(chip);
+    return next;
+  }, []);
+  const addComposerChip = (0, import_react40.useCallback)((chip) => {
+    if (!onComposerChipsChange) return;
+    onComposerChipsChange(mergeComposerChip(resolvedComposerChips, chip));
+  }, [
+    mergeComposerChip,
+    onComposerChipsChange,
+    resolvedComposerChips
+  ]);
+  const removeComposerChip = (0, import_react40.useCallback)((chipId) => {
+    if (!onComposerChipsChange) return;
+    onComposerChipsChange(resolvedComposerChips.filter((chip) => chip.chipId !== chipId));
+  }, [
+    onComposerChipsChange,
+    resolvedComposerChips
+  ]);
+  const buildOutlierChip = (0, import_react40.useCallback)((outlier) => {
+    const chipId = `verify:${selectedSessionId ?? ""}:${outlier.key}`;
+    const runId = outlier.maxRunId ?? outlier.minRunId;
+    const score = outlier.maxScore ?? outlier.minScore ?? void 0;
+    const agreementText = outlier.agreementRate === null ? "agreement unavailable" : `agreement ${Math.round(outlier.agreementRate * 100)}%`;
+    const deltaText = outlier.scoreDelta === null ? "delta unavailable" : `delta ${outlier.scoreDelta}`;
+    return {
+      chipId,
+      source: "verify_outlier",
+      workspaceId: selectedSessionId ?? void 0,
+      runId,
+      capturedAt: (/* @__PURE__ */ new Date()).toISOString(),
+      outlierKey: outlier.key,
+      instability: outlier.instability,
+      score,
+      message: `Verify outlier ${outlier.label}: ${agreementText}, ${deltaText}, samples ${outlier.sampleSize}`,
+      enabled: true
+    };
+  }, [
+    selectedSessionId
+  ]);
+  (0, import_react40.useEffect)(() => {
+    if (!setNavActions) return;
+    setNavActions(null);
+    return () => setNavActions(null);
+  }, [
+    setNavActions
+  ]);
+  (0, import_react40.useEffect)(() => {
+    return () => {
+      batchSeqRef.current += 1;
+    };
+  }, []);
+  return /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(PageShell, {
+    className: "verify-shell",
+    children: /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)(PageGrid, {
+      as: "main",
+      className: "verify-layout",
+      children: [
+        /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)(Panel_default, {
+          className: "verify-controls",
+          children: [
+            /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+              className: "verify-controls-header",
+              children: [
+                /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("strong", {
+                  children: "Verify consistency"
+                }),
+                /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("span", {
+                  className: "secondary-note",
+                  children: "Run repeated grading checks against one grader and scenario."
+                })
+              ]
+            }),
+            scenarioRunOptions.length > 0 && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Listbox, {
+              label: "Scenario run",
+              value: selectedScenarioRunId ?? NO_SCENARIO_RUN_VALUE,
+              onChange: (runId) => {
+                if (runId === NO_SCENARIO_RUN_VALUE) {
+                  setSelectedScenarioRunId(null);
+                  workspaceRouting.setTestRunId(null);
+                  return;
+                }
+                setSelectedScenarioRunId(runId);
+                workspaceRouting.setTestRunId(runId);
+              },
+              options: [
+                {
+                  value: NO_SCENARIO_RUN_VALUE,
+                  label: "Current workspace context",
+                  meta: "Run without a prior scenario run binding"
+                },
+                ...scenarioRunOptions.map((entry) => ({
+                  value: entry.scenarioRunId,
+                  label: getScenarioTitle3(entry),
+                  meta: [
+                    entry.updatedAt ? formatTimestampShort(entry.updatedAt) : null,
+                    entry.scenarioRunId
+                  ].filter(Boolean).join(" \xB7 ")
+                }))
+              ],
+              placeholder: "Select scenario run"
+            }),
+            /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Listbox, {
+              label: "Grader",
+              value: selectedGraderId ?? "",
+              onChange: (value) => setSelectedGraderId(value.length ? value : null),
+              options: graders.map((grader) => ({
+                value: grader.id,
+                label: grader.label,
+                meta: botFilename(grader.path ?? null) ?? void 0
+              })),
+              placeholder: "Select grader",
+              disabled: graders.length === 0
+            }),
+            /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+              className: "verify-number-grid",
+              children: [
+                /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("label", {
+                  className: "verify-number-field",
+                  children: [
+                    "Batch size",
+                    /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("input", {
+                      type: "number",
+                      min: 1,
+                      max: MAX_BATCH_SIZE,
+                      value: batchSize,
+                      onChange: (event) => setBatchSize(clampInt(Number(event.target.value), 1, MAX_BATCH_SIZE))
+                    })
+                  ]
+                }),
+                /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("label", {
+                  className: "verify-number-field",
+                  children: [
+                    "Concurrency",
+                    /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("input", {
+                      type: "number",
+                      min: 1,
+                      max: MAX_BATCH_CONCURRENCY,
+                      value: batchConcurrency,
+                      onChange: (event) => setBatchConcurrency(clampInt(Number(event.target.value), 1, MAX_BATCH_CONCURRENCY))
+                    })
+                  ]
+                })
+              ]
+            }),
+            /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Button, {
+              variant: "primary",
+              onClick: runConsistencyBatch,
+              disabled: !canRun,
+              children: batchState.status === "running" ? "Running consistency batch\u2026" : "Run consistency batch"
+            }),
+            /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Callout, {
+              variant: "emphasis",
+              title: "Build assistant stays available",
+              children: "Use the chat drawer toggle in the top-right corner to investigate and iterate while this page remains open."
+            }),
+            sessions.length === 0 && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Callout, {
+              children: "No workspaces found yet. Run a Test scenario first so Verify has evidence to grade."
+            }),
+            graders.length === 0 && /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)(Callout, {
+              children: [
+                "No graders are available. Add ",
+                /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("code", {
+                  children: "[[graders]]"
+                }),
+                " ",
+                "entries to the active root deck."
+              ]
+            }),
+            selectedGrader?.description && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Callout, {
+              children: selectedGrader.description
+            })
+          ]
+        }),
+        /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)(Panel_default, {
+          className: "verify-results",
+          children: [
+            error && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+              className: "error",
+              children: error
+            }),
+            loading && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+              className: "editor-status",
+              children: "Loading verify data\u2026"
+            }),
+            !loading && /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)(import_jsx_runtime43.Fragment, {
+              children: [
+                /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                  className: "verify-status-row",
+                  children: [
+                    /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                      className: "verify-status-main flex-row items-center gap-8",
+                      children: [
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("strong", {
+                          children: "Batch status"
+                        }),
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Badge, {
+                          status: batchState.status,
+                          children: batchState.status
+                        })
+                      ]
+                    }),
+                    consistencyReport.sampleSize > 0 && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("span", {
+                      className: classNames("verify-verdict-badge", `verify-verdict-badge--${consistencyReport.verdict.toLowerCase()}`),
+                      children: consistencyReport.verdict
+                    })
+                  ]
+                }),
+                batchState.status === "idle" && consistencyReport.sampleSize === 0 && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Callout, {
+                  children: "Run a consistency batch to compute agreement, spread, and instability for the selected grader."
+                }),
+                /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                  className: "verify-metric-grid",
+                  children: [
+                    /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                      className: "verify-metric-card",
+                      children: /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                        className: "verify-sample-size-row",
+                        children: [
+                          /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                            className: "verify-sample-size-copy",
+                            children: [
+                              /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                                className: "verify-metric-label",
+                                children: "Sample size"
+                              }),
+                              /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                                className: "verify-metric-value",
+                                children: consistencyReport.sampleSize
+                              })
+                            ]
+                          }),
+                          /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                            className: "verify-sample-scope-select",
+                            children: /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Listbox, {
+                              value: reportScope,
+                              onChange: (value) => setReportScope(value),
+                              size: "small",
+                              popoverMatchTriggerWidth: false,
+                              popoverMinWidth: 320,
+                              popoverAlign: "right",
+                              options: [
+                                {
+                                  value: "current_batch",
+                                  label: `Current batch (${completedBatchRuns.length})`,
+                                  triggerLabel: "Current batch",
+                                  triggerMeta: null,
+                                  meta: "Only runs from the latest batch launch"
+                                },
+                                {
+                                  value: "all_matching",
+                                  label: `All matching runs (${historicalCompletedRuns.length})`,
+                                  triggerLabel: "All matching runs",
+                                  triggerMeta: null,
+                                  meta: "All runs matching selected scenario + grader"
+                                }
+                              ]
+                            })
+                          })
+                        ]
+                      })
+                    }),
+                    /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                      className: "verify-metric-card",
+                      children: [
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                          className: "verify-metric-label",
+                          children: "Agreement rate"
+                        }),
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                          className: "verify-metric-value",
+                          children: consistencyReport.agreementRate === null ? "\u2014" : `${Math.round(consistencyReport.agreementRate * 100)}%`
+                        })
+                      ]
+                    }),
+                    /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                      className: "verify-metric-card",
+                      children: [
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                          className: "verify-metric-label",
+                          children: "Score spread (min/median/max)"
+                        }),
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                          className: "verify-metric-value verify-metric-value--compact",
+                          children: consistencyReport.scoreSpreadMin === null ? "\u2014" : `${consistencyReport.scoreSpreadMin} / ${consistencyReport.scoreSpreadMedian ?? "\u2014"} / ${consistencyReport.scoreSpreadMax ?? "\u2014"}`
+                        })
+                      ]
+                    }),
+                    /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                      className: "verify-metric-card",
+                      children: [
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                          className: "verify-metric-label",
+                          children: "Instability count"
+                        }),
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                          className: "verify-metric-value",
+                          children: consistencyReport.instabilityCount
+                        })
+                      ]
+                    })
+                  ]
+                }),
+                consistencyReport.sampleSize > 0 && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Callout, {
+                  variant: consistencyReport.verdict === "FAIL" ? "danger" : consistencyReport.verdict === "WARN" ? "emphasis" : "muted",
+                  title: `Verdict: ${consistencyReport.verdict}`,
+                  children: consistencyReport.verdictReason
+                }),
+                /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)(Callout, {
+                  title: "Thresholds in code",
+                  children: [
+                    "Min sample size: ",
+                    VERIFY_CONSISTENCY_THRESHOLDS.minSampleSize,
+                    " ",
+                    "\xB7 PASS requires agreement \u2265 ",
+                    Math.round(VERIFY_CONSISTENCY_THRESHOLDS.pass.agreementMin * 100),
+                    "%, spread \u2264",
+                    " ",
+                    VERIFY_CONSISTENCY_THRESHOLDS.pass.maxSpread,
+                    ", instability \u2264",
+                    " ",
+                    VERIFY_CONSISTENCY_THRESHOLDS.pass.maxInstabilityCount,
+                    " ",
+                    "\xB7 WARN allows agreement \u2265 ",
+                    Math.round(VERIFY_CONSISTENCY_THRESHOLDS.warn.agreementMin * 100),
+                    "%, spread \u2264",
+                    " ",
+                    VERIFY_CONSISTENCY_THRESHOLDS.warn.maxSpread,
+                    ", instability \u2264",
+                    " ",
+                    VERIFY_CONSISTENCY_THRESHOLDS.warn.maxInstabilityCount,
+                    "."
+                  ]
+                }),
+                /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                  className: "verify-section",
+                  children: [
+                    /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                      className: "verify-section-header",
+                      children: [
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("strong", {
+                          children: "Examples"
+                        }),
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                          className: "verify-section-controls",
+                          children: [
+                            /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Button, {
+                              variant: inconsistentOnly ? "primary-deemph" : "secondary",
+                              size: "small",
+                              onClick: () => setInconsistentOnly((prev) => !prev),
+                              children: "Inconsistent"
+                            }),
+                            /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                              className: "verify-section-sort",
+                              children: /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Listbox, {
+                                value: exampleSort,
+                                onChange: (value) => setExampleSort(value),
+                                size: "small",
+                                options: [
+                                  {
+                                    value: "default",
+                                    label: "Sort: default"
+                                  },
+                                  {
+                                    value: "delta_desc",
+                                    label: "Sort: score delta"
+                                  },
+                                  {
+                                    value: "agreement_asc",
+                                    label: "Sort: agreement"
+                                  },
+                                  {
+                                    value: "samples_desc",
+                                    label: "Sort: sample size"
+                                  },
+                                  {
+                                    value: "label_asc",
+                                    label: "Sort: label"
+                                  }
+                                ]
+                              })
+                            })
+                          ]
+                        })
+                      ]
+                    }),
+                    displayedOutliers.length === 0 ? /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Callout, {
+                      children: consistencyReport.outliers.length === 0 ? `Examples will appear here as soon as at least one completed run is available in ${reportScopeLabel}.` : `No examples match the current filters in ${reportScopeLabel}.`
+                    }) : /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                      className: "verify-outlier-list",
+                      children: displayedOutliers.map((outlier) => {
+                        const runLinks = (() => {
+                          if (!selectedSessionId) return [];
+                          const ids = [
+                            outlier.maxRunId,
+                            outlier.minRunId
+                          ].filter((value) => Boolean(value));
+                          return [
+                            ...new Set(ids)
+                          ];
+                        })();
+                        return /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                          className: "verify-outlier-card",
+                          children: [
+                            /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                              className: "verify-outlier-header",
+                              children: [
+                                /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("strong", {
+                                  children: outlier.label
+                                }),
+                                /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                                  style: {
+                                    display: "flex",
+                                    alignItems: "center",
+                                    gap: "8px"
+                                  },
+                                  children: [
+                                    outlier.minScore === outlier.maxScore ? /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Badge, {
+                                      variant: scoreBadgeVariant(outlier.minScore),
+                                      children: formatSignedScore(outlier.minScore)
+                                    }) : /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                                      style: {
+                                        display: "flex",
+                                        alignItems: "center"
+                                      },
+                                      children: [
+                                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Badge, {
+                                          variant: scoreBadgeVariant(outlier.minScore),
+                                          style: {
+                                            borderTopRightRadius: 0,
+                                            borderBottomRightRadius: 0
+                                          },
+                                          children: formatSignedScore(outlier.minScore)
+                                        }),
+                                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Badge, {
+                                          variant: scoreBadgeVariant(outlier.maxScore),
+                                          style: {
+                                            marginLeft: "-1px",
+                                            borderTopLeftRadius: 0,
+                                            borderBottomLeftRadius: 0
+                                          },
+                                          children: formatSignedScore(outlier.maxScore)
+                                        })
+                                      ]
+                                    }),
+                                    /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Badge, {
+                                      variant: outlier.instability ? "error" : "completed",
+                                      children: outlier.instability ? "Unstable" : "Stable"
+                                    })
+                                  ]
+                                })
+                              ]
+                            }),
+                            /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                              className: "flex-row items-center",
+                              children: [
+                                /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                                  className: "flex-1 flex-column",
+                                  children: [
+                                    /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                                      className: "verify-outlier-meta",
+                                      children: [
+                                        "agreement ",
+                                        outlier.agreementRate === null ? "\u2014" : `${Math.round(outlier.agreementRate * 100)}%`,
+                                        " \xB7 delta ",
+                                        outlier.scoreDelta ?? "\u2014",
+                                        " ",
+                                        "\xB7 samples ",
+                                        outlier.sampleSize,
+                                        outlier.passFlip ? " \xB7 pass/fail flip" : "",
+                                        outlier.messageRefId ? ` \xB7 ref ${outlier.messageRefId}` : ""
+                                      ]
+                                    }),
+                                    runLinks.length > 0 && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                                      className: "verify-outlier-links",
+                                      children: runLinks.map((runId) => {
+                                        if (!selectedSessionId) return null;
+                                        const href = buildGradePath(selectedSessionId, runId);
+                                        return /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("a", {
+                                          href,
+                                          onClick: (event) => handleInternalLinkClick(event, href),
+                                          children: [
+                                            "Open grade run ",
+                                            runId
+                                          ]
+                                        }, runId);
+                                      })
+                                    })
+                                  ]
+                                }),
+                                (() => {
+                                  const outlierChip = buildOutlierChip(outlier);
+                                  const inChat = composerChipIds.has(outlierChip.chipId);
+                                  return /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                                    className: "workbench-summary-actions",
+                                    children: /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Button, {
+                                      variant: "secondary",
+                                      size: "small",
+                                      onClick: () => inChat ? removeComposerChip(outlierChip.chipId) : addComposerChip(outlierChip),
+                                      disabled: !onComposerChipsChange,
+                                      children: inChat ? "Remove from chat" : "Add to chat"
+                                    })
+                                  });
+                                })()
+                              ]
+                            })
+                          ]
+                        }, outlier.key);
+                      })
+                    })
+                  ]
+                }),
+                batchState.requests.length > 0 && /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                  className: "verify-section",
+                  children: [
+                    /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("strong", {
+                      children: "Batch requests"
+                    }),
+                    (batchState.startedAt || batchState.finishedAt) && /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                      className: "verify-status-meta",
+                      children: [
+                        batchState.startedAt ? `started ${formatTimestampShort(batchState.startedAt)}` : "",
+                        batchState.finishedAt ? ` \xB7 finished ${formatTimestampShort(batchState.finishedAt)}` : ""
+                      ]
+                    }),
+                    batchState.requested > 0 && /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                      className: "verify-progress-row",
+                      children: [
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("span", {
+                          children: [
+                            "Queued: ",
+                            queuedCount
+                          ]
+                        }),
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("span", {
+                          children: [
+                            "Running: ",
+                            batchState.active
+                          ]
+                        }),
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("span", {
+                          children: [
+                            "Completed: ",
+                            batchState.completed
+                          ]
+                        }),
+                        /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("span", {
+                          children: [
+                            "Failed: ",
+                            batchState.failed
+                          ]
+                        })
+                      ]
+                    }),
+                    /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("ul", {
+                      className: "verify-request-list",
+                      children: batchState.requests.map((request, index) => /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("li", {
+                        className: "verify-request-row",
+                        children: [
+                          /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("span", {
+                            className: "verify-request-index",
+                            children: [
+                              "#",
+                              index + 1
+                            ]
+                          }),
+                          /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Badge, {
+                            status: request.status === "queued" ? "idle" : request.status,
+                            children: request.status
+                          }),
+                          selectedSessionId && request.runId ? (() => {
+                            const href = buildGradePath(selectedSessionId, request.runId);
+                            return /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("a", {
+                              href,
+                              onClick: (event) => handleInternalLinkClick(event, href),
+                              children: request.runId
+                            });
+                          })() : request.runId ? /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("code", {
+                            children: request.runId
+                          }) : null,
+                          request.error && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("span", {
+                            className: "verify-request-error",
+                            children: request.error
+                          })
+                        ]
+                      }, request.requestId))
+                    })
+                  ]
+                })
+              ]
+            })
+          ]
+        })
+      ]
+    })
+  });
+}
+var VerifyPage_default = VerifyPage;
+
+// simulator-ui/src/main.tsx
+var globalStyleEl = document.createElement("style");
+globalStyleEl.textContent = globalStyles;
+document.head.appendChild(globalStyleEl);
+function useSimulator() {
+  const [connectionStatus, setConnectionStatus] = (0, import_react41.useState)("connecting");
+  const [savedState, setSavedState] = (0, import_react41.useState)(null);
+  const [traceEvents, setTraceEvents] = (0, import_react41.useState)([]);
+  const [errors, setErrors] = (0, import_react41.useState)([]);
+  const [streamText, setStreamText] = (0, import_react41.useState)("");
+  const [isRunning, setIsRunning] = (0, import_react41.useState)(false);
+  const [connectSeq, setConnectSeq] = (0, import_react41.useState)(0);
+  (0, import_react41.useEffect)(() => {
+    const streamId = SIMULATOR_STREAM_ID;
+    const streamUrl = buildDurableStreamUrl(streamId, getDurableStreamOffset(streamId));
+    const source = new EventSource(streamUrl);
+    setConnectionStatus("connecting");
+    source.onopen = () => {
+      setConnectionStatus("connected");
+      setErrors([]);
+    };
+    source.onerror = () => {
+      setConnectionStatus("error");
+      setErrors((prev) => prev.includes("Stream connection error") ? prev : [
+        ...prev,
+        "Stream connection error"
+      ]);
+      setIsRunning(false);
+      setStreamText("");
+    };
+    const handleMessage = (event) => {
+      let msg = null;
+      try {
+        msg = JSON.parse(event.data);
+      } catch (err) {
+        console.error("[sim] failed to parse stream event payload", err);
+        return;
+      }
+      const parsedOffset = Number(event.lastEventId);
+      if (Number.isFinite(parsedOffset)) {
+        setDurableStreamOffset(streamId, parsedOffset + 1);
+      }
+      if (!msg || typeof msg !== "object") return;
+      if (msg.type === "state") {
+        setSavedState(msg.state);
+        if (Array.isArray(msg.state.traces)) {
+          setTraceEvents(msg.state.traces);
+        }
+      } else if (msg.type === "stream") {
+        if (typeof msg.chunk === "string" && msg.chunk.length > 0) {
+          setStreamText((prev) => prev + msg.chunk);
+        }
+      } else if (msg.type === "result") {
+        setIsRunning(false);
+        setStreamText("");
+      } else if (msg.type === "trace" && msg.event) {
+        setTraceEvents((prev) => [
+          ...prev,
+          msg.event
+        ].slice(-200));
+      } else if (msg.type === "error") {
+        const message = msg.message ?? "Unknown error";
+        setErrors((prev) => [
+          ...prev,
+          message
+        ]);
+        if (msg.runId || message !== "Run already in progress") {
+          setIsRunning(false);
+        }
+        setStreamText("");
+      }
+    };
+    const simulatorEventTypes = [
+      "ready",
+      "pong",
+      "stream",
+      "result",
+      "trace",
+      "state",
+      "error"
+    ];
+    for (const type of simulatorEventTypes) {
+      source.addEventListener(type, handleMessage);
+    }
+    return () => {
+      for (const type of simulatorEventTypes) {
+        source.removeEventListener(type, handleMessage);
+      }
+      source.close();
+      setConnectionStatus("closed");
+      setIsRunning(false);
+      setStreamText("");
+    };
+  }, [
     connectSeq
   ]);
-  const run = (0, import_react40.useCallback)(async (opts) => {
+  const run = (0, import_react41.useCallback)(async (opts) => {
     setIsRunning(true);
     setStreamText("");
     const sessionId = opts.resetState ? void 0 : savedState?.meta?.sessionId;
@@ -29278,7 +31248,7 @@ function useSimulator() {
   }, [
     savedState?.meta?.sessionId
   ]);
-  const sendFeedback = (0, import_react40.useCallback)(async (messageRefId, score, reason) => {
+  const sendFeedback = (0, import_react41.useCallback)(async (messageRefId, score, reason) => {
     const sessionId = savedState?.meta?.sessionId;
     if (!sessionId) return;
     try {
@@ -29305,7 +31275,7 @@ function useSimulator() {
   }, [
     savedState?.meta?.sessionId
   ]);
-  const loadSession = (0, import_react40.useCallback)(async (sessionId) => {
+  const loadSession = (0, import_react41.useCallback)(async (sessionId) => {
     try {
       const res = await fetch("/api/simulator/load-session", {
         method: "POST",
@@ -29334,7 +31304,7 @@ function useSimulator() {
       ]);
     }
   }, []);
-  const saveNotes = (0, import_react40.useCallback)(async (text) => {
+  const saveNotes = (0, import_react41.useCallback)(async (text) => {
     const sessionId = savedState?.meta?.sessionId;
     if (!sessionId) return;
     try {
@@ -29359,7 +31329,7 @@ function useSimulator() {
   }, [
     savedState?.meta?.sessionId
   ]);
-  const saveSessionScore = (0, import_react40.useCallback)(async (score) => {
+  const saveSessionScore = (0, import_react41.useCallback)(async (score) => {
     const sessionId = savedState?.meta?.sessionId;
     if (!sessionId) return;
     try {
@@ -29384,10 +31354,10 @@ function useSimulator() {
   }, [
     savedState?.meta?.sessionId
   ]);
-  const reconnect = (0, import_react40.useCallback)(() => {
+  const reconnect = (0, import_react41.useCallback)(() => {
     setConnectSeq((prev) => prev + 1);
   }, []);
-  const resetLocal = (0, import_react40.useCallback)(() => {
+  const resetLocal = (0, import_react41.useCallback)(() => {
     setSavedState(null);
     setTraceEvents([]);
     setErrors([]);
@@ -29410,10 +31380,10 @@ function useSimulator() {
   };
 }
 function useWorkspaces() {
-  const [workspaces, setWorkspaces] = (0, import_react40.useState)([]);
-  const [loading, setLoading] = (0, import_react40.useState)(false);
-  const [error, setError] = (0, import_react40.useState)(null);
-  const byNewest = (0, import_react40.useCallback)((items) => {
+  const [workspaces, setWorkspaces] = (0, import_react41.useState)([]);
+  const [loading, setLoading] = (0, import_react41.useState)(false);
+  const [error, setError] = (0, import_react41.useState)(null);
+  const byNewest = (0, import_react41.useCallback)((items) => {
     return [
       ...items
     ].sort((a, b) => {
@@ -29424,7 +31394,7 @@ function useWorkspaces() {
       return safeBTime - safeATime;
     });
   }, []);
-  const refresh = (0, import_react40.useCallback)(async () => {
+  const refresh = (0, import_react41.useCallback)(async () => {
     setLoading(true);
     setError(null);
     try {
@@ -29440,7 +31410,7 @@ function useWorkspaces() {
   }, [
     byNewest
   ]);
-  const deleteWorkspace = (0, import_react40.useCallback)(async (workspaceId) => {
+  const deleteWorkspace = (0, import_react41.useCallback)(async (workspaceId) => {
     setLoading(true);
     setError(null);
     try {
@@ -29463,7 +31433,7 @@ function useWorkspaces() {
   }, [
     refresh
   ]);
-  const deleteAll = (0, import_react40.useCallback)(async (scope) => {
+  const deleteAll = (0, import_react41.useCallback)(async (scope) => {
     setLoading(true);
     setError(null);
     try {
@@ -29504,44 +31474,44 @@ function useWorkspaces() {
 function RecentSessionsEmptyState(props) {
   const { sessions, loading, error, onSelect, onOpenAll } = props;
   const preview = sessions.slice(0, 4);
-  return /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+  return /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
     className: "empty-state",
     children: [
-      /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("p", {
+      /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("p", {
         children: "Start a new chat or load a previous session to review feedback."
       }),
-      loading && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("p", {
+      loading && /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("p", {
         children: "Loading recent sessions\u2026"
       }),
-      error && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("p", {
+      error && /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("p", {
         className: "error",
         children: error
       }),
-      !loading && !error && preview.length === 0 && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("p", {
+      !loading && !error && preview.length === 0 && /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("p", {
         children: "No saved sessions yet."
       }),
-      !loading && !error && preview.length > 0 && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+      !loading && !error && preview.length > 0 && /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("div", {
         className: "recent-sessions",
-        children: preview.map((session) => /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("button", {
+        children: preview.map((session) => /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("button", {
           type: "button",
           className: "recent-session-button",
           onClick: () => onSelect(session.id),
           children: [
-            /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("strong", {
+            /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("strong", {
               children: session.testBotName ?? session.deckSlug ?? session.deck ?? "session"
             }),
-            /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("span", {
+            /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("span", {
               children: formatTimestamp(session.createdAt)
             }),
-            /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("code", {
+            /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("code", {
               children: session.id
             })
           ]
         }, session.id))
       }),
-      /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+      /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("div", {
         className: "empty-state-actions",
-        children: /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Button, {
+        children: /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(Button, {
           variant: "ghost",
           onClick: onOpenAll,
           children: "View all sessions"
@@ -29560,35 +31530,35 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
   const sessionBasePath = rootPath === "/" ? WORKSPACES_BASE_PATH : rootPath;
   const normalizedSessionBase = normalizeBasePath(sessionBasePath);
   const newSessionPath = `${normalizedSessionBase === "" ? WORKSPACES_BASE_PATH : normalizedSessionBase}/new`.replace(/\/{2,}/g, "/");
-  const buildSessionUrl = (0, import_react40.useCallback)((sessionId2) => `${normalizedSessionBase === "" ? WORKSPACES_BASE_PATH : normalizedSessionBase}/${encodeURIComponent(sessionId2)}/debug`.replace(/\/{2,}/g, "/"), [
+  const buildSessionUrl = (0, import_react41.useCallback)((sessionId2) => `${normalizedSessionBase === "" ? WORKSPACES_BASE_PATH : normalizedSessionBase}/${encodeURIComponent(sessionId2)}/debug`.replace(/\/{2,}/g, "/"), [
     normalizedSessionBase
   ]);
-  const [message, setMessage] = (0, import_react40.useState)("");
-  const [pendingReset, setPendingReset] = (0, import_react40.useState)(false);
-  const [initValue, setInitValue] = (0, import_react40.useState)(void 0);
-  const [initDirty, setInitDirty] = (0, import_react40.useState)(false);
-  const [initMode, setInitMode] = (0, import_react40.useState)("form");
-  const [initJsonText, setInitJsonText] = (0, import_react40.useState)("");
-  const [initJsonError, setInitJsonError] = (0, import_react40.useState)(null);
-  const [initOpen, setInitOpen] = (0, import_react40.useState)(false);
-  const [jsonErrors, setJsonErrors] = (0, import_react40.useState)({});
-  const [pendingSessionId, setPendingSessionId] = (0, import_react40.useState)(null);
-  const appliedSessionIdRef = (0, import_react40.useRef)(null);
-  const externalSessionIdRef = (0, import_react40.useRef)(null);
-  const initializedRef = (0, import_react40.useRef)(false);
-  const [noteDraft, setNoteDraft] = (0, import_react40.useState)("");
-  const [noteStatus, setNoteStatus] = (0, import_react40.useState)("idle");
-  const pendingNoteRef = (0, import_react40.useRef)(null);
-  const [scoreStatus, setScoreStatus] = (0, import_react40.useState)("idle");
-  const pendingScoreRef = (0, import_react40.useRef)(null);
-  (0, import_react40.useEffect)(() => {
+  const [message, setMessage] = (0, import_react41.useState)("");
+  const [pendingReset, setPendingReset] = (0, import_react41.useState)(false);
+  const [initValue, setInitValue] = (0, import_react41.useState)(void 0);
+  const [initDirty, setInitDirty] = (0, import_react41.useState)(false);
+  const [initMode, setInitMode] = (0, import_react41.useState)("form");
+  const [initJsonText, setInitJsonText] = (0, import_react41.useState)("");
+  const [initJsonError, setInitJsonError] = (0, import_react41.useState)(null);
+  const [initOpen, setInitOpen] = (0, import_react41.useState)(false);
+  const [jsonErrors, setJsonErrors] = (0, import_react41.useState)({});
+  const [pendingSessionId, setPendingSessionId] = (0, import_react41.useState)(null);
+  const appliedSessionIdRef = (0, import_react41.useRef)(null);
+  const externalSessionIdRef = (0, import_react41.useRef)(null);
+  const initializedRef = (0, import_react41.useRef)(false);
+  const [noteDraft, setNoteDraft] = (0, import_react41.useState)("");
+  const [noteStatus, setNoteStatus] = (0, import_react41.useState)("idle");
+  const pendingNoteRef = (0, import_react41.useRef)(null);
+  const [scoreStatus, setScoreStatus] = (0, import_react41.useState)("idle");
+  const pendingScoreRef = (0, import_react41.useRef)(null);
+  (0, import_react41.useEffect)(() => {
     if (simulator.connectionStatus === "connecting") {
       appliedSessionIdRef.current = null;
     }
   }, [
     simulator.connectionStatus
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     refresh();
   }, [
     refresh
@@ -29598,7 +31568,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
   const schemaError = httpSchema.schemaResponse?.error ?? httpSchema.error ?? void 0;
   const conversationStarted = Boolean(simulator.savedState && simulator.savedState.messages.length > 0);
   const initEditable = Boolean(schema) && (!conversationStarted || pendingReset);
-  const lockedInitValue = (0, import_react40.useMemo)(() => {
+  const lockedInitValue = (0, import_react41.useMemo)(() => {
     const fromTraces = extractInitFromTraces(simulator.savedState?.traces);
     if (fromTraces !== void 0) return fromTraces;
     return schemaDefaults ?? deriveInitialFromSchema(schema);
@@ -29607,7 +31577,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     schemaDefaults,
     schema
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (!schema) return;
     if (initDirty) return;
     if (schemaDefaults !== void 0) {
@@ -29620,7 +31590,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     schemaDefaults,
     initDirty
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (initMode !== "json") return;
     if (initDirty) return;
     try {
@@ -29633,19 +31603,19 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     initValue,
     initDirty
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (initEditable) {
       setInitOpen(true);
     }
   }, [
     initEditable
   ]);
-  const messages = (0, import_react40.useMemo)(() => {
+  const messages = (0, import_react41.useMemo)(() => {
     return buildConversationEntries(simulator.savedState);
   }, [
     simulator.savedState
   ]);
-  const missingRequired = (0, import_react40.useMemo)(() => {
+  const missingRequired = (0, import_react41.useMemo)(() => {
     if (!schema || !initEditable) return [];
     return findMissingRequiredFields(schema, initValue);
   }, [
@@ -29653,13 +31623,13 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     initEditable,
     initValue
   ]);
-  const jsonErrorCount = (0, import_react40.useMemo)(() => {
+  const jsonErrorCount = (0, import_react41.useMemo)(() => {
     return Object.values(jsonErrors).filter((v) => typeof v === "string" && v).length;
   }, [
     jsonErrors
   ]);
   const canStartWithInit = Boolean(schema) && initEditable && missingRequired.length === 0 && jsonErrorCount === 0;
-  const resetInitValue = (0, import_react40.useCallback)(() => {
+  const resetInitValue = (0, import_react41.useCallback)(() => {
     setInitJsonError(null);
     setJsonErrors((prev) => ({
       ...prev,
@@ -29677,7 +31647,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     schema,
     schemaDefaults
   ]);
-  const startNewChat = (0, import_react40.useCallback)((opts) => {
+  const startNewChat = (0, import_react41.useCallback)((opts) => {
     const shouldPush = opts?.pushHistory ?? true;
     if (shouldPush) {
       if (opts?.replace) {
@@ -29707,14 +31677,14 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     resetInitValue,
     newSessionPath
   ]);
-  const adoptSessionFromPath = (0, import_react40.useCallback)((sessionId2) => {
+  const adoptSessionFromPath = (0, import_react41.useCallback)((sessionId2) => {
     appliedSessionIdRef.current = null;
     setPendingSessionId(sessionId2);
     setPendingReset(false);
     setInitOpen(false);
     setInitDirty(true);
   }, []);
-  const navigateToSession = (0, import_react40.useCallback)((sessionId2, opts) => {
+  const navigateToSession = (0, import_react41.useCallback)((sessionId2, opts) => {
     const url = buildSessionUrl(sessionId2);
     if (opts?.replace) {
       window.history.replaceState({}, "", url);
@@ -29726,7 +31696,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     adoptSessionFromPath,
     buildSessionUrl
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (initializedRef.current) return;
     initializedRef.current = true;
     const initialSession = getWorkspaceIdFromPath(void 0, sessionBasePath) ?? getWorkspaceIdFromPath();
@@ -29744,7 +31714,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     startNewChat,
     sessionBasePath
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (!activeWorkspaceId) {
       externalSessionIdRef.current = null;
       return;
@@ -29756,7 +31726,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     activeWorkspaceId,
     adoptSessionFromPath
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     const handler = () => {
       const sessionFromPath = getWorkspaceIdFromPath(void 0, sessionBasePath) ?? getWorkspaceIdFromPath();
       if (sessionFromPath) {
@@ -29774,7 +31744,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     startNewChat,
     sessionBasePath
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (!pendingSessionId) return;
     if (simulator.connectionStatus !== "connected") return;
     if (appliedSessionIdRef.current === pendingSessionId) return;
@@ -29789,7 +31759,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
   const serverNotesUpdatedAt = simulator.savedState?.notes?.updatedAt;
   const serverScore = simulator.savedState?.conversationScore?.score ?? null;
   const serverScoreUpdatedAt = simulator.savedState?.conversationScore?.updatedAt;
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (pendingNoteRef.current !== null) {
       if (serverNotesText === pendingNoteRef.current) {
         pendingNoteRef.current = null;
@@ -29802,7 +31772,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
   }, [
     serverNotesText
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (noteStatus !== "dirty") return;
     const handle = window.setTimeout(() => {
       setNoteStatus("saving");
@@ -29815,7 +31785,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     noteDraft,
     simulator
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (pendingScoreRef.current !== null) {
       if (serverScore === pendingScoreRef.current) {
         pendingScoreRef.current = null;
@@ -29831,7 +31801,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
   }, [
     serverScore
   ]);
-  const handleSend = (0, import_react40.useCallback)(() => {
+  const handleSend = (0, import_react41.useCallback)(() => {
     const trimmed = message.trim();
     if (schema && initEditable) {
       if (!canStartWithInit) return;
@@ -29871,12 +31841,12 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     canStartWithInit,
     initValue
   ]);
-  const handleScore = (0, import_react40.useCallback)((refId, score, reason) => {
+  const handleScore = (0, import_react41.useCallback)((refId, score, reason) => {
     simulator.sendFeedback(refId, score, reason);
   }, [
     simulator
   ]);
-  const handleReason = (0, import_react40.useCallback)((refId, score, reason) => {
+  const handleReason = (0, import_react41.useCallback)((refId, score, reason) => {
     simulator.sendFeedback(refId, score, reason);
   }, [
     simulator
@@ -29886,16 +31856,16 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
   const sessionPermalink = sessionId ? buildSessionUrl(sessionId) : null;
   const sessionStatePath = typeof runMeta.sessionStatePath === "string" ? runMeta.sessionStatePath : typeof runMeta.sessionDir === "string" ? `${runMeta.sessionDir}/state.json` : void 0;
   const currentSessionScore = pendingScoreRef.current !== null ? pendingScoreRef.current : serverScore;
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (!setNavActions) return;
-    setNavActions(/* @__PURE__ */ (0, import_jsx_runtime43.jsxs)(import_jsx_runtime43.Fragment, {
+    setNavActions(/* @__PURE__ */ (0, import_jsx_runtime44.jsxs)(import_jsx_runtime44.Fragment, {
       children: [
-        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Button, {
+        /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(Button, {
           variant: "secondary",
           onClick: () => startNewChat(),
           children: "New chat"
         }),
-        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+        /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("div", {
           className: `status-indicator ${simulator.connectionStatus}`,
           children: simulator.connectionStatus
         })
@@ -29909,7 +31879,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     simulator.connectionStatus,
     startNewChat
   ]);
-  const deckSessions = (0, import_react40.useMemo)(() => {
+  const deckSessions = (0, import_react41.useMemo)(() => {
     return workspaces.filter((session) => {
       if (!session) return false;
       if (typeof session.deck === "string") {
@@ -29928,22 +31898,22 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
     normalizedDeckPath,
     repoRootPath
   ]);
-  const recentSessionsEmpty = /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(RecentSessionsEmptyState, {
+  const recentSessionsEmpty = /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(RecentSessionsEmptyState, {
     sessions: deckSessions,
     loading: workspacesLoading,
     error: workspacesError,
     onSelect: (id) => navigateToSession(id),
     onOpenAll: onOpenSessionsDrawer
   });
-  return /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)(PageShell, {
+  return /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)(PageShell, {
     children: [
-      /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)(PageGrid, {
+      /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)(PageGrid, {
         as: "main",
         className: "app-main",
         children: [
-          /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(ConversationView, {
+          /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(ConversationView, {
             messages,
-            header: schema ? /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(InitPanel, {
+            header: schema ? /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(InitPanel, {
               schema,
               value: initValue,
               lockedValue: lockedInitValue,
@@ -30003,18 +31973,18 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
             onScore: handleScore,
             onReasonChange: handleReason
           }),
-          /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(TraceList, {
+          /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(TraceList, {
             traces: simulator.traceEvents
           })
         ]
       }),
-      /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("footer", {
+      /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("footer", {
         className: "composer",
         children: [
-          /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+          /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
             className: "composer-inputs",
             children: [
-              /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("textarea", {
+              /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("textarea", {
                 className: "message-input",
                 "data-testid": "debug-message-input",
                 placeholder: schema && initEditable ? "Optional first message (init will be sent too)" : "Optional message (assistant can start)",
@@ -30027,23 +31997,23 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
                   }
                 }
               }),
-              /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+              /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
                 className: "notes-inline",
                 children: [
-                  /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("header", {
+                  /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("header", {
                     children: [
-                      /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("label", {
+                      /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("label", {
                         htmlFor: "session-notes",
                         children: "Session notes"
                       }),
-                      /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                      /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
                         className: "rating-controls",
                         children: [
-                          /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("span", {
+                          /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("span", {
                             className: "rating-label",
                             children: "Overall score"
                           }),
-                          SCORE_VALUES.map((value) => /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("button", {
+                          SCORE_VALUES.map((value) => /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("button", {
                             type: "button",
                             className: classNames("rating-button", currentSessionScore === value && "active"),
                             onClick: () => {
@@ -30057,7 +32027,7 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
                       })
                     ]
                   }),
-                  /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("textarea", {
+                  /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("textarea", {
                     id: "session-notes",
                     value: noteDraft,
                     onChange: (e) => {
@@ -30066,14 +32036,14 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
                     },
                     placeholder: "Add context or TODOs..."
                   }),
-                  /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                  /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("div", {
                     className: "notes-inline-status",
-                    children: /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("span", {
+                    children: /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("span", {
                       className: classNames("state", noteStatus === "saving" && "saving", noteStatus === "dirty" && "unsaved", noteStatus === "idle" && "idle", noteStatus === "saved" && "saved"),
                       children: noteStatus === "saving" ? "Saving\u2026" : noteStatus === "dirty" ? "Unsaved changes\u2026" : noteStatus === "saved" ? serverNotesUpdatedAt ? `Saved ${formatTimestamp(serverNotesUpdatedAt)}` : "Saved" : "No notes yet."
                     })
                   }),
-                  /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                  /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
                     className: "rating-status",
                     children: [
                       "Overall score:",
@@ -30087,52 +32057,52 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
               })
             ]
           }),
-          pendingReset && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+          pendingReset && /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("div", {
             className: "reset-note",
             children: "Next message will start a new chat."
           }),
-          /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+          /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
             className: "composer-actions",
             children: [
-              /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Button, {
+              /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(Button, {
                 variant: "primary",
                 onClick: handleSend,
                 disabled: schema && initEditable && !canStartWithInit,
                 "data-testid": "debug-send",
                 children: schema && initEditable ? "Start chat" : "Send"
               }),
-              /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Button, {
+              /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(Button, {
                 variant: "secondary",
                 onClick: simulator.reconnect,
                 children: "Reconnect"
               })
             ]
           }),
-          simulator.errors.map((err, idx) => /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+          simulator.errors.map((err, idx) => /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("div", {
             className: "error",
             children: err
           }, idx)),
-          sessionId && /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+          sessionId && /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
             className: "session-meta",
             children: [
               "Session: ",
-              sessionPermalink ? /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("a", {
+              sessionPermalink ? /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("a", {
                 href: sessionPermalink,
                 className: "session-link",
                 title: "Open session permalink",
-                children: /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("code", {
+                children: /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("code", {
                   children: sessionId
                 })
-              }) : /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("code", {
+              }) : /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("code", {
                 children: sessionId
               })
             ]
           }),
-          sessionStatePath && /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+          sessionStatePath && /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
             className: "session-meta session-path",
             children: [
               "State file: ",
-              /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("code", {
+              /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("code", {
                 children: sessionStatePath
               })
             ]
@@ -30144,18 +32114,18 @@ function SimulatorApp({ basePath, setNavActions, workspacesApi, onOpenSessionsDr
 }
 function App() {
   const simulatorBasePath = WORKSPACES_BASE_PATH;
-  const [path, setPath] = (0, import_react40.useState)(() => normalizeAppPath(window.location.pathname));
-  const [bundleStamp, setBundleStamp] = (0, import_react40.useState)(null);
-  const [navActions, setNavActions] = (0, import_react40.useState)(null);
-  const [sessionsDrawerOpen, setSessionsDrawerOpen] = (0, import_react40.useState)(false);
-  const [workbenchDrawerOpen, setWorkbenchDrawerOpen] = (0, import_react40.useState)(true);
-  const [workbenchComposerChips, setWorkbenchComposerChips] = (0, import_react40.useState)([]);
-  const [workspaceRunIds, setWorkspaceRunIds] = (0, import_react40.useState)({
+  const [path, setPath] = (0, import_react41.useState)(() => normalizeAppPath(window.location.pathname));
+  const [bundleStamp, setBundleStamp] = (0, import_react41.useState)(null);
+  const [navActions, setNavActions] = (0, import_react41.useState)(null);
+  const [sessionsDrawerOpen, setSessionsDrawerOpen] = (0, import_react41.useState)(false);
+  const [workbenchDrawerOpen, setWorkbenchDrawerOpen] = (0, import_react41.useState)(true);
+  const [workbenchComposerChips, setWorkbenchComposerChips] = (0, import_react41.useState)([]);
+  const [workspaceRunIds, setWorkspaceRunIds] = (0, import_react41.useState)({
     testRunId: null,
     gradeRunId: null
   });
   const workspacesApi = useWorkspaces();
-  const [testBotResetToken, setTestBotResetToken] = (0, import_react40.useState)(0);
+  const [testBotResetToken, setTestBotResetToken] = (0, import_react41.useState)(0);
   const pathRoute = getWorkspaceRouteFromPath(path);
   const livePath = window.location.pathname.replace(/\/+$/, "") || "/";
   const liveRoute = getWorkspaceRouteFromPath(livePath);
@@ -30164,23 +32134,23 @@ function App() {
   const activeWorkspaceId = routeRequestsNewWorkspace ? null : routeState?.workspaceId ?? workspaceIdFromWindow;
   const requestedTestRunId = routeState?.tab === "test" ? routeState.testRunId ?? null : null;
   const requestedGradeRunId = routeState?.tab === "grade" ? routeState.gradeRunId ?? null : null;
-  const lastWorkspaceIdRef = (0, import_react40.useRef)(null);
-  const [workbenchSessionDetail, setWorkbenchSessionDetail] = (0, import_react40.useState)(null);
-  const [workbenchSessionDetailError, setWorkbenchSessionDetailError] = (0, import_react40.useState)(null);
-  const [workbenchSessionDetailLoading, setWorkbenchSessionDetailLoading] = (0, import_react40.useState)(false);
-  const workbenchSessionDetailRequestRef = (0, import_react40.useRef)(0);
-  const workbenchSessionRetryRef = (0, import_react40.useRef)({});
-  const workbenchRefreshTimeoutRef = (0, import_react40.useRef)(null);
-  const activeWorkspaceIdRef = (0, import_react40.useRef)(activeWorkspaceId);
-  const workspaceInitRef = (0, import_react40.useRef)(false);
-  (0, import_react40.useEffect)(() => {
+  const lastWorkspaceIdRef = (0, import_react41.useRef)(null);
+  const [workbenchSessionDetail, setWorkbenchSessionDetail] = (0, import_react41.useState)(null);
+  const [workbenchSessionDetailError, setWorkbenchSessionDetailError] = (0, import_react41.useState)(null);
+  const [workbenchSessionDetailLoading, setWorkbenchSessionDetailLoading] = (0, import_react41.useState)(false);
+  const workbenchSessionDetailRequestRef = (0, import_react41.useRef)(0);
+  const workbenchSessionRetryRef = (0, import_react41.useRef)({});
+  const workbenchRefreshTimeoutRef = (0, import_react41.useRef)(null);
+  const activeWorkspaceIdRef = (0, import_react41.useRef)(activeWorkspaceId);
+  const workspaceInitRef = (0, import_react41.useRef)(false);
+  (0, import_react41.useEffect)(() => {
     if (activeWorkspaceId) {
       lastWorkspaceIdRef.current = activeWorkspaceId;
     }
   }, [
     activeWorkspaceId
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (!activeWorkspaceId) {
       setWorkbenchComposerChips([]);
       return;
@@ -30189,7 +32159,7 @@ function App() {
   }, [
     activeWorkspaceId
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     const syncPath = () => setPath(normalizeAppPath(window.location.pathname));
     const historyObj = window.history;
     const originalPushState = historyObj.pushState.bind(historyObj);
@@ -30211,7 +32181,7 @@ function App() {
       window.removeEventListener("locationchange", syncPath);
     };
   }, []);
-  const loadWorkbenchSessionDetail = (0, import_react40.useCallback)(async (sessionId) => {
+  const loadWorkbenchSessionDetail = (0, import_react41.useCallback)(async (sessionId) => {
     const requestId = ++workbenchSessionDetailRequestRef.current;
     const shouldApply = () => requestId === workbenchSessionDetailRequestRef.current && activeWorkspaceIdRef.current === sessionId;
     try {
@@ -30253,7 +32223,7 @@ function App() {
       }
     }
   }, []);
-  const scheduleWorkbenchRefresh = (0, import_react40.useCallback)(() => {
+  const scheduleWorkbenchRefresh = (0, import_react41.useCallback)(() => {
     if (!activeWorkspaceId) return;
     if (workbenchRefreshTimeoutRef.current) {
       window.clearTimeout(workbenchRefreshTimeoutRef.current);
@@ -30268,7 +32238,7 @@ function App() {
     activeWorkspaceId,
     loadWorkbenchSessionDetail
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     activeWorkspaceIdRef.current = activeWorkspaceId;
     if (workbenchRefreshTimeoutRef.current) {
       window.clearTimeout(workbenchRefreshTimeoutRef.current);
@@ -30277,14 +32247,14 @@ function App() {
   }, [
     activeWorkspaceId
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     return () => {
       if (workbenchRefreshTimeoutRef.current) {
         window.clearTimeout(workbenchRefreshTimeoutRef.current);
       }
     };
   }, []);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (!activeWorkspaceId) {
       setWorkbenchSessionDetail(null);
       setWorkbenchSessionDetailError(null);
@@ -30297,14 +32267,14 @@ function App() {
     activeWorkspaceId,
     loadWorkbenchSessionDetail
   ]);
-  const handleFeedbackPersisted = (0, import_react40.useCallback)((workspaceId) => {
+  const handleFeedbackPersisted = (0, import_react41.useCallback)((workspaceId) => {
     if (!workspaceId) return;
     if (activeWorkspaceIdRef.current !== workspaceId) return;
     scheduleWorkbenchRefresh();
   }, [
     scheduleWorkbenchRefresh
   ]);
-  const applyFlagsUpdate = (0, import_react40.useCallback)((flags) => {
+  const applyFlagsUpdate = (0, import_react41.useCallback)((flags) => {
     setWorkbenchSessionDetail((prev) => {
       if (!prev) return prev;
       return {
@@ -30319,7 +32289,7 @@ function App() {
   }, [
     scheduleWorkbenchRefresh
   ]);
-  const optimisticToggleFlag = (0, import_react40.useCallback)((item) => {
+  const optimisticToggleFlag = (0, import_react41.useCallback)((item) => {
     setWorkbenchSessionDetail((prev) => {
       if (!prev) return prev;
       const meta = prev.meta ?? {};
@@ -30344,7 +32314,7 @@ function App() {
       };
     });
   }, []);
-  const optimisticFlagReason = (0, import_react40.useCallback)((refId, reason) => {
+  const optimisticFlagReason = (0, import_react41.useCallback)((refId, reason) => {
     setWorkbenchSessionDetail((prev) => {
       if (!prev) return prev;
       const meta = prev.meta ?? {};
@@ -30362,7 +32332,7 @@ function App() {
       };
     });
   }, []);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     const loadBundleStamp = async () => {
       try {
         const res = await fetch("/ui/bundle.js", {
@@ -30378,36 +32348,36 @@ function App() {
     };
     loadBundleStamp();
   }, []);
-  const navigate = (0, import_react40.useCallback)((next) => {
+  const navigate = (0, import_react41.useCallback)((next) => {
     if (next === path) return;
     window.history.pushState({}, "", next);
     setPath(next);
   }, [
     path
   ]);
-  const replacePath = (0, import_react40.useCallback)((next) => {
+  const replacePath = (0, import_react41.useCallback)((next) => {
     if (next === path) return;
     window.history.replaceState({}, "", next);
     setPath(next);
   }, [
     path
   ]);
-  const handleAppPathChange = (0, import_react40.useCallback)((next) => {
+  const handleAppPathChange = (0, import_react41.useCallback)((next) => {
     setPath(normalizeAppPath(next));
   }, []);
-  const handleReplaceTestBotSession = (0, import_react40.useCallback)((workspaceId, runId) => replacePath(buildTestPath(workspaceId, runId)), [
+  const handleReplaceTestBotSession = (0, import_react41.useCallback)((workspaceId, runId) => replacePath(buildTestPath(workspaceId, runId)), [
     replacePath
   ]);
-  const handleResetTestBotSession = (0, import_react40.useCallback)(() => replacePath(buildTestPath(activeWorkspaceId ?? void 0)), [
+  const handleResetTestBotSession = (0, import_react41.useCallback)(() => replacePath(buildTestPath(activeWorkspaceId ?? void 0)), [
     activeWorkspaceId,
     replacePath
   ]);
-  const handleWorkspaceChange = (0, import_react40.useCallback)((workspaceId) => {
+  const handleWorkspaceChange = (0, import_react41.useCallback)((workspaceId) => {
     replacePath(buildWorkspacePath("build", workspaceId));
   }, [
     replacePath
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (!buildTabEnabled && path === "/build") {
       replacePath(DOCS_PATH);
     }
@@ -30415,17 +32385,42 @@ function App() {
     path,
     replacePath
   ]);
+  (0, import_react41.useEffect)(() => {
+    if (verifyTabEnabled) return;
+    const route = getWorkspaceRouteFromPath(path);
+    if (route?.tab !== "verify") return;
+    const fallback = route.workspaceId ? buildGradePath(route.workspaceId) : DEFAULT_GRADE_PATH;
+    replacePath(fallback);
+  }, [
+    path,
+    replacePath,
+    verifyTabEnabled
+  ]);
   const isDocs = path === DOCS_PATH;
   const routeTab = liveRoute?.tab ?? pathRoute?.tab;
   const isBuild = buildTabEnabled && (routeTab === "build" || path === "/build");
   const isTestBot = !isDocs && (routeTab === "test" || /\/test$/.test(path));
   const isGrade = !isDocs && routeTab === "grade";
-  const currentPage = isDocs ? "docs" : isBuild ? "build" : isTestBot ? "test" : isGrade ? "grade" : "debug";
-  (0, import_react40.useEffect)(() => {
+  const isVerify = !isDocs && verifyTabEnabled && routeTab === "verify";
+  let currentPage;
+  if (isDocs) {
+    currentPage = "docs";
+  } else if (isBuild) {
+    currentPage = "build";
+  } else if (isTestBot) {
+    currentPage = "test";
+  } else if (isGrade) {
+    currentPage = "grade";
+  } else if (isVerify) {
+    currentPage = "verify";
+  } else {
+    currentPage = "debug";
+  }
+  (0, import_react41.useEffect)(() => {
     if (activeWorkspaceId || lastWorkspaceIdRef.current) return;
     if (!routeRequestsNewWorkspace) return;
     if (workspaceInitRef.current) return;
-    if (currentPage !== "build" && currentPage !== "test" && currentPage !== "grade") {
+    if (currentPage !== "build" && currentPage !== "test" && currentPage !== "grade" && currentPage !== "verify") {
       return;
     }
     workspaceInitRef.current = true;
@@ -30434,7 +32429,22 @@ function App() {
     }).then(async (res) => {
       const data = await res.json().catch(() => ({}));
       if (!res.ok || typeof data.workspaceId !== "string") return;
-      const nextPath = currentPage === "test" ? buildWorkspacePath("test", data.workspaceId) : currentPage === "grade" ? buildGradePath(data.workspaceId) : buildWorkspacePath("build", data.workspaceId);
+      let nextPath;
+      switch (currentPage) {
+        case "test":
+          nextPath = buildWorkspacePath("test", data.workspaceId);
+          break;
+        case "grade":
+          nextPath = buildGradePath(data.workspaceId);
+          break;
+        case "verify":
+          nextPath = buildWorkspacePath("verify", data.workspaceId);
+          break;
+        case "build":
+        default:
+          nextPath = buildWorkspacePath("build", data.workspaceId);
+          break;
+      }
       replacePath(nextPath);
     }).finally(() => {
       workspaceInitRef.current = false;
@@ -30445,9 +32455,9 @@ function App() {
     replacePath,
     routeRequestsNewWorkspace
   ]);
-  const resolveNavWorkspaceId = (0, import_react40.useCallback)(() => {
+  const resolveNavWorkspaceId = (0, import_react41.useCallback)(() => {
     const normalizedPath = (window.location.pathname || "/").replace(/\/+$/, "") || "/";
-    const directMatch = normalizedPath.match(/^\/workspaces\/([^/]+)\/(?:debug|build|test|grade)(?:\/[^/]+)?$/);
+    const directMatch = normalizedPath.match(/^\/workspaces\/([^/]+)\/(?:debug|build|test|grade|verify)(?:\/[^/]+)?$/);
     if (directMatch?.[1] && directMatch[1] !== "new") {
       return decodeURIComponent(directMatch[1]);
     }
@@ -30455,7 +32465,7 @@ function App() {
   }, [
     activeWorkspaceId
   ]);
-  const resolveNavPath = (0, import_react40.useCallback)((next) => {
+  const resolveNavPath = (0, import_react41.useCallback)((next) => {
     if (next === "docs") return DOCS_PATH;
     const workspaceId = resolveNavWorkspaceId();
     if (next === "build") return buildWorkspacePath("build", workspaceId);
@@ -30465,21 +32475,44 @@ function App() {
     if (next === "grade") {
       return workspaceId ? buildGradePath(workspaceId, workspaceRunIds.gradeRunId ?? void 0) : DEFAULT_GRADE_PATH;
     }
+    if (next === "verify") {
+      return workspaceId ? buildVerifyPath(workspaceId) : DEFAULT_VERIFY_PATH;
+    }
     return buildWorkspacePath("debug", workspaceId);
   }, [
+    DEFAULT_VERIFY_PATH,
     resolveNavWorkspaceId,
     workspaceRunIds.gradeRunId,
     workspaceRunIds.testRunId
   ]);
-  const handleSelectSession = (0, import_react40.useCallback)((sessionId) => {
-    const nextPath = currentPage === "test" || currentPage === "docs" ? buildWorkspacePath("test", sessionId) : currentPage === "grade" ? buildGradePath(sessionId) : currentPage === "build" ? buildWorkspacePath("build", sessionId) : buildWorkspacePath("debug", sessionId);
+  const handleSelectSession = (0, import_react41.useCallback)((sessionId) => {
+    let nextPath;
+    switch (currentPage) {
+      case "docs":
+      case "test":
+        nextPath = buildWorkspacePath("test", sessionId);
+        break;
+      case "grade":
+        nextPath = buildGradePath(sessionId);
+        break;
+      case "verify":
+        nextPath = buildVerifyPath(sessionId);
+        break;
+      case "build":
+        nextPath = buildWorkspacePath("build", sessionId);
+        break;
+      case "debug":
+      default:
+        nextPath = buildWorkspacePath("debug", sessionId);
+        break;
+    }
     navigate(nextPath);
     setSessionsDrawerOpen(false);
   }, [
     currentPage,
     navigate
   ]);
-  (0, import_react40.useEffect)(() => {
+  (0, import_react41.useEffect)(() => {
     if (sessionsDrawerOpen) {
       workspacesApi.refresh();
     }
@@ -30487,7 +32520,7 @@ function App() {
     workspacesApi.refresh,
     sessionsDrawerOpen
   ]);
-  const deckSessions = (0, import_react40.useMemo)(() => {
+  const deckSessions = (0, import_react41.useMemo)(() => {
     if (workspaceIdFromWindow) {
       return workspacesApi.workspaces;
     }
@@ -30510,7 +32543,7 @@ function App() {
     repoRootPath,
     workspaceIdFromWindow
   ]);
-  const handleDeleteAll = (0, import_react40.useCallback)(async () => {
+  const handleDeleteAll = (0, import_react41.useCallback)(async () => {
     await workspacesApi.deleteAll(deckSessions);
     setTestBotResetToken((prev) => prev + 1);
     setSessionsDrawerOpen(false);
@@ -30519,7 +32552,7 @@ function App() {
     deckSessions,
     workspacesApi.deleteAll
   ]);
-  const handleDeleteSession = (0, import_react40.useCallback)(async (sessionId) => {
+  const handleDeleteSession = (0, import_react41.useCallback)(async (sessionId) => {
     await workspacesApi.deleteWorkspace(sessionId);
     if (sessionId === activeWorkspaceId) {
       window.location.assign(DEFAULT_TEST_PATH);
@@ -30528,11 +32561,12 @@ function App() {
     activeWorkspaceId,
     workspacesApi.deleteWorkspace
   ]);
-  const handleAddScenarioErrorToWorkbench = (0, import_react40.useCallback)((payload) => {
+  const handleAddErrorToWorkbench = (0, import_react41.useCallback)((payload) => {
     const resolvedWorkspaceId = payload.workspaceId ?? activeWorkspaceId ?? void 0;
+    const source = payload.source ?? "scenario_run_error";
     const nextChip = {
-      chipId: `error:${resolvedWorkspaceId ?? ""}:${payload.runId ?? ""}:${payload.error}`,
-      source: "scenario_run_error",
+      chipId: `error:${source}:${resolvedWorkspaceId ?? ""}:${payload.runId ?? ""}:${payload.error}`,
+      source,
       workspaceId: resolvedWorkspaceId,
       runId: payload.runId,
       error: payload.error,
@@ -30559,30 +32593,97 @@ function App() {
   }, [
     activeWorkspaceId
   ]);
-  return /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(WorkspaceProvider, {
+  const pageContent = (0, import_react41.useMemo)(() => {
+    switch (currentPage) {
+      case "docs":
+        return /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(DocsPage, {});
+      case "build":
+        return /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(BuildPage, {
+          setNavActions
+        });
+      case "debug":
+        return /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(SimulatorApp, {
+          basePath: simulatorBasePath,
+          setNavActions,
+          workspacesApi,
+          onOpenSessionsDrawer: () => setSessionsDrawerOpen(true),
+          activeWorkspaceId
+        });
+      case "test":
+        return /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(TestBotPage, {
+          onReplaceTestBotSession: handleReplaceTestBotSession,
+          onResetTestBotSession: handleResetTestBotSession,
+          activeWorkspaceId,
+          requestedRunId: requestedTestRunId,
+          resetToken: testBotResetToken,
+          setNavActions,
+          onFeedbackPersisted: handleFeedbackPersisted,
+          onAddErrorToWorkbench: handleAddErrorToWorkbench
+        });
+      case "verify":
+        return /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(VerifyPage_default, {
+          setNavActions,
+          onAppPathChange: handleAppPathChange,
+          activeWorkspaceId,
+          composerChips: workbenchComposerChips,
+          onComposerChipsChange: setWorkbenchComposerChips
+        });
+      case "grade":
+      default:
+        return /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(GradePage_default, {
+          setNavActions,
+          onAppPathChange: handleAppPathChange,
+          activeWorkspaceId,
+          requestedGradeRunId,
+          onFlagsUpdate: applyFlagsUpdate,
+          onOptimisticToggleFlag: optimisticToggleFlag,
+          onOptimisticFlagReason: optimisticFlagReason,
+          onAddErrorToWorkbench: handleAddErrorToWorkbench
+        });
+    }
+  }, [
+    activeWorkspaceId,
+    applyFlagsUpdate,
+    currentPage,
+    handleAddErrorToWorkbench,
+    handleAppPathChange,
+    handleFeedbackPersisted,
+    handleReplaceTestBotSession,
+    handleResetTestBotSession,
+    optimisticFlagReason,
+    optimisticToggleFlag,
+    requestedGradeRunId,
+    requestedTestRunId,
+    setNavActions,
+    simulatorBasePath,
+    testBotResetToken,
+    workbenchComposerChips,
+    workspacesApi
+  ]);
+  return /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(WorkspaceProvider, {
     workspaceId: activeWorkspaceId,
     onWorkspaceChange: handleWorkspaceChange,
     requestedTestRunId: routeState?.tab === "test" ? requestedTestRunId : void 0,
     requestedGradeRunId: routeState?.tab === "grade" ? requestedGradeRunId : void 0,
     onRoutingStateChange: setWorkspaceRunIds,
-    children: /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)(import_jsx_runtime43.Fragment, {
+    children: /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)(import_jsx_runtime44.Fragment, {
       children: [
-        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+        /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("div", {
           className: "app-frame",
-          children: /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+          children: /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
             className: "app-root",
             children: [
-              /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+              /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
                 className: "top-nav",
                 children: [
-                  /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                  /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("div", {
                     className: "top-nav-left",
-                    children: /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Button, {
+                    children: /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(Button, {
                       variant: sessionsDrawerOpen ? "primary-deemph" : "ghost",
                       className: classNames("sessions-toggle", sessionsDrawerOpen && "active"),
                       onClick: () => setSessionsDrawerOpen(true),
                       "data-testid": "nav-sessions",
-                      children: /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Icon, {
+                      children: /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(Icon, {
                         name: "hamburgerMenu",
                         size: 17,
                         style: {
@@ -30591,7 +32692,7 @@ function App() {
                       })
                     })
                   }),
-                  /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Tabs, {
+                  /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(Tabs, {
                     className: "top-nav-buttons",
                     activeId: currentPage,
                     onChange: (next) => navigate(resolveNavPath(next)),
@@ -30622,6 +32723,14 @@ function App() {
                         testId: "nav-grade",
                         href: resolveNavPath("grade")
                       },
+                      ...verifyTabEnabled ? [
+                        {
+                          id: "verify",
+                          label: "Verify",
+                          testId: "nav-verify",
+                          href: resolveNavPath("verify")
+                        }
+                      ] : [],
                       {
                         id: "debug",
                         label: "Debug",
@@ -30630,27 +32739,27 @@ function App() {
                       }
                     ]
                   }),
-                  /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                  /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("div", {
                     className: "top-nav-center",
-                    children: /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("span", {
+                    children: /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("span", {
                       className: "top-nav-deck",
                       title: deckPath,
                       children: deckLabel
                     })
                   }),
-                  /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                  /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("div", {
                     className: "top-nav-right",
-                    children: /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+                    children: /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
                       className: "top-nav-actions",
                       children: [
                         navActions,
-                        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Button, {
+                        /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(Button, {
                           variant: workbenchDrawerOpen ? "primary-deemph" : "ghost",
                           className: classNames("workbench-toggle", workbenchDrawerOpen && "active"),
                           onClick: () => setWorkbenchDrawerOpen((prev) => !prev),
                           "aria-label": workbenchDrawerOpen ? "Close workbench drawer" : "Open workbench drawer",
                           "data-testid": "nav-workbench",
-                          children: /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(Icon, {
+                          children: /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(Icon, {
                             name: "chat",
                             size: 16,
                             style: {
@@ -30663,39 +32772,14 @@ function App() {
                   })
                 ]
               }),
-              /* @__PURE__ */ (0, import_jsx_runtime43.jsxs)("div", {
+              /* @__PURE__ */ (0, import_jsx_runtime44.jsxs)("div", {
                 className: "app-content-frame",
                 children: [
-                  /* @__PURE__ */ (0, import_jsx_runtime43.jsx)("div", {
+                  /* @__PURE__ */ (0, import_jsx_runtime44.jsx)("div", {
                     className: "page-shell",
-                    children: currentPage === "docs" ? /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(DocsPage, {}) : currentPage === "build" ? /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(BuildPage, {
-                      setNavActions
-                    }) : currentPage === "debug" ? /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(SimulatorApp, {
-                      basePath: simulatorBasePath,
-                      setNavActions,
-                      workspacesApi,
-                      onOpenSessionsDrawer: () => setSessionsDrawerOpen(true),
-                      activeWorkspaceId
-                    }) : currentPage === "test" ? /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(TestBotPage, {
-                      onReplaceTestBotSession: handleReplaceTestBotSession,
-                      onResetTestBotSession: handleResetTestBotSession,
-                      activeWorkspaceId,
-                      requestedRunId: requestedTestRunId,
-                      resetToken: testBotResetToken,
-                      setNavActions,
-                      onFeedbackPersisted: handleFeedbackPersisted,
-                      onAddScenarioErrorToWorkbench: handleAddScenarioErrorToWorkbench
-                    }) : /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(GradePage_default, {
-                      setNavActions,
-                      onAppPathChange: handleAppPathChange,
-                      activeWorkspaceId,
-                      requestedGradeRunId,
-                      onFlagsUpdate: applyFlagsUpdate,
-                      onOptimisticToggleFlag: optimisticToggleFlag,
-                      onOptimisticFlagReason: optimisticFlagReason
-                    })
+                    children: pageContent
                   }),
-                  workbenchDrawerOpen && /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(WorkbenchDrawer, {
+                  workbenchDrawerOpen && /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(WorkbenchDrawer, {
                     open: workbenchDrawerOpen,
                     onClose: () => setWorkbenchDrawerOpen(false),
                     loading: workbenchSessionDetailLoading,
@@ -30710,7 +32794,7 @@ function App() {
             ]
           })
         }),
-        /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(AppDrawer, {
+        /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(AppDrawer, {
           open: sessionsDrawerOpen,
           workspaces: deckSessions,
           loading: workspacesApi.loading,
@@ -30727,8 +32811,8 @@ function App() {
     })
   });
 }
-(0, import_client.createRoot)(document.getElementById("root")).render(/* @__PURE__ */ (0, import_jsx_runtime43.jsx)(import_react40.default.StrictMode, {
-  children: /* @__PURE__ */ (0, import_jsx_runtime43.jsx)(App, {})
+(0, import_client.createRoot)(document.getElementById("root")).render(/* @__PURE__ */ (0, import_jsx_runtime44.jsx)(import_react41.default.StrictMode, {
+  children: /* @__PURE__ */ (0, import_jsx_runtime44.jsx)(App, {})
 }));
 /**
  * @license React