feat(h5): inject direct-chat history on agent escalation and defer ambiguous routing.

Phase B appends snapshot context when leaving h5direct sessions so Goose no longer starts blind. Phase C lets the 0.72 fallback return null on gated chat text so the existing LLM router can run, flags default off.

Co-authored-by: Cursor <cursoragent@cursor.com>

agent-run-gateway.mjs

@@ -45,6 +45,10 @@ import {
 import { applyCursorFirstAgentExecution } from './cursor-page-routing.mjs';
 import { cursorDeepseekFallbackEnabled } from './cursor-agent-launch.mjs';
 import { buildCursorBillingTokenState } from './cursor-agent-usage.mjs';
+import {
+  isDirectEscalationContextEnabledForUser,
+  resolveDirectEscalationContextPolicy,
+} from './chat-task-intent-config.mjs';
 
 const DEFAULT_RUN_RETRY_DELAYS_MS = [1500, 5000, 15000];
 const TERMINAL_STATUSES = new Set(['succeeded', 'failed']);
@@ -260,6 +264,26 @@ function appendFreshSessionContext(
   return { ...message, content };
 }
 
+function maybeInjectDirectEscalationContext({
+  userMessage,
+  conversation,
+  userId,
+  policy,
+} = {}) {
+  if (!isDirectEscalationContextEnabledForUser(userId, policy)) {
+    return { userMessage, injected: false, messageCount: 0 };
+  }
+  const context = buildFreshSessionContext(conversation);
+  if (!context) {
+    return { userMessage, injected: false, messageCount: 0 };
+  }
+  return {
+    userMessage: appendFreshSessionContext(userMessage, conversation),
+    injected: true,
+    messageCount: Array.isArray(conversation) ? conversation.length : 0,
+  };
+}
+
 const FRESH_SESSION_RECOVERY_CODES = new Set([
   'SESSION_TOOL_HISTORY_FRESH_SESSION_REQUIRED',
   'SESSION_REASONING_CONTENT_POISONED',
@@ -871,9 +895,12 @@ export function createAgentRunGateway({
   ),
   goalRunService = null,
   workerIdentity = null,
+  directEscalationContextPolicy = null,
 }) {
   const worker = normalizeAgentRunWorkerIdentity(workerIdentity ?? {});
   const sessionStore = resolveSessionAccess({ userAuth, sessionAccess });
+  const escalationContextPolicy = directEscalationContextPolicy
+    ?? resolveDirectEscalationContextPolicy();
   const inFlight = new Set();
   const queuedDispatches = [];
   const queuedDispatchSet = new Set();
@@ -2044,6 +2071,7 @@ export function createAgentRunGateway({
       forceDeepReasoning: runOptions.forceDeepReasoning,
     });
     let escalatedDirectSessionId = null;
+    let escalationContextMessages = null;
     if (isDirectChatSessionId(sessionId)) {
       escalatedDirectSessionId = sessionId;
       await appendEvent(runId, 'direct_session_escalated_to_deep_reasoning', {
@@ -2069,7 +2097,7 @@ export function createAgentRunGateway({
       });
       await appendRunSnapshot(runId);
       if (escalatedDirectSessionId) {
-        const priorMessages = await loadSnapshotMessages(
+        escalationContextMessages = await loadSnapshotMessages(
           sessionSnapshotService,
           escalatedDirectSessionId,
         );
@@ -2077,7 +2105,7 @@ export function createAgentRunGateway({
           conversationMemoryService,
           sessionId,
           userId: row.user_id,
-          messages: priorMessages,
+          messages: escalationContextMessages,
         });
         if (persisted.saved > 0) {
           await appendEvent(runId, 'direct_session_transcript_persisted', {
@@ -2102,6 +2130,30 @@ export function createAgentRunGateway({
         sessionId,
         saved: transcriptPersisted.saved,
       });
+      if (!escalationContextMessages?.length) {
+        escalationContextMessages = await loadSnapshotMessages(
+          sessionSnapshotService,
+          escalatedDirectSessionId ?? sessionId,
+        );
+      }
+    }
+    if (escalationContextMessages?.length) {
+      const escalationContext = maybeInjectDirectEscalationContext({
+        userMessage,
+        conversation: escalationContextMessages,
+        userId: row.user_id,
+        policy: escalationContextPolicy,
+      });
+      if (escalationContext.injected) {
+        userMessage = escalationContext.userMessage;
+        await appendEvent(runId, 'direct_escalation_context_injected', {
+          sessionId,
+          previousSessionId: escalatedDirectSessionId ?? null,
+          messageCount: escalationContext.messageCount,
+          retainedContextMessages: 16,
+          retainedContextChars: 12_000,
+        });
+      }
     }
     if (typeof tkmindProxy.compactSessionConversationForUser === 'function') {
       try {