[pull] main from CodebuffAI:main

.claude/settings.json

@@ -0,0 +1,5 @@
+{
+  "permissions": {
+    "defaultMode": "auto"
+  }
+}

cli/src/app.tsx

@@ -285,17 +285,6 @@ export const App = ({
     )
   }
 
-  // Render chat history screen when requested
-  if (showChatHistory) {
-    return (
-      <ChatHistoryScreen
-        onSelectChat={handleResumeChat}
-        onCancel={closeChatHistory}
-        onNewChat={handleNewChat}
-      />
-    )
-  }
-
   // Use key to force remount when resuming a different chat from history
   const chatKey = resumeChatId ?? 'current'
 
@@ -316,6 +305,10 @@ export const App = ({
       initialMode={initialMode}
       gitRoot={gitRoot}
       onSwitchToGitRoot={handleSwitchToGitRoot}
+      showChatHistory={showChatHistory}
+      onSelectChat={handleResumeChat}
+      onCancelChatHistory={closeChatHistory}
+      onNewChat={handleNewChat}
     />
   )
 }
@@ -336,6 +329,10 @@ interface AuthedSurfaceProps {
   initialMode: AgentMode | undefined
   gitRoot: string | null | undefined
   onSwitchToGitRoot: () => void
+  showChatHistory: boolean
+  onSelectChat: (chatId: string) => void
+  onCancelChatHistory: () => void
+  onNewChat: () => void
 }
 
 /**
@@ -359,6 +356,10 @@ const AuthedSurface = ({
   initialMode,
   gitRoot,
   onSwitchToGitRoot,
+  showChatHistory,
+  onSelectChat,
+  onCancelChatHistory,
+  onNewChat,
 }: AuthedSurfaceProps) => {
   const { session, error: sessionError } = useFreebuffSession()
 
@@ -388,6 +389,20 @@ const AuthedSurface = ({
     return <WaitingRoomScreen session={session} error={sessionError} />
   }
 
+  // Chat history renders inside AuthedSurface so the freebuff session stays
+  // mounted while the user browses history. Unmounting this surface would
+  // DELETE the session row and drop the user back into the waiting room on
+  // return.
+  if (showChatHistory) {
+    return (
+      <ChatHistoryScreen
+        onSelectChat={onSelectChat}
+        onCancel={onCancelChatHistory}
+        onNewChat={onNewChat}
+      />
+    )
+  }
+
   return (
     <Chat
       key={chatKey}

cli/src/chat.tsx

@@ -1473,15 +1473,17 @@ export const Chat = ({
         )}
 
         {reviewMode ? (
-          // Review takes precedence over the session-ended banner: during the
-          // grace window the agent may still be asking to run tools, and
-          // those approvals must be reachable for the run to finish.
+          // Review and ask_user take precedence over the session-ended banner:
+          // during the grace window the agent may still be asking to run tools
+          // or asking the user a question, and those approvals/answers must be
+          // reachable for the run to finish — otherwise the agent hangs
+          // waiting for input that can never be given.
           <ReviewScreen
             onSelectOption={handleReviewOptionSelect}
             onCustom={handleReviewCustom}
             onCancel={handleCloseReviewScreen}
           />
-        ) : isFreebuffSessionOver ? (
+        ) : isFreebuffSessionOver && !askUserState ? (
           <SessionEndedBanner
             isStreaming={isStreaming || isWaitingForResponse}
           />

cli/src/hooks/helpers/send-message.ts

@@ -510,10 +510,16 @@ function handleFreebuffGateError(
   switch (kind) {
     case 'session_expired':
     case 'waiting_room_required':
-      // Our seat is gone mid-chat. Flip to `ended` instead of auto re-queuing:
-      // the Chat surface stays mounted so any in-flight agent work can finish
-      // under the server-side grace period, and the session-ended banner
-      // prompts the user to press Enter when they're ready to rejoin.
+      // Our seat is gone mid-chat. Finalize the AI message so its streaming
+      // indicator stops — otherwise `isComplete` stays false and the message
+      // keeps rendering a blinking cursor forever, making the user think the
+      // agent is still working even though the SessionEndedBanner is visible
+      // and actionable. Also disposes the batched-updater flush interval.
+      updater.markComplete()
+      // Flip to `ended` instead of auto re-queuing: the Chat surface stays
+      // mounted so any in-flight agent work can finish under the server-side
+      // grace period, and the session-ended banner prompts the user to press
+      // Enter when they're ready to rejoin.
       markFreebuffSessionEnded()
       return
     case 'waiting_room_queued':

docs/error-schema.md

@@ -34,7 +34,7 @@ Used for errors that the client needs to identify programmatically:
 
 | Status | `error` code | Example `message` |
 |--------|-------------|-------------------|
-| 403 | `account_suspended` | `"Your account has been suspended due to billing issues. Please contact support@codebuff.com to resolve this."` |
+| 403 | `account_suspended` | `"Your account has been suspended. Please contact support@codebuff.com if you did not expect this."` |
 | 403 | `free_mode_unavailable` | `"Free mode is not available in your country."` (Freebuff: `"Freebuff is not available in your country."`) |
 | 429 | `rate_limit_exceeded` | `"Subscription weekly limit reached. Your limit resets in 2 hours. Enable 'Continue with credits' in the CLI to use a-la-carte credits."` |
 

packages/agent-runtime/src/__tests__/main-prompt.test.ts

@@ -375,6 +375,7 @@ describe('mainPrompt', () => {
   it('should update consecutiveAssistantMessages when new prompt is received', async () => {
     const sessionState = getInitialSessionState(mockFileContext)
     sessionState.mainAgentState.stepsRemaining = 12
+    const initialStepsRemaining = sessionState.mainAgentState.stepsRemaining
 
     const action = {
       type: 'prompt' as const,
@@ -394,7 +395,7 @@ describe('mainPrompt', () => {
 
     // When there's a new prompt, consecutiveAssistantMessages should be set to 1
     expect(newSessionState.mainAgentState.stepsRemaining).toBe(
-      sessionState.mainAgentState.stepsRemaining - 1,
+      initialStepsRemaining - 1,
     )
   })
 

packages/agent-runtime/src/run-agent-step.ts

@@ -536,6 +536,17 @@ export const runAgentStep = async (
   }
 }
 
+/**
+ * Runs the agent loop.
+ *
+ * IMPORTANT: This function mutates `params.agentState` in place throughout the
+ * run (not just at return time). Fields like `messageHistory`, `systemPrompt`,
+ * `toolDefinitions`, `creditsUsed`, and `output` are updated as work progresses
+ * so that callers holding a reference to the same object (e.g. the SDK's
+ * `sessionState.mainAgentState`) see in-progress work immediately — which
+ * matters when an error is thrown mid-run and the normal return path is
+ * skipped.
+ */
 export async function loopAgentSteps(
   params: {
     addAgentStep: AddAgentStepFn
@@ -800,12 +811,13 @@ export async function loopAgentSteps(
     return cachedAdditionalToolDefinitions
   }
 
-  let currentAgentState: AgentState = {
-    ...initialAgentState,
-    messageHistory: initialMessages,
-    systemPrompt: system,
-    toolDefinitions,
-  }
+  // Mutate initialAgentState so that in-progress work propagates back to the
+  // caller's shared reference (e.g. SDK's sessionState.mainAgentState) even if
+  // an error is thrown before we return.
+  initialAgentState.messageHistory = initialMessages
+  initialAgentState.systemPrompt = system
+  initialAgentState.toolDefinitions = toolDefinitions
+  let currentAgentState: AgentState = initialAgentState
 
   // Convert tool definitions to Anthropic format for accurate token counting
   // Tool definitions are stored as { [name]: { description, inputSchema } }
@@ -908,7 +920,8 @@ export async function loopAgentSteps(
         } = programmaticResult
         n = generateN
 
-        currentAgentState = programmaticAgentState
+        Object.assign(initialAgentState, programmaticAgentState)
+        currentAgentState = initialAgentState
         totalSteps = stepNumber
 
         shouldEndTurn = endTurn
@@ -989,7 +1002,8 @@ export async function loopAgentSteps(
         logger.error('No runId found for agent state after finishing agent run')
       }
 
-      currentAgentState = newAgentState
+      Object.assign(initialAgentState, newAgentState)
+      currentAgentState = initialAgentState
       shouldEndTurn = llmShouldEndTurn
       nResponses = generatedResponses