letta-code/src/cli/App.tsx

// src/cli/App.tsx

import { existsSync, readFileSync, renameSync, writeFileSync } from "node:fs";
import { homedir, tmpdir } from "node:os";
import { join } from "node:path";
import { APIError, APIUserAbortError } from "@letta-ai/letta-client/core/error";
import type {
  AgentState,
  MessageCreate,
} from "@letta-ai/letta-client/resources/agents/agents";
import type {
  ApprovalCreate,
  Message,
} from "@letta-ai/letta-client/resources/agents/messages";
import type { LlmConfig } from "@letta-ai/letta-client/resources/models/models";
import type { StopReasonType } from "@letta-ai/letta-client/resources/runs/runs";
import { Box, Static } from "ink";
import {
  useCallback,
  useEffect,
  useMemo,
  useRef,
  useState,
  useSyncExternalStore,
} from "react";
import {
  type ApprovalResult,
  executeAutoAllowedTools,
  getDisplayableToolReturn,
} from "../agent/approval-execution";
import {
  fetchRunErrorDetail,
  isApprovalPendingError,
  isConversationBusyError,
  isInvalidToolCallIdsError,
} from "../agent/approval-recovery";
import { prefetchAvailableModelHandles } from "../agent/available-models";
import { getResumeData } from "../agent/check-approval";
import { getClient, getServerUrl } from "../agent/client";
import { getCurrentAgentId, setCurrentAgentId } from "../agent/context";
import { type AgentProvenance, createAgent } from "../agent/create";
import { getLettaCodeHeaders } from "../agent/http-headers";

import { ISOLATED_BLOCK_LABELS } from "../agent/memory";
import {
  ensureMemoryFilesystemDirs,
  getMemoryFilesystemRoot,
} from "../agent/memoryFilesystem";
import { sendMessageStream } from "../agent/message";
import { getModelInfo, getModelShortName } from "../agent/model";
import { INTERRUPT_RECOVERY_ALERT } from "../agent/promptAssets";
import { SessionStats } from "../agent/stats";
import {
  INTERRUPTED_BY_USER,
  MEMFS_CONFLICT_CHECK_INTERVAL,
  SYSTEM_ALERT_CLOSE,
  SYSTEM_ALERT_OPEN,
  SYSTEM_REMINDER_CLOSE,
  SYSTEM_REMINDER_OPEN,
} from "../constants";
import {
  runNotificationHooks,
  runPreCompactHooks,
  runSessionEndHooks,
  runSessionStartHooks,
  runStopHooks,
  runUserPromptSubmitHooks,
} from "../hooks";
import type { ApprovalContext } from "../permissions/analyzer";
import { type PermissionMode, permissionMode } from "../permissions/mode";
import {
  DEFAULT_COMPLETION_PROMISE,
  type RalphState,
  ralphMode,
} from "../ralph/mode";
import { updateProjectSettings } from "../settings";
import { settingsManager } from "../settings-manager";
import { telemetry } from "../telemetry";
import {
  analyzeToolApproval,
  checkToolPermission,
  executeTool,
  isGeminiModel,
  isOpenAIModel,
  savePermissionRule,
  type ToolExecutionResult,
} from "../tools/manager";
import { debugLog, debugWarn } from "../utils/debug";
import {
  handleMcpAdd,
  handleMcpUsage,
  type McpCommandContext,
  setActiveCommandId as setActiveMcpCommandId,
} from "./commands/mcp";
import {
  addCommandResult,
  handlePin,
  handleProfileDelete,
  handleProfileSave,
  handleProfileUsage,
  handleUnpin,
  type ProfileCommandContext,
  setActiveCommandId as setActiveProfileCommandId,
  validateProfileLoad,
} from "./commands/profile";
import { type CommandHandle, createCommandRunner } from "./commands/runner";
import { AgentSelector } from "./components/AgentSelector";
// ApprovalDialog removed - all approvals now render inline
import { ApprovalPreview } from "./components/ApprovalPreview";
import { ApprovalSwitch } from "./components/ApprovalSwitch";
import { AssistantMessage } from "./components/AssistantMessageRich";
import { BashCommandMessage } from "./components/BashCommandMessage";
import { CommandMessage } from "./components/CommandMessage";
import { ConversationSelector } from "./components/ConversationSelector";
import { colors } from "./components/colors";
// EnterPlanModeDialog removed - now using InlineEnterPlanModeApproval
import { ErrorMessage } from "./components/ErrorMessageRich";
import { EventMessage } from "./components/EventMessage";
import { FeedbackDialog } from "./components/FeedbackDialog";
import { HelpDialog } from "./components/HelpDialog";
import { HooksManager } from "./components/HooksManager";
import { Input } from "./components/InputRich";
import { McpConnectFlow } from "./components/McpConnectFlow";
import { McpSelector } from "./components/McpSelector";
import { MemfsTreeViewer } from "./components/MemfsTreeViewer";
import { MemoryTabViewer } from "./components/MemoryTabViewer";
import { MessageSearch } from "./components/MessageSearch";
import { ModelSelector } from "./components/ModelSelector";
import { NewAgentDialog } from "./components/NewAgentDialog";
import { PendingApprovalStub } from "./components/PendingApprovalStub";
import { PinDialog, validateAgentName } from "./components/PinDialog";
import { ProviderSelector } from "./components/ProviderSelector";
import { ReasoningMessage } from "./components/ReasoningMessageRich";
import { formatDuration, formatUsageStats } from "./components/SessionStats";
import { SleeptimeSelector } from "./components/SleeptimeSelector";
// InlinePlanApproval kept for easy rollback if needed
// import { InlinePlanApproval } from "./components/InlinePlanApproval";
import { StatusMessage } from "./components/StatusMessage";
import { SubagentGroupDisplay } from "./components/SubagentGroupDisplay";
import { SubagentGroupStatic } from "./components/SubagentGroupStatic";
import { SubagentManager } from "./components/SubagentManager";
import { SystemPromptSelector } from "./components/SystemPromptSelector";
import { Text } from "./components/Text";
import { ToolCallMessage } from "./components/ToolCallMessageRich";
import { ToolsetSelector } from "./components/ToolsetSelector";
import { TrajectorySummary } from "./components/TrajectorySummary";
import { UserMessage } from "./components/UserMessageRich";
import { WelcomeScreen } from "./components/WelcomeScreen";
import { AnimationProvider } from "./contexts/AnimationContext";
import {
  appendStreamingOutput,
  type Buffers,
  createBuffers,
  type Line,
  markIncompleteToolsAsCancelled,
  onChunk,
  setToolCallsRunning,
  toLines,
} from "./helpers/accumulator";
import { classifyApprovals } from "./helpers/approvalClassification";
import { backfillBuffers } from "./helpers/backfill";
import {
  type ContextWindowOverview,
  renderContextUsage,
} from "./helpers/contextChart";
import {
  createContextTracker,
  resetContextHistory,
} from "./helpers/contextTracker";
import {
  type AdvancedDiffSuccess,
  computeAdvancedDiff,
  parsePatchToAdvancedDiff,
} from "./helpers/diff";
import { setErrorContext } from "./helpers/errorContext";
import {
  formatErrorDetails,
  isEncryptedContentError,
} from "./helpers/errorFormatter";
import { formatCompact } from "./helpers/format";
import { parsePatchOperations } from "./helpers/formatArgsDisplay";
import {
  buildCompactionMemoryReminder,
  buildMemoryReminder,
  getReflectionSettings,
  parseMemoryPreference,
  type ReflectionSettings,
  reflectionSettingsToLegacyMode,
  shouldFireStepCountTrigger,
} from "./helpers/memoryReminder";
import {
  type QueuedMessage,
  setMessageQueueAdder,
} from "./helpers/messageQueueBridge";
import {
  buildMessageContentFromDisplay,
  clearPlaceholdersInText,
  resolvePlaceholders,
} from "./helpers/pasteRegistry";
import { generatePlanFilePath } from "./helpers/planName";
import {
  buildQueuedContentParts,
  buildQueuedUserText,
  getQueuedNotificationSummaries,
} from "./helpers/queuedMessageParts";
import { safeJsonParseOr } from "./helpers/safeJsonParse";
import { getDeviceType, getLocalTime } from "./helpers/sessionContext";
import { resolveStatusLineConfig } from "./helpers/statusLineConfig";
import { formatStatusLineHelp } from "./helpers/statusLineHelp";
import { buildStatusLinePayload } from "./helpers/statusLinePayload";
import { executeStatusLineCommand } from "./helpers/statusLineRuntime";
import { type ApprovalRequest, drainStreamWithResume } from "./helpers/stream";
import {
  collectFinishedTaskToolCalls,
  createSubagentGroupItem,
  hasInProgressTaskToolCalls,
} from "./helpers/subagentAggregation";
import {
  clearCompletedSubagents,
  clearSubagentsByIds,
  getSubagentByToolCallId,
  getSnapshot as getSubagentSnapshot,
  hasActiveSubagents,
  interruptActiveSubagents,
  subscribe as subscribeToSubagents,
} from "./helpers/subagentState";
import {
  flushEligibleLinesBeforeReentry,
  shouldClearCompletedSubagentsOnTurnStart,
} from "./helpers/subagentTurnStart";
import { extractTaskNotificationsForDisplay } from "./helpers/taskNotifications";
import {
  getRandomPastTenseVerb,
  getRandomThinkingVerb,
} from "./helpers/thinkingMessages";
import {
  isFileEditTool,
  isFileWriteTool,
  isPatchTool,
  isShellTool,
} from "./helpers/toolNameMapping";
import {
  alwaysRequiresUserInput,
  isTaskTool,
} from "./helpers/toolNameMapping.js";
import { useConfigurableStatusLine } from "./hooks/useConfigurableStatusLine";
import { useSuspend } from "./hooks/useSuspend/useSuspend.ts";
import { useSyncedState } from "./hooks/useSyncedState";
import { useTerminalRows, useTerminalWidth } from "./hooks/useTerminalWidth";

// Used only for terminal resize, not for dialog dismissal (see PR for details)
const CLEAR_SCREEN_AND_HOME = "\u001B[2J\u001B[H";
const MIN_RESIZE_DELTA = 2;
const RESIZE_SETTLE_MS = 250;
const MIN_CLEAR_INTERVAL_MS = 750;
const STABLE_WIDTH_SETTLE_MS = 180;
const TOOL_CALL_COMMIT_DEFER_MS = 50;
const ANIMATION_RESUME_HYSTERESIS_ROWS = 2;

// Eager approval checking is now CONDITIONAL (LET-7101):
// - Enabled when resuming a session (--resume, --continue, or startupApprovals exist)
// - Disabled for normal messages (lazy recovery handles edge cases)
// This saves ~2s latency per message in the common case.

// Feature flag: Eagerly cancel streams client-side when user presses ESC
// When true (default), immediately abort the stream after calling .cancel()
// This provides instant feedback to the user without waiting for backend acknowledgment
// When false, wait for backend to send "cancelled" stop_reason (useful for testing backend behavior)
const EAGER_CANCEL = true;

// Maximum retries for transient LLM API errors (matches headless.ts)
const LLM_API_ERROR_MAX_RETRIES = 3;

// Retry config for 409 "conversation busy" errors (exponential backoff)
const CONVERSATION_BUSY_MAX_RETRIES = 3; // 2.5s -> 5s -> 10s
const CONVERSATION_BUSY_RETRY_BASE_DELAY_MS = 2500; // 2.5 seconds

// Message shown when user interrupts the stream
const INTERRUPT_MESSAGE =
  "Interrupted – tell the agent what to do differently. Something went wrong? Use /feedback to report issues.";

// Hint shown after errors to encourage feedback
const ERROR_FEEDBACK_HINT =
  "Something went wrong? Use /feedback to report issues.";

// Hint shown when Anthropic Opus 4.5 hits llm_api_error and Bedrock is available
const OPUS_BEDROCK_FALLBACK_HINT =
  "Downstream provider issues? Use /model to switch to Bedrock Opus 4.5";

// Generic hint for llm_api_error when specific model suggestion not applicable
const PROVIDER_FALLBACK_HINT =
  "Downstream provider issues? Use /model to switch to another provider";

// Helper to get appropriate error hint based on stop reason and current model
function getErrorHintForStopReason(
  stopReason: StopReasonType | null,
  currentModelId: string | null,
): string {
  if (
    currentModelId === "opus" &&
    stopReason === "llm_api_error" &&
    getModelInfo("bedrock-opus")
  ) {
    return OPUS_BEDROCK_FALLBACK_HINT;
  }
  if (stopReason === "llm_api_error") {
    return PROVIDER_FALLBACK_HINT;
  }
  return ERROR_FEEDBACK_HINT;
}

// Interactive slash commands that open overlays immediately (bypass queueing)
// These commands let users browse/view while the agent is working
// Any changes made in the overlay will be queued until end_turn
const INTERACTIVE_SLASH_COMMANDS = new Set([
  "/model",
  "/toolset",
  "/system",
  "/subagents",
  "/memory",
  "/sleeptime",
  "/mcp",
  "/help",
  "/agents",
  "/resume",
  "/pinned",
  "/profiles",
  "/search",
  "/feedback",
  "/pin",
  "/pin-local",
  "/conversations",
  "/profile",
]);

// Non-state commands that should run immediately while the agent is busy
// These don't modify agent state, so they should bypass queueing
const NON_STATE_COMMANDS = new Set([
  "/ade",
  "/bg",
  "/usage",
  "/help",
  "/hooks",
  "/search",
  "/memory",
  "/feedback",
  "/export",
  "/download",
  "/statusline",
]);

// Check if a command is interactive (opens overlay, should not be queued)
function isInteractiveCommand(msg: string): boolean {
  const trimmed = msg.trim().toLowerCase();
  // Check exact matches first
  if (INTERACTIVE_SLASH_COMMANDS.has(trimmed)) return true;
  // Check prefix matches for commands with arguments
  for (const cmd of INTERACTIVE_SLASH_COMMANDS) {
    if (trimmed.startsWith(`${cmd} `)) return true;
  }
  return false;
}

function isNonStateCommand(msg: string): boolean {
  const trimmed = msg.trim().toLowerCase();
  if (NON_STATE_COMMANDS.has(trimmed)) return true;
  for (const cmd of NON_STATE_COMMANDS) {
    if (trimmed.startsWith(`${cmd} `)) return true;
  }
  return false;
}

const APPROVAL_OPTIONS_HEIGHT = 8;
const APPROVAL_PREVIEW_BUFFER = 4;
const MIN_WRAP_WIDTH = 10;
const TEXT_WRAP_GUTTER = 6;
const DIFF_WRAP_GUTTER = 12;

function countWrappedLines(text: string, width: number): number {
  if (!text) return 0;
  const wrapWidth = Math.max(1, width);
  return text.split(/\r?\n/).reduce((sum, line) => {
    const len = line.length;
    const wrapped = Math.max(1, Math.ceil(len / wrapWidth));
    return sum + wrapped;
  }, 0);
}

function countWrappedLinesFromList(lines: string[], width: number): number {
  if (!lines.length) return 0;
  const wrapWidth = Math.max(1, width);
  return lines.reduce((sum, line) => {
    const len = line.length;
    const wrapped = Math.max(1, Math.ceil(len / wrapWidth));
    return sum + wrapped;
  }, 0);
}

function estimateAdvancedDiffLines(
  diff: AdvancedDiffSuccess,
  width: number,
): number {
  const wrapWidth = Math.max(1, width);
  let total = 0;
  for (const hunk of diff.hunks) {
    for (const line of hunk.lines) {
      const raw = line.raw || "";
      if (raw.startsWith("\\")) continue;
      const text = raw.slice(1);
      total += Math.max(1, Math.ceil(text.length / wrapWidth));
    }
  }
  return total;
}

// tiny helper for unique ids (avoid overwriting prior user lines)
function uid(prefix: string) {
  return `${prefix}-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 8)}`;
}

// Send desktop notification via terminal bell
// Modern terminals (iTerm2, Ghostty, WezTerm, Kitty) convert this to a desktop
// notification when the terminal is not focused
function sendDesktopNotification(
  message = "Awaiting your input",
  level: "info" | "warning" | "error" = "info",
) {
  // Send terminal bell for native notification
  process.stdout.write("\x07");
  // Run Notification hooks (fire-and-forget, don't block)
  runNotificationHooks(message, level).catch((error) => {
    debugLog("hooks", "Notification hook error", error);
  });
}

// Check if error is retriable based on stop reason and run metadata
async function isRetriableError(
  stopReason: StopReasonType,
  lastRunId: string | null | undefined,
): Promise<boolean> {
  // Primary check: backend sets stop_reason=llm_api_error for LLMError exceptions
  if (stopReason === "llm_api_error") return true;

  // Early exit for stop reasons that should never be retried
  const nonRetriableReasons: StopReasonType[] = [
    "cancelled",
    "requires_approval",
    "max_steps",
    "max_tokens_exceeded",
    "context_window_overflow_in_system_prompt",
    "end_turn",
    "tool_rule",
    "no_tool_call",
  ];
  if (nonRetriableReasons.includes(stopReason)) return false;

  // Fallback check: for error-like stop_reasons, check metadata for retriable patterns
  // This handles cases where the backend sends a generic error stop_reason but the
  // underlying cause is a transient LLM/network issue that should be retried
  if (lastRunId) {
    try {
      const client = await getClient();
      const run = await client.runs.retrieve(lastRunId);
      const metaError = run.metadata?.error as
        | {
            error_type?: string;
            detail?: string;
            // Handle nested error structure (error.error) that can occur in some edge cases
            error?: { error_type?: string; detail?: string };
          }
        | undefined;

      // Check for llm_error at top level or nested (handles error.error nesting)
      const errorType = metaError?.error_type ?? metaError?.error?.error_type;
      const detail = metaError?.detail ?? metaError?.error?.detail ?? "";

      // Don't retry 4xx client errors (validation, auth, malformed requests)
      // These are not transient and won't succeed on retry
      const is4xxError = /Error code: 4\d{2}/.test(detail);

      if (errorType === "llm_error" && !is4xxError) return true;

      // Fallback: detect LLM provider errors from detail even if misclassified
      // This handles edge cases where streaming errors weren't properly converted to LLMError
      // Patterns are derived from handle_llm_error() message formats in the backend
      const llmProviderPatterns = [
        "Anthropic API error", // anthropic_client.py:759
        "OpenAI API error", // openai_client.py:1034
        "Google Vertex API error", // google_vertex_client.py:848
        "overloaded", // anthropic_client.py:753 - used for LLMProviderOverloaded
        "api_error", // Anthropic SDK error type field
        "Network error", // Transient network failures during streaming
        "Connection error during Anthropic streaming", // Peer disconnections, incomplete chunked reads
      ];
      if (
        llmProviderPatterns.some((pattern) => detail.includes(pattern)) &&
        !is4xxError
      ) {
        return true;
      }

      return false;
    } catch {
      return false;
    }
  }
  return false;
}

// Save current agent as lastAgent before exiting
// This ensures subagent overwrites during the session don't persist
function saveLastAgentBeforeExit() {
  try {
    const currentAgentId = getCurrentAgentId();
    settingsManager.updateLocalProjectSettings({ lastAgent: currentAgentId });
    settingsManager.updateSettings({ lastAgent: currentAgentId });
  } catch {
    // Ignore if no agent context set
  }
}

// Get plan mode system reminder if in plan mode
function getPlanModeReminder(): string {
  if (permissionMode.getMode() !== "plan") {
    return "";
  }

  const planFilePath = permissionMode.getPlanFilePath();

  // Generate dynamic reminder with plan file path
  return `${SYSTEM_REMINDER_OPEN}
      Plan mode is active. The user indicated that they do not want you to execute yet -- you MUST NOT make any edits (with the exception of the plan file mentioned below), run any non-readonly tools (including changing configs or making commits), or otherwise make any changes to the system. This supersedes any other instructions you have received.

## Plan File Info:
${planFilePath ? `No plan file exists yet. You should create your plan at ${planFilePath} using a write tool (e.g. Write, ApplyPatch, etc. depending on your toolset).` : "No plan file path assigned."}

You should build your plan incrementally by writing to or editing this file. NOTE that this is the only file you are allowed to edit - other than this you are only allowed to take READ-ONLY actions.

**Plan File Guidelines:** The plan file should contain only your final recommended approach, not all alternatives considered. Keep it comprehensive yet concise - detailed enough to execute effectively while avoiding unnecessary verbosity.

## Enhanced Planning Workflow

### Phase 1: Initial Understanding
Goal: Gain a comprehensive understanding of the user's request by reading through code and asking them questions.

1. Understand the user's request thoroughly
2. Explore the codebase to understand existing patterns and relevant code
3. Use AskUserQuestion tool to clarify ambiguities in the user request up front.

### Phase 2: Planning
Goal: Come up with an approach to solve the problem identified in phase 1.

- Provide any background context that may help with the task without prescribing the exact design itself
- Create a detailed plan

### Phase 3: Synthesis
Goal: Synthesize the perspectives from Phase 2, and ensure that it aligns with the user's intentions by asking them questions.

1. Collect all findings from exploration
2. Keep track of critical files that should be read before implementing the plan
3. Use AskUserQuestion to ask the user questions about trade offs.

### Phase 4: Final Plan
Once you have all the information you need, ensure that the plan file has been updated with your synthesized recommendation including:

- Recommended approach with rationale
- Key insights from different perspectives
- Critical files that need modification

### Phase 5: Call ExitPlanMode
At the very end of your turn, once you have asked the user questions and are happy with your final plan file - you should always call ExitPlanMode to indicate to the user that you are done planning.

This is critical - your turn should only end with either asking the user a question or calling ExitPlanMode. Do not stop unless it's for these 2 reasons.

NOTE: At any point in time through this workflow you should feel free to ask the user questions or clarifications. Don't make large assumptions about user intent. The goal is to present a well researched plan to the user, and tie any loose ends before implementation begins.
${SYSTEM_REMINDER_CLOSE}
`;
}

// Check if plan file exists
function planFileExists(): boolean {
  const planFilePath = permissionMode.getPlanFilePath();
  return !!planFilePath && existsSync(planFilePath);
}

// Read plan content from the plan file
function _readPlanFile(): string {
  const planFilePath = permissionMode.getPlanFilePath();
  if (!planFilePath) {
    return "No plan file path set.";
  }
  if (!existsSync(planFilePath)) {
    return `Plan file not found at ${planFilePath}`;
  }
  try {
    return readFileSync(planFilePath, "utf-8");
  } catch {
    return `Failed to read plan file at ${planFilePath}`;
  }
}

// Extract questions from AskUserQuestion tool args
function getQuestionsFromApproval(approval: ApprovalRequest) {
  const parsed = safeJsonParseOr<Record<string, unknown>>(
    approval.toolArgs,
    {},
  );
  return (
    (parsed.questions as Array<{
      question: string;
      header: string;
      options: Array<{ label: string; description: string }>;
      multiSelect: boolean;
    }>) || []
  );
}

// Parse /ralph or /yolo-ralph command arguments
function parseRalphArgs(input: string): {
  prompt: string | null;
  completionPromise: string | null | undefined; // undefined = use default, null = no promise
  maxIterations: number;
} {
  let rest = input.replace(/^\/(yolo-)?ralph\s*/, "");

  // Extract --completion-promise "value" or --completion-promise 'value'
  // Also handles --completion-promise "" or none for opt-out
  let completionPromise: string | null | undefined;
  const promiseMatch = rest.match(/--completion-promise\s+["']([^"']*)["']/);
  if (promiseMatch) {
    const val = promiseMatch[1] ?? "";
    completionPromise = val === "" || val.toLowerCase() === "none" ? null : val;
    rest = rest.replace(/--completion-promise\s+["'][^"']*["']\s*/, "");
  }

  // Extract --max-iterations N
  const maxMatch = rest.match(/--max-iterations\s+(\d+)/);
  const maxIterations = maxMatch?.[1] ? parseInt(maxMatch[1], 10) : 0;
  rest = rest.replace(/--max-iterations\s+\d+\s*/, "");

  // Remaining text is the inline prompt (may be quoted)
  const prompt = rest.trim().replace(/^["']|["']$/g, "") || null;
  return { prompt, completionPromise, maxIterations };
}

// Build Ralph first-turn reminder (when activating)
// Uses exact wording from claude-code/plugins/ralph-wiggum/scripts/setup-ralph-loop.sh
function buildRalphFirstTurnReminder(state: RalphState): string {
  const iterInfo =
    state.maxIterations > 0
      ? `${state.currentIteration}/${state.maxIterations}`
      : `${state.currentIteration}`;

  let reminder = `${SYSTEM_REMINDER_OPEN}
🔄 Ralph Wiggum mode activated (iteration ${iterInfo})
`;

  if (state.completionPromise) {
    reminder += `
═══════════════════════════════════════════════════════════
RALPH LOOP COMPLETION PROMISE
═══════════════════════════════════════════════════════════

To complete this loop, output this EXACT text:
  <promise>${state.completionPromise}</promise>

STRICT REQUIREMENTS (DO NOT VIOLATE):
  ✓ Use <promise> XML tags EXACTLY as shown above
  ✓ The statement MUST be completely and unequivocally TRUE
  ✓ Do NOT output false statements to exit the loop
  ✓ Do NOT lie even if you think you should exit

IMPORTANT - Do not circumvent the loop:
  Even if you believe you're stuck, the task is impossible,
  or you've been running too long - you MUST NOT output a
  false promise statement. The loop is designed to continue
  until the promise is GENUINELY TRUE. Trust the process.

  If the loop should stop, the promise statement will become
  true naturally. Do not force it by lying.
═══════════════════════════════════════════════════════════
`;
  } else {
    reminder += `
No completion promise set - loop runs until --max-iterations or ESC/Shift+Tab to exit.
`;
  }

  reminder += SYSTEM_REMINDER_CLOSE;
  return reminder;
}

// Build Ralph continuation reminder (on subsequent iterations)
// Exact format from claude-code/plugins/ralph-wiggum/hooks/stop-hook.sh line 160
function buildRalphContinuationReminder(state: RalphState): string {
  const iterInfo =
    state.maxIterations > 0
      ? `${state.currentIteration}/${state.maxIterations}`
      : `${state.currentIteration}`;

  if (state.completionPromise) {
    return `${SYSTEM_REMINDER_OPEN}
🔄 Ralph iteration ${iterInfo} | To stop: output <promise>${state.completionPromise}</promise> (ONLY when statement is TRUE - do not lie to exit!)
${SYSTEM_REMINDER_CLOSE}`;
  } else {
    return `${SYSTEM_REMINDER_OPEN}
🔄 Ralph iteration ${iterInfo} | No completion promise set - loop runs infinitely
${SYSTEM_REMINDER_CLOSE}`;
  }
}

function stripSystemReminders(text: string): string {
  return text
    .replace(
      new RegExp(
        `${SYSTEM_REMINDER_OPEN}[\\s\\S]*?${SYSTEM_REMINDER_CLOSE}`,
        "g",
      ),
      "",
    )
    .replace(
      new RegExp(`${SYSTEM_ALERT_OPEN}[\\s\\S]*?${SYSTEM_ALERT_CLOSE}`, "g"),
      "",
    )
    .trim();
}

function formatReflectionSettings(settings: ReflectionSettings): string {
  if (settings.trigger === "off") {
    return "Off";
  }
  const behaviorLabel =
    settings.behavior === "auto-launch" ? "auto-launch" : "reminder";
  if (settings.trigger === "compaction-event") {
    return `Compaction event (${behaviorLabel})`;
  }
  return `Step count (every ${settings.stepCount} turns, ${behaviorLabel})`;
}

const AUTO_REFLECTION_DESCRIPTION = "Reflect on recent conversations";
const AUTO_REFLECTION_PROMPT =
  "Review recent conversation history and update memory files with important information worth preserving.";

function hasActiveReflectionSubagent(): boolean {
  const snapshot = getSubagentSnapshot();
  return snapshot.agents.some(
    (agent) =>
      agent.type.toLowerCase() === "reflection" &&
      (agent.status === "pending" || agent.status === "running"),
  );
}

function buildTextParts(
  ...parts: Array<string | undefined | null>
): Array<{ type: "text"; text: string }> {
  const out: Array<{ type: "text"; text: string }> = [];
  for (const part of parts) {
    if (!part) continue;
    out.push({ type: "text", text: part });
  }
  return out;
}

// Items that have finished rendering and no longer change
type StaticItem =
  | {
      kind: "welcome";
      id: string;
      snapshot: {
        continueSession: boolean;
        agentState?: AgentState | null;
        agentProvenance?: AgentProvenance | null;
        terminalWidth: number;
      };
    }
  | {
      kind: "subagent_group";
      id: string;
      agents: Array<{
        id: string;
        type: string;
        description: string;
        status: "completed" | "error" | "running";
        toolCount: number;
        totalTokens: number;
        agentURL: string | null;
        error?: string;
      }>;
    }
  | {
      // Preview content committed early during approval to enable flicker-free UI
      // When an approval's content is tall enough to overflow the viewport,
      // we commit the preview to static and only show small approval options in dynamic
      kind: "approval_preview";
      id: string;
      toolCallId: string;
      toolName: string;
      toolArgs: string;
      // Optional precomputed/cached data for rendering
      precomputedDiff?: AdvancedDiffSuccess;
      planContent?: string; // For ExitPlanMode
      planFilePath?: string; // For ExitPlanMode
    }
  | Line;

export default function App({
  agentId: initialAgentId,
  agentState: initialAgentState,
  conversationId: initialConversationId,
  loadingState = "ready",
  continueSession = false,
  startupApproval = null,
  startupApprovals = [],
  messageHistory = [],
  resumedExistingConversation = false,
  tokenStreaming = false,
  showCompactions = false,
  agentProvenance = null,
  releaseNotes = null,
}: {
  agentId: string;
  agentState?: AgentState | null;
  conversationId: string; // Required: created at startup
  loadingState?:
    | "assembling"
    | "importing"
    | "initializing"
    | "checking"
    | "ready";
  continueSession?: boolean;
  startupApproval?: ApprovalRequest | null; // Deprecated: use startupApprovals
  startupApprovals?: ApprovalRequest[];
  messageHistory?: Message[];
  resumedExistingConversation?: boolean; // True if we explicitly resumed via --resume
  tokenStreaming?: boolean;
  showCompactions?: boolean;
  agentProvenance?: AgentProvenance | null;
  releaseNotes?: string | null; // Markdown release notes to display above header
}) {
  // Warm the model-access cache in the background so /model is fast on first open.
  useEffect(() => {
    prefetchAvailableModelHandles();
  }, []);

  // Track current agent (can change when swapping)
  const [agentId, setAgentId] = useState(initialAgentId);
  const [agentState, setAgentState] = useState(initialAgentState);

  // Helper to update agent name (updates agentState, which is the single source of truth)
  const updateAgentName = useCallback((name: string) => {
    setAgentState((prev) => (prev ? { ...prev, name } : prev));
  }, []);

  const projectDirectory = process.cwd();

  // Track current conversation (always created fresh on startup)
  const [conversationId, setConversationId] = useState(initialConversationId);

  // Keep a ref to the current agentId for use in callbacks that need the latest value
  const agentIdRef = useRef(agentId);
  useEffect(() => {
    agentIdRef.current = agentId;
    telemetry.setCurrentAgentId(agentId);
  }, [agentId]);

  // Keep a ref to the current conversationId for use in callbacks
  const conversationIdRef = useRef(conversationId);
  useEffect(() => {
    conversationIdRef.current = conversationId;
  }, [conversationId]);

  const resumeKey = useSuspend();

  // Pending conversation switch context — consumed on first message after a switch
  const pendingConversationSwitchRef = useRef<
    import("./helpers/conversationSwitchAlert").ConversationSwitchContext | null
  >(null);

  // Track previous prop values to detect actual prop changes (not internal state changes)
  const prevInitialAgentIdRef = useRef(initialAgentId);
  const prevInitialAgentStateRef = useRef(initialAgentState);
  const prevInitialConversationIdRef = useRef(initialConversationId);

  // Sync with prop changes (e.g., when parent updates from "loading" to actual ID)
  // Only sync when the PROP actually changes, not when internal state changes
  useEffect(() => {
    if (initialAgentId !== prevInitialAgentIdRef.current) {
      prevInitialAgentIdRef.current = initialAgentId;
      agentIdRef.current = initialAgentId;
      setAgentId(initialAgentId);
    }
  }, [initialAgentId]);

  useEffect(() => {
    if (initialAgentState !== prevInitialAgentStateRef.current) {
      prevInitialAgentStateRef.current = initialAgentState;
      setAgentState(initialAgentState);
    }
  }, [initialAgentState]);

  useEffect(() => {
    if (initialConversationId !== prevInitialConversationIdRef.current) {
      prevInitialConversationIdRef.current = initialConversationId;
      conversationIdRef.current = initialConversationId;
      setConversationId(initialConversationId);
    }
  }, [initialConversationId]);

  // Set agent context for tools (especially Task tool)
  useEffect(() => {
    if (agentId) {
      setCurrentAgentId(agentId);
    }
  }, [agentId]);

  // Set terminal title to "{Agent Name} | Letta Code"
  useEffect(() => {
    const title = agentState?.name
      ? `${agentState.name} | Letta Code`
      : "Letta Code";
    process.stdout.write(`\x1b]0;${title}\x07`);
  }, [agentState?.name]);

  // Whether a stream is in flight (disables input)
  // Uses synced state to keep ref in sync for reliable async checks
  const [streaming, setStreaming, streamingRef] = useSyncedState(false);
  const [networkPhase, setNetworkPhase] = useState<
    "upload" | "download" | "error" | null
  >(null);
  // Track permission mode changes for UI updates
  const [uiPermissionMode, setUiPermissionMode] = useState(
    permissionMode.getMode(),
  );
  const statusLineTriggerVersionRef = useRef(0);
  const [statusLineTriggerVersion, setStatusLineTriggerVersion] = useState(0);

  useEffect(() => {
    if (!streaming) {
      setNetworkPhase(null);
    }
  }, [streaming]);

  const triggerStatusLineRefresh = useCallback(() => {
    statusLineTriggerVersionRef.current += 1;
    setStatusLineTriggerVersion(statusLineTriggerVersionRef.current);
  }, []);

  // Guard ref for preventing concurrent processConversation calls
  // Separate from streaming state which may be set early for UI responsiveness
  // Tracks depth to allow intentional reentry while blocking parallel calls
  const processingConversationRef = useRef(0);

  // Generation counter - incremented on each ESC interrupt.
  // Allows processConversation to detect if it's been superseded.
  const conversationGenerationRef = useRef(0);

  // Whether an interrupt has been requested for the current stream
  const [interruptRequested, setInterruptRequested] = useState(false);

  // Whether a command is running (disables input but no streaming UI)
  // Uses synced state to keep ref in sync for reliable async checks
  const [commandRunning, setCommandRunning, commandRunningRef] =
    useSyncedState(false);

  // Profile load confirmation - when loading a profile and current agent is unsaved
  const [profileConfirmPending, setProfileConfirmPending] = useState<{
    name: string;
    agentId: string;
    cmdId: string;
  } | null>(null);

  // If we have approval requests, we should show the approval dialog instead of the input area
  const [pendingApprovals, setPendingApprovals] = useState<ApprovalRequest[]>(
    [],
  );
  const [approvalContexts, setApprovalContexts] = useState<ApprovalContext[]>(
    [],
  );

  // Sequential approval: track results as user reviews each approval
  const [approvalResults, setApprovalResults] = useState<
    Array<
      | { type: "approve"; approval: ApprovalRequest }
      | { type: "deny"; approval: ApprovalRequest; reason: string }
    >
  >([]);
  const [isExecutingTool, setIsExecutingTool] = useState(false);
  const [queuedApprovalResults, setQueuedApprovalResults] = useState<
    ApprovalResult[] | null
  >(null);
  const toolAbortControllerRef = useRef<AbortController | null>(null);

  // Bash mode state - track running commands for input locking and ESC cancellation
  const [bashRunning, setBashRunning] = useState(false);
  const bashAbortControllerRef = useRef<AbortController | null>(null);

  // Eager approval checking: only enabled when resuming a session (LET-7101)
  // After first successful message, we disable it since any new approvals are from our own turn
  const [needsEagerApprovalCheck, setNeedsEagerApprovalCheck] = useState(
    () => resumedExistingConversation || startupApprovals.length > 0,
  );

  // Track auto-handled results to combine with user decisions
  const [autoHandledResults, setAutoHandledResults] = useState<
    Array<{
      toolCallId: string;
      result: ToolExecutionResult;
    }>
  >([]);
  const [autoDeniedApprovals, setAutoDeniedApprovals] = useState<
    Array<{
      approval: ApprovalRequest;
      reason: string;
    }>
  >([]);
  const executingToolCallIdsRef = useRef<string[]>([]);
  const interruptQueuedRef = useRef(false);
  // Prevents interrupt handler from queueing results while approvals are in-flight.
  const toolResultsInFlightRef = useRef(false);
  const autoAllowedExecutionRef = useRef<{
    toolCallIds: string[];
    results: ApprovalResult[] | null;
    conversationId: string;
    generation: number;
  } | null>(null);
  const queuedApprovalMetadataRef = useRef<{
    conversationId: string;
    generation: number;
  } | null>(null);

  const queueApprovalResults = useCallback(
    (
      results: ApprovalResult[] | null,
      metadata?: { conversationId: string; generation: number },
    ) => {
      setQueuedApprovalResults(results);
      if (results) {
        queuedApprovalMetadataRef.current = metadata ?? {
          conversationId: conversationIdRef.current,
          generation: conversationGenerationRef.current,
        };
      } else {
        queuedApprovalMetadataRef.current = null;
      }
    },
    [],
  );

  // Bash mode: cache bash commands to prefix next user message
  // Use ref instead of state to avoid stale closure issues in onSubmit
  const bashCommandCacheRef = useRef<Array<{ input: string; output: string }>>(
    [],
  );

  // Ralph Wiggum mode: config waiting for next message to capture as prompt
  const [pendingRalphConfig, setPendingRalphConfig] = useState<{
    completionPromise: string | null | undefined;
    maxIterations: number;
    isYolo: boolean;
  } | null>(null);

  // Track ralph mode for UI updates (singleton state doesn't trigger re-renders)
  const [uiRalphActive, setUiRalphActive] = useState(
    ralphMode.getState().isActive,
  );

  // Derive current approval from pending approvals and results
  // This is the approval currently being shown to the user
  const currentApproval = pendingApprovals[approvalResults.length];
  const currentApprovalContext = approvalContexts[approvalResults.length];
  const activeApprovalId = currentApproval?.toolCallId ?? null;

  // Build Sets/Maps for three approval states (excluding the active one):
  // - pendingIds: undecided approvals (index > approvalResults.length)
  // - queuedIds: decided but not yet executed (index < approvalResults.length)
  // Used to render appropriate stubs while one approval is active
  const {
    pendingIds,
    queuedIds,
    approvalMap,
    stubDescriptions,
    queuedDecisions,
  } = useMemo(() => {
    const pending = new Set<string>();
    const queued = new Set<string>();
    const map = new Map<string, ApprovalRequest>();
    const descriptions = new Map<string, string>();
    const decisions = new Map<
      string,
      { type: "approve" | "deny"; reason?: string }
    >();

    // Helper to compute stub description - called once per approval during memo
    const computeStubDescription = (
      approval: ApprovalRequest,
    ): string | undefined => {
      try {
        const args = JSON.parse(approval.toolArgs || "{}");

        if (
          isFileEditTool(approval.toolName) ||
          isFileWriteTool(approval.toolName)
        ) {
          return args.file_path || undefined;
        }
        if (isShellTool(approval.toolName)) {
          const cmd =
            typeof args.command === "string"
              ? args.command
              : Array.isArray(args.command)
                ? args.command.join(" ")
                : "";
          return cmd.length > 50 ? `${cmd.slice(0, 50)}...` : cmd || undefined;
        }
        if (isPatchTool(approval.toolName)) {
          return "patch operation";
        }
        return undefined;
      } catch {
        return undefined;
      }
    };

    const activeIndex = approvalResults.length;

    for (let i = 0; i < pendingApprovals.length; i++) {
      const approval = pendingApprovals[i];
      if (!approval?.toolCallId || approval.toolCallId === activeApprovalId) {
        continue;
      }

      const id = approval.toolCallId;
      map.set(id, approval);

      const desc = computeStubDescription(approval);
      if (desc) {
        descriptions.set(id, desc);
      }

      if (i < activeIndex) {
        // Decided but not yet executed
        queued.add(id);
        const result = approvalResults[i];
        if (result) {
          decisions.set(id, {
            type: result.type,
            reason: result.type === "deny" ? result.reason : undefined,
          });
        }
      } else {
        // Undecided (waiting in queue)
        pending.add(id);
      }
    }

    return {
      pendingIds: pending,
      queuedIds: queued,
      approvalMap: map,
      stubDescriptions: descriptions,
      queuedDecisions: decisions,
    };
  }, [pendingApprovals, approvalResults, activeApprovalId]);

  // Overlay/selector state - only one can be open at a time
  type ActiveOverlay =
    | "model"
    | "sleeptime"
    | "toolset"
    | "system"
    | "agent"
    | "resume"
    | "conversations"
    | "search"
    | "subagent"
    | "feedback"
    | "memory"
    | "memfs-sync"
    | "pin"
    | "new"
    | "mcp"
    | "mcp-connect"
    | "help"
    | "hooks"
    | "connect"
    | null;
  const [activeOverlay, setActiveOverlay] = useState<ActiveOverlay>(null);
  const pendingOverlayCommandRef = useRef<{
    overlay: ActiveOverlay;
    command: CommandHandle;
    openingOutput: string;
    dismissOutput: string;
  } | null>(null);
  const memoryFilesystemInitializedRef = useRef(false);
  const memfsWatcherRef = useRef<ReturnType<
    typeof import("node:fs").watch
  > | null>(null);
  const memfsGitCheckInFlightRef = useRef(false);
  const pendingGitReminderRef = useRef<{
    dirty: boolean;
    aheadOfRemote: boolean;
    summary: string;
  } | null>(null);
  const [feedbackPrefill, setFeedbackPrefill] = useState("");
  const [searchQuery, setSearchQuery] = useState("");
  const [modelSelectorOptions, setModelSelectorOptions] = useState<{
    filterProvider?: string;
    forceRefresh?: boolean;
  }>({});
  const closeOverlay = useCallback(() => {
    const pending = pendingOverlayCommandRef.current;
    if (pending && pending.overlay === activeOverlay) {
      pending.command.finish(pending.dismissOutput, true);
      pendingOverlayCommandRef.current = null;
    }
    setActiveOverlay(null);
    setFeedbackPrefill("");
    setSearchQuery("");
    setModelSelectorOptions({});
  }, [activeOverlay]);

  // Queued overlay action - executed after end_turn when user makes a selection
  // while agent is busy (streaming/executing tools)
  type QueuedOverlayAction =
    | { type: "switch_agent"; agentId: string; commandId?: string }
    | { type: "switch_model"; modelId: string; commandId?: string }
    | {
        type: "set_sleeptime";
        settings: ReflectionSettings;
        commandId?: string;
      }
    | {
        type: "switch_conversation";
        conversationId: string;
        commandId?: string;
      }
    | {
        type: "switch_toolset";
        toolsetId:
          | "codex"
          | "codex_snake"
          | "default"
          | "gemini"
          | "gemini_snake"
          | "none";
        commandId?: string;
      }
    | { type: "switch_system"; promptId: string; commandId?: string }
    | null;
  const [queuedOverlayAction, setQueuedOverlayAction] =
    useState<QueuedOverlayAction>(null);

  // Pin dialog state
  const [pinDialogLocal, setPinDialogLocal] = useState(false);

  // Derived: check if any selector/overlay is open (blocks queue processing and hides input)
  const anySelectorOpen = activeOverlay !== null;

  // Other model/agent state
  const [currentSystemPromptId, setCurrentSystemPromptId] = useState<
    string | null
  >("default");
  const [currentToolset, setCurrentToolset] = useState<
    | "codex"
    | "codex_snake"
    | "default"
    | "gemini"
    | "gemini_snake"
    | "none"
    | null
  >(null);
  const [llmConfig, setLlmConfig] = useState<LlmConfig | null>(null);
  const llmConfigRef = useRef(llmConfig);
  useEffect(() => {
    llmConfigRef.current = llmConfig;
  }, [llmConfig]);
  const [currentModelId, setCurrentModelId] = useState<string | null>(null);
  // Full model handle for API calls (e.g., "anthropic/claude-sonnet-4-5-20251101")
  const [_currentModelHandle, setCurrentModelHandle] = useState<string | null>(
    null,
  );
  // Derive agentName from agentState (single source of truth)
  const agentName = agentState?.name ?? null;
  const [agentDescription, setAgentDescription] = useState<string | null>(null);
  const [agentLastRunAt, setAgentLastRunAt] = useState<string | null>(null);
  const currentModelLabel =
    llmConfig?.model_endpoint_type && llmConfig?.model
      ? `${llmConfig.model_endpoint_type}/${llmConfig.model}`
      : (llmConfig?.model ?? null);
  const currentModelDisplay = currentModelLabel
    ? (getModelShortName(currentModelLabel) ??
      currentModelLabel.split("/").pop())
    : null;
  const currentModelProvider = llmConfig?.provider_name ?? null;

  // Billing tier for conditional UI and error context (fetched once on mount)
  const [billingTier, setBillingTier] = useState<string | null>(null);

  // Update error context when model or billing tier changes
  useEffect(() => {
    setErrorContext({
      modelDisplayName: currentModelDisplay ?? undefined,
      billingTier: billingTier ?? undefined,
    });
  }, [currentModelDisplay, billingTier]);

  // Fetch billing tier once on mount
  useEffect(() => {
    (async () => {
      try {
        const settings = settingsManager.getSettings();
        const baseURL =
          process.env.LETTA_BASE_URL ||
          settings.env?.LETTA_BASE_URL ||
          "https://api.letta.com";
        const apiKey = process.env.LETTA_API_KEY || settings.env?.LETTA_API_KEY;

        const response = await fetch(`${baseURL}/v1/metadata/balance`, {
          headers: getLettaCodeHeaders(apiKey),
        });

        if (response.ok) {
          const data = (await response.json()) as { billing_tier?: string };
          if (data.billing_tier) {
            setBillingTier(data.billing_tier);
          }
        }
      } catch {
        // Silently ignore - billing tier is optional context
      }
    })();
  }, []);

  // Token streaming preference (can be toggled at runtime)
  const [tokenStreamingEnabled, setTokenStreamingEnabled] =
    useState(tokenStreaming);

  // Show compaction messages preference (can be toggled at runtime)
  const [showCompactionsEnabled, _setShowCompactionsEnabled] =
    useState(showCompactions);

  // Live, approximate token counter (resets each turn)
  const [tokenCount, setTokenCount] = useState(0);

  // Trajectory token/time bases (accumulated across runs)
  const [trajectoryTokenBase, setTrajectoryTokenBase] = useState(0);
  const [trajectoryElapsedBaseMs, setTrajectoryElapsedBaseMs] = useState(0);
  const trajectoryRunTokenStartRef = useRef(0);
  const trajectoryTokenDisplayRef = useRef(0);
  const trajectorySegmentStartRef = useRef<number | null>(null);

  // Current thinking message (rotates each turn)
  const [thinkingMessage, setThinkingMessage] = useState(
    getRandomThinkingVerb(),
  );

  // Session stats tracking
  const sessionStatsRef = useRef(new SessionStats());
  const sessionStartTimeRef = useRef(Date.now());
  const sessionHooksRanRef = useRef(false);

  const syncTrajectoryTokenBase = useCallback(() => {
    const snapshot = sessionStatsRef.current.getTrajectorySnapshot();
    setTrajectoryTokenBase(snapshot?.tokens ?? 0);
  }, []);

  const openTrajectorySegment = useCallback(() => {
    if (trajectorySegmentStartRef.current === null) {
      trajectorySegmentStartRef.current = performance.now();
      sessionStatsRef.current.startTrajectory();
    }
  }, []);

  const closeTrajectorySegment = useCallback(() => {
    const start = trajectorySegmentStartRef.current;
    if (start !== null) {
      const segmentMs = performance.now() - start;
      sessionStatsRef.current.accumulateTrajectory({ wallMs: segmentMs });
      trajectorySegmentStartRef.current = null;
    }
  }, []);

  const syncTrajectoryElapsedBase = useCallback(() => {
    const snapshot = sessionStatsRef.current.getTrajectorySnapshot();
    setTrajectoryElapsedBaseMs(snapshot?.wallMs ?? 0);
  }, []);

  const resetTrajectoryBases = useCallback(() => {
    sessionStatsRef.current.resetTrajectory();
    setTrajectoryTokenBase(0);
    setTrajectoryElapsedBaseMs(0);
    trajectoryRunTokenStartRef.current = 0;
    trajectoryTokenDisplayRef.current = 0;
    trajectorySegmentStartRef.current = null;
  }, []);

  // Wire up session stats to telemetry for safety net handlers
  useEffect(() => {
    telemetry.setSessionStatsGetter(() =>
      sessionStatsRef.current.getSnapshot(),
    );

    // Cleanup on unmount (defensive, prevents potential memory leak)
    return () => {
      telemetry.setSessionStatsGetter(undefined);
    };
  }, []);

  // Track trajectory wall time based on streaming state (matches InputRich timer)
  useEffect(() => {
    if (streaming) {
      openTrajectorySegment();
      return;
    }
    closeTrajectorySegment();
    syncTrajectoryElapsedBase();
  }, [
    streaming,
    openTrajectorySegment,
    closeTrajectorySegment,
    syncTrajectoryElapsedBase,
  ]);

  // SessionStart hook feedback to prepend to first user message
  const sessionStartFeedbackRef = useRef<string[]>([]);

  // Run SessionStart hooks when agent becomes available (not the "loading" placeholder)
  useEffect(() => {
    if (agentId && agentId !== "loading" && !sessionHooksRanRef.current) {
      sessionHooksRanRef.current = true;
      // Determine if this is a new session or resumed
      const isNewSession = !initialConversationId;
      runSessionStartHooks(
        isNewSession,
        agentId,
        agentName ?? undefined,
        conversationIdRef.current ?? undefined,
      )
        .then((result) => {
          // Store feedback to prepend to first user message
          if (result.feedback.length > 0) {
            sessionStartFeedbackRef.current = result.feedback;
          }
        })
        .catch(() => {
          // Silently ignore hook errors
        });
    }
  }, [agentId, agentName, initialConversationId]);

  // Run SessionEnd hooks helper
  const runEndHooks = useCallback(async () => {
    const durationMs = Date.now() - sessionStartTimeRef.current;
    try {
      await runSessionEndHooks(
        durationMs,
        undefined,
        undefined,
        agentIdRef.current ?? undefined,
        conversationIdRef.current ?? undefined,
      );
    } catch {
      // Silently ignore hook errors
    }
  }, []);

  // Show exit stats on exit (double Ctrl+C)
  const [showExitStats, setShowExitStats] = useState(false);

  // Track if we've sent the session context for this CLI session
  const hasSentSessionContextRef = useRef(false);

  // Track if we've set the conversation summary for this new conversation
  // Initialized to true for resumed conversations (they already have context)
  const hasSetConversationSummaryRef = useRef(resumedExistingConversation);
  // Store first user query for conversation summary
  const firstUserQueryRef = useRef<string | null>(null);

  // Track skills injection state (LET-7353)
  const discoveredSkillsRef = useRef<import("../agent/skills").Skill[] | null>(
    null,
  );
  const hasInjectedSkillsRef = useRef(false);

  // Track conversation turn count for periodic memory reminders
  const turnCountRef = useRef(0);

  // Track last notified permission mode to detect changes
  const lastNotifiedModeRef = useRef<PermissionMode>("default");

  // Static items (things that are done rendering and can be frozen)
  const [staticItems, setStaticItems] = useState<StaticItem[]>([]);

  // Track committed ids to avoid duplicates
  const emittedIdsRef = useRef<Set<string>>(new Set());

  // Guard to append welcome snapshot only once
  const welcomeCommittedRef = useRef(false);

  // AbortController for stream cancellation
  const abortControllerRef = useRef<AbortController | null>(null);

  // Track if user wants to cancel (persists across state updates)
  const userCancelledRef = useRef(false);

  // Retry counter for transient LLM API errors (ref for synchronous access in loop)
  const llmApiErrorRetriesRef = useRef(0);

  // Retry counter for 409 "conversation busy" errors
  const conversationBusyRetriesRef = useRef(0);

  // Message queue state for queueing messages during streaming
  const [messageQueue, setMessageQueue] = useState<QueuedMessage[]>([]);

  const messageQueueRef = useRef<QueuedMessage[]>([]); // For synchronous access
  useEffect(() => {
    messageQueueRef.current = messageQueue;
  }, [messageQueue]);

  // Override content parts for queued submissions (to preserve part boundaries)
  const overrideContentPartsRef = useRef<MessageCreate["content"] | null>(null);

  // Set up message queue bridge for background tasks
  // This allows non-React code (Task.ts) to add notifications to messageQueue
  useEffect(() => {
    // Provide a queue adder that adds to messageQueue and bumps dequeueEpoch
    setMessageQueueAdder((message: QueuedMessage) => {
      setMessageQueue((q) => [...q, message]);
      setDequeueEpoch((e) => e + 1);
    });
    return () => setMessageQueueAdder(null);
  }, []);

  const waitingForQueueCancelRef = useRef(false);
  const queueSnapshotRef = useRef<QueuedMessage[]>([]);
  const [restoreQueueOnCancel, setRestoreQueueOnCancel] = useState(false);
  const restoreQueueOnCancelRef = useRef(restoreQueueOnCancel);
  useEffect(() => {
    restoreQueueOnCancelRef.current = restoreQueueOnCancel;
  }, [restoreQueueOnCancel]);

  // Cache last sent input - cleared on successful completion, remains if interrupted
  const lastSentInputRef = useRef<Array<MessageCreate | ApprovalCreate> | null>(
    null,
  );
  // Non-null only when the previous turn was explicitly interrupted by the user.
  // Used to gate recovery alert injection to true user-interrupt retries.
  const pendingInterruptRecoveryConversationIdRef = useRef<string | null>(null);

  // Epoch counter to force dequeue effect re-run when refs change but state doesn't
  // Incremented when userCancelledRef is reset while messages are queued
  const [dequeueEpoch, setDequeueEpoch] = useState(0);

  // Track last dequeued message for restoration on error
  // If an error occurs after dequeue, we restore this to the input field (if input is empty)
  const lastDequeuedMessageRef = useRef<string | null>(null);

  // Restored input value - set when we need to restore a message to the input after error
  const [restoredInput, setRestoredInput] = useState<string | null>(null);

  // Helper to check if agent is busy (streaming, executing tool, or running command)
  // Uses refs for synchronous access outside React's closure system
  // biome-ignore lint/correctness/useExhaustiveDependencies: refs are stable objects, .current is read dynamically
  const isAgentBusy = useCallback(() => {
    return (
      streamingRef.current ||
      isExecutingTool ||
      commandRunningRef.current ||
      abortControllerRef.current !== null
    );
  }, [isExecutingTool]);

  const appendTaskNotificationEvents = useCallback(
    (summaries: string[]): boolean => {
      if (summaries.length === 0) return false;
      for (const summary of summaries) {
        const eventId = uid("event");
        buffersRef.current.byId.set(eventId, {
          kind: "event",
          id: eventId,
          eventType: "task_notification",
          eventData: {},
          phase: "finished",
          summary,
        });
        buffersRef.current.order.push(eventId);
      }
      return true;
    },
    [],
  );

  // Consume queued messages for appending to tool results (clears queue)
  const consumeQueuedMessages = useCallback((): QueuedMessage[] | null => {
    if (messageQueueRef.current.length === 0) return null;
    const messages = [...messageQueueRef.current];
    setMessageQueue([]);
    return messages;
  }, []);

  // Helper to wrap async handlers that need to close overlay and lock input
  // Closes overlay and sets commandRunning before executing, releases lock in finally
  const withCommandLock = useCallback(
    async (asyncFn: () => Promise<void>) => {
      setActiveOverlay(null);
      setCommandRunning(true);
      try {
        await asyncFn();
      } finally {
        setCommandRunning(false);
      }
    },
    [setCommandRunning],
  );

  // Track terminal dimensions for layout and overflow detection
  const rawColumns = useTerminalWidth();
  const terminalRows = useTerminalRows();
  const [stableColumns, setStableColumns] = useState(rawColumns);
  const stableColumnsTimeoutRef = useRef<ReturnType<typeof setTimeout> | null>(
    null,
  );
  const prevColumnsRef = useRef(rawColumns);
  const lastClearedColumnsRef = useRef(rawColumns);
  const pendingResizeRef = useRef(false);
  const pendingResizeColumnsRef = useRef<number | null>(null);
  const [staticRenderEpoch, setStaticRenderEpoch] = useState(0);
  const resizeClearTimeout = useRef<ReturnType<typeof setTimeout> | null>(null);
  const lastClearAtRef = useRef(0);
  const isInitialResizeRef = useRef(true);
  const columns = stableColumns;
  const debugFlicker = process.env.LETTA_DEBUG_FLICKER === "1";

  useEffect(() => {
    if (rawColumns === stableColumns) {
      if (stableColumnsTimeoutRef.current) {
        clearTimeout(stableColumnsTimeoutRef.current);
        stableColumnsTimeoutRef.current = null;
      }
      return;
    }

    const delta = Math.abs(rawColumns - stableColumns);
    if (delta >= MIN_RESIZE_DELTA) {
      if (stableColumnsTimeoutRef.current) {
        clearTimeout(stableColumnsTimeoutRef.current);
        stableColumnsTimeoutRef.current = null;
      }
      setStableColumns(rawColumns);
      return;
    }

    if (stableColumnsTimeoutRef.current) {
      clearTimeout(stableColumnsTimeoutRef.current);
    }
    stableColumnsTimeoutRef.current = setTimeout(() => {
      stableColumnsTimeoutRef.current = null;
      setStableColumns(rawColumns);
    }, STABLE_WIDTH_SETTLE_MS);
  }, [rawColumns, stableColumns]);

  const clearAndRemount = useCallback(
    (targetColumns: number) => {
      if (debugFlicker) {
        // eslint-disable-next-line no-console
        console.error(
          `[debug:flicker:clear-remount] target=${targetColumns} previousCleared=${lastClearedColumnsRef.current} raw=${prevColumnsRef.current}`,
        );
      }

      if (
        typeof process !== "undefined" &&
        process.stdout &&
        "write" in process.stdout &&
        process.stdout.isTTY
      ) {
        process.stdout.write(CLEAR_SCREEN_AND_HOME);
      }
      setStaticRenderEpoch((epoch) => epoch + 1);
      lastClearedColumnsRef.current = targetColumns;
      lastClearAtRef.current = Date.now();
    },
    [debugFlicker],
  );

  const scheduleResizeClear = useCallback(
    (targetColumns: number) => {
      if (targetColumns === lastClearedColumnsRef.current) {
        return;
      }

      if (resizeClearTimeout.current) {
        clearTimeout(resizeClearTimeout.current);
        resizeClearTimeout.current = null;
      }

      const elapsedSinceClear = Date.now() - lastClearAtRef.current;
      const rateLimitDelay =
        elapsedSinceClear >= MIN_CLEAR_INTERVAL_MS
          ? 0
          : MIN_CLEAR_INTERVAL_MS - elapsedSinceClear;
      const delay = Math.max(RESIZE_SETTLE_MS, rateLimitDelay);
      if (debugFlicker) {
        // eslint-disable-next-line no-console
        console.error(
          `[debug:flicker:resize-schedule] target=${targetColumns} delay=${delay}ms elapsedSinceClear=${elapsedSinceClear}ms`,
        );
      }

      resizeClearTimeout.current = setTimeout(() => {
        resizeClearTimeout.current = null;

        // If resize changed again while waiting, let the latest schedule win.
        if (prevColumnsRef.current !== targetColumns) {
          if (debugFlicker) {
            // eslint-disable-next-line no-console
            console.error(
              `[debug:flicker:resize-skip] stale target=${targetColumns} currentRaw=${prevColumnsRef.current}`,
            );
          }
          return;
        }

        if (targetColumns === lastClearedColumnsRef.current) {
          if (debugFlicker) {
            // eslint-disable-next-line no-console
            console.error(
              `[debug:flicker:resize-skip] already-cleared target=${targetColumns}`,
            );
          }
          return;
        }

        if (debugFlicker) {
          // eslint-disable-next-line no-console
          console.error(
            `[debug:flicker:resize-fire] clear target=${targetColumns}`,
          );
        }
        clearAndRemount(targetColumns);
      }, delay);
    },
    [clearAndRemount, debugFlicker],
  );

  useEffect(() => {
    const prev = prevColumnsRef.current;
    if (rawColumns === prev) return;

    // Clear pending debounced operation on any resize
    if (resizeClearTimeout.current) {
      clearTimeout(resizeClearTimeout.current);
      resizeClearTimeout.current = null;
    }

    // Skip initial mount - no clearing needed on first render
    if (isInitialResizeRef.current) {
      isInitialResizeRef.current = false;
      prevColumnsRef.current = rawColumns;
      lastClearedColumnsRef.current = rawColumns;
      return;
    }

    const delta = Math.abs(rawColumns - prev);
    const isMinorJitter = delta > 0 && delta < MIN_RESIZE_DELTA;
    if (isMinorJitter) {
      prevColumnsRef.current = rawColumns;
      return;
    }

    if (streaming) {
      // Defer clear/remount until streaming ends to avoid Ghostty flicker.
      pendingResizeRef.current = true;
      pendingResizeColumnsRef.current = rawColumns;
      prevColumnsRef.current = rawColumns;
      return;
    }

    if (rawColumns === lastClearedColumnsRef.current) {
      pendingResizeRef.current = false;
      pendingResizeColumnsRef.current = null;
      prevColumnsRef.current = rawColumns;
      return;
    }

    // Debounce to avoid flicker from rapid resize events (e.g., drag resize, Ghostty focus)
    // and keep clear frequency bounded to prevent flash storms.
    scheduleResizeClear(rawColumns);

    prevColumnsRef.current = rawColumns;
  }, [rawColumns, streaming, scheduleResizeClear]);

  useEffect(() => {
    if (streaming) {
      if (resizeClearTimeout.current) {
        clearTimeout(resizeClearTimeout.current);
        resizeClearTimeout.current = null;
        pendingResizeRef.current = true;
        pendingResizeColumnsRef.current = rawColumns;
      }
      return;
    }

    if (!pendingResizeRef.current) return;

    const pendingColumns = pendingResizeColumnsRef.current;
    pendingResizeRef.current = false;
    pendingResizeColumnsRef.current = null;

    if (pendingColumns === null) return;
    if (pendingColumns === lastClearedColumnsRef.current) return;

    scheduleResizeClear(pendingColumns);
  }, [rawColumns, streaming, scheduleResizeClear]);

  useEffect(() => {
    return () => {
      if (resizeClearTimeout.current) {
        clearTimeout(resizeClearTimeout.current);
        resizeClearTimeout.current = null;
      }
      if (stableColumnsTimeoutRef.current) {
        clearTimeout(stableColumnsTimeoutRef.current);
        stableColumnsTimeoutRef.current = null;
      }
    };
  }, []);

  const deferredToolCallCommitsRef = useRef<Map<string, number>>(new Map());
  const [deferredCommitAt, setDeferredCommitAt] = useState<number | null>(null);
  const resetDeferredToolCallCommits = useCallback(() => {
    deferredToolCallCommitsRef.current.clear();
    setDeferredCommitAt(null);
  }, []);

  // Commit immutable/finished lines into the historical log
  const commitEligibleLines = useCallback(
    (b: Buffers, opts?: { deferToolCalls?: boolean }) => {
      const deferToolCalls = opts?.deferToolCalls !== false;
      const newlyCommitted: StaticItem[] = [];
      let firstTaskIndex = -1;
      const deferredCommits = deferredToolCallCommitsRef.current;
      const now = Date.now();
      let blockedByDeferred = false;
      // If we eagerly committed a tall preview for file tools, don't also
      // commit the successful tool_call line (preview already represents it).
      const shouldSkipCommittedToolCall = (ln: Line): boolean => {
        if (ln.kind !== "tool_call") return false;
        if (!ln.toolCallId || !ln.name) return false;
        if (ln.phase !== "finished" || ln.resultOk === false) return false;
        if (!eagerCommittedPreviewsRef.current.has(ln.toolCallId)) return false;
        return (
          isFileEditTool(ln.name) ||
          isFileWriteTool(ln.name) ||
          isPatchTool(ln.name)
        );
      };
      if (!deferToolCalls && deferredCommits.size > 0) {
        deferredCommits.clear();
        setDeferredCommitAt(null);
      }

      // Check if there are any in-progress Task tool_calls
      const hasInProgress = hasInProgressTaskToolCalls(
        b.order,
        b.byId,
        emittedIdsRef.current,
      );

      // Collect finished Task tool_calls for grouping
      const finishedTaskToolCalls = collectFinishedTaskToolCalls(
        b.order,
        b.byId,
        emittedIdsRef.current,
        hasInProgress,
      );

      // Commit regular lines (non-Task tools)
      for (const id of b.order) {
        if (emittedIdsRef.current.has(id)) continue;
        const ln = b.byId.get(id);
        if (!ln) continue;
        if (
          ln.kind === "user" ||
          ln.kind === "error" ||
          ln.kind === "status" ||
          ln.kind === "trajectory_summary"
        ) {
          emittedIdsRef.current.add(id);
          newlyCommitted.push({ ...ln });
          continue;
        }
        // Events only commit when finished (they have running/finished phases)
        if (ln.kind === "event" && ln.phase === "finished") {
          emittedIdsRef.current.add(id);
          newlyCommitted.push({ ...ln });
          continue;
        }
        // Commands with phase should only commit when finished
        if (ln.kind === "command" || ln.kind === "bash_command") {
          if (!ln.phase || ln.phase === "finished") {
            emittedIdsRef.current.add(id);
            newlyCommitted.push({ ...ln });
          }
          continue;
        }
        // Handle Task tool_calls specially - track position but don't add individually
        // (unless there's no subagent data, in which case commit as regular tool call)
        if (ln.kind === "tool_call" && ln.name && isTaskTool(ln.name)) {
          if (hasInProgress && ln.toolCallId) {
            const subagent = getSubagentByToolCallId(ln.toolCallId);
            if (subagent) {
              if (firstTaskIndex === -1) {
                firstTaskIndex = newlyCommitted.length;
              }
              continue;
            }
          }
          // Check if this specific Task tool has subagent data (will be grouped)
          const hasSubagentData = finishedTaskToolCalls.some(
            (tc) => tc.lineId === id,
          );
          if (hasSubagentData) {
            // Has subagent data - will be grouped later
            if (firstTaskIndex === -1) {
              firstTaskIndex = newlyCommitted.length;
            }
            continue;
          }
          // No subagent data (e.g., backfilled from history) - commit as regular tool call
          if (ln.phase === "finished") {
            emittedIdsRef.current.add(id);
            newlyCommitted.push({ ...ln });
          }
          continue;
        }
        if ("phase" in ln && ln.phase === "finished") {
          if (shouldSkipCommittedToolCall(ln)) {
            deferredCommits.delete(id);
            emittedIdsRef.current.add(id);
            continue;
          }
          if (
            deferToolCalls &&
            ln.kind === "tool_call" &&
            (!ln.name || !isTaskTool(ln.name))
          ) {
            const commitAt = deferredCommits.get(id);
            if (commitAt === undefined) {
              const nextCommitAt = now + TOOL_CALL_COMMIT_DEFER_MS;
              deferredCommits.set(id, nextCommitAt);
              setDeferredCommitAt(nextCommitAt);
              blockedByDeferred = true;
              break;
            }
            if (commitAt > now) {
              setDeferredCommitAt(commitAt);
              blockedByDeferred = true;
              break;
            }
            deferredCommits.delete(id);
          }
          emittedIdsRef.current.add(id);
          newlyCommitted.push({ ...ln });
          // Note: We intentionally don't cleanup precomputedDiffs here because
          // the Static area renders AFTER this function returns (on next React tick),
          // and the diff needs to be available for ToolCallMessage to render.
          // The diffs will be cleaned up when the session ends or on next session start.
        }
      }

      // If we collected Task tool_calls (all are finished), create a subagent_group
      if (!blockedByDeferred && finishedTaskToolCalls.length > 0) {
        // Mark all as emitted
        for (const tc of finishedTaskToolCalls) {
          emittedIdsRef.current.add(tc.lineId);
        }

        const groupItem = createSubagentGroupItem(finishedTaskToolCalls);

        // Insert at the position of the first Task tool_call
        newlyCommitted.splice(
          firstTaskIndex >= 0 ? firstTaskIndex : newlyCommitted.length,
          0,
          groupItem,
        );

        // Clear these agents from the subagent store
        clearSubagentsByIds(groupItem.agents.map((a) => a.id));
      }

      if (deferredCommits.size === 0) {
        setDeferredCommitAt(null);
      }

      if (newlyCommitted.length > 0) {
        setStaticItems((prev) => [...prev, ...newlyCommitted]);
      }
    },
    [],
  );

  // Render-ready transcript
  const [lines, setLines] = useState<Line[]>([]);

  // Canonical buffers stored in a ref (mutated by onChunk), PERSISTED for session
  const buffersRef = useRef(createBuffers());

  // Context-window token tracking, decoupled from streaming buffers
  const contextTrackerRef = useRef(createContextTracker());

  // Track whether we've already backfilled history (should only happen once)
  const hasBackfilledRef = useRef(false);

  // Keep buffers in sync with tokenStreamingEnabled state for aggressive static promotion
  useEffect(() => {
    buffersRef.current.tokenStreamingEnabled = tokenStreamingEnabled;
  }, [tokenStreamingEnabled]);

  // Configurable status line hook
  const sessionStatsSnapshot = sessionStatsRef.current.getSnapshot();
  const contextWindowSize = llmConfigRef.current?.context_window;
  const statusLine = useConfigurableStatusLine({
    modelId: llmConfigRef.current?.model ?? null,
    modelDisplayName: currentModelDisplay,
    currentDirectory: process.cwd(),
    projectDirectory,
    sessionId: conversationId,
    agentName,
    totalDurationMs: sessionStatsSnapshot.totalWallMs,
    totalApiDurationMs: sessionStatsSnapshot.totalApiMs,
    totalInputTokens: sessionStatsSnapshot.usage.promptTokens,
    totalOutputTokens: sessionStatsSnapshot.usage.completionTokens,
    contextWindowSize,
    usedContextTokens: contextTrackerRef.current.lastContextTokens,
    permissionMode: uiPermissionMode,
    networkPhase,
    terminalWidth: columns,
    triggerVersion: statusLineTriggerVersion,
  });

  const previousStreamingForStatusLineRef = useRef(streaming);
  useEffect(() => {
    // Trigger status line when an assistant stream completes.
    if (previousStreamingForStatusLineRef.current && !streaming) {
      triggerStatusLineRefresh();
    }
    previousStreamingForStatusLineRef.current = streaming;
  }, [streaming, triggerStatusLineRefresh]);

  const statusLineRefreshIdentity = `${conversationId}|${currentModelDisplay ?? ""}|${currentModelProvider ?? ""}|${agentName ?? ""}|${columns}|${contextWindowSize ?? ""}`;

  // Trigger status line when key session identity/display state changes.
  useEffect(() => {
    void statusLineRefreshIdentity;
    triggerStatusLineRefresh();
  }, [statusLineRefreshIdentity, triggerStatusLineRefresh]);

  // Keep buffers in sync with agentId for server-side tool hooks
  useEffect(() => {
    buffersRef.current.agentId = agentState?.id;
  }, [agentState?.id]);

  // Cache precomputed diffs from approval dialogs for tool return rendering
  // Key: toolCallId or "toolCallId:filePath" for Patch operations
  const precomputedDiffsRef = useRef<Map<string, AdvancedDiffSuccess>>(
    new Map(),
  );

  // Store the last plan file path for post-approval rendering
  // (needed because plan mode is exited before rendering the result)
  const lastPlanFilePathRef = useRef<string | null>(null);

  // Track which approval tool call IDs have had their previews eagerly committed
  // This prevents double-committing when the approval changes
  const eagerCommittedPreviewsRef = useRef<Set<string>>(new Set());

  const estimateApprovalPreviewLines = useCallback(
    (approval: ApprovalRequest): number => {
      const toolName = approval.toolName;
      if (!toolName) return 0;
      const args = safeJsonParseOr<Record<string, unknown>>(
        approval.toolArgs || "{}",
        {},
      );
      const wrapWidth = Math.max(MIN_WRAP_WIDTH, columns - TEXT_WRAP_GUTTER);
      const diffWrapWidth = Math.max(
        MIN_WRAP_WIDTH,
        columns - DIFF_WRAP_GUTTER,
      );

      if (isShellTool(toolName)) {
        const t = toolName.toLowerCase();
        let command = "(no command)";
        let description = "";

        if (t === "shell") {
          const cmdVal = args.command;
          command = Array.isArray(cmdVal)
            ? cmdVal.join(" ")
            : typeof cmdVal === "string"
              ? cmdVal
              : "(no command)";
          description =
            typeof args.justification === "string" ? args.justification : "";
        } else {
          command =
            typeof args.command === "string" ? args.command : "(no command)";
          description =
            typeof args.description === "string"
              ? args.description
              : typeof args.justification === "string"
                ? args.justification
                : "";
        }

        let lines = 3; // solid line + header + blank line
        lines += countWrappedLines(command, wrapWidth);
        if (description) {
          lines += countWrappedLines(description, wrapWidth);
        }
        return lines;
      }

      if (
        isFileEditTool(toolName) ||
        isFileWriteTool(toolName) ||
        isPatchTool(toolName)
      ) {
        const headerLines = 4; // solid line + header + dotted lines
        let diffLines = 0;
        const toolCallId = approval.toolCallId;

        if (isPatchTool(toolName) && typeof args.input === "string") {
          const operations = parsePatchOperations(args.input);
          operations.forEach((op, idx) => {
            if (idx > 0) diffLines += 1; // blank line between operations
            diffLines += 1; // filename line

            const diffKey = toolCallId ? `${toolCallId}:${op.path}` : undefined;
            const opDiff =
              diffKey && precomputedDiffsRef.current.has(diffKey)
                ? precomputedDiffsRef.current.get(diffKey)
                : undefined;

            if (opDiff) {
              diffLines += estimateAdvancedDiffLines(opDiff, diffWrapWidth);
              return;
            }

            if (op.kind === "add") {
              diffLines += countWrappedLines(op.content, wrapWidth);
              return;
            }
            if (op.kind === "update") {
              if (op.patchLines?.length) {
                diffLines += countWrappedLinesFromList(
                  op.patchLines,
                  wrapWidth,
                );
              } else {
                diffLines += countWrappedLines(op.oldString || "", wrapWidth);
                diffLines += countWrappedLines(op.newString || "", wrapWidth);
              }
              return;
            }

            diffLines += 1; // delete placeholder
          });

          return headerLines + diffLines;
        }

        const diff =
          toolCallId && precomputedDiffsRef.current.has(toolCallId)
            ? precomputedDiffsRef.current.get(toolCallId)
            : undefined;

        if (diff) {
          diffLines += estimateAdvancedDiffLines(diff, diffWrapWidth);
          return headerLines + diffLines;
        }

        if (Array.isArray(args.edits)) {
          for (const edit of args.edits) {
            if (!edit || typeof edit !== "object") continue;
            const oldString =
              typeof edit.old_string === "string" ? edit.old_string : "";
            const newString =
              typeof edit.new_string === "string" ? edit.new_string : "";
            diffLines += countWrappedLines(oldString, wrapWidth);
            diffLines += countWrappedLines(newString, wrapWidth);
          }
          return headerLines + diffLines;
        }

        if (typeof args.content === "string") {
          diffLines += countWrappedLines(args.content, wrapWidth);
          return headerLines + diffLines;
        }

        const oldString =
          typeof args.old_string === "string" ? args.old_string : "";
        const newString =
          typeof args.new_string === "string" ? args.new_string : "";
        diffLines += countWrappedLines(oldString, wrapWidth);
        diffLines += countWrappedLines(newString, wrapWidth);
        return headerLines + diffLines;
      }

      return 0;
    },
    [columns],
  );

  const shouldEagerCommitApprovalPreview = useCallback(
    (approval: ApprovalRequest): boolean => {
      if (!terminalRows) return false;
      const previewLines = estimateApprovalPreviewLines(approval);
      if (previewLines === 0) return false;
      return (
        previewLines + APPROVAL_OPTIONS_HEIGHT + APPROVAL_PREVIEW_BUFFER >=
        terminalRows
      );
    },
    [estimateApprovalPreviewLines, terminalRows],
  );

  const currentApprovalShouldCommitPreview = useMemo(() => {
    if (!currentApproval) return false;
    if (currentApproval.toolName === "ExitPlanMode") return false;
    return shouldEagerCommitApprovalPreview(currentApproval);
  }, [currentApproval, shouldEagerCommitApprovalPreview]);

  // Recompute UI state from buffers after each streaming chunk
  const refreshDerived = useCallback(() => {
    const b = buffersRef.current;
    setTokenCount(b.tokenCount);
    const newLines = toLines(b);
    setLines(newLines);
    commitEligibleLines(b);
  }, [commitEligibleLines]);

  const commandRunner = useMemo(
    () =>
      createCommandRunner({
        buffersRef,
        refreshDerived,
        createId: uid,
      }),
    [refreshDerived],
  );

  const startOverlayCommand = useCallback(
    (
      overlay: ActiveOverlay,
      input: string,
      openingOutput: string,
      dismissOutput: string,
    ) => {
      const pending = pendingOverlayCommandRef.current;
      if (pending && pending.overlay === overlay) {
        pending.openingOutput = openingOutput;
        pending.dismissOutput = dismissOutput;
        return pending.command;
      }
      const command = commandRunner.start(input, openingOutput);
      pendingOverlayCommandRef.current = {
        overlay,
        command,
        openingOutput,
        dismissOutput,
      };
      return command;
    },
    [commandRunner],
  );

  const consumeOverlayCommand = useCallback((overlay: ActiveOverlay) => {
    const pending = pendingOverlayCommandRef.current;
    if (!pending || pending.overlay !== overlay) {
      return null;
    }
    pendingOverlayCommandRef.current = null;
    return pending.command;
  }, []);

  useEffect(() => {
    const pending = pendingOverlayCommandRef.current;
    if (!pending || pending.overlay !== activeOverlay) {
      return;
    }
    pending.command.update({
      output: pending.openingOutput,
      phase: "waiting",
      dimOutput: true,
    });
  }, [activeOverlay]);

  useEffect(() => {
    if (deferredCommitAt === null) return;
    const delay = Math.max(0, deferredCommitAt - Date.now());
    const timer = setTimeout(() => {
      setDeferredCommitAt(null);
      refreshDerived();
    }, delay);
    return () => clearTimeout(timer);
  }, [deferredCommitAt, refreshDerived]);

  // Trailing-edge debounce for bash streaming output (100ms = max 10 updates/sec)
  // Unlike refreshDerivedThrottled, this REPLACES pending updates to always show latest state
  const streamingRefreshTimeoutRef = useRef<ReturnType<
    typeof setTimeout
  > | null>(null);
  const refreshDerivedStreaming = useCallback(() => {
    // Cancel any pending refresh - we want the LATEST state
    if (streamingRefreshTimeoutRef.current) {
      clearTimeout(streamingRefreshTimeoutRef.current);
    }
    streamingRefreshTimeoutRef.current = setTimeout(() => {
      streamingRefreshTimeoutRef.current = null;
      if (!buffersRef.current.interrupted) {
        refreshDerived();
      }
    }, 100);
  }, [refreshDerived]);

  // Cleanup streaming refresh on unmount
  useEffect(() => {
    return () => {
      if (streamingRefreshTimeoutRef.current) {
        clearTimeout(streamingRefreshTimeoutRef.current);
      }
    };
  }, []);

  // Helper to update streaming output for bash/shell tools
  const updateStreamingOutput = useCallback(
    (toolCallId: string, chunk: string, isStderr = false) => {
      const lineId = buffersRef.current.toolCallIdToLineId.get(toolCallId);
      if (!lineId) return;

      const entry = buffersRef.current.byId.get(lineId);
      if (!entry || entry.kind !== "tool_call") return;

      // Immutable update with tail buffer
      const newStreaming = appendStreamingOutput(
        entry.streaming,
        chunk,
        entry.streaming?.startTime || Date.now(),
        isStderr,
      );

      buffersRef.current.byId.set(lineId, {
        ...entry,
        streaming: newStreaming,
      });

      refreshDerivedStreaming();
    },
    [refreshDerivedStreaming],
  );

  // Throttled version for streaming updates (~60fps max)
  const refreshDerivedThrottled = useCallback(() => {
    // Use a ref to track pending refresh
    if (!buffersRef.current.pendingRefresh) {
      buffersRef.current.pendingRefresh = true;
      // Capture the current generation to detect if resume invalidates this refresh
      const capturedGeneration = buffersRef.current.commitGeneration || 0;
      setTimeout(() => {
        buffersRef.current.pendingRefresh = false;
        // Skip refresh if stream was interrupted - prevents stale updates appearing
        // after user cancels. Normal stream completion still renders (interrupted=false).
        // Also skip if commitGeneration changed - this means a resume is in progress and
        // committing now would lock in the stale "Interrupted by user" state.
        if (
          !buffersRef.current.interrupted &&
          (buffersRef.current.commitGeneration || 0) === capturedGeneration
        ) {
          refreshDerived();
        }
      }, 16); // ~60fps
    }
  }, [refreshDerived]);

  // Restore pending approval from startup when ready
  // All approvals (including fancy UI tools) go through pendingApprovals
  // The render logic determines which UI to show based on tool name
  useEffect(() => {
    // Use new plural field if available, otherwise wrap singular in array for backward compat
    const approvals =
      startupApprovals?.length > 0
        ? startupApprovals
        : startupApproval
          ? [startupApproval]
          : [];

    if (loadingState === "ready" && approvals.length > 0) {
      // All approvals go through the same flow - UI rendering decides which dialog to show
      setPendingApprovals(approvals);

      // Analyze approval contexts for all restored approvals
      const analyzeStartupApprovals = async () => {
        try {
          const contexts = await Promise.all(
            approvals.map(async (approval) => {
              const parsedArgs = safeJsonParseOr<Record<string, unknown>>(
                approval.toolArgs,
                {},
              );
              return await analyzeToolApproval(approval.toolName, parsedArgs);
            }),
          );
          setApprovalContexts(contexts);
        } catch (error) {
          // If analysis fails, leave context as null (will show basic options)
          console.error("Failed to analyze startup approvals:", error);
        }
      };

      analyzeStartupApprovals();
    }
  }, [loadingState, startupApproval, startupApprovals]);

  // Eager commit for ExitPlanMode: Always commit plan preview to staticItems
  // This keeps the dynamic area small (just approval options) to avoid flicker
  useEffect(() => {
    if (!currentApproval) return;
    if (currentApproval.toolName !== "ExitPlanMode") return;

    const toolCallId = currentApproval.toolCallId;
    if (!toolCallId) return;

    // Already committed preview for this approval?
    if (eagerCommittedPreviewsRef.current.has(toolCallId)) return;

    const planFilePath = permissionMode.getPlanFilePath();
    if (!planFilePath) return;

    try {
      const { readFileSync, existsSync } = require("node:fs");
      if (!existsSync(planFilePath)) return;

      const planContent = readFileSync(planFilePath, "utf-8");

      // Commit preview to static area
      const previewItem: StaticItem = {
        kind: "approval_preview",
        id: `approval-preview-${toolCallId}`,
        toolCallId,
        toolName: currentApproval.toolName,
        toolArgs: currentApproval.toolArgs || "{}",
        planContent,
        planFilePath,
      };

      setStaticItems((prev) => [...prev, previewItem]);
      eagerCommittedPreviewsRef.current.add(toolCallId);

      // Also capture plan file path for post-approval rendering
      lastPlanFilePathRef.current = planFilePath;
    } catch {
      // Failed to read plan, don't commit preview
    }
  }, [currentApproval]);

  // Eager commit for large approval previews (bash/file edits) to avoid flicker
  useEffect(() => {
    if (!currentApproval) return;
    if (currentApproval.toolName === "ExitPlanMode") return;

    const toolCallId = currentApproval.toolCallId;
    if (!toolCallId) return;
    if (eagerCommittedPreviewsRef.current.has(toolCallId)) return;
    if (!currentApprovalShouldCommitPreview) return;

    const previewItem: StaticItem = {
      kind: "approval_preview",
      id: `approval-preview-${toolCallId}`,
      toolCallId,
      toolName: currentApproval.toolName,
      toolArgs: currentApproval.toolArgs || "{}",
    };

    if (
      (isFileEditTool(currentApproval.toolName) ||
        isFileWriteTool(currentApproval.toolName)) &&
      precomputedDiffsRef.current.has(toolCallId)
    ) {
      previewItem.precomputedDiff = precomputedDiffsRef.current.get(toolCallId);
    }

    setStaticItems((prev) => [...prev, previewItem]);
    eagerCommittedPreviewsRef.current.add(toolCallId);
  }, [currentApproval, currentApprovalShouldCommitPreview]);

  // Backfill message history when resuming (only once)
  useEffect(() => {
    if (
      loadingState === "ready" &&
      messageHistory.length > 0 &&
      !hasBackfilledRef.current
    ) {
      // Set flag FIRST to prevent double-execution in strict mode
      hasBackfilledRef.current = true;
      // Append welcome snapshot FIRST so it appears above history
      if (!welcomeCommittedRef.current) {
        welcomeCommittedRef.current = true;
        setStaticItems((prev) => [
          ...prev,
          {
            kind: "welcome",
            id: `welcome-${Date.now().toString(36)}`,
            snapshot: {
              continueSession,
              agentState,
              agentProvenance,
              terminalWidth: columns,
            },
          },
        ]);
      }
      // Use backfillBuffers to properly populate the transcript from history
      backfillBuffers(buffersRef.current, messageHistory);

      // Add combined status at the END so user sees it without scrolling
      const statusId = `status-resumed-${Date.now().toString(36)}`;

      // Check if agent is pinned (locally or globally)
      const isPinned = agentState?.id
        ? settingsManager.getLocalPinnedAgents().includes(agentState.id) ||
          settingsManager.getGlobalPinnedAgents().includes(agentState.id)
        : false;

      // Build status message
      const agentName = agentState?.name || "Unnamed Agent";
      const isResumingConversation =
        resumedExistingConversation || messageHistory.length > 0;
      if (process.env.DEBUG) {
        console.log(
          `[DEBUG] Header: resumedExistingConversation=${resumedExistingConversation}, messageHistory.length=${messageHistory.length}`,
        );
      }
      const headerMessage = isResumingConversation
        ? `Resuming conversation with **${agentName}**`
        : `Starting new conversation with **${agentName}**`;

      // Command hints - vary based on agent state:
      // - Resuming: show /new (they may want a fresh conversation)
      // - New session + unpinned: show /pin (they should save their agent)
      // - New session + pinned: show /memory (they're already saved)
      const commandHints = isResumingConversation
        ? [
            "→ **/agents**    list all agents",
            "→ **/resume**    browse all conversations",
            "→ **/new**       start a new conversation",
            "→ **/init**      initialize your agent's memory",
            "→ **/remember**  teach your agent",
          ]
        : isPinned
          ? [
              "→ **/agents**    list all agents",
              "→ **/resume**    resume a previous conversation",
              "→ **/memory**    view your agent's memory blocks",
              "→ **/init**      initialize your agent's memory",
              "→ **/remember**  teach your agent",
            ]
          : [
              "→ **/agents**    list all agents",
              "→ **/resume**    resume a previous conversation",
              "→ **/pin**       save + name your agent",
              "→ **/init**      initialize your agent's memory",
              "→ **/remember**  teach your agent",
            ];

      // Build status lines with optional release notes above header
      const statusLines: string[] = [];

      // Add release notes first (above everything) - same styling as rest of status block
      if (releaseNotes) {
        statusLines.push(releaseNotes);
        statusLines.push(""); // blank line separator
      }

      statusLines.push(headerMessage);
      statusLines.push(...commandHints);

      buffersRef.current.byId.set(statusId, {
        kind: "status",
        id: statusId,
        lines: statusLines,
      });
      buffersRef.current.order.push(statusId);

      refreshDerived();
      commitEligibleLines(buffersRef.current, { deferToolCalls: false });
    }
  }, [
    loadingState,
    messageHistory,
    refreshDerived,
    commitEligibleLines,
    continueSession,
    columns,
    agentState,
    agentProvenance,
    resumedExistingConversation,
    releaseNotes,
  ]);

  // Fetch llmConfig when agent is ready
  useEffect(() => {
    if (loadingState === "ready" && agentId && agentId !== "loading") {
      const fetchConfig = async () => {
        try {
          const { getClient } = await import("../agent/client");
          const client = await getClient();
          const agent = await client.agents.retrieve(agentId);
          setAgentState(agent);
          setLlmConfig(agent.llm_config);
          setAgentDescription(agent.description ?? null);
          // Get last message timestamp from agent state if available
          const lastRunCompletion = (agent as { last_run_completion?: string })
            .last_run_completion;
          setAgentLastRunAt(lastRunCompletion ?? null);

          // Derive model ID from llm_config for ModelSelector
          const agentModelHandle =
            agent.llm_config.model_endpoint_type && agent.llm_config.model
              ? `${agent.llm_config.model_endpoint_type}/${agent.llm_config.model}`
              : agent.llm_config.model;
          const { getModelInfoForLlmConfig } = await import("../agent/model");
          const modelInfo = getModelInfoForLlmConfig(
            agentModelHandle || "",
            agent.llm_config as unknown as {
              reasoning_effort?: string | null;
              enable_reasoner?: boolean | null;
            },
          );
          if (modelInfo) {
            setCurrentModelId(modelInfo.id);
          } else {
            setCurrentModelId(agentModelHandle || null);
          }
          // Store full handle for API calls (e.g., compaction)
          setCurrentModelHandle(agentModelHandle || null);

          // Derive toolset from agent's model (not persisted, computed on resume)
          if (agentModelHandle) {
            const derivedToolset = isOpenAIModel(agentModelHandle)
              ? "codex"
              : isGeminiModel(agentModelHandle)
                ? "gemini"
                : "default";
            setCurrentToolset(derivedToolset);
          }
        } catch (error) {
          console.error("Error fetching agent config:", error);
        }
      };
      fetchConfig();
    }
  }, [loadingState, agentId]);

  // Helper to append an error to the transcript
  // Also tracks the error in telemetry so we know an error was shown
  const appendError = useCallback(
    (message: string, skipTelemetry = false) => {
      // Defensive: ensure message is always a string (guards against [object Object])
      const text =
        typeof message === "string"
          ? message
          : message != null
            ? JSON.stringify(message)
            : "[Unknown error]";

      const id = uid("err");
      buffersRef.current.byId.set(id, {
        kind: "error",
        id,
        text,
      });
      buffersRef.current.order.push(id);
      refreshDerived();

      // Track error in telemetry (unless explicitly skipped for user-initiated actions)
      if (!skipTelemetry) {
        telemetry.trackError("ui_error", text, "error_display", {
          modelId: currentModelId || undefined,
        });
      }
    },
    [refreshDerived, currentModelId],
  );

  const updateMemorySyncCommand = useCallback(
    (
      commandId: string,
      output: string,
      success: boolean,
      input = "/memfs sync",
      keepRunning = false, // If true, keep phase as "running" (for conflict dialogs)
    ) => {
      buffersRef.current.byId.set(commandId, {
        kind: "command",
        id: commandId,
        input,
        output,
        phase: keepRunning ? "running" : "finished",
        success,
      });
      refreshDerived();
    },
    [refreshDerived],
  );

  const maybeCheckMemoryGitStatus = useCallback(async () => {
    // Only check if memfs is enabled for this agent
    if (!agentId || agentId === "loading") return;
    if (!settingsManager.isMemfsEnabled(agentId)) return;

    // Git-backed memory: check status periodically (fire-and-forget).
    // Runs every N turns to detect uncommitted changes or unpushed commits.
    const isIntervalTurn =
      turnCountRef.current > 0 &&
      turnCountRef.current % MEMFS_CONFLICT_CHECK_INTERVAL === 0;

    if (isIntervalTurn && !memfsGitCheckInFlightRef.current) {
      memfsGitCheckInFlightRef.current = true;

      import("../agent/memoryGit")
        .then(({ getMemoryGitStatus }) => getMemoryGitStatus(agentId))
        .then((status) => {
          pendingGitReminderRef.current =
            status.dirty || status.aheadOfRemote ? status : null;
        })
        .catch(() => {})
        .finally(() => {
          memfsGitCheckInFlightRef.current = false;
        });
    }
  }, [agentId]);

  useEffect(() => {
    if (loadingState !== "ready") {
      return;
    }
    if (!agentId || agentId === "loading") {
      return;
    }
    if (memoryFilesystemInitializedRef.current) {
      return;
    }
    // Only run startup sync if memfs is enabled for this agent
    if (!settingsManager.isMemfsEnabled(agentId)) {
      return;
    }

    memoryFilesystemInitializedRef.current = true;

    // Git-backed memory: clone or pull on startup
    (async () => {
      try {
        const { isGitRepo, cloneMemoryRepo, pullMemory } = await import(
          "../agent/memoryGit"
        );
        if (!isGitRepo(agentId)) {
          await cloneMemoryRepo(agentId);
        } else {
          await pullMemory(agentId);
        }
      } catch (err) {
        const errMsg = err instanceof Error ? err.message : String(err);
        debugWarn("memfs-git", `Startup sync failed: ${errMsg}`);
        // Warn user visually
        appendError(`Memory git sync failed: ${errMsg}`);
        // Inject reminder so the agent also knows memory isn't synced
        pendingGitReminderRef.current = {
          dirty: false,
          aheadOfRemote: false,
          summary: `Git memory sync failed on startup: ${errMsg}\nMemory may be stale. Try running: git -C ~/.letta/agents/${agentId}/memory pull`,
        };
      }
    })();
  }, [agentId, loadingState, appendError]);

  // Set up fs.watch on the memory directory to detect external file edits.
  // When a change is detected, set a dirty flag — the actual conflict check
  // runs on the next turn (debounced, non-blocking).
  useEffect(() => {
    if (!agentId || agentId === "loading") return;
    if (!settingsManager.isMemfsEnabled(agentId)) return;

    let watcher: ReturnType<typeof import("node:fs").watch> | null = null;

    (async () => {
      try {
        const { watch } = await import("node:fs");
        const { existsSync } = await import("node:fs");
        const memRoot = getMemoryFilesystemRoot(agentId);
        if (!existsSync(memRoot)) return;

        watcher = watch(memRoot, { recursive: true }, () => {
          // Git-backed memory: no auto-sync on file changes.
          // Agent handles commit/push. Status checked on interval.
        });
        memfsWatcherRef.current = watcher;
        debugLog("memfs", `Watching memory directory: ${memRoot}`);

        watcher.on("error", (err) => {
          debugWarn(
            "memfs",
            "fs.watch error (falling back to interval check)",
            err,
          );
        });
      } catch (err) {
        debugWarn(
          "memfs",
          "Failed to set up fs.watch (falling back to interval check)",
          err,
        );
      }
    })();

    return () => {
      if (watcher) {
        watcher.close();
      }
      if (memfsWatcherRef.current) {
        memfsWatcherRef.current = null;
      }
    };
  }, [agentId]);

  // Note: Old memFS conflict resolution overlay (handleMemorySyncConflictSubmit/Cancel)
  // removed. Git-backed memory uses standard git merge conflict resolution via the agent.

  // Core streaming function - iterative loop that processes conversation turns
  const processConversation = useCallback(
    async (
      initialInput: Array<MessageCreate | ApprovalCreate>,
      options?: { allowReentry?: boolean; submissionGeneration?: number },
    ): Promise<void> => {
      // Reset per-run approval tracking used by streaming UI.
      buffersRef.current.approvalsPending = false;
      if (buffersRef.current.serverToolCalls.size > 0) {
        let didPromote = false;
        for (const [toolCallId, toolInfo] of buffersRef.current
          .serverToolCalls) {
          const lineId = buffersRef.current.toolCallIdToLineId.get(toolCallId);
          if (!lineId) continue;
          const line = buffersRef.current.byId.get(lineId);
          if (!line || line.kind !== "tool_call" || line.phase === "finished") {
            continue;
          }
          const argsCandidate = toolInfo.toolArgs ?? "";
          const trimmed = argsCandidate.trim();
          let argsComplete = false;
          if (trimmed.length === 0) {
            argsComplete = true;
          } else {
            try {
              JSON.parse(argsCandidate);
              argsComplete = true;
            } catch {
              // Args still incomplete.
            }
          }
          if (argsComplete && line.phase !== "running") {
            const nextLine = {
              ...line,
              phase: "running" as const,
              argsText: line.argsText ?? argsCandidate,
            };
            buffersRef.current.byId.set(lineId, nextLine);
            didPromote = true;
          }
        }
        if (didPromote) {
          refreshDerived();
        }
      }
      // Helper function for Ralph Wiggum mode continuation
      // Defined here to have access to buffersRef, processConversation via closure
      const handleRalphContinuation = () => {
        const ralphState = ralphMode.getState();

        // Extract LAST assistant message from buffers to check for promise
        // (We only want to check the most recent response, not the entire transcript)
        const lines = toLines(buffersRef.current);
        const assistantLines = lines.filter(
          (l): l is Line & { kind: "assistant" } => l.kind === "assistant",
        );
        const lastAssistantText =
          assistantLines.length > 0
            ? (assistantLines[assistantLines.length - 1]?.text ?? "")
            : "";

        // Check for completion promise
        if (ralphMode.checkForPromise(lastAssistantText)) {
          // Promise matched - exit ralph mode
          const wasYolo = ralphState.isYolo;
          ralphMode.deactivate();
          setUiRalphActive(false);
          if (wasYolo) {
            permissionMode.setMode("default");
          }

          // Add completion status to transcript
          const statusId = uid("status");
          buffersRef.current.byId.set(statusId, {
            kind: "status",
            id: statusId,
            lines: [
              `✅ Ralph loop complete: promise detected after ${ralphState.currentIteration} iteration(s)`,
            ],
          });
          buffersRef.current.order.push(statusId);
          refreshDerived();
          return;
        }

        // Check iteration limit
        if (!ralphMode.shouldContinue()) {
          // Max iterations reached - exit ralph mode
          const wasYolo = ralphState.isYolo;
          ralphMode.deactivate();
          setUiRalphActive(false);
          if (wasYolo) {
            permissionMode.setMode("default");
          }

          // Add status to transcript
          const statusId = uid("status");
          buffersRef.current.byId.set(statusId, {
            kind: "status",
            id: statusId,
            lines: [
              `🛑 Ralph loop: Max iterations (${ralphState.maxIterations}) reached`,
            ],
          });
          buffersRef.current.order.push(statusId);
          refreshDerived();
          return;
        }

        // Continue loop - increment iteration and re-send prompt
        ralphMode.incrementIteration();
        const newState = ralphMode.getState();
        const systemMsg = buildRalphContinuationReminder(newState);

        // Re-inject original prompt with ralph reminder prepended
        // Use setTimeout to avoid blocking the current render cycle
        setTimeout(() => {
          processConversation(
            [
              {
                type: "message",
                role: "user",
                content: `${systemMsg}\n\n${newState.originalPrompt}`,
              },
            ],
            { allowReentry: true },
          );
        }, 0);
      };

      // Copy so we can safely mutate for retry recovery flows
      let currentInput = [...initialInput];
      const allowReentry = options?.allowReentry ?? false;

      // Use provided generation (from onSubmit) or capture current
      // This allows detecting if ESC was pressed during async work before this function was called
      const myGeneration =
        options?.submissionGeneration ?? conversationGenerationRef.current;

      // Check if we're already stale (ESC was pressed while we were queued in onSubmit).
      // This can happen if ESC was pressed during async work before processConversation was called.
      // We check early to avoid setting state (streaming, etc.) for stale conversations.
      if (myGeneration !== conversationGenerationRef.current) {
        return;
      }

      // Guard against concurrent processConversation calls
      // This can happen if user submits two messages in quick succession
      // Uses dedicated ref (not streamingRef) since streaming may be set early for UI responsiveness
      if (processingConversationRef.current > 0 && !allowReentry) {
        return;
      }
      processingConversationRef.current += 1;

      // Reset retry counters for new conversation turns (fresh budget per user message)
      if (!allowReentry) {
        llmApiErrorRetriesRef.current = 0;
        conversationBusyRetriesRef.current = 0;
      }

      // Track last run ID for error reporting (accessible in catch block)
      let currentRunId: string | undefined;

      try {
        // Check if user hit escape before we started
        if (userCancelledRef.current) {
          userCancelledRef.current = false; // Reset for next time
          return;
        }

        // Double-check we haven't become stale between entry and try block
        if (myGeneration !== conversationGenerationRef.current) {
          return;
        }

        setStreaming(true);
        openTrajectorySegment();
        setNetworkPhase("upload");
        abortControllerRef.current = new AbortController();

        // Recover interrupted message only after explicit user interrupt:
        // if cache contains ONLY user messages, prepend them.
        // Note: type="message" is a local discriminator (not in SDK types) to distinguish from approvals
        const originalInput = currentInput;
        const cacheIsAllUserMsgs = lastSentInputRef.current?.every(
          (m) => m.type === "message" && m.role === "user",
        );
        const canInjectInterruptRecovery =
          pendingInterruptRecoveryConversationIdRef.current !== null &&
          pendingInterruptRecoveryConversationIdRef.current ===
            conversationIdRef.current;
        if (
          cacheIsAllUserMsgs &&
          lastSentInputRef.current &&
          canInjectInterruptRecovery
        ) {
          currentInput = [
            ...lastSentInputRef.current,
            ...currentInput.map((m) =>
              m.type === "message" && m.role === "user"
                ? {
                    ...m,
                    content: [
                      { type: "text" as const, text: INTERRUPT_RECOVERY_ALERT },
                      ...(typeof m.content === "string"
                        ? [{ type: "text" as const, text: m.content }]
                        : m.content),
                    ],
                  }
                : m,
            ),
          ];
          pendingInterruptRecoveryConversationIdRef.current = null;
          // Cache old + new for chained recovery
          lastSentInputRef.current = [
            ...lastSentInputRef.current,
            ...originalInput,
          ];
        } else {
          pendingInterruptRecoveryConversationIdRef.current = null;
          lastSentInputRef.current = originalInput;
        }

        // Clear any stale pending tool calls from previous turns
        // If we're sending a new message, old pending state is no longer relevant
        // Pass false to avoid setting interrupted=true, which causes race conditions
        // with concurrent processConversation calls reading the flag
        // IMPORTANT: Skip this when allowReentry=true (continuing after tool execution)
        // because server-side tools (like memory) may still be pending and their results
        // will arrive in this stream. Cancelling them prematurely shows "Cancelled" in UI.
        if (!allowReentry) {
          markIncompleteToolsAsCancelled(
            buffersRef.current,
            false,
            "internal_cancel",
          );
        }
        // Reset interrupted flag since we're starting a fresh stream
        buffersRef.current.interrupted = false;

        // Clear completed subagents only on true new turns.
        if (
          shouldClearCompletedSubagentsOnTurnStart(
            allowReentry,
            hasActiveSubagents(),
          )
        ) {
          clearCompletedSubagents();
        }

        while (true) {
          // Capture the signal BEFORE any async operations
          // This prevents a race where handleInterrupt nulls the ref during await
          const signal = abortControllerRef.current?.signal;

          // Check if cancelled before starting new stream
          if (signal?.aborted) {
            const isStaleAtAbort =
              myGeneration !== conversationGenerationRef.current;
            // Only set streaming=false if this is the current generation.
            // If stale, a newer processConversation might be running and we shouldn't affect its UI.
            if (!isStaleAtAbort) {
              setStreaming(false);
            }
            return;
          }

          // Inject queued skill content as user message parts (LET-7353)
          // This centralizes skill content injection so all approval-send paths
          // automatically get skill SKILL.md content alongside tool results.
          {
            const { consumeQueuedSkillContent } = await import(
              "../tools/impl/skillContentRegistry"
            );
            const skillContents = consumeQueuedSkillContent();
            if (skillContents.length > 0) {
              currentInput = [
                ...currentInput,
                {
                  role: "user",
                  content: skillContents.map((sc) => ({
                    type: "text" as const,
                    text: sc.content,
                  })),
                },
              ];
            }
          }

          // Stream one turn - use ref to always get the latest conversationId
          // Wrap in try-catch to handle pre-stream desync errors (when sendMessageStream
          // throws before streaming begins, e.g., retry after LLM error when backend
          // already cleared the approval)
          let stream: Awaited<ReturnType<typeof sendMessageStream>>;
          try {
            stream = await sendMessageStream(
              conversationIdRef.current,
              currentInput,
              { agentId: agentIdRef.current },
            );
          } catch (preStreamError) {
            // Extract error detail from APIError (handles both direct and nested structures)
            // Direct: e.error.detail | Nested: e.error.error.detail (matches formatErrorDetails)
            let errorDetail = "";
            if (
              preStreamError instanceof APIError &&
              preStreamError.error &&
              typeof preStreamError.error === "object"
            ) {
              const errObj = preStreamError.error as Record<string, unknown>;
              // Check nested structure first: e.error.error.detail
              if (
                errObj.error &&
                typeof errObj.error === "object" &&
                "detail" in errObj.error
              ) {
                const nested = errObj.error as Record<string, unknown>;
                errorDetail =
                  typeof nested.detail === "string" ? nested.detail : "";
              }
              // Fallback to direct structure: e.error.detail
              if (!errorDetail && typeof errObj.detail === "string") {
                errorDetail = errObj.detail;
              }
            }
            // Final fallback: use Error.message
            if (!errorDetail && preStreamError instanceof Error) {
              errorDetail = preStreamError.message;
            }

            // Check for 409 "conversation busy" error - retry once with delay
            if (
              isConversationBusyError(errorDetail) &&
              conversationBusyRetriesRef.current < CONVERSATION_BUSY_MAX_RETRIES
            ) {
              conversationBusyRetriesRef.current += 1;
              const retryDelayMs =
                CONVERSATION_BUSY_RETRY_BASE_DELAY_MS *
                2 ** (conversationBusyRetriesRef.current - 1);

              // Show status message
              const statusId = uid("status");
              buffersRef.current.byId.set(statusId, {
                kind: "status",
                id: statusId,
                lines: ["Conversation is busy, waiting and retrying…"],
              });
              buffersRef.current.order.push(statusId);
              refreshDerived();

              // Wait with abort checking (same pattern as LLM API error retry)
              let cancelled = false;
              const startTime = Date.now();
              while (Date.now() - startTime < retryDelayMs) {
                if (
                  abortControllerRef.current?.signal.aborted ||
                  userCancelledRef.current
                ) {
                  cancelled = true;
                  break;
                }
                await new Promise((resolve) => setTimeout(resolve, 100));
              }

              // Remove status message
              buffersRef.current.byId.delete(statusId);
              buffersRef.current.order = buffersRef.current.order.filter(
                (id) => id !== statusId,
              );
              refreshDerived();

              if (!cancelled) {
                // Reset interrupted flag so retry stream chunks are processed
                buffersRef.current.interrupted = false;
                continue;
              }
              // User pressed ESC - fall through to error handling
            }

            // Reset conversation busy retry counter on non-busy error
            conversationBusyRetriesRef.current = 0;

            // Check if this is a pre-stream approval desync error
            const hasApprovalInPayload = currentInput.some(
              (item) => item?.type === "approval",
            );

            if (hasApprovalInPayload) {
              // "Invalid tool call IDs" means server HAS pending approvals but with different IDs.
              // We need to fetch the actual pending approvals and show them to the user.
              if (isInvalidToolCallIdsError(errorDetail)) {
                try {
                  const client = await getClient();
                  const agent = await client.agents.retrieve(
                    agentIdRef.current,
                  );
                  const { pendingApprovals: serverApprovals } =
                    await getResumeData(
                      client,
                      agent,
                      conversationIdRef.current,
                    );

                  if (serverApprovals && serverApprovals.length > 0) {
                    // Preserve user message from current input (if any)
                    // Filter out system reminders to avoid re-injecting them
                    const userMessage = currentInput.find(
                      (item) => item?.type === "message",
                    );
                    if (userMessage && "content" in userMessage) {
                      const content = userMessage.content;
                      let textToRestore = "";
                      if (typeof content === "string") {
                        textToRestore = stripSystemReminders(content);
                      } else if (Array.isArray(content)) {
                        // Extract text parts, filtering out system reminders
                        textToRestore = content
                          .filter(
                            (c): c is { type: "text"; text: string } =>
                              typeof c === "object" &&
                              c !== null &&
                              "type" in c &&
                              c.type === "text" &&
                              "text" in c &&
                              typeof c.text === "string" &&
                              !c.text.includes(SYSTEM_REMINDER_OPEN) &&
                              !c.text.includes(SYSTEM_ALERT_OPEN),
                          )
                          .map((c) => c.text)
                          .join("\n");
                      }
                      if (textToRestore.trim()) {
                        setRestoredInput(textToRestore);
                      }
                    }

                    // Clear all stale approval state before setting new approvals
                    setApprovalResults([]);
                    setAutoHandledResults([]);
                    setAutoDeniedApprovals([]);
                    setApprovalContexts([]);
                    queueApprovalResults(null);

                    // Set up approval UI with fetched approvals
                    setPendingApprovals(serverApprovals);

                    // Analyze approval contexts (same logic as /resume)
                    try {
                      const contexts = await Promise.all(
                        serverApprovals.map(async (approval) => {
                          const parsedArgs = safeJsonParseOr<
                            Record<string, unknown>
                          >(approval.toolArgs, {});
                          return await analyzeToolApproval(
                            approval.toolName,
                            parsedArgs,
                          );
                        }),
                      );
                      setApprovalContexts(contexts);
                    } catch {
                      // If analysis fails, contexts remain empty (will show basic options)
                    }

                    // Stop streaming and exit - user needs to approve/deny
                    // (finally block will decrement processingConversationRef)
                    setStreaming(false);
                    sendDesktopNotification("Approval needed");
                    return;
                  }
                  // No approvals found - fall through to error handling below
                } catch {
                  // Fetch failed - fall through to error handling below
                }
              }
            }

            // Not a recoverable desync - re-throw to outer catch
            throw preStreamError;
          }

          // Check again after network call - user may have pressed Escape during sendMessageStream
          if (signal?.aborted) {
            const isStaleAtAbort =
              myGeneration !== conversationGenerationRef.current;
            // Only set streaming=false if this is the current generation.
            // If stale, a newer processConversation might be running and we shouldn't affect its UI.
            if (!isStaleAtAbort) {
              setStreaming(false);
            }
            return;
          }

          // Define callback to sync agent state on first message chunk
          // This ensures the UI shows the correct model as early as possible
          const syncAgentState = async () => {
            try {
              const client = await getClient();
              const agent = await client.agents.retrieve(agentIdRef.current);

              // Keep model UI in sync with the agent configuration.
              // Note: many tiers share the same handle (e.g. gpt-5.2-none/high), so we
              // must also treat reasoning settings as model-affecting.
              const currentModel = llmConfigRef.current?.model;
              const currentEndpoint = llmConfigRef.current?.model_endpoint_type;
              const currentEffort = llmConfigRef.current?.reasoning_effort;
              const currentEnableReasoner = (
                llmConfigRef.current as unknown as {
                  enable_reasoner?: boolean | null;
                }
              )?.enable_reasoner;

              const agentModel = agent.llm_config.model;
              const agentEndpoint = agent.llm_config.model_endpoint_type;
              const agentEffort = agent.llm_config.reasoning_effort;
              const agentEnableReasoner = (
                agent.llm_config as unknown as {
                  enable_reasoner?: boolean | null;
                }
              )?.enable_reasoner;

              if (
                currentModel !== agentModel ||
                currentEndpoint !== agentEndpoint ||
                currentEffort !== agentEffort ||
                currentEnableReasoner !== agentEnableReasoner
              ) {
                // Model has changed - update local state
                setLlmConfig(agent.llm_config);

                // Derive model ID from llm_config for ModelSelector
                // Try to find matching model by handle in models.json
                const { getModelInfoForLlmConfig } = await import(
                  "../agent/model"
                );
                const agentModelHandle =
                  agent.llm_config.model_endpoint_type && agent.llm_config.model
                    ? `${agent.llm_config.model_endpoint_type}/${agent.llm_config.model}`
                    : agent.llm_config.model;

                const modelInfo = getModelInfoForLlmConfig(
                  agentModelHandle || "",
                  agent.llm_config as unknown as {
                    reasoning_effort?: string | null;
                    enable_reasoner?: boolean | null;
                  },
                );
                if (modelInfo) {
                  setCurrentModelId(modelInfo.id);
                } else {
                  // Model not in models.json (e.g., BYOK model) - use handle as ID
                  setCurrentModelId(agentModelHandle || null);
                }

                // Also update agent state if other fields changed
                setAgentState(agent);
                setAgentDescription(agent.description ?? null);
                const lastRunCompletion = (
                  agent as { last_run_completion?: string }
                ).last_run_completion;
                setAgentLastRunAt(lastRunCompletion ?? null);
              }
            } catch (error) {
              // Silently fail - don't interrupt the conversation flow
              console.error("Failed to sync agent state:", error);
            }
          };

          const handleFirstMessage = () => {
            setNetworkPhase("download");
            void syncAgentState();
          };

          const runTokenStart = buffersRef.current.tokenCount;
          trajectoryRunTokenStartRef.current = runTokenStart;
          sessionStatsRef.current.startTrajectory();

          // Only bump turn counter for actual user messages, not approval continuations.
          // This ensures all LLM steps within one user "turn" are counted as one.
          const hasUserMessage = currentInput.some(
            (item) => item.type === "message",
          );
          if (hasUserMessage) {
            contextTrackerRef.current.currentTurnId++;
          }

          const {
            stopReason,
            approval,
            approvals,
            apiDurationMs,
            lastRunId,
            fallbackError,
          } = await drainStreamWithResume(
            stream,
            buffersRef.current,
            refreshDerivedThrottled,
            signal, // Use captured signal, not ref (which may be nulled by handleInterrupt)
            handleFirstMessage,
            undefined,
            contextTrackerRef.current,
          );

          // Update currentRunId for error reporting in catch block
          currentRunId = lastRunId ?? undefined;

          // Track API duration and trajectory deltas
          sessionStatsRef.current.endTurn(apiDurationMs);
          const usageDelta = sessionStatsRef.current.updateUsageFromBuffers(
            buffersRef.current,
          );
          const tokenDelta = Math.max(
            0,
            buffersRef.current.tokenCount - runTokenStart,
          );
          sessionStatsRef.current.accumulateTrajectory({
            apiDurationMs,
            usageDelta,
            tokenDelta,
          });
          syncTrajectoryTokenBase();

          const wasInterrupted = !!buffersRef.current.interrupted;
          const wasAborted = !!signal?.aborted;
          let stopReasonToHandle = wasAborted ? "cancelled" : stopReason;

          // Check if this conversation became stale while the stream was running.
          // If stale, a newer processConversation is running and we shouldn't modify UI state.
          const isStaleAfterDrain =
            myGeneration !== conversationGenerationRef.current;

          // If this conversation is stale, exit without modifying UI state.
          // A newer conversation is running and should control the UI.
          if (isStaleAfterDrain) {
            return;
          }

          // Immediate refresh after stream completes to show final state unless
          // the user already cancelled (handleInterrupt rendered the UI).
          if (!wasInterrupted) {
            refreshDerived();
          }

          // If the turn was interrupted client-side but the backend had already emitted
          // requires_approval, treat it as a cancel. This avoids re-entering approval flow
          // and keeps queue-cancel flags consistent with the normal cancel branch below.
          if (wasInterrupted && stopReasonToHandle === "requires_approval") {
            stopReasonToHandle = "cancelled";
          }

          // Case 1: Turn ended normally
          if (stopReasonToHandle === "end_turn") {
            setStreaming(false);
            const liveElapsedMs = (() => {
              const snapshot = sessionStatsRef.current.getTrajectorySnapshot();
              const base = snapshot?.wallMs ?? 0;
              const segmentStart = trajectorySegmentStartRef.current;
              if (segmentStart === null) {
                return base;
              }
              return base + (performance.now() - segmentStart);
            })();
            closeTrajectorySegment();
            llmApiErrorRetriesRef.current = 0; // Reset retry counter on success
            conversationBusyRetriesRef.current = 0;
            lastDequeuedMessageRef.current = null; // Clear - message was processed successfully
            lastSentInputRef.current = null; // Clear - no recovery needed
            pendingInterruptRecoveryConversationIdRef.current = null;

            // Get last assistant message, user message, and reasoning for Stop hook
            const lastAssistant = Array.from(
              buffersRef.current.byId.values(),
            ).findLast((item) => item.kind === "assistant" && "text" in item);
            const assistantMessage =
              lastAssistant && "text" in lastAssistant
                ? lastAssistant.text
                : undefined;
            const lastUser = Array.from(
              buffersRef.current.byId.values(),
            ).findLast((item) => item.kind === "user" && "text" in item);
            const userMessage =
              lastUser && "text" in lastUser ? lastUser.text : undefined;
            const precedingReasoning = buffersRef.current.lastReasoning;
            buffersRef.current.lastReasoning = undefined; // Clear after use

            // Run Stop hooks - if blocked/errored, continue the conversation with feedback
            const stopHookResult = await runStopHooks(
              stopReasonToHandle,
              buffersRef.current.order.length,
              Array.from(buffersRef.current.byId.values()).filter(
                (item) => item.kind === "tool_call",
              ).length,
              undefined, // workingDirectory (uses default)
              precedingReasoning,
              assistantMessage,
              userMessage,
            );

            // If hook blocked (exit 2), inject stderr feedback and continue conversation
            if (stopHookResult.blocked) {
              const stderrOutput = stopHookResult.results
                .map((r) => r.stderr)
                .filter(Boolean)
                .join("\n");
              const feedback = stderrOutput || "Stop hook blocked";
              const hookMessage = `<stop-hook>\n${feedback}\n</stop-hook>`;

              // Add status to transcript so user sees what's happening
              const statusId = uid("status");
              buffersRef.current.byId.set(statusId, {
                kind: "status",
                id: statusId,
                lines: ["Stop hook blocked, continuing conversation."],
              });
              buffersRef.current.order.push(statusId);
              refreshDerived();

              // Continue conversation with the hook feedback
              setTimeout(() => {
                processConversation(
                  [
                    {
                      type: "message",
                      role: "user",
                      content: hookMessage,
                    },
                  ],
                  { allowReentry: true },
                );
              }, 0);
              return;
            }

            // Disable eager approval check after first successful message (LET-7101)
            // Any new approvals from here on are from our own turn, not orphaned
            if (needsEagerApprovalCheck) {
              setNeedsEagerApprovalCheck(false);
            }

            // Set conversation summary from first user query for new conversations
            if (
              !hasSetConversationSummaryRef.current &&
              firstUserQueryRef.current &&
              conversationIdRef.current !== "default"
            ) {
              hasSetConversationSummaryRef.current = true;
              const client = await getClient();
              client.conversations
                .update(conversationIdRef.current, {
                  summary: firstUserQueryRef.current,
                })
                .catch((err) => {
                  // Silently ignore - not critical
                  if (process.env.DEBUG) {
                    console.error(
                      "[DEBUG] Failed to set conversation summary:",
                      err,
                    );
                  }
                });
            }

            const trajectorySnapshot = sessionStatsRef.current.endTrajectory();
            setTrajectoryTokenBase(0);
            setTrajectoryElapsedBaseMs(0);
            trajectoryRunTokenStartRef.current = 0;
            trajectoryTokenDisplayRef.current = 0;
            if (trajectorySnapshot) {
              const summaryWallMs = Math.max(
                liveElapsedMs,
                trajectorySnapshot.wallMs,
              );
              const shouldShowSummary =
                (trajectorySnapshot.stepCount > 3 && summaryWallMs > 10000) ||
                summaryWallMs > 60000;
              if (shouldShowSummary) {
                const summaryId = uid("trajectory-summary");
                buffersRef.current.byId.set(summaryId, {
                  kind: "trajectory_summary",
                  id: summaryId,
                  durationMs: summaryWallMs,
                  stepCount: trajectorySnapshot.stepCount,
                  verb: getRandomPastTenseVerb(),
                });
                buffersRef.current.order.push(summaryId);
                refreshDerived();
              }
            }

            // Send desktop notification when turn completes
            // and we're not about to auto-send another queued message
            if (!waitingForQueueCancelRef.current) {
              sendDesktopNotification("Turn completed, awaiting your input");
            }

            // Check if we were waiting for cancel but stream finished naturally
            if (waitingForQueueCancelRef.current) {
              // Queue-cancel completed - let dequeue effect handle the messages
              // We don't call onSubmit here because isAgentBusy() would return true
              // (abortControllerRef is still set until finally block), causing re-queue
              debugLog(
                "queue",
                "Queue-cancel completed (end_turn): messages will be processed by dequeue effect",
              );
              if (restoreQueueOnCancelRef.current) {
                setRestoreQueueOnCancel(false);
              }

              // Reset flags - dequeue effect will fire when streaming=false commits
              waitingForQueueCancelRef.current = false;
              queueSnapshotRef.current = [];
            }

            await maybeCheckMemoryGitStatus();

            // === RALPH WIGGUM CONTINUATION CHECK ===
            // Check if ralph mode is active and should auto-continue
            // This happens at the very end, right before we'd release input
            if (ralphMode.getState().isActive) {
              handleRalphContinuation();
              return;
            }

            return;
          }

          // Case 1.5: Stream was cancelled by user
          if (stopReasonToHandle === "cancelled") {
            setStreaming(false);
            closeTrajectorySegment();
            syncTrajectoryElapsedBase();

            // Check if this cancel was triggered by queue threshold
            if (waitingForQueueCancelRef.current) {
              // Queue-cancel completed - let dequeue effect handle the messages
              // We don't call onSubmit here because isAgentBusy() would return true
              // (abortControllerRef is still set until finally block), causing re-queue
              debugLog(
                "queue",
                "Queue-cancel completed (cancelled): messages will be processed by dequeue effect",
              );
              if (restoreQueueOnCancelRef.current) {
                setRestoreQueueOnCancel(false);
              }

              // Reset flags - dequeue effect will fire when streaming=false commits
              waitingForQueueCancelRef.current = false;
              queueSnapshotRef.current = [];
            } else {
              // Regular user cancellation - show error
              if (!EAGER_CANCEL) {
                appendError(INTERRUPT_MESSAGE, true);
              }

              // In ralph mode, ESC interrupts but does NOT exit ralph
              // User can type additional instructions, which will get ralph prefix prepended
              // (Similar to how plan mode works)
              if (ralphMode.getState().isActive) {
                // Add status to transcript showing ralph is paused
                const statusId = uid("status");
                buffersRef.current.byId.set(statusId, {
                  kind: "status",
                  id: statusId,
                  lines: [
                    `⏸️ Ralph loop paused - type to continue or shift+tab to exit`,
                  ],
                });
                buffersRef.current.order.push(statusId);
                refreshDerived();
              }
            }

            return;
          }

          // Case 2: Requires approval
          if (stopReasonToHandle === "requires_approval") {
            // Clear stale state immediately to prevent ID mismatch bugs
            setAutoHandledResults([]);
            setAutoDeniedApprovals([]);
            lastSentInputRef.current = null; // Clear - message was received by server
            pendingInterruptRecoveryConversationIdRef.current = null;

            // Use new approvals array, fallback to legacy approval for backward compat
            const approvalsToProcess =
              approvals && approvals.length > 0
                ? approvals
                : approval
                  ? [approval]
                  : [];

            if (approvalsToProcess.length === 0) {
              appendError(
                `Unexpected empty approvals with stop reason: ${stopReason}`,
              );
              setStreaming(false);
              closeTrajectorySegment();
              syncTrajectoryElapsedBase();
              return;
            }

            // If in quietCancel mode (user queued messages), auto-reject all approvals
            // and send denials + queued messages together
            if (waitingForQueueCancelRef.current) {
              // Create denial results for all approvals
              const denialResults = approvalsToProcess.map((approvalItem) => ({
                type: "approval" as const,
                tool_call_id: approvalItem.toolCallId,
                approve: false,
                reason: "User cancelled - new message queued",
              }));

              // Update buffers to show tools as cancelled
              for (const approvalItem of approvalsToProcess) {
                onChunk(buffersRef.current, {
                  message_type: "tool_return_message",
                  id: "dummy",
                  date: new Date().toISOString(),
                  tool_call_id: approvalItem.toolCallId,
                  tool_return: "Cancelled - user sent new message",
                  status: "error",
                });
              }
              refreshDerived();

              // Queue denial results - dequeue effect will pick them up via onSubmit
              queueApprovalResults(denialResults);

              debugLog(
                "queue",
                `Queue-cancel completed (requires_approval): ${denialResults.length} denial(s) queued, messages will be processed by dequeue effect`,
              );

              if (restoreQueueOnCancelRef.current) {
                setRestoreQueueOnCancel(false);
              }

              // Reset flags - dequeue effect will fire when streaming=false commits
              waitingForQueueCancelRef.current = false;
              queueSnapshotRef.current = [];
              setStreaming(false);
              closeTrajectorySegment();
              syncTrajectoryElapsedBase();
              return;
            }

            // Check if user cancelled before starting permission checks
            if (
              userCancelledRef.current ||
              abortControllerRef.current?.signal.aborted
            ) {
              setStreaming(false);
              closeTrajectorySegment();
              syncTrajectoryElapsedBase();
              markIncompleteToolsAsCancelled(
                buffersRef.current,
                true,
                "user_interrupt",
              );
              refreshDerived();
              return;
            }

            // Check permissions for all approvals (including fancy UI tools)
            const { needsUserInput, autoAllowed, autoDenied } =
              await classifyApprovals(approvalsToProcess, {
                getContext: analyzeToolApproval,
                alwaysRequiresUserInput,
                missingNameReason:
                  "Tool call incomplete - missing name or arguments",
              });

            // Precompute diffs for file edit tools before execution (both auto-allowed and needs-user-input)
            // This is needed for inline approval UI to show diffs, and for post-approval rendering
            for (const ac of [...autoAllowed, ...needsUserInput]) {
              const toolName = ac.approval.toolName;
              const toolCallId = ac.approval.toolCallId;
              try {
                const args = JSON.parse(ac.approval.toolArgs || "{}");

                if (isFileWriteTool(toolName)) {
                  const filePath = args.file_path as string | undefined;
                  if (filePath) {
                    const result = computeAdvancedDiff({
                      kind: "write",
                      filePath,
                      content: (args.content as string) || "",
                    });
                    if (result.mode === "advanced") {
                      precomputedDiffsRef.current.set(toolCallId, result);
                    }
                  }
                } else if (isFileEditTool(toolName)) {
                  const filePath = args.file_path as string | undefined;
                  if (filePath) {
                    // Check if it's a multi-edit (has edits array) or single edit
                    if (args.edits && Array.isArray(args.edits)) {
                      const result = computeAdvancedDiff({
                        kind: "multi_edit",
                        filePath,
                        edits: args.edits as Array<{
                          old_string: string;
                          new_string: string;
                          replace_all?: boolean;
                        }>,
                      });
                      if (result.mode === "advanced") {
                        precomputedDiffsRef.current.set(toolCallId, result);
                      }
                    } else {
                      const result = computeAdvancedDiff({
                        kind: "edit",
                        filePath,
                        oldString: (args.old_string as string) || "",
                        newString: (args.new_string as string) || "",
                        replaceAll: args.replace_all as boolean | undefined,
                      });
                      if (result.mode === "advanced") {
                        precomputedDiffsRef.current.set(toolCallId, result);
                      }
                    }
                  }
                } else if (isPatchTool(toolName) && args.input) {
                  // Patch tools - parse hunks directly (patches ARE diffs)
                  const operations = parsePatchOperations(args.input as string);
                  for (const op of operations) {
                    const key = `${toolCallId}:${op.path}`;
                    if (op.kind === "add" || op.kind === "update") {
                      const result = parsePatchToAdvancedDiff(
                        op.patchLines,
                        op.path,
                      );
                      if (result) {
                        precomputedDiffsRef.current.set(key, result);
                      }
                    }
                    // Delete operations don't need diffs
                  }
                }
              } catch {
                // Ignore errors in diff computation for auto-allowed tools
              }
            }

            const autoAllowedToolCallIds = autoAllowed.map(
              (ac) => ac.approval.toolCallId,
            );
            const autoAllowedAbortController =
              abortControllerRef.current ?? new AbortController();
            const shouldTrackAutoAllowed = autoAllowedToolCallIds.length > 0;
            let autoAllowedResults: Array<{
              toolCallId: string;
              result: ToolExecutionResult;
            }> = [];
            let autoDeniedResults: Array<{
              approval: ApprovalRequest;
              reason: string;
            }> = [];

            if (shouldTrackAutoAllowed) {
              setIsExecutingTool(true);
              executingToolCallIdsRef.current = autoAllowedToolCallIds;
              toolAbortControllerRef.current = autoAllowedAbortController;
              autoAllowedExecutionRef.current = {
                toolCallIds: autoAllowedToolCallIds,
                results: null,
                conversationId: conversationIdRef.current,
                generation: conversationGenerationRef.current,
              };
            }

            try {
              if (autoAllowedToolCallIds.length > 0) {
                // Set phase to "running" for auto-allowed tools
                setToolCallsRunning(buffersRef.current, autoAllowedToolCallIds);
                refreshDerived();
              }

              // Execute auto-allowed tools (sequential for writes, parallel for reads)
              autoAllowedResults =
                autoAllowed.length > 0
                  ? await executeAutoAllowedTools(
                      autoAllowed,
                      (chunk) => onChunk(buffersRef.current, chunk),
                      {
                        abortSignal: autoAllowedAbortController.signal,
                        onStreamingOutput: updateStreamingOutput,
                      },
                    )
                  : [];

              // Create denial results for auto-denied tools and update buffers
              autoDeniedResults = autoDenied.map((ac) => {
                // Prefer the detailed reason over the short matchedRule name
                // (e.g., reason contains plan file path info, matchedRule is just "plan mode")
                const reason = ac.permission.reason
                  ? `Permission denied: ${ac.permission.reason}`
                  : "matchedRule" in ac.permission && ac.permission.matchedRule
                    ? `Permission denied by rule: ${ac.permission.matchedRule}`
                    : "Permission denied: Unknown reason";

                // Update buffers with tool rejection for UI
                onChunk(buffersRef.current, {
                  message_type: "tool_return_message",
                  id: "dummy",
                  date: new Date().toISOString(),
                  tool_call_id: ac.approval.toolCallId,
                  tool_return: `Error: request to call tool denied. User reason: ${reason}`,
                  status: "error",
                  stdout: null,
                  stderr: null,
                });

                return {
                  approval: ac.approval,
                  reason,
                };
              });

              const allResults = [
                ...autoAllowedResults.map((ar) => ({
                  type: "tool" as const,
                  tool_call_id: ar.toolCallId,
                  tool_return: ar.result.toolReturn,
                  status: ar.result.status,
                  stdout: ar.result.stdout,
                  stderr: ar.result.stderr,
                })),
                ...autoDeniedResults.map((ad) => ({
                  type: "approval" as const,
                  tool_call_id: ad.approval.toolCallId,
                  approve: false,
                  reason: ad.reason,
                })),
              ];

              if (autoAllowedExecutionRef.current) {
                autoAllowedExecutionRef.current.results = allResults;
              }
              const autoAllowedMetadata = autoAllowedExecutionRef.current
                ? {
                    conversationId:
                      autoAllowedExecutionRef.current.conversationId,
                    generation: conversationGenerationRef.current,
                  }
                : undefined;

              // If all are auto-handled, continue immediately without showing dialog
              if (needsUserInput.length === 0) {
                // Check if user cancelled before continuing
                if (
                  userCancelledRef.current ||
                  abortControllerRef.current?.signal.aborted ||
                  interruptQueuedRef.current
                ) {
                  if (allResults.length > 0) {
                    queueApprovalResults(allResults, autoAllowedMetadata);
                  }
                  setStreaming(false);
                  closeTrajectorySegment();
                  syncTrajectoryElapsedBase();
                  markIncompleteToolsAsCancelled(
                    buffersRef.current,
                    true,
                    "user_interrupt",
                  );
                  refreshDerived();
                  return;
                }

                // Append queued messages if any (from 15s append mode)
                const queuedItemsToAppend = consumeQueuedMessages();
                const queuedNotifications = queuedItemsToAppend
                  ? getQueuedNotificationSummaries(queuedItemsToAppend)
                  : [];
                const hadNotifications =
                  appendTaskNotificationEvents(queuedNotifications);
                const queuedUserText = queuedItemsToAppend
                  ? buildQueuedUserText(queuedItemsToAppend)
                  : "";

                if (queuedUserText) {
                  const userId = uid("user");
                  buffersRef.current.byId.set(userId, {
                    kind: "user",
                    id: userId,
                    text: queuedUserText,
                  });
                  buffersRef.current.order.push(userId);
                }

                if (queuedItemsToAppend && queuedItemsToAppend.length > 0) {
                  const queuedContentParts =
                    buildQueuedContentParts(queuedItemsToAppend);
                  setThinkingMessage(getRandomThinkingVerb());
                  refreshDerived();
                  toolResultsInFlightRef.current = true;
                  await processConversation(
                    [
                      { type: "approval", approvals: allResults },
                      {
                        type: "message",
                        role: "user",
                        content: queuedContentParts,
                      },
                    ],
                    { allowReentry: true },
                  );
                  toolResultsInFlightRef.current = false;
                  return;
                }
                if (hadNotifications || queuedUserText.length > 0) {
                  refreshDerived();
                }

                // Cancel mode - queue results and let dequeue effect handle
                if (waitingForQueueCancelRef.current) {
                  // Queue results - dequeue effect will pick them up via onSubmit
                  if (allResults.length > 0) {
                    queueApprovalResults(allResults, autoAllowedMetadata);
                  }

                  debugLog(
                    "queue",
                    `Queue-cancel completed (auto-allowed): ${allResults.length} result(s) queued, messages will be processed by dequeue effect`,
                  );

                  if (restoreQueueOnCancelRef.current) {
                    setRestoreQueueOnCancel(false);
                  }

                  // Reset flags - dequeue effect will fire when streaming=false commits
                  waitingForQueueCancelRef.current = false;
                  queueSnapshotRef.current = [];
                  setStreaming(false);
                  closeTrajectorySegment();
                  syncTrajectoryElapsedBase();
                  return;
                }

                setThinkingMessage(getRandomThinkingVerb());
                refreshDerived();

                toolResultsInFlightRef.current = true;
                await processConversation(
                  [
                    {
                      type: "approval",
                      approvals: allResults,
                    },
                  ],
                  { allowReentry: true },
                );
                toolResultsInFlightRef.current = false;
                return;
              }

              // Check again if user queued messages during auto-allowed tool execution
              if (waitingForQueueCancelRef.current) {
                // Create denial results for tools that need user input
                const denialResults = needsUserInput.map((ac) => ({
                  type: "approval" as const,
                  tool_call_id: ac.approval.toolCallId,
                  approve: false,
                  reason: "User cancelled - new message queued",
                }));

                // Update buffers to show tools as cancelled
                for (const ac of needsUserInput) {
                  onChunk(buffersRef.current, {
                    message_type: "tool_return_message",
                    id: "dummy",
                    date: new Date().toISOString(),
                    tool_call_id: ac.approval.toolCallId,
                    tool_return: "Cancelled - user sent new message",
                    status: "error",
                  });
                }
                refreshDerived();

                // Combine with auto-handled results and queue for sending
                const queuedResults = [...allResults, ...denialResults];
                if (queuedResults.length > 0) {
                  queueApprovalResults(queuedResults, autoAllowedMetadata);
                }

                debugLog(
                  "queue",
                  `Queue-cancel completed (auto-allowed+approvals): ${queuedResults.length} result(s) queued, messages will be processed by dequeue effect`,
                );

                if (restoreQueueOnCancelRef.current) {
                  setRestoreQueueOnCancel(false);
                }

                // Reset flags - dequeue effect will fire when streaming=false commits
                waitingForQueueCancelRef.current = false;
                queueSnapshotRef.current = [];
                setStreaming(false);
                closeTrajectorySegment();
                syncTrajectoryElapsedBase();
                return;
              }
            } finally {
              if (shouldTrackAutoAllowed) {
                setIsExecutingTool(false);
                toolAbortControllerRef.current = null;
                executingToolCallIdsRef.current = [];
                autoAllowedExecutionRef.current = null;
                toolResultsInFlightRef.current = false;
              }
            }

            // Check if user cancelled before showing dialog
            if (
              userCancelledRef.current ||
              abortControllerRef.current?.signal.aborted
            ) {
              setStreaming(false);
              closeTrajectorySegment();
              syncTrajectoryElapsedBase();
              markIncompleteToolsAsCancelled(
                buffersRef.current,
                true,
                "user_interrupt",
              );
              refreshDerived();
              return;
            }

            // Show approval dialog for tools that need user input
            setPendingApprovals(needsUserInput.map((ac) => ac.approval));
            setApprovalContexts(
              needsUserInput
                .map((ac) => ac.context)
                .filter((ctx): ctx is ApprovalContext => ctx !== null),
            );
            setAutoHandledResults(autoAllowedResults);
            setAutoDeniedApprovals(autoDeniedResults);
            setStreaming(false);
            closeTrajectorySegment();
            syncTrajectoryElapsedBase();
            // Notify user that approval is needed
            sendDesktopNotification("Approval needed");
            return;
          }

          // Unexpected stop reason (error, llm_api_error, etc.)
          // Cache desync detection and last failure for consistent handling
          // Check if payload contains approvals (could be approval-only or mixed with user message)
          const hasApprovalInPayload = currentInput.some(
            (item) => item?.type === "approval",
          );

          // Capture the most recent error text in this turn (if any)
          let latestErrorText: string | null = null;
          for (let i = buffersRef.current.order.length - 1; i >= 0; i -= 1) {
            const id = buffersRef.current.order[i];
            if (!id) continue;
            const entry = buffersRef.current.byId.get(id);
            if (entry?.kind === "error" && typeof entry.text === "string") {
              latestErrorText = entry.text;
              break;
            }
          }

          // Check for "Invalid tool call IDs" error - server HAS pending approvals but with different IDs.
          // Fetch the actual pending approvals and show them to the user.
          const detailFromRun = await fetchRunErrorDetail(lastRunId);
          const invalidIdsDetected =
            isInvalidToolCallIdsError(detailFromRun) ||
            isInvalidToolCallIdsError(latestErrorText);

          if (hasApprovalInPayload && invalidIdsDetected) {
            try {
              const client = await getClient();
              const agent = await client.agents.retrieve(agentIdRef.current);
              const { pendingApprovals: serverApprovals } = await getResumeData(
                client,
                agent,
                conversationIdRef.current,
              );

              if (serverApprovals && serverApprovals.length > 0) {
                // Preserve user message from current input (if any)
                // Filter out system reminders to avoid re-injecting them
                const userMessage = currentInput.find(
                  (item) => item?.type === "message",
                );
                if (userMessage && "content" in userMessage) {
                  const content = userMessage.content;
                  let textToRestore = "";
                  if (typeof content === "string") {
                    textToRestore = stripSystemReminders(content);
                  } else if (Array.isArray(content)) {
                    // Extract text parts, filtering out system reminders
                    textToRestore = content
                      .filter(
                        (c): c is { type: "text"; text: string } =>
                          typeof c === "object" &&
                          c !== null &&
                          "type" in c &&
                          c.type === "text" &&
                          "text" in c &&
                          typeof c.text === "string" &&
                          !c.text.includes(SYSTEM_REMINDER_OPEN) &&
                          !c.text.includes(SYSTEM_ALERT_OPEN),
                      )
                      .map((c) => c.text)
                      .join("\n");
                  }
                  if (textToRestore.trim()) {
                    setRestoredInput(textToRestore);
                  }
                }

                // Clear all stale approval state before setting new approvals
                setApprovalResults([]);
                setAutoHandledResults([]);
                setAutoDeniedApprovals([]);
                setApprovalContexts([]);
                queueApprovalResults(null);

                // Set up approval UI with fetched approvals
                setPendingApprovals(serverApprovals);

                // Analyze approval contexts
                try {
                  const contexts = await Promise.all(
                    serverApprovals.map(async (approval) => {
                      const parsedArgs = safeJsonParseOr<
                        Record<string, unknown>
                      >(approval.toolArgs, {});
                      return await analyzeToolApproval(
                        approval.toolName,
                        parsedArgs,
                      );
                    }),
                  );
                  setApprovalContexts(contexts);
                } catch {
                  // If analysis fails, contexts remain empty (will show basic options)
                }

                // Stop streaming and exit - user needs to approve/deny
                // (finally block will decrement processingConversationRef)
                setStreaming(false);
                sendDesktopNotification("Approval needed");
                return;
              }
              // No approvals found - fall through to error handling below
            } catch {
              // Fetch failed - fall through to error handling below
            }
          }

          // Check for approval pending error (sent user message while approval waiting)
          // This is the lazy recovery path for when needsEagerApprovalCheck is false
          const approvalPendingDetected =
            isApprovalPendingError(detailFromRun) ||
            isApprovalPendingError(latestErrorText);

          if (
            !hasApprovalInPayload &&
            approvalPendingDetected &&
            llmApiErrorRetriesRef.current < LLM_API_ERROR_MAX_RETRIES
          ) {
            llmApiErrorRetriesRef.current += 1;

            try {
              // Fetch pending approvals and auto-deny them
              const client = await getClient();
              const agent = await client.agents.retrieve(agentIdRef.current);
              const { pendingApprovals: existingApprovals } =
                await getResumeData(client, agent, conversationIdRef.current);

              if (existingApprovals && existingApprovals.length > 0) {
                // Create denial results for all stale approvals
                // Use the same format as handleCancelApprovals (lines 6390-6395)
                const denialResults = existingApprovals.map((approval) => ({
                  type: "approval" as const,
                  tool_call_id: approval.toolCallId,
                  approve: false,
                  reason:
                    "Auto-denied: stale approval from interrupted session",
                }));

                // Prepend approval denials to the current input (keeps user message)
                const approvalPayload: ApprovalCreate = {
                  type: "approval",
                  approvals: denialResults,
                };
                currentInput.unshift(approvalPayload);
              }
              // else: No approvals found - server state may have cleared, just retry
            } catch {
              // If we can't fetch approvals, just retry the original message
            }

            // Reset interrupted flag so retry stream chunks are processed
            buffersRef.current.interrupted = false;
            continue;
          }

          // Check if this is a retriable error (transient LLM API error)
          const retriable = await isRetriableError(
            stopReasonToHandle,
            lastRunId,
          );

          if (
            retriable &&
            llmApiErrorRetriesRef.current < LLM_API_ERROR_MAX_RETRIES
          ) {
            llmApiErrorRetriesRef.current += 1;
            const attempt = llmApiErrorRetriesRef.current;
            const delayMs = 1000 * 2 ** (attempt - 1); // 1s, 2s, 4s

            // Show subtle grey status message
            const statusId = uid("status");
            const statusLines = [
              "Unexpected downstream LLM API error, retrying...",
            ];
            buffersRef.current.byId.set(statusId, {
              kind: "status",
              id: statusId,
              lines: statusLines,
            });
            buffersRef.current.order.push(statusId);
            refreshDerived();

            // Wait before retry (check abort signal periodically for ESC cancellation)
            let cancelled = false;
            const startTime = Date.now();
            while (Date.now() - startTime < delayMs) {
              if (
                abortControllerRef.current?.signal.aborted ||
                userCancelledRef.current
              ) {
                cancelled = true;
                break;
              }
              await new Promise((resolve) => setTimeout(resolve, 100)); // Check every 100ms
            }

            // Remove status message
            buffersRef.current.byId.delete(statusId);
            buffersRef.current.order = buffersRef.current.order.filter(
              (id) => id !== statusId,
            );
            refreshDerived();

            if (!cancelled) {
              // Reset interrupted flag so retry stream chunks are processed
              buffersRef.current.interrupted = false;
              // Retry by continuing the while loop (same currentInput)
              continue;
            }
            // User pressed ESC - fall through to error handling
          }

          // Reset retry counters on non-retriable error (or max retries exceeded)
          llmApiErrorRetriesRef.current = 0;
          conversationBusyRetriesRef.current = 0;

          // Mark incomplete tool calls as finished to prevent stuck blinking UI
          markIncompleteToolsAsCancelled(
            buffersRef.current,
            true,
            "stream_error",
          );

          // Track the error in telemetry
          telemetry.trackError(
            fallbackError
              ? "FallbackError"
              : stopReasonToHandle || "unknown_stop_reason",
            fallbackError ||
              `Stream stopped with reason: ${stopReasonToHandle}`,
            "message_stream",
            {
              modelId: currentModelId || undefined,
              runId: lastRunId ?? undefined,
            },
          );

          // If we have a client-side stream error (e.g., JSON parse error), show it directly
          // Fallback error: no run_id available, show whatever error message we have
          if (fallbackError) {
            setNetworkPhase("error");
            const errorMsg = lastRunId
              ? `Stream error: ${fallbackError}\n(run_id: ${lastRunId})`
              : `Stream error: ${fallbackError}`;
            appendError(errorMsg, true); // Skip telemetry - already tracked above
            appendError(ERROR_FEEDBACK_HINT, true);

            // Restore dequeued message to input on error
            if (lastDequeuedMessageRef.current) {
              setRestoredInput(lastDequeuedMessageRef.current);
              lastDequeuedMessageRef.current = null;
            }
            // Clear any remaining queue on error
            setMessageQueue([]);

            setStreaming(false);
            sendDesktopNotification("Stream error", "error"); // Notify user of error
            refreshDerived();
            resetTrajectoryBases();
            return;
          }

          // Fetch error details from the run if available (server-side errors)
          if (lastRunId) {
            try {
              const client = await getClient();
              const run = await client.runs.retrieve(lastRunId);

              // Check if run has error information in metadata
              if (run.metadata?.error) {
                const errorData = run.metadata.error as {
                  type?: string;
                  message?: string;
                  detail?: string;
                };

                // Pass structured error data to our formatter
                const errorObject = {
                  error: {
                    error: errorData,
                    run_id: lastRunId,
                  },
                };
                const errorDetails = formatErrorDetails(
                  errorObject,
                  agentIdRef.current,
                );

                // Encrypted content errors are self-explanatory (include /clear advice)
                // — skip the generic "Something went wrong?" hint
                appendError(errorDetails, true); // Skip telemetry - already tracked above

                if (!isEncryptedContentError(errorObject)) {
                  // Show appropriate error hint based on stop reason
                  appendError(
                    getErrorHintForStopReason(
                      stopReasonToHandle,
                      currentModelId,
                    ),
                    true,
                  );
                }
              } else {
                // No error metadata, show generic error with run info
                appendError(
                  `An error occurred during agent execution\n(run_id: ${lastRunId}, stop_reason: ${stopReason})`,
                  true, // Skip telemetry - already tracked above
                );

                // Show appropriate error hint based on stop reason
                appendError(
                  getErrorHintForStopReason(stopReasonToHandle, currentModelId),
                  true,
                );
              }
            } catch (_e) {
              // If we can't fetch error details, show generic error
              appendError(
                `An error occurred during agent execution\n(run_id: ${lastRunId}, stop_reason: ${stopReason})\n(Unable to fetch additional error details from server)`,
                true, // Skip telemetry - already tracked above
              );

              // Show appropriate error hint based on stop reason
              appendError(
                getErrorHintForStopReason(stopReasonToHandle, currentModelId),
                true,
              );

              // Restore dequeued message to input on error
              if (lastDequeuedMessageRef.current) {
                setRestoredInput(lastDequeuedMessageRef.current);
                lastDequeuedMessageRef.current = null;
              }
              // Clear any remaining queue on error
              setMessageQueue([]);

              setStreaming(false);
              sendDesktopNotification();
              refreshDerived();
              resetTrajectoryBases();
              return;
            }
          } else {
            // No run_id available - but this is unusual since errors should have run_ids
            appendError(
              `An error occurred during agent execution\n(stop_reason: ${stopReason})`,
              true, // Skip telemetry - already tracked above
            );

            // Show appropriate error hint based on stop reason
            appendError(
              getErrorHintForStopReason(stopReasonToHandle, currentModelId),
              true,
            );
          }

          // Restore dequeued message to input on error
          if (lastDequeuedMessageRef.current) {
            setRestoredInput(lastDequeuedMessageRef.current);
            lastDequeuedMessageRef.current = null;
          }
          // Clear any remaining queue on error
          setMessageQueue([]);

          setStreaming(false);
          sendDesktopNotification("Execution error", "error"); // Notify user of error
          refreshDerived();
          resetTrajectoryBases();
          return;
        }
      } catch (e) {
        // Mark incomplete tool calls as cancelled to prevent stuck blinking UI
        markIncompleteToolsAsCancelled(
          buffersRef.current,
          true,
          e instanceof APIUserAbortError ? "user_interrupt" : "stream_error",
        );

        // If using eager cancel and this is an abort error, silently ignore it
        // The user already got "Stream interrupted by user" feedback from handleInterrupt
        if (EAGER_CANCEL && e instanceof APIUserAbortError) {
          setStreaming(false);
          refreshDerived();
          return;
        }

        // Track error with enhanced context
        const errorType =
          e instanceof Error ? e.constructor.name : "UnknownError";
        const errorMessage = e instanceof Error ? e.message : String(e);

        // Extract HTTP status code if available (API errors often have this)
        const httpStatus =
          e &&
          typeof e === "object" &&
          "status" in e &&
          typeof e.status === "number"
            ? e.status
            : undefined;

        telemetry.trackError(errorType, errorMessage, "message_stream", {
          httpStatus,
          modelId: currentModelId || undefined,
          runId: currentRunId,
        });

        // Use comprehensive error formatting
        const errorDetails = formatErrorDetails(e, agentIdRef.current);
        appendError(errorDetails, true); // Skip telemetry - already tracked above with more context
        appendError(ERROR_FEEDBACK_HINT, true);

        // Restore dequeued message to input on error (Input component will only use if empty)
        if (lastDequeuedMessageRef.current) {
          setRestoredInput(lastDequeuedMessageRef.current);
          lastDequeuedMessageRef.current = null;
        }
        // Clear any remaining queue on error
        setMessageQueue([]);

        setStreaming(false);
        sendDesktopNotification("Processing error", "error"); // Notify user of error
        refreshDerived();
        resetTrajectoryBases();
      } finally {
        // Check if this conversation was superseded by an ESC interrupt
        const isStale = myGeneration !== conversationGenerationRef.current;

        abortControllerRef.current = null;

        // Trigger dequeue effect now that processConversation is no longer active.
        // The dequeue effect checks abortControllerRef (a ref, not state), so it
        // won't re-run on its own — bump dequeueEpoch to force re-evaluation.
        // Only bump for normal completions — if stale (ESC was pressed), the user
        // cancelled and queued messages should NOT be auto-submitted.
        if (!isStale && messageQueueRef.current.length > 0) {
          setDequeueEpoch((e) => e + 1);
        }

        // Only decrement ref if this conversation is still current.
        // If stale (ESC was pressed), handleInterrupt already reset ref to 0.
        if (!isStale) {
          processingConversationRef.current = Math.max(
            0,
            processingConversationRef.current - 1,
          );
        }
      }
    },
    [
      appendError,
      refreshDerived,
      refreshDerivedThrottled,
      setStreaming,
      currentModelId,
      updateStreamingOutput,
      needsEagerApprovalCheck,
      queueApprovalResults,
      consumeQueuedMessages,
      appendTaskNotificationEvents,
      maybeCheckMemoryGitStatus,
      openTrajectorySegment,
      syncTrajectoryTokenBase,
      syncTrajectoryElapsedBase,
      closeTrajectorySegment,
      resetTrajectoryBases,
    ],
  );

  const handleExit = useCallback(async () => {
    saveLastAgentBeforeExit();

    // Run SessionEnd hooks
    await runEndHooks();

    // Track session end explicitly (before exit) with stats
    const stats = sessionStatsRef.current.getSnapshot();
    telemetry.trackSessionEnd(stats, "exit_command");

    // Flush telemetry before exit
    await telemetry.flush();

    setShowExitStats(true);
    // Give React time to render the stats, then exit
    setTimeout(() => {
      process.exit(0);
    }, 100);
  }, [runEndHooks]);

  // Handler when user presses UP/ESC to load queue into input for editing
  const handleEnterQueueEditMode = useCallback(() => {
    setMessageQueue([]);
  }, []);

  // Handle paste errors (e.g., image too large)
  const handlePasteError = useCallback(
    (message: string) => {
      const statusId = uid("status");
      buffersRef.current.byId.set(statusId, {
        kind: "status",
        id: statusId,
        lines: [`⚠️ ${message}`],
      });
      buffersRef.current.order.push(statusId);
      refreshDerived();
    },
    [refreshDerived],
  );

  const handleInterrupt = useCallback(async () => {
    // If we're executing client-side tools, abort them AND the main stream
    const hasTrackedTools =
      executingToolCallIdsRef.current.length > 0 ||
      autoAllowedExecutionRef.current?.results;
    if (
      isExecutingTool &&
      toolAbortControllerRef.current &&
      hasTrackedTools &&
      !toolResultsInFlightRef.current
    ) {
      toolAbortControllerRef.current.abort();

      // Mark any in-flight conversation as stale, consistent with EAGER_CANCEL.
      // Increment before tagging queued results so they are tied to the post-interrupt state.
      conversationGenerationRef.current += 1;
      processingConversationRef.current = 0;

      const autoAllowedResults = autoAllowedExecutionRef.current?.results;
      const autoAllowedMetadata = autoAllowedExecutionRef.current
        ? {
            conversationId: autoAllowedExecutionRef.current.conversationId,
            generation: conversationGenerationRef.current,
          }
        : undefined;
      if (autoAllowedResults && autoAllowedResults.length > 0) {
        queueApprovalResults(autoAllowedResults, autoAllowedMetadata);
        interruptQueuedRef.current = true;
      } else if (executingToolCallIdsRef.current.length > 0) {
        const interruptedResults = executingToolCallIdsRef.current.map(
          (toolCallId) => ({
            type: "tool" as const,
            tool_call_id: toolCallId,
            tool_return: INTERRUPTED_BY_USER,
            status: "error" as const,
          }),
        );
        queueApprovalResults(interruptedResults);
        interruptQueuedRef.current = true;
      }
      executingToolCallIdsRef.current = [];
      autoAllowedExecutionRef.current = null;

      // ALSO abort the main stream - don't leave it running
      buffersRef.current.abortGeneration =
        (buffersRef.current.abortGeneration || 0) + 1;
      const toolsCancelled = markIncompleteToolsAsCancelled(
        buffersRef.current,
        true,
        "user_interrupt",
      );

      // Mark any running subagents as interrupted
      interruptActiveSubagents(INTERRUPTED_BY_USER);

      // Show interrupt feedback (yellow message if no tools were cancelled)
      if (!toolsCancelled) {
        appendError(INTERRUPT_MESSAGE, true);
      }

      if (abortControllerRef.current) {
        abortControllerRef.current.abort();
        abortControllerRef.current = null;
      }

      pendingInterruptRecoveryConversationIdRef.current =
        conversationIdRef.current;
      userCancelledRef.current = true; // Prevent dequeue
      setStreaming(false);
      setIsExecutingTool(false);
      toolResultsInFlightRef.current = false;
      refreshDerived();

      // Delay flag reset to ensure React has flushed state updates before dequeue can fire.
      // Use setTimeout(50) instead of setTimeout(0) - the longer delay ensures React's
      // batched state updates have been fully processed before we allow the dequeue effect.
      setTimeout(() => {
        userCancelledRef.current = false;
      }, 50);

      return;
    }

    if (!streaming || interruptRequested) {
      return;
    }

    // If we're in the middle of queue cancel, set flag to restore instead of auto-send
    if (waitingForQueueCancelRef.current) {
      setRestoreQueueOnCancel(true);
      // Don't reset flags - let the cancel complete naturally
    }

    // If EAGER_CANCEL is enabled, immediately stop everything client-side first
    if (EAGER_CANCEL) {
      // Prevent multiple handleInterrupt calls while state updates are pending
      setInterruptRequested(true);

      // Set interrupted flag FIRST, before abort() triggers any async work.
      // This ensures onChunk and other guards see interrupted=true immediately.
      buffersRef.current.abortGeneration =
        (buffersRef.current.abortGeneration || 0) + 1;
      const toolsCancelled = markIncompleteToolsAsCancelled(
        buffersRef.current,
        true,
        "user_interrupt",
      );

      // Mark any running subagents as interrupted
      interruptActiveSubagents(INTERRUPTED_BY_USER);

      // NOW abort the stream - interrupted flag is already set
      if (abortControllerRef.current) {
        abortControllerRef.current.abort();
        abortControllerRef.current = null; // Clear ref so isAgentBusy() returns false
      }

      // Set cancellation flag to prevent processConversation from starting
      pendingInterruptRecoveryConversationIdRef.current =
        conversationIdRef.current;
      userCancelledRef.current = true;

      // Increment generation to mark any in-flight processConversation as stale.
      // The stale processConversation will check this and exit quietly without
      // decrementing the ref (since we reset it here).
      conversationGenerationRef.current += 1;

      // Reset the processing guard so the next message can start a new conversation.
      processingConversationRef.current = 0;

      // Stop streaming and show error message (unless tool calls were cancelled,
      // since the tool result will show "Interrupted by user")
      setStreaming(false);
      toolResultsInFlightRef.current = false;
      if (!toolsCancelled) {
        appendError(INTERRUPT_MESSAGE, true);
      }
      refreshDerived();

      // Cache pending approvals, plus any auto-handled results, for the next message.
      const denialResults = pendingApprovals.map((approval) => ({
        type: "approval" as const,
        tool_call_id: approval.toolCallId,
        approve: false,
        reason: "User interrupted the stream",
      }));
      const autoHandledSnapshot = [...autoHandledResults];
      const autoDeniedSnapshot = [...autoDeniedApprovals];
      const queuedResults = [
        ...autoHandledSnapshot.map((ar) => ({
          type: "tool" as const,
          tool_call_id: ar.toolCallId,
          tool_return: ar.result.toolReturn,
          status: ar.result.status,
          stdout: ar.result.stdout,
          stderr: ar.result.stderr,
        })),
        ...autoDeniedSnapshot.map((ad) => ({
          type: "approval" as const,
          tool_call_id: ad.approval.toolCallId,
          approve: false,
          reason: ad.reason,
        })),
        ...denialResults,
      ];
      if (queuedResults.length > 0) {
        queueApprovalResults(queuedResults);
      }

      // Clear local approval state
      setPendingApprovals([]);
      setApprovalContexts([]);
      setApprovalResults([]);
      setAutoHandledResults([]);
      setAutoDeniedApprovals([]);

      // Send cancel request to backend asynchronously (fire-and-forget)
      // Don't wait for it or show errors since user already got feedback
      getClient()
        .then((client) => {
          // Use agents API for "default" conversation (primary message history)
          if (conversationIdRef.current === "default") {
            return client.agents.messages.cancel(agentIdRef.current);
          }
          return client.conversations.cancel(conversationIdRef.current);
        })
        .catch(() => {
          // Silently ignore - cancellation already happened client-side
        });

      // Reset cancellation flags after cleanup is complete.
      // Use setTimeout(50) instead of setTimeout(0) to ensure React has fully processed
      // the streaming=false state before we allow the dequeue effect to start a new conversation.
      // This prevents the "Maximum update depth exceeded" infinite render loop.
      setTimeout(() => {
        userCancelledRef.current = false;
        setInterruptRequested(false);
      }, 50);

      return;
    } else {
      setInterruptRequested(true);
      try {
        const client = await getClient();
        // Use agents API for "default" conversation (primary message history)
        if (conversationIdRef.current === "default") {
          await client.agents.messages.cancel(agentIdRef.current);
        } else {
          await client.conversations.cancel(conversationIdRef.current);
        }

        if (abortControllerRef.current) {
          abortControllerRef.current.abort();
        }
        pendingInterruptRecoveryConversationIdRef.current =
          conversationIdRef.current;
      } catch (e) {
        const errorDetails = formatErrorDetails(e, agentId);
        appendError(`Failed to interrupt stream: ${errorDetails}`);
        setInterruptRequested(false);
      }
    }
  }, [
    agentId,
    streaming,
    interruptRequested,
    appendError,
    isExecutingTool,
    refreshDerived,
    setStreaming,
    pendingApprovals,
    autoHandledResults,
    autoDeniedApprovals,
    queueApprovalResults,
  ]);

  // Keep ref to latest processConversation to avoid circular deps in useEffect
  const processConversationRef = useRef(processConversation);
  useEffect(() => {
    processConversationRef.current = processConversation;
  }, [processConversation]);

  const handleAgentSelect = useCallback(
    async (
      targetAgentId: string,
      opts?: {
        profileName?: string;
        conversationId?: string;
        commandId?: string;
      },
    ) => {
      const overlayCommand = opts?.commandId
        ? commandRunner.getHandle(opts.commandId, "/agents")
        : consumeOverlayCommand("resume");

      // Close selector immediately
      setActiveOverlay(null);

      // Skip if already on this agent (no async work needed, queue can proceed)
      if (targetAgentId === agentId) {
        const label = agentName || targetAgentId.slice(0, 12);
        const cmd =
          overlayCommand ??
          commandRunner.start("/agents", `Already on "${label}"`);
        cmd.finish(`Already on "${label}"`, true);
        return;
      }

      // If agent is busy, queue the switch for after end_turn
      if (isAgentBusy()) {
        const cmd =
          overlayCommand ??
          commandRunner.start(
            "/agents",
            "Agent switch queued – will switch after current task completes",
          );
        cmd.update({
          output:
            "Agent switch queued – will switch after current task completes",
          phase: "running",
        });
        setQueuedOverlayAction({
          type: "switch_agent",
          agentId: targetAgentId,
          commandId: cmd.id,
        });
        return;
      }

      // Lock input for async operation (set before any await to prevent queue processing)
      setCommandRunning(true);

      // Show loading indicator while switching
      const cmd =
        overlayCommand ?? commandRunner.start("/agents", "Switching agent...");
      cmd.update({ output: "Switching agent...", phase: "running" });

      try {
        const client = await getClient();
        // Fetch new agent
        const agent = await client.agents.retrieve(targetAgentId);

        // Use specified conversation or default to the agent's default conversation
        const targetConversationId = opts?.conversationId ?? "default";

        // Update project settings with new agent
        await updateProjectSettings({ lastAgent: targetAgentId });

        // Save the session (agent + conversation) to settings
        settingsManager.setLocalLastSession(
          { agentId: targetAgentId, conversationId: targetConversationId },
          process.cwd(),
        );
        settingsManager.setGlobalLastSession({
          agentId: targetAgentId,
          conversationId: targetConversationId,
        });

        // Clear current transcript and static items
        buffersRef.current.byId.clear();
        buffersRef.current.order = [];
        buffersRef.current.tokenCount = 0;
        emittedIdsRef.current.clear();
        resetDeferredToolCallCommits();
        setStaticItems([]);
        setStaticRenderEpoch((e) => e + 1);
        resetTrajectoryBases();

        // Reset turn counter for memory reminders when switching agents
        turnCountRef.current = 0;

        // Update agent state - also update ref immediately for any code that runs before re-render
        agentIdRef.current = targetAgentId;
        setAgentId(targetAgentId);
        setAgentState(agent);
        setLlmConfig(agent.llm_config);
        setConversationId(targetConversationId);

        // Set conversation switch context for agent switch
        {
          const { getModelDisplayName } = await import("../agent/model");
          const modelHandle =
            agent.model ||
            (agent.llm_config?.model_endpoint_type && agent.llm_config?.model
              ? `${agent.llm_config.model_endpoint_type}/${agent.llm_config.model}`
              : null);
          const modelLabel =
            (modelHandle && getModelDisplayName(modelHandle)) ||
            modelHandle ||
            "unknown";
          pendingConversationSwitchRef.current = {
            origin: "agent-switch",
            conversationId: targetConversationId,
            isDefault: targetConversationId === "default",
            agentSwitchContext: {
              name: agent.name || targetAgentId,
              description: agent.description ?? undefined,
              model: modelLabel,
              blockCount: agent.blocks?.length ?? 0,
            },
          };
        }

        // Reset context token tracking for new agent
        resetContextHistory(contextTrackerRef.current);

        // Build success message
        const agentLabel = agent.name || targetAgentId;
        const isSpecificConv =
          opts?.conversationId && opts.conversationId !== "default";
        const successOutput = isSpecificConv
          ? [
              `Switched to **${agentLabel}**`,
              `⎿  Conversation: ${opts.conversationId}`,
            ].join("\n")
          : [
              `Resumed the default conversation with **${agentLabel}**.`,
              `⎿  Type /resume to browse all conversations`,
              `⎿  Type /new to start a new conversation`,
            ].join("\n");
        const separator = {
          kind: "separator" as const,
          id: uid("sep"),
        };
        setStaticItems([separator]);
        cmd.finish(successOutput, true);
      } catch (error) {
        const errorDetails = formatErrorDetails(error, agentId);
        cmd.fail(`Failed: ${errorDetails}`);
      } finally {
        setCommandRunning(false);
      }
    },
    [
      agentId,
      agentName,
      commandRunner,
      consumeOverlayCommand,
      setCommandRunning,
      isAgentBusy,
      resetDeferredToolCallCommits,
      resetTrajectoryBases,
    ],
  );

  // Handle creating a new agent and switching to it
  const handleCreateNewAgent = useCallback(
    async (name: string) => {
      // Close dialog immediately
      setActiveOverlay(null);

      // Lock input for async operation
      setCommandRunning(true);

      const inputCmd = "/new";
      const cmd = commandRunner.start(inputCmd, `Creating agent "${name}"...`);

      try {
        // Create the new agent
        const { agent } = await createAgent(name);

        // Update project settings with new agent
        await updateProjectSettings({ lastAgent: agent.id });

        // Build success message with hints
        const agentUrl = `https://app.letta.com/projects/default-project/agents/${agent.id}`;
        const successOutput = [
          `Created **${agent.name || agent.id}** (use /pin to save)`,
          `⎿  ${agentUrl}`,
          `⎿  Tip: use /init to initialize your agent's memory system!`,
        ].join("\n");
        cmd.finish(successOutput, true);
        const successItem: StaticItem = {
          kind: "command",
          id: cmd.id,
          input: cmd.input,
          output: successOutput,
          phase: "finished",
          success: true,
        };

        // Clear current transcript and static items
        buffersRef.current.byId.clear();
        buffersRef.current.order = [];
        buffersRef.current.tokenCount = 0;
        emittedIdsRef.current.clear();
        resetDeferredToolCallCommits();
        setStaticItems([]);
        setStaticRenderEpoch((e) => e + 1);
        resetTrajectoryBases();

        // Reset turn counter for memory reminders
        turnCountRef.current = 0;

        // Update agent state
        agentIdRef.current = agent.id;
        setAgentId(agent.id);
        setAgentState(agent);
        setLlmConfig(agent.llm_config);

        // Reset context token tracking for new agent
        resetContextHistory(contextTrackerRef.current);

        const separator = {
          kind: "separator" as const,
          id: uid("sep"),
        };

        setStaticItems([separator, successItem]);
        // Sync lines display after clearing buffers
        setLines(toLines(buffersRef.current));
      } catch (error) {
        const errorDetails = formatErrorDetails(error, agentId);
        cmd.fail(`Failed to create agent: ${errorDetails}`);
      } finally {
        setCommandRunning(false);
      }
    },
    [
      agentId,
      commandRunner,
      setCommandRunning,
      resetDeferredToolCallCommits,
      resetTrajectoryBases,
    ],
  );

  // Handle bash mode command submission
  // Expands aliases from shell config files, then runs with spawnCommand
  // Implements input locking and ESC cancellation (LET-7199)
  const handleBashSubmit = useCallback(
    async (command: string) => {
      // Input locking - prevent multiple concurrent bash commands
      if (bashRunning) return;

      const cmdId = uid("bash");
      const startTime = Date.now();

      // Set up state for input locking and cancellation
      setBashRunning(true);
      bashAbortControllerRef.current = new AbortController();

      // Add running bash_command line with streaming state
      buffersRef.current.byId.set(cmdId, {
        kind: "bash_command",
        id: cmdId,
        input: command,
        output: "",
        phase: "running",
        streaming: {
          tailLines: [],
          partialLine: "",
          partialIsStderr: false,
          totalLineCount: 0,
          startTime,
        },
      });
      buffersRef.current.order.push(cmdId);
      refreshDerived();

      try {
        // Expand aliases before running
        const { expandAliases } = await import("./helpers/shellAliases");
        const expanded = expandAliases(command);

        // If command uses a shell function, prepend the function definition
        const finalCommand = expanded.functionDef
          ? `${expanded.functionDef}\n${expanded.command}`
          : expanded.command;

        // Use spawnCommand for actual execution
        const { spawnCommand } = await import("../tools/impl/Bash.js");
        const { getShellEnv } = await import("../tools/impl/shellEnv.js");

        const result = await spawnCommand(finalCommand, {
          cwd: process.cwd(),
          env: getShellEnv(),
          timeout: 0, // No timeout - user must ESC to interrupt (LET-7199)
          signal: bashAbortControllerRef.current.signal,
          onOutput: (chunk, stream) => {
            const entry = buffersRef.current.byId.get(cmdId);
            if (entry && entry.kind === "bash_command") {
              const newStreaming = appendStreamingOutput(
                entry.streaming,
                chunk,
                startTime,
                stream === "stderr",
              );
              buffersRef.current.byId.set(cmdId, {
                ...entry,
                streaming: newStreaming,
              });
              refreshDerivedStreaming();
            }
          },
        });

        // Combine stdout and stderr for output
        const output = (result.stdout + result.stderr).trim();
        const success = result.exitCode === 0;

        // Update line with output, clear streaming state
        const displayOutput =
          output ||
          (success
            ? "(Command completed with no output)"
            : `Exit code: ${result.exitCode}`);
        buffersRef.current.byId.set(cmdId, {
          kind: "bash_command",
          id: cmdId,
          input: command,
          output: displayOutput,
          phase: "finished",
          success,
          streaming: undefined,
        });

        // Cache for next user message
        bashCommandCacheRef.current.push({
          input: command,
          output: displayOutput,
        });
      } catch (error: unknown) {
        // Check if this was an abort (user pressed ESC)
        const err = error as { name?: string; code?: string; message?: string };
        const isAbort =
          bashAbortControllerRef.current?.signal.aborted ||
          err.code === "ABORT_ERR" ||
          err.name === "AbortError" ||
          err.message === "The operation was aborted";

        let errOutput: string;
        if (isAbort) {
          errOutput = INTERRUPTED_BY_USER;
        } else {
          // Handle command errors (timeout, other failures)
          errOutput =
            error instanceof Error
              ? (error as { stderr?: string; stdout?: string }).stderr ||
                (error as { stdout?: string }).stdout ||
                error.message
              : String(error);
        }

        buffersRef.current.byId.set(cmdId, {
          kind: "bash_command",
          id: cmdId,
          input: command,
          output: errOutput,
          phase: "finished",
          success: false,
          streaming: undefined,
        });

        // Still cache for next user message (even failures are visible to agent)
        bashCommandCacheRef.current.push({ input: command, output: errOutput });
      } finally {
        // Clean up state
        setBashRunning(false);
        bashAbortControllerRef.current = null;
      }

      refreshDerived();
    },
    [bashRunning, refreshDerived, refreshDerivedStreaming],
  );

  // Handle ESC interrupt for bash mode commands (LET-7199)
  const handleBashInterrupt = useCallback(() => {
    if (bashAbortControllerRef.current) {
      bashAbortControllerRef.current.abort();
    }
  }, []);

  /**
   * Check and handle any pending approvals before sending a slash command.
   * Returns true if approvals need user input (caller should return { submitted: false }).
   * Returns false if no approvals or all auto-handled (caller can proceed).
   */
  const checkPendingApprovalsForSlashCommand = useCallback(async (): Promise<
    { blocked: true } | { blocked: false }
  > => {
    // Only check eagerly when resuming a session (LET-7101)
    if (!needsEagerApprovalCheck) {
      return { blocked: false };
    }

    try {
      const client = await getClient();
      const agent = await client.agents.retrieve(agentId);
      const { pendingApprovals: existingApprovals } = await getResumeData(
        client,
        agent,
        conversationIdRef.current,
      );

      if (!existingApprovals || existingApprovals.length === 0) {
        return { blocked: false };
      }

      // There are pending approvals - check permissions (respects yolo mode)
      const { needsUserInput, autoAllowed, autoDenied } =
        await classifyApprovals(existingApprovals, {
          getContext: analyzeToolApproval,
          alwaysRequiresUserInput,
          missingNameReason: "Tool call incomplete - missing name",
        });

      // If any approvals need user input, show dialog
      if (needsUserInput.length > 0) {
        setPendingApprovals(needsUserInput.map((ac) => ac.approval));
        setApprovalContexts(
          needsUserInput
            .map((ac) => ac.context)
            .filter((ctx): ctx is ApprovalContext => ctx !== null),
        );
        return { blocked: true };
      }

      // All approvals can be auto-handled - execute them before proceeding
      const allResults: ApprovalResult[] = [];

      const autoAllowedToolCallIds = autoAllowed.map(
        (ac) => ac.approval.toolCallId,
      );
      const autoAllowedAbortController =
        abortControllerRef.current ?? new AbortController();
      const shouldTrackAutoAllowed = autoAllowedToolCallIds.length > 0;
      let autoAllowedResults: Array<{
        toolCallId: string;
        result: ToolExecutionResult;
      }> = [];

      if (shouldTrackAutoAllowed) {
        setIsExecutingTool(true);
        executingToolCallIdsRef.current = autoAllowedToolCallIds;
        toolAbortControllerRef.current = autoAllowedAbortController;
        autoAllowedExecutionRef.current = {
          toolCallIds: autoAllowedToolCallIds,
          results: null,
          conversationId: conversationIdRef.current,
          generation: conversationGenerationRef.current,
        };
      }

      try {
        // Execute auto-allowed tools
        if (autoAllowed.length > 0) {
          // Set phase to "running" for auto-allowed tools
          setToolCallsRunning(buffersRef.current, autoAllowedToolCallIds);
          refreshDerived();

          autoAllowedResults = await executeAutoAllowedTools(
            autoAllowed,
            (chunk) => onChunk(buffersRef.current, chunk),
            {
              abortSignal: autoAllowedAbortController.signal,
              onStreamingOutput: updateStreamingOutput,
            },
          );
          // Map to ApprovalResult format (ToolReturn)
          allResults.push(
            ...autoAllowedResults.map((ar) => ({
              type: "tool" as const,
              tool_call_id: ar.toolCallId,
              tool_return: ar.result.toolReturn,
              status: ar.result.status,
              stdout: ar.result.stdout,
              stderr: ar.result.stderr,
            })),
          );
        }

        // Create denial results for auto-denied
        for (const ac of autoDenied) {
          const reason = ac.permission.reason || "Permission denied";
          // Update UI with denial
          onChunk(buffersRef.current, {
            message_type: "tool_return_message",
            id: "dummy",
            date: new Date().toISOString(),
            tool_call_id: ac.approval.toolCallId,
            tool_return: `Error: request to call tool denied. User reason: ${reason}`,
            status: "error",
            stdout: null,
            stderr: null,
          });
          // Map to ApprovalResult format (ApprovalReturn)
          allResults.push({
            type: "approval" as const,
            tool_call_id: ac.approval.toolCallId,
            approve: false,
            reason,
          });
        }

        if (autoAllowedExecutionRef.current) {
          autoAllowedExecutionRef.current.results = allResults;
        }
        const autoAllowedMetadata = autoAllowedExecutionRef.current
          ? {
              conversationId: autoAllowedExecutionRef.current.conversationId,
              generation: conversationGenerationRef.current,
            }
          : undefined;

        if (
          userCancelledRef.current ||
          autoAllowedAbortController.signal.aborted ||
          interruptQueuedRef.current
        ) {
          if (allResults.length > 0) {
            queueApprovalResults(allResults, autoAllowedMetadata);
          }
          return { blocked: false };
        }

        // Send all results to server if any
        if (allResults.length > 0) {
          toolResultsInFlightRef.current = true;
          await processConversation([
            { type: "approval", approvals: allResults },
          ]);
          toolResultsInFlightRef.current = false;

          // Clear any stale queued results from previous interrupts.
          queueApprovalResults(null);
        }
      } finally {
        if (shouldTrackAutoAllowed) {
          setIsExecutingTool(false);
          toolAbortControllerRef.current = null;
          executingToolCallIdsRef.current = [];
          autoAllowedExecutionRef.current = null;
          toolResultsInFlightRef.current = false;
        }
      }

      return { blocked: false };
    } catch {
      // If check fails, proceed anyway (don't block user)
      return { blocked: false };
    }
  }, [
    agentId,
    processConversation,
    refreshDerived,
    updateStreamingOutput,
    needsEagerApprovalCheck,
    queueApprovalResults,
  ]);

  // biome-ignore lint/correctness/useExhaustiveDependencies: refs read .current dynamically, complex callback with intentional deps
  const onSubmit = useCallback(
    async (message?: string): Promise<{ submitted: boolean }> => {
      const msg = message?.trim() ?? "";
      const overrideContentParts = overrideContentPartsRef.current;
      if (overrideContentParts) {
        overrideContentPartsRef.current = null;
      }
      const { notifications: taskNotifications, cleanedText } =
        extractTaskNotificationsForDisplay(msg);
      const userTextForInput = cleanedText.trim();
      const isSystemOnly =
        taskNotifications.length > 0 && userTextForInput.length === 0;

      // Handle profile load confirmation (Enter to continue)
      if (profileConfirmPending && !msg) {
        // User pressed Enter with empty input - proceed with loading
        const { name, agentId: targetAgentId, cmdId } = profileConfirmPending;
        const cmd = commandRunner.getHandle(cmdId, `/profile load ${name}`);
        cmd.update({ output: "Loading profile...", phase: "running" });
        setProfileConfirmPending(null);
        await handleAgentSelect(targetAgentId, {
          profileName: name,
          commandId: cmdId,
        });
        return { submitted: true };
      }

      // Cancel profile confirmation if user types something else
      if (profileConfirmPending && msg) {
        const { cmdId, name } = profileConfirmPending;
        const cmd = commandRunner.getHandle(cmdId, `/profile load ${name}`);
        cmd.fail("Cancelled");
        setProfileConfirmPending(null);
        // Continue processing the new message
      }

      if (!msg) return { submitted: false };

      // Run UserPromptSubmit hooks - can block the prompt from being processed
      const isCommand = userTextForInput.startsWith("/");
      const hookResult = isSystemOnly
        ? { blocked: false, feedback: [] as string[] }
        : await runUserPromptSubmitHooks(
            userTextForInput,
            isCommand,
            agentId,
            conversationIdRef.current,
          );
      if (!isSystemOnly && hookResult.blocked) {
        // Show feedback from hook in the transcript
        const feedbackId = uid("status");
        const feedback = hookResult.feedback.join("\n") || "Blocked by hook";
        buffersRef.current.byId.set(feedbackId, {
          kind: "status",
          id: feedbackId,
          lines: [
            `<user-prompt-submit-hook>${feedback}</user-prompt-submit-hook>`,
          ],
        });
        buffersRef.current.order.push(feedbackId);
        refreshDerived();
        return { submitted: false };
      }

      // Capture successful hook feedback to inject into agent context
      const userPromptSubmitHookFeedback =
        hookResult.feedback.length > 0
          ? `${SYSTEM_REMINDER_OPEN}\n${hookResult.feedback.join("\n")}\n${SYSTEM_REMINDER_CLOSE}`
          : "";

      // Capture the generation at submission time, BEFORE any async work.
      // This allows detecting if ESC was pressed during async operations.
      const submissionGeneration = conversationGenerationRef.current;

      // Track user input (agent_id automatically added from telemetry.currentAgentId)
      if (!isSystemOnly && userTextForInput.length > 0) {
        telemetry.trackUserInput(
          userTextForInput,
          "user",
          currentModelId || "unknown",
        );
      }

      // Capture first user query for conversation summary (before any async work)
      // Only for new conversations, non-commands, and if we haven't captured yet
      if (
        !hasSetConversationSummaryRef.current &&
        firstUserQueryRef.current === null &&
        !isSystemOnly &&
        userTextForInput.length > 0 &&
        !userTextForInput.startsWith("/")
      ) {
        firstUserQueryRef.current = userTextForInput.slice(0, 100);
      }

      // Block submission if waiting for explicit user action (approvals)
      // In this case, input is hidden anyway, so this shouldn't happen
      if (pendingApprovals.length > 0) {
        return { submitted: false };
      }

      // Queue message if agent is busy (streaming, executing tool, or running command)
      // This allows messages to queue up while agent is working

      // Reset cancellation flag before queue check - this ensures queued messages
      // can be dequeued even if the user just cancelled. The dequeue effect checks
      // userCancelledRef.current, so we must clear it here to prevent blocking.
      userCancelledRef.current = false;

      // If there are queued messages and agent is not busy, bump epoch to trigger
      // dequeue effect. Without this, the effect won't re-run because refs aren't
      // in its deps array (only state values are).
      if (!isAgentBusy() && messageQueue.length > 0) {
        debugLog(
          "queue",
          `Bumping dequeueEpoch: userCancelledRef was reset, ${messageQueue.length} message(s) queued, agent not busy`,
        );
        setDequeueEpoch((e) => e + 1);
      }

      // Interactive slash commands (like /memory, /model, /agents) bypass queueing
      // so users can browse/view while the agent is working.
      // Changes made in these overlays will be queued until end_turn.
      const shouldBypassQueue =
        isInteractiveCommand(userTextForInput) ||
        isNonStateCommand(userTextForInput);

      if (isAgentBusy() && !shouldBypassQueue) {
        setMessageQueue((prev) => {
          const newQueue: QueuedMessage[] = [
            ...prev,
            { kind: "user", text: msg },
          ];

          // Regular messages: queue and wait for tool completion

          return newQueue;
        });
        return { submitted: true }; // Clears input
      }

      // Note: userCancelledRef.current was already reset above before the queue check
      // to ensure the dequeue effect isn't blocked by a stale cancellation flag.

      // Handle pending Ralph config - activate ralph mode but let message flow through normal path
      // This ensures session context and other reminders are included
      // Track if we just activated so we can use first turn reminder vs continuation
      let justActivatedRalph = false;
      if (pendingRalphConfig && !msg.startsWith("/")) {
        const { completionPromise, maxIterations, isYolo } = pendingRalphConfig;
        ralphMode.activate(msg, completionPromise, maxIterations, isYolo);
        setUiRalphActive(true);
        setPendingRalphConfig(null);
        justActivatedRalph = true;
        if (isYolo) {
          permissionMode.setMode("bypassPermissions");
        }

        const ralphState = ralphMode.getState();

        // Add status to transcript
        const statusId = uid("status");
        const promiseDisplay = ralphState.completionPromise
          ? `"${ralphState.completionPromise.slice(0, 50)}${ralphState.completionPromise.length > 50 ? "..." : ""}"`
          : "(none)";
        buffersRef.current.byId.set(statusId, {
          kind: "status",
          id: statusId,
          lines: [
            `🔄 ${isYolo ? "yolo-ralph" : "ralph"} mode started (iter 1/${maxIterations || "∞"})`,
            `Promise: ${promiseDisplay}`,
          ],
        });
        buffersRef.current.order.push(statusId);
        refreshDerived();

        // Don't return - let message flow through normal path which will:
        // 1. Add session context reminder (if first message)
        // 2. Add ralph mode reminder (since ralph is now active)
        // 3. Add other reminders (skill unload, memory, etc.)
      }

      let aliasedMsg = msg;
      if (msg === "exit" || msg === "quit") {
        aliasedMsg = "/exit";
      }

      // Handle commands (messages starting with "/")
      if (aliasedMsg.startsWith("/")) {
        const trimmed = aliasedMsg.trim();

        // Special handling for /model command - opens selector
        if (trimmed === "/model") {
          startOverlayCommand(
            "model",
            "/model",
            "Opening model selector...",
            "Models dialog dismissed",
          );
          setModelSelectorOptions({}); // Clear any filters from previous connection
          setActiveOverlay("model");
          return { submitted: true };
        }

        // Special handling for /sleeptime command - opens reflection settings
        if (trimmed === "/sleeptime") {
          startOverlayCommand(
            "sleeptime",
            "/sleeptime",
            "Opening sleeptime settings...",
            "Sleeptime settings dismissed",
          );
          setActiveOverlay("sleeptime");
          return { submitted: true };
        }

        // Special handling for /toolset command - opens selector
        if (trimmed === "/toolset") {
          startOverlayCommand(
            "toolset",
            "/toolset",
            "Opening toolset selector...",
            "Toolset dialog dismissed",
          );
          setActiveOverlay("toolset");
          return { submitted: true };
        }

        // Special handling for /ade command - open agent in browser
        if (trimmed === "/ade") {
          const adeUrl = `https://app.letta.com/agents/${agentId}?conversation=${conversationIdRef.current}`;
          const cmd = commandRunner.start("/ade", "Opening ADE...");

          // Fire-and-forget browser open
          import("open")
            .then(({ default: open }) => open(adeUrl, { wait: false }))
            .catch(() => {
              // Silently ignore - user can use the URL from the output
            });

          // Always show the URL in case browser doesn't open
          cmd.finish(`Opening ADE...\n→ ${adeUrl}`, true);
          return { submitted: true };
        }

        // Special handling for /system command - opens system prompt selector
        if (trimmed === "/system") {
          startOverlayCommand(
            "system",
            "/system",
            "Opening system prompt selector...",
            "System prompt dialog dismissed",
          );
          setActiveOverlay("system");
          return { submitted: true };
        }

        // Special handling for /subagents command - opens subagent manager
        if (trimmed === "/subagents") {
          startOverlayCommand(
            "subagent",
            "/subagents",
            "Opening subagent manager...",
            "Subagent manager dismissed",
          );
          setActiveOverlay("subagent");
          return { submitted: true };
        }

        // Special handling for /memory command - opens memory viewer
        if (trimmed === "/memory") {
          startOverlayCommand(
            "memory",
            "/memory",
            "Opening memory viewer...",
            "Memory viewer dismissed",
          );
          setActiveOverlay("memory");
          return { submitted: true };
        }

        // Special handling for /mcp command - manage MCP servers
        if (msg.trim().startsWith("/mcp")) {
          const mcpCtx: McpCommandContext = {
            buffersRef,
            refreshDerived,
            setCommandRunning,
          };

          // Check for subcommand by looking at the first word after /mcp
          const afterMcp = msg.trim().slice(4).trim(); // Remove "/mcp" prefix
          const firstWord = afterMcp.split(/\s+/)[0]?.toLowerCase();

          // /mcp - open MCP server selector
          if (!firstWord) {
            startOverlayCommand(
              "mcp",
              "/mcp",
              "Opening MCP server manager...",
              "MCP dialog dismissed",
            );
            setActiveOverlay("mcp");
            return { submitted: true };
          }

          // /mcp add --transport <type> <name> <url/command> [options]
          if (firstWord === "add") {
            // Pass the full command string after "add" to preserve quotes
            const afterAdd = afterMcp.slice(firstWord.length).trim();
            const cmd = commandRunner.start(msg, "Adding MCP server...");
            setActiveMcpCommandId(cmd.id);
            try {
              await handleMcpAdd(mcpCtx, msg, afterAdd);
            } finally {
              setActiveMcpCommandId(null);
            }
            return { submitted: true };
          }

          // /mcp connect - interactive TUI for connecting with OAuth
          if (firstWord === "connect") {
            startOverlayCommand(
              "mcp-connect",
              "/mcp connect",
              "Opening MCP connect flow...",
              "MCP connect dismissed",
            );
            setActiveOverlay("mcp-connect");
            return { submitted: true };
          }

          // Unknown subcommand
          {
            const cmd = commandRunner.start(msg, "Checking MCP usage...");
            setActiveMcpCommandId(cmd.id);
            try {
              handleMcpUsage(mcpCtx, msg);
            } finally {
              setActiveMcpCommandId(null);
            }
          }
          return { submitted: true };
        }

        // Special handling for /connect command - opens provider selector
        if (msg.trim() === "/connect") {
          startOverlayCommand(
            "connect",
            "/connect",
            "Opening provider selector...",
            "Connect dialog dismissed",
          );
          setActiveOverlay("connect");
          return { submitted: true };
        }

        // /connect codex - direct OAuth flow (kept for backwards compatibility)
        if (msg.trim().startsWith("/connect codex")) {
          const cmd = commandRunner.start(msg, "Starting connection...");
          const {
            handleConnect,
            setActiveCommandId: setActiveConnectCommandId,
          } = await import("./commands/connect");
          setActiveConnectCommandId(cmd.id);
          try {
            await handleConnect(
              {
                buffersRef,
                refreshDerived,
                setCommandRunning,
                onCodexConnected: () => {
                  setModelSelectorOptions({
                    filterProvider: "chatgpt-plus-pro",
                    forceRefresh: true,
                  });
                  startOverlayCommand(
                    "model",
                    "/model",
                    "Opening model selector...",
                    "Models dialog dismissed",
                  );
                  setActiveOverlay("model");
                },
              },
              msg,
            );
          } finally {
            setActiveConnectCommandId(null);
          }
          return { submitted: true };
        }

        // Special handling for /disconnect command - remove OAuth connection
        if (msg.trim().startsWith("/disconnect")) {
          const cmd = commandRunner.start(msg, "Disconnecting...");
          const {
            handleDisconnect,
            setActiveCommandId: setActiveConnectCommandId,
          } = await import("./commands/connect");
          setActiveConnectCommandId(cmd.id);
          try {
            await handleDisconnect(
              {
                buffersRef,
                refreshDerived,
                setCommandRunning,
              },
              msg,
            );
          } finally {
            setActiveConnectCommandId(null);
          }
          return { submitted: true };
        }

        // Special handling for /help command - opens help dialog
        if (trimmed === "/help") {
          startOverlayCommand(
            "help",
            "/help",
            "Opening help...",
            "Help dialog dismissed",
          );
          setActiveOverlay("help");
          return { submitted: true };
        }

        // Special handling for /hooks command - opens hooks manager
        if (trimmed === "/hooks") {
          startOverlayCommand(
            "hooks",
            "/hooks",
            "Opening hooks manager...",
            "Hooks manager dismissed",
          );
          setActiveOverlay("hooks");
          return { submitted: true };
        }

        // Special handling for /statusline command
        if (trimmed === "/statusline" || trimmed.startsWith("/statusline ")) {
          const rawArgs = trimmed.slice("/statusline".length).trim();
          const spaceIdx = rawArgs.indexOf(" ");
          const sub =
            spaceIdx === -1 ? rawArgs || "show" : rawArgs.slice(0, spaceIdx);
          const rest =
            spaceIdx === -1 ? "" : rawArgs.slice(spaceIdx + 1).trim();
          const cmd = commandRunner.start(trimmed, "Managing status line...");

          (async () => {
            try {
              const wd = process.cwd();
              if (sub === "help") {
                cmd.finish(formatStatusLineHelp(), true, true);
              } else if (sub === "show") {
                // Display config from all levels + resolved effective
                const lines: string[] = [];
                try {
                  const global = settingsManager.getSettings().statusLine;
                  lines.push(
                    `Global: ${global?.command ? `command="${global.command}" refreshInterval=${global.refreshIntervalMs ?? "off"} timeout=${global.timeout ?? "default"} debounce=${global.debounceMs ?? "default"} padding=${global.padding ?? 0} disabled=${global.disabled ?? false}` : "(not set)"}`,
                  );
                } catch {
                  lines.push("Global: (unavailable)");
                }
                try {
                  const project =
                    settingsManager.getProjectSettings(wd)?.statusLine;
                  lines.push(
                    `Project: ${project?.command ? `command="${project.command}"` : "(not set)"}`,
                  );
                } catch {
                  lines.push("Project: (not loaded)");
                }
                try {
                  const local =
                    settingsManager.getLocalProjectSettings(wd)?.statusLine;
                  lines.push(
                    `Local: ${local?.command ? `command="${local.command}"` : "(not set)"}`,
                  );
                } catch {
                  lines.push("Local: (not loaded)");
                }
                const effective = resolveStatusLineConfig(wd);
                lines.push(
                  `Effective: ${effective ? `command="${effective.command}" refreshInterval=${effective.refreshIntervalMs ?? "off"} timeout=${effective.timeout}ms debounce=${effective.debounceMs}ms padding=${effective.padding}` : "(inactive)"}`,
                );
                cmd.finish(lines.join("\n"), true);
              } else if (sub === "set") {
                if (!rest) {
                  cmd.finish("Usage: /statusline set <command> [-l|-p]", false);
                  return;
                }
                const scopeMatch = rest.match(/\s+-(l|p)$/);
                const command = scopeMatch
                  ? rest.slice(0, scopeMatch.index)
                  : rest;
                const isLocal = scopeMatch?.[1] === "l";
                const isProject = scopeMatch?.[1] === "p";
                const config = { command };
                if (isLocal) {
                  settingsManager.updateLocalProjectSettings(
                    { statusLine: config },
                    wd,
                  );
                  cmd.finish(`Status line set (local): ${command}`, true);
                } else if (isProject) {
                  await settingsManager.loadProjectSettings(wd);
                  settingsManager.updateProjectSettings(
                    { statusLine: config },
                    wd,
                  );
                  cmd.finish(`Status line set (project): ${command}`, true);
                } else {
                  settingsManager.updateSettings({ statusLine: config });
                  cmd.finish(`Status line set (global): ${command}`, true);
                }
              } else if (sub === "clear") {
                const isLocal = rest === "-l";
                const isProject = rest === "-p";
                if (isLocal) {
                  settingsManager.updateLocalProjectSettings(
                    { statusLine: undefined },
                    wd,
                  );
                  cmd.finish("Status line cleared (local)", true);
                } else if (isProject) {
                  await settingsManager.loadProjectSettings(wd);
                  settingsManager.updateProjectSettings(
                    { statusLine: undefined },
                    wd,
                  );
                  cmd.finish("Status line cleared (project)", true);
                } else {
                  settingsManager.updateSettings({ statusLine: undefined });
                  cmd.finish("Status line cleared (global)", true);
                }
              } else if (sub === "test") {
                const config = resolveStatusLineConfig(wd);
                if (!config) {
                  cmd.finish("No status line configured", false);
                  return;
                }
                const stats = sessionStatsRef.current.getSnapshot();
                const result = await executeStatusLineCommand(
                  config.command,
                  buildStatusLinePayload({
                    modelId: llmConfigRef.current?.model ?? null,
                    modelDisplayName: currentModelDisplay,
                    currentDirectory: wd,
                    projectDirectory,
                    sessionId: conversationIdRef.current,
                    agentName,
                    totalDurationMs: stats.totalWallMs,
                    totalApiDurationMs: stats.totalApiMs,
                    totalInputTokens: stats.usage.promptTokens,
                    totalOutputTokens: stats.usage.completionTokens,
                    contextWindowSize: llmConfigRef.current?.context_window,
                    usedContextTokens:
                      contextTrackerRef.current.lastContextTokens,
                    permissionMode: uiPermissionMode,
                    networkPhase,
                    terminalWidth: columns,
                  }),
                  { timeout: config.timeout, workingDirectory: wd },
                );
                if (result.ok) {
                  cmd.finish(
                    `Output: ${result.text} (${result.durationMs}ms)`,
                    true,
                  );
                } else {
                  cmd.finish(
                    `Error: ${result.error} (${result.durationMs}ms)`,
                    false,
                  );
                }
              } else if (sub === "disable") {
                settingsManager.updateSettings({
                  statusLine: {
                    ...settingsManager.getSettings().statusLine,
                    command:
                      settingsManager.getSettings().statusLine?.command ?? "",
                    disabled: true,
                  },
                });
                cmd.finish("Status line disabled", true);
              } else if (sub === "enable") {
                const current = settingsManager.getSettings().statusLine;
                if (!current?.command) {
                  cmd.finish(
                    "No status line configured. Use /statusline set <command> first.",
                    false,
                  );
                } else {
                  settingsManager.updateSettings({
                    statusLine: { ...current, disabled: false },
                  });
                  cmd.finish("Status line enabled", true);
                }
              } else {
                cmd.finish(
                  `Unknown subcommand: ${sub}. Use help|show|set|clear|test|enable|disable`,
                  false,
                );
              }
            } catch (error) {
              cmd.finish(
                `Error: ${error instanceof Error ? error.message : String(error)}`,
                false,
              );
            }
          })();

          triggerStatusLineRefresh();
          return { submitted: true };
        }

        // Special handling for /usage command - show session stats
        if (trimmed === "/usage") {
          const cmd = commandRunner.start(
            trimmed,
            "Fetching usage statistics...",
          );

          // Fetch balance and display stats asynchronously
          (async () => {
            try {
              const stats = sessionStatsRef.current.getSnapshot();

              // Try to fetch balance info (only works for Letta Cloud)
              // Silently skip if endpoint not available (not deployed yet or self-hosted)
              let balance:
                | {
                    total_balance: number;
                    monthly_credit_balance: number;
                    purchased_credit_balance: number;
                    billing_tier: string;
                  }
                | undefined;

              try {
                const settings = settingsManager.getSettings();
                const baseURL =
                  process.env.LETTA_BASE_URL ||
                  settings.env?.LETTA_BASE_URL ||
                  "https://api.letta.com";
                const apiKey =
                  process.env.LETTA_API_KEY || settings.env?.LETTA_API_KEY;

                const balanceResponse = await fetch(
                  `${baseURL}/v1/metadata/balance`,
                  {
                    headers: getLettaCodeHeaders(apiKey),
                  },
                );

                if (balanceResponse.ok) {
                  balance = (await balanceResponse.json()) as {
                    total_balance: number;
                    monthly_credit_balance: number;
                    purchased_credit_balance: number;
                    billing_tier: string;
                  };
                }
              } catch {
                // Silently skip balance info if endpoint not available
              }

              const output = formatUsageStats({
                stats,
                balance,
              });

              cmd.finish(output, true, true);
            } catch (error) {
              cmd.fail(
                `Error fetching usage: ${error instanceof Error ? error.message : String(error)}`,
              );
            }
          })();

          return { submitted: true };
        }

        // Special handling for /context command - show context window usage
        if (trimmed === "/context") {
          const contextWindow = llmConfigRef.current?.context_window ?? 0;
          const model = llmConfigRef.current?.model ?? "unknown";

          // Use most recent total tokens from usage_statistics as context size (after turn)
          const usedTokens = contextTrackerRef.current.lastContextTokens;
          const history = contextTrackerRef.current.contextTokensHistory;

          const cmd = commandRunner.start(
            trimmed,
            "Fetching context breakdown...",
          );

          // Fetch breakdown (5s timeout)
          let breakdown: ContextWindowOverview | undefined;
          try {
            const settings =
              await settingsManager.getSettingsWithSecureTokens();
            const apiKey =
              process.env.LETTA_API_KEY || settings.env?.LETTA_API_KEY;
            const baseUrl = getServerUrl();

            const controller = new AbortController();
            const timeoutId = setTimeout(() => controller.abort(), 5000);

            const res = await fetch(
              `${baseUrl}/v1/agents/${agentIdRef.current}/context`,
              {
                headers: { Authorization: `Bearer ${apiKey}` },
                signal: controller.signal,
              },
            );
            clearTimeout(timeoutId);

            if (res.ok) {
              breakdown = (await res.json()) as ContextWindowOverview;
            }
          } catch {
            // Timeout or network error — proceed without breakdown
          }

          // Render the full chart once, directly into the finished output
          cmd.finish(
            renderContextUsage({
              usedTokens,
              contextWindow,
              model,
              history,
              ...(breakdown && { breakdown }),
            }),
            true,
            false,
            true,
          );

          return { submitted: true };
        }

        // Special handling for /exit command - exit without stats
        if (trimmed === "/exit") {
          const cmd = commandRunner.start(trimmed, "See ya!");
          cmd.finish("See ya!", true);
          handleExit();
          return { submitted: true };
        }

        // Special handling for /logout command - clear credentials and exit
        if (trimmed === "/logout") {
          const cmd = commandRunner.start(msg.trim(), "Logging out...");

          setCommandRunning(true);

          try {
            const { settingsManager } = await import("../settings-manager");
            const currentSettings =
              await settingsManager.getSettingsWithSecureTokens();

            // Revoke refresh token on server if we have one
            if (currentSettings.refreshToken) {
              const { revokeToken } = await import("../auth/oauth");
              await revokeToken(currentSettings.refreshToken);
            }

            // Clear all credentials including secrets
            await settingsManager.logout();

            cmd.finish(
              "✓ Logged out successfully. Run 'letta' to re-authenticate.",
              true,
            );

            saveLastAgentBeforeExit();

            // Track session end explicitly (before exit) with stats
            const stats = sessionStatsRef.current.getSnapshot();
            telemetry.trackSessionEnd(stats, "logout");

            // Flush telemetry before exit
            await telemetry.flush();

            // Exit after a brief delay to show the message
            setTimeout(() => process.exit(0), 500);
          } catch (error) {
            let errorOutput = formatErrorDetails(error, agentId);

            // Add helpful tip for summarization failures
            if (errorOutput.includes("Summarization failed")) {
              errorOutput +=
                "\n\nTip: Use /clear instead to clear the current message buffer.";
            }

            cmd.fail(`Failed: ${errorOutput}`);
          } finally {
            setCommandRunning(false);
          }
          return { submitted: true };
        }

        // Special handling for /ralph and /yolo-ralph commands - Ralph Wiggum mode
        if (trimmed.startsWith("/yolo-ralph") || trimmed.startsWith("/ralph")) {
          const isYolo = trimmed.startsWith("/yolo-ralph");
          const { prompt, completionPromise, maxIterations } =
            parseRalphArgs(trimmed);

          const cmd = commandRunner.start(trimmed, "Activating ralph mode...");

          if (prompt) {
            // Inline prompt - activate immediately and send
            ralphMode.activate(
              prompt,
              completionPromise,
              maxIterations,
              isYolo,
            );
            setUiRalphActive(true);
            if (isYolo) {
              permissionMode.setMode("bypassPermissions");
            }

            const ralphState = ralphMode.getState();
            const promiseDisplay = ralphState.completionPromise
              ? `"${ralphState.completionPromise.slice(0, 50)}${ralphState.completionPromise.length > 50 ? "..." : ""}"`
              : "(none)";

            cmd.finish(
              `🔄 ${isYolo ? "yolo-ralph" : "ralph"} mode activated (iter 1/${maxIterations || "∞"})\nPromise: ${promiseDisplay}`,
              true,
            );

            // Send the prompt with ralph reminder prepended
            const systemMsg = buildRalphFirstTurnReminder(ralphState);
            processConversation([
              {
                type: "message",
                role: "user",
                content: buildTextParts(systemMsg, prompt),
              },
            ]);
          } else {
            // No inline prompt - wait for next message
            setPendingRalphConfig({ completionPromise, maxIterations, isYolo });

            const defaultPromisePreview = DEFAULT_COMPLETION_PROMISE.slice(
              0,
              40,
            );

            cmd.finish(
              `🔄 ${isYolo ? "yolo-ralph" : "ralph"} mode ready (waiting for task)\nMax iterations: ${maxIterations || "unlimited"}\nPromise: ${completionPromise === null ? "(none)" : (completionPromise ?? `"${defaultPromisePreview}..." (default)`)}\n\nType your task to begin the loop.`,
              true,
            );
          }
          return { submitted: true };
        }

        // Special handling for /stream command - toggle and save
        if (msg.trim() === "/stream") {
          const newValue = !tokenStreamingEnabled;

          // Immediately add command to transcript with "running" phase and loading message
          const cmd = commandRunner.start(
            msg.trim(),
            `${newValue ? "Enabling" : "Disabling"} token streaming...`,
          );

          // Lock input during async operation
          setCommandRunning(true);

          try {
            setTokenStreamingEnabled(newValue);

            // Save to settings
            const { settingsManager } = await import("../settings-manager");
            settingsManager.updateSettings({ tokenStreaming: newValue });

            // Update the same command with final result
            cmd.finish(
              `Token streaming ${newValue ? "enabled" : "disabled"}`,
              true,
            );
          } catch (error) {
            // Mark command as failed
            const errorDetails = formatErrorDetails(error, agentId);
            cmd.fail(`Failed: ${errorDetails}`);
          } finally {
            // Unlock input
            setCommandRunning(false);
          }
          return { submitted: true };
        }

        // Special handling for /new command - start new conversation
        if (msg.trim() === "/new") {
          const cmd = commandRunner.start(
            msg.trim(),
            "Starting new conversation...",
          );

          setCommandRunning(true);

          // Run SessionEnd hooks for current session before starting new one
          await runEndHooks();

          try {
            const client = await getClient();

            // Create a new conversation for the current agent
            const conversation = await client.conversations.create({
              agent_id: agentId,
              isolated_block_labels: [...ISOLATED_BLOCK_LABELS],
            });

            // Update conversationId state
            setConversationId(conversation.id);

            pendingConversationSwitchRef.current = {
              origin: "new",
              conversationId: conversation.id,
              isDefault: false,
            };

            // Save the new session to settings
            settingsManager.setLocalLastSession(
              { agentId, conversationId: conversation.id },
              process.cwd(),
            );
            settingsManager.setGlobalLastSession({
              agentId,
              conversationId: conversation.id,
            });

            // Reset context tokens for new conversation
            resetContextHistory(contextTrackerRef.current);

            // Reset turn counter for memory reminders
            turnCountRef.current = 0;

            // Re-run SessionStart hooks for new conversation
            sessionHooksRanRef.current = false;
            runSessionStartHooks(
              true, // isNewSession
              agentId,
              agentName ?? undefined,
              conversation.id,
            )
              .then((result) => {
                if (result.feedback.length > 0) {
                  sessionStartFeedbackRef.current = result.feedback;
                }
              })
              .catch(() => {});
            sessionHooksRanRef.current = true;

            // Update command with success
            cmd.finish(
              "Started new conversation (use /resume to change convos)",
              true,
            );
          } catch (error) {
            const errorDetails = formatErrorDetails(error, agentId);
            cmd.fail(`Failed: ${errorDetails}`);
          } finally {
            setCommandRunning(false);
          }
          return { submitted: true };
        }

        // Special handling for /clear command - reset all agent messages (destructive)
        if (msg.trim() === "/clear") {
          const cmd = commandRunner.start(
            msg.trim(),
            "Clearing in-context messages...",
          );

          setCommandRunning(true);

          // Run SessionEnd hooks for current session before clearing
          await runEndHooks();

          try {
            const client = await getClient();

            // Reset all messages on the agent (destructive operation)
            await client.agents.messages.reset(agentId, {
              add_default_initial_messages: false,
            });

            // Also create a new conversation since messages were cleared
            const conversation = await client.conversations.create({
              agent_id: agentId,
              isolated_block_labels: [...ISOLATED_BLOCK_LABELS],
            });
            setConversationId(conversation.id);

            pendingConversationSwitchRef.current = {
              origin: "clear",
              conversationId: conversation.id,
              isDefault: false,
            };

            settingsManager.setLocalLastSession(
              { agentId, conversationId: conversation.id },
              process.cwd(),
            );
            settingsManager.setGlobalLastSession({
              agentId,
              conversationId: conversation.id,
            });

            // Reset context tokens for new conversation
            resetContextHistory(contextTrackerRef.current);

            // Reset turn counter for memory reminders
            turnCountRef.current = 0;

            // Re-run SessionStart hooks for new conversation
            sessionHooksRanRef.current = false;
            runSessionStartHooks(
              true, // isNewSession
              agentId,
              agentName ?? undefined,
              conversation.id,
            )
              .then((result) => {
                if (result.feedback.length > 0) {
                  sessionStartFeedbackRef.current = result.feedback;
                }
              })
              .catch(() => {});
            sessionHooksRanRef.current = true;

            // Update command with success
            cmd.finish(
              "Agent's in-context messages cleared & moved to conversation history",
              true,
            );
          } catch (error) {
            const errorDetails = formatErrorDetails(error, agentId);
            cmd.fail(`Failed: ${errorDetails}`);
          } finally {
            setCommandRunning(false);
          }
          return { submitted: true };
        }

        // Special handling for /compact command - summarize conversation history
        // Supports: /compact, /compact all, /compact sliding_window
        if (msg.trim().startsWith("/compact")) {
          const parts = msg.trim().split(/\s+/);
          const modeArg = parts[1] as "all" | "sliding_window" | undefined;
          const validModes = ["all", "sliding_window"];

          // Validate mode if provided
          if (modeArg && !validModes.includes(modeArg)) {
            const cmd = commandRunner.start(
              msg.trim(),
              `Invalid mode "${modeArg}". Valid modes: ${validModes.join(", ")}`,
            );
            cmd.fail(
              `Invalid mode "${modeArg}". Valid modes: ${validModes.join(", ")}`,
            );
            return { submitted: true };
          }

          const modeDisplay = modeArg ? ` (mode: ${modeArg})` : "";
          const cmd = commandRunner.start(
            msg.trim(),
            `Compacting conversation history${modeDisplay}...`,
          );

          setCommandRunning(true);

          try {
            // Run PreCompact hooks - can block the compact operation
            const preCompactResult = await runPreCompactHooks(
              undefined, // context_length - not available here
              undefined, // max_context_length - not available here
              agentId,
              conversationIdRef.current,
            );
            if (preCompactResult.blocked) {
              const feedback =
                preCompactResult.feedback.join("\n") || "Blocked by hook";
              cmd.fail(`Compact blocked: ${feedback}`);
              setCommandRunning(false);
              return { submitted: true };
            }

            const client = await getClient();

            // Compute model handle from llmConfig
            const modelHandle =
              llmConfig?.model_endpoint_type && llmConfig?.model
                ? `${llmConfig.model_endpoint_type}/${llmConfig.model}`
                : llmConfig?.model || null;

            // Build compaction settings if mode was specified
            // Pass mode-specific prompt to override any agent defaults
            const compactParams =
              modeArg && modelHandle
                ? {
                    compaction_settings: {
                      mode: modeArg,
                      model: modelHandle,
                    },
                  }
                : undefined;

            // Use agent-level compact API for "default" conversation,
            // otherwise use conversation-level API
            const result =
              conversationIdRef.current === "default"
                ? await client.agents.messages.compact(agentId, compactParams)
                : await client.conversations.messages.compact(
                    conversationIdRef.current,
                    compactParams,
                  );

            // Format success message with before/after counts and summary
            const outputLines = [
              `Compaction completed${modeDisplay}. Message buffer length reduced from ${result.num_messages_before} to ${result.num_messages_after}.`,
              "",
              `Summary: ${result.summary}`,
            ];

            // Update command with success
            cmd.finish(outputLines.join("\n"), true);

            // Manual /compact bypasses stream compaction events, so trigger
            // post-compaction reminder/skills reinjection on the next user turn.
            contextTrackerRef.current.pendingReflectionTrigger = true;
            contextTrackerRef.current.pendingSkillsReinject = true;
          } catch (error) {
            let errorOutput: string;

            // Check for summarization failure - format it cleanly
            const apiError = error as {
              status?: number;
              error?: { detail?: string };
            };
            const detail = apiError?.error?.detail;
            if (
              apiError?.status === 400 &&
              detail?.includes("Summarization failed")
            ) {
              // Clean format for this specific error, but preserve raw JSON
              const cleanDetail = detail.replace(/^\d{3}:\s*/, "");
              const rawJson = JSON.stringify(apiError.error);
              errorOutput = [
                `Request failed (code=400)`,
                `Raw: ${rawJson}`,
                `Detail: ${cleanDetail}`,
                "",
                "Tip: Use /clear instead to clear the current message buffer.",
              ].join("\n");
            } else {
              errorOutput = formatErrorDetails(error, agentId);
            }

            cmd.fail(`Failed: ${errorOutput}`);
          } finally {
            setCommandRunning(false);
          }
          return { submitted: true };
        }

        // Special handling for /rename command - rename agent or conversation
        if (msg.trim().startsWith("/rename")) {
          const parts = msg.trim().split(/\s+/);
          const subcommand = parts[1]?.toLowerCase();
          const cmd = commandRunner.start(msg.trim(), "Processing rename...");

          if (
            !subcommand ||
            (subcommand !== "agent" && subcommand !== "convo")
          ) {
            cmd.fail("Usage: /rename agent <name> or /rename convo <summary>");
            return { submitted: true };
          }

          const newValue = parts.slice(2).join(" ");
          if (!newValue) {
            cmd.fail(
              subcommand === "convo"
                ? "Please provide a summary: /rename convo <summary>"
                : "Please provide a name: /rename agent <name>",
            );
            return { submitted: true };
          }

          if (subcommand === "convo") {
            cmd.update({
              output: `Renaming conversation to "${newValue}"...`,
              phase: "running",
            });

            setCommandRunning(true);

            try {
              const client = await getClient();
              await client.conversations.update(conversationId, {
                summary: newValue,
              });

              cmd.finish(`Conversation renamed to "${newValue}"`, true);
            } catch (error) {
              const errorDetails = formatErrorDetails(error, agentId);
              cmd.fail(`Failed: ${errorDetails}`);
            } finally {
              setCommandRunning(false);
            }
            return { submitted: true };
          }

          // Rename agent (default behavior)
          const validationError = validateAgentName(newValue);
          if (validationError) {
            cmd.fail(validationError);
            return { submitted: true };
          }

          cmd.update({
            output: `Renaming agent to "${newValue}"...`,
            phase: "running",
          });

          setCommandRunning(true);

          try {
            const client = await getClient();
            await client.agents.update(agentId, { name: newValue });
            updateAgentName(newValue);

            cmd.finish(`Agent renamed to "${newValue}"`, true);
          } catch (error) {
            const errorDetails = formatErrorDetails(error, agentId);
            cmd.fail(`Failed: ${errorDetails}`);
          } finally {
            setCommandRunning(false);
          }
          return { submitted: true };
        }

        // Special handling for /description command - update agent description
        if (msg.trim().startsWith("/description")) {
          const parts = msg.trim().split(/\s+/);
          const newDescription = parts.slice(1).join(" ");
          const cmd = commandRunner.start(
            msg.trim(),
            "Updating description...",
          );

          if (!newDescription) {
            cmd.fail("Please provide a description: /description <text>");
            return { submitted: true };
          }

          cmd.update({ output: "Updating description...", phase: "running" });

          setCommandRunning(true);

          try {
            const client = await getClient();
            await client.agents.update(agentId, {
              description: newDescription,
            });

            cmd.finish(`Description updated to "${newDescription}"`, true);
          } catch (error) {
            const errorDetails = formatErrorDetails(error, agentId);
            cmd.fail(`Failed: ${errorDetails}`);
          } finally {
            setCommandRunning(false);
          }
          return { submitted: true };
        }

        // Special handling for /agents command - show agent browser
        // /pinned, /profiles are hidden aliases
        if (
          msg.trim() === "/agents" ||
          msg.trim() === "/pinned" ||
          msg.trim() === "/profiles"
        ) {
          startOverlayCommand(
            "resume",
            "/agents",
            "Opening agent browser...",
            "Agent browser dismissed",
          );
          setActiveOverlay("resume");
          return { submitted: true };
        }

        // Special handling for /resume command - show conversation selector or switch directly
        if (msg.trim().startsWith("/resume")) {
          const parts = msg.trim().split(/\s+/);
          const targetConvId = parts[1]; // Optional conversation ID

          if (targetConvId) {
            const cmd = commandRunner.start(
              msg.trim(),
              "Switching conversation...",
            );
            // Direct switch to specified conversation
            if (targetConvId === conversationId) {
              cmd.finish("Already on this conversation", true);
              return { submitted: true };
            }

            // Lock input and show loading
            setCommandRunning(true);

            try {
              // Validate conversation exists BEFORE updating state
              // (getResumeData throws 404/422 for non-existent conversations)
              if (agentState) {
                const client = await getClient();
                const resumeData = await getResumeData(
                  client,
                  agentState,
                  targetConvId,
                );

                // Only update state after validation succeeds
                setConversationId(targetConvId);

                pendingConversationSwitchRef.current = {
                  origin: "resume-direct",
                  conversationId: targetConvId,
                  isDefault: targetConvId === "default",
                  messageCount: resumeData.messageHistory.length,
                  messageHistory: resumeData.messageHistory,
                };

                settingsManager.setLocalLastSession(
                  { agentId, conversationId: targetConvId },
                  process.cwd(),
                );
                settingsManager.setGlobalLastSession({
                  agentId,
                  conversationId: targetConvId,
                });

                // Build success message
                const currentAgentName = agentState.name || "Unnamed Agent";
                const successLines =
                  resumeData.messageHistory.length > 0
                    ? [
                        `Resumed conversation with "${currentAgentName}"`,
                        `⎿  Agent: ${agentId}`,
                        `⎿  Conversation: ${targetConvId}`,
                      ]
                    : [
                        `Switched to conversation with "${currentAgentName}"`,
                        `⎿  Agent: ${agentId}`,
                        `⎿  Conversation: ${targetConvId} (empty)`,
                      ];
                const successOutput = successLines.join("\n");
                cmd.finish(successOutput, true);
                const successItem: StaticItem = {
                  kind: "command",
                  id: cmd.id,
                  input: cmd.input,
                  output: successOutput,
                  phase: "finished",
                  success: true,
                };

                // Clear current transcript and static items
                buffersRef.current.byId.clear();
                buffersRef.current.order = [];
                buffersRef.current.tokenCount = 0;
                resetContextHistory(contextTrackerRef.current);
                emittedIdsRef.current.clear();
                resetDeferredToolCallCommits();
                setStaticItems([]);
                setStaticRenderEpoch((e) => e + 1);
                resetTrajectoryBases();

                // Backfill message history
                if (resumeData.messageHistory.length > 0) {
                  hasBackfilledRef.current = false;
                  backfillBuffers(
                    buffersRef.current,
                    resumeData.messageHistory,
                  );
                  const backfilledItems: StaticItem[] = [];
                  for (const id of buffersRef.current.order) {
                    const ln = buffersRef.current.byId.get(id);
                    if (!ln) continue;
                    emittedIdsRef.current.add(id);
                    backfilledItems.push({ ...ln } as StaticItem);
                  }
                  const separator = {
                    kind: "separator" as const,
                    id: uid("sep"),
                  };
                  setStaticItems([separator, ...backfilledItems, successItem]);
                  setLines(toLines(buffersRef.current));
                  hasBackfilledRef.current = true;
                } else {
                  const separator = {
                    kind: "separator" as const,
                    id: uid("sep"),
                  };
                  setStaticItems([separator, successItem]);
                  setLines(toLines(buffersRef.current));
                }

                // Restore pending approvals if any (fixes #540 for /resume command)
                if (resumeData.pendingApprovals.length > 0) {
                  setPendingApprovals(resumeData.pendingApprovals);

                  // Analyze approval contexts (same logic as startup)
                  try {
                    const contexts = await Promise.all(
                      resumeData.pendingApprovals.map(async (approval) => {
                        const parsedArgs = safeJsonParseOr<
                          Record<string, unknown>
                        >(approval.toolArgs, {});
                        return await analyzeToolApproval(
                          approval.toolName,
                          parsedArgs,
                        );
                      }),
                    );
                    setApprovalContexts(contexts);
                  } catch (approvalError) {
                    // If analysis fails, leave context as null (will show basic options)
                    console.error(
                      "Failed to analyze resume approvals:",
                      approvalError,
                    );
                  }
                }
              }
            } catch (error) {
              // Update existing loading message instead of creating new one
              // Format error message to be user-friendly (avoid raw JSON/internal details)
              let errorMsg = "Unknown error";
              if (error instanceof APIError) {
                if (error.status === 404) {
                  errorMsg = "Conversation not found";
                } else if (error.status === 422) {
                  errorMsg = "Invalid conversation ID";
                } else {
                  errorMsg = error.message;
                }
              } else if (error instanceof Error) {
                errorMsg = error.message;
              }
              cmd.fail(`Failed to switch conversation: ${errorMsg}`);
            } finally {
              setCommandRunning(false);
            }
            return { submitted: true };
          }

          // No conversation ID provided - show selector
          startOverlayCommand(
            "conversations",
            "/resume",
            "Opening conversation selector...",
            "Conversation selector dismissed",
          );
          setActiveOverlay("conversations");
          return { submitted: true };
        }

        // Special handling for /search command - show message search
        if (trimmed.startsWith("/search")) {
          // Extract optional query after /search
          const [, ...rest] = trimmed.split(/\s+/);
          const query = rest.join(" ").trim();
          setSearchQuery(query);
          startOverlayCommand(
            "search",
            "/search",
            "Opening message search...",
            "Message search dismissed",
          );
          setActiveOverlay("search");
          return { submitted: true };
        }

        // Special handling for /profile command - manage local profiles
        if (msg.trim().startsWith("/profile")) {
          const parts = msg.trim().split(/\s+/);
          const subcommand = parts[1]?.toLowerCase();
          const profileName = parts.slice(2).join(" ");

          const profileCtx: ProfileCommandContext = {
            buffersRef,
            refreshDerived,
            agentId,
            agentName: agentName || "",
            setCommandRunning,
            updateAgentName,
          };

          // /profile - open agent browser (now points to /agents)
          if (!subcommand) {
            startOverlayCommand(
              "resume",
              "/profile",
              "Opening agent browser...",
              "Agent browser dismissed",
            );
            setActiveOverlay("resume");
            return { submitted: true };
          }

          const cmd = commandRunner.start(
            msg.trim(),
            "Running profile command...",
          );
          setActiveProfileCommandId(cmd.id);
          const clearProfileCommandId = () => setActiveProfileCommandId(null);

          // /profile save <name>
          if (subcommand === "save") {
            await handleProfileSave(profileCtx, msg, profileName);
            clearProfileCommandId();
            return { submitted: true };
          }

          // /profile load <name>
          if (subcommand === "load") {
            const validation = validateProfileLoad(
              profileCtx,
              msg,
              profileName,
            );
            if (validation.errorMessage) {
              clearProfileCommandId();
              return { submitted: true };
            }

            if (validation.needsConfirmation && validation.targetAgentId) {
              // Show warning and wait for confirmation
              const cmdId = addCommandResult(
                buffersRef,
                refreshDerived,
                msg,
                "Warning: Current agent is not saved to any profile.\nPress Enter to continue, or type anything to cancel.",
                false,
                "running",
              );
              setProfileConfirmPending({
                name: profileName,
                agentId: validation.targetAgentId,
                cmdId,
              });
              clearProfileCommandId();
              return { submitted: true };
            }

            // Current agent is saved, proceed with loading
            if (validation.targetAgentId) {
              await handleAgentSelect(validation.targetAgentId, {
                profileName,
                commandId: cmd.id,
              });
            }
            clearProfileCommandId();
            return { submitted: true };
          }

          // /profile delete <name>
          if (subcommand === "delete") {
            handleProfileDelete(profileCtx, msg, profileName);
            clearProfileCommandId();
            return { submitted: true };
          }

          // Unknown subcommand
          handleProfileUsage(profileCtx, msg);
          clearProfileCommandId();
          return { submitted: true };
        }

        // Special handling for /new command - create new agent dialog
        // Special handling for /pin command - pin current agent to project (or globally with -g)
        if (msg.trim() === "/pin" || msg.trim().startsWith("/pin ")) {
          const argsStr = msg.trim().slice(4).trim();

          // Parse args to check if name was provided
          const parts = argsStr.split(/\s+/).filter(Boolean);
          let hasNameArg = false;
          let isLocal = false;

          for (const part of parts) {
            if (part === "-l" || part === "--local") {
              isLocal = true;
            } else {
              hasNameArg = true;
            }
          }

          // If no name provided, show the pin dialog
          if (!hasNameArg) {
            setPinDialogLocal(isLocal);
            startOverlayCommand(
              "pin",
              "/pin",
              "Opening pin dialog...",
              "Pin dialog dismissed",
            );
            setActiveOverlay("pin");
            return { submitted: true };
          }

          // Name was provided, use existing behavior
          const profileCtx: ProfileCommandContext = {
            buffersRef,
            refreshDerived,
            agentId,
            agentName: agentName || "",
            setCommandRunning,
            updateAgentName,
          };
          {
            const cmd = commandRunner.start(msg.trim(), "Pinning agent...");
            setActiveProfileCommandId(cmd.id);
            try {
              await handlePin(profileCtx, msg, argsStr);
            } finally {
              setActiveProfileCommandId(null);
            }
          }
          return { submitted: true };
        }

        // Special handling for /unpin command - unpin current agent from project (or globally with -g)
        if (msg.trim() === "/unpin" || msg.trim().startsWith("/unpin ")) {
          const profileCtx: ProfileCommandContext = {
            buffersRef,
            refreshDerived,
            agentId,
            agentName: agentName || "",
            setCommandRunning,
            updateAgentName,
          };
          const argsStr = msg.trim().slice(6).trim();
          {
            const cmd = commandRunner.start(msg.trim(), "Unpinning agent...");
            setActiveProfileCommandId(cmd.id);
            try {
              handleUnpin(profileCtx, msg, argsStr);
            } finally {
              setActiveProfileCommandId(null);
            }
          }
          return { submitted: true };
        }

        // Special handling for /bg command - show background shell processes
        if (msg.trim() === "/bg") {
          const { backgroundProcesses } = await import(
            "../tools/impl/process_manager"
          );
          const cmd = commandRunner.start(
            msg.trim(),
            "Checking background processes...",
          );

          let output: string;
          if (backgroundProcesses.size === 0) {
            output = "No background processes running";
          } else {
            const lines = ["Background processes:"];
            for (const [id, proc] of backgroundProcesses) {
              const status =
                proc.status === "running"
                  ? "running"
                  : proc.status === "completed"
                    ? `completed (exit ${proc.exitCode})`
                    : `failed (exit ${proc.exitCode})`;
              lines.push(`  ${id}: ${proc.command} [${status}]`);
            }
            output = lines.join("\n");
          }

          cmd.finish(output, true);
          return { submitted: true };
        }

        // Special handling for /export command (also accepts legacy /download)
        if (msg.trim() === "/export" || msg.trim() === "/download") {
          const cmd = commandRunner.start(
            msg.trim(),
            "Exporting agent file...",
          );

          setCommandRunning(true);

          try {
            const client = await getClient();

            // Build export parameters (include conversation_id if in specific conversation)
            const exportParams: { conversation_id?: string } = {};
            if (conversationId !== "default") {
              exportParams.conversation_id = conversationId;
            }

            // Package skills from agent/project/global directories
            const { packageSkills } = await import("../agent/export");
            const skills = await packageSkills(agentId);

            // Export agent with skills
            let fileContent: unknown;
            if (skills.length > 0) {
              // Use raw fetch with auth from settings
              const { settingsManager } = await import("../settings-manager");
              const { getServerUrl } = await import("../agent/client");
              const settings =
                await settingsManager.getSettingsWithSecureTokens();
              const apiKey =
                process.env.LETTA_API_KEY || settings.env?.LETTA_API_KEY;
              const baseUrl = getServerUrl();

              const body: Record<string, unknown> = {
                ...exportParams,
                skills,
              };

              const response = await fetch(
                `${baseUrl}/v1/agents/${agentId}/export`,
                {
                  method: "POST",
                  headers: {
                    Authorization: `Bearer ${apiKey}`,
                    "Content-Type": "application/json",
                  },
                  body: JSON.stringify(body),
                },
              );

              if (!response.ok) {
                throw new Error(`Export failed: ${response.statusText}`);
              }

              fileContent = await response.json();
            } else {
              // No skills to include, use SDK
              fileContent = await client.agents.exportFile(
                agentId,
                exportParams,
              );
            }

            // Generate filename
            const fileName = exportParams.conversation_id
              ? `${exportParams.conversation_id}.af`
              : `${agentId}.af`;

            writeFileSync(fileName, JSON.stringify(fileContent, null, 2));

            // Build success message
            let summary = `AgentFile exported to ${fileName}`;
            if (skills.length > 0) {
              summary += `\n📦 Included ${skills.length} skill(s): ${skills.map((s) => s.name).join(", ")}`;
            }

            cmd.finish(summary, true);
          } catch (error) {
            const errorDetails = formatErrorDetails(error, agentId);
            cmd.fail(`Failed: ${errorDetails}`);
          } finally {
            setCommandRunning(false);
          }
          return { submitted: true };
        }

        // Special handling for /memfs command - manage filesystem-backed memory
        if (trimmed.startsWith("/memfs")) {
          const [, subcommand] = trimmed.split(/\s+/);
          const cmd = commandRunner.start(
            msg.trim(),
            "Processing memfs command...",
          );
          const cmdId = cmd.id;

          if (!subcommand || subcommand === "help") {
            const output = [
              "memfs commands:",
              "- /memfs status  — show status",
              "- /memfs enable  — enable filesystem-backed memory",
              "- /memfs disable — disable filesystem-backed memory",
              "- /memfs sync    — sync blocks and files now",
              "- /memfs reset   — move local memfs to /tmp and recreate dirs",
            ].join("\n");
            cmd.finish(output, true);
            return { submitted: true };
          }

          if (subcommand === "status") {
            // Show status
            const enabled = settingsManager.isMemfsEnabled(agentId);
            let output: string;
            if (enabled) {
              const memoryDir = getMemoryFilesystemRoot(agentId);
              output = `Memory filesystem is enabled.\nPath: ${memoryDir}`;
            } else {
              output =
                "Memory filesystem is disabled. Run `/memfs enable` to enable.";
            }
            cmd.finish(output, true);
            return { submitted: true };
          }

          if (subcommand === "enable") {
            // memfs requires Letta Cloud (git memfs not supported on self-hosted)
            const serverUrl = getServerUrl();
            if (!serverUrl.includes("api.letta.com")) {
              cmd.fail(
                "Memory filesystem is only available on Letta Cloud (api.letta.com).",
              );
              return { submitted: true };
            }

            updateMemorySyncCommand(
              cmdId,
              "Enabling memory filesystem...",
              true,
              msg,
              true,
            );
            setCommandRunning(true);

            try {
              // 1. Detach memory tools from agent
              const { detachMemoryTools } = await import("../tools/toolset");
              await detachMemoryTools(agentId);

              // 2. Update settings
              settingsManager.setMemfsEnabled(agentId, true);

              // 3. Update system prompt to include memfs section
              const { updateAgentSystemPromptMemfs } = await import(
                "../agent/modify"
              );
              await updateAgentSystemPromptMemfs(agentId, true);

              // 4. Add git-memory-enabled tag and clone repo
              const { addGitMemoryTag, isGitRepo, cloneMemoryRepo } =
                await import("../agent/memoryGit");
              await addGitMemoryTag(agentId);
              if (!isGitRepo(agentId)) {
                await cloneMemoryRepo(agentId);
              }
              const memoryDir = getMemoryFilesystemRoot(agentId);
              updateMemorySyncCommand(
                cmdId,
                `Memory filesystem enabled (git-backed).\nPath: ${memoryDir}`,
                true,
                msg,
              );
            } catch (error) {
              const errorText =
                error instanceof Error ? error.message : String(error);
              updateMemorySyncCommand(
                cmdId,
                `Failed to enable memfs: ${errorText}`,
                false,
                msg,
              );
            } finally {
              setCommandRunning(false);
            }

            return { submitted: true };
          }

          if (subcommand === "sync") {
            // Check if memfs is enabled for this agent
            if (!settingsManager.isMemfsEnabled(agentId)) {
              cmd.fail(
                "Memory filesystem is disabled. Run `/memfs enable` first.",
              );
              return { submitted: true };
            }

            updateMemorySyncCommand(
              cmdId,
              "Pulling latest memory from server...",
              true,
              msg,
              true,
            );

            setCommandRunning(true);

            try {
              const { pullMemory } = await import("../agent/memoryGit");
              const result = await pullMemory(agentId);
              updateMemorySyncCommand(cmdId, result.summary, true, msg);
            } catch (error) {
              const errorText =
                error instanceof Error ? error.message : String(error);
              updateMemorySyncCommand(cmdId, `Failed: ${errorText}`, false);
            } finally {
              setCommandRunning(false);
            }

            return { submitted: true };
          }

          if (subcommand === "reset") {
            updateMemorySyncCommand(
              cmdId,
              "Resetting memory filesystem...",
              true,
              msg,
              true,
            );
            setCommandRunning(true);

            try {
              const memoryDir = getMemoryFilesystemRoot(agentId);
              if (!existsSync(memoryDir)) {
                updateMemorySyncCommand(
                  cmdId,
                  "No local memory filesystem found to reset.",
                  true,
                  msg,
                );
                return { submitted: true };
              }

              const backupDir = join(
                tmpdir(),
                `letta-memfs-reset-${agentId}-${Date.now()}`,
              );
              renameSync(memoryDir, backupDir);

              ensureMemoryFilesystemDirs(agentId);

              updateMemorySyncCommand(
                cmdId,
                `Memory filesystem reset.\nBackup moved to ${backupDir}\nRun \`/memfs sync\` to repopulate from API.`,
                true,
                msg,
              );
            } catch (error) {
              const errorText =
                error instanceof Error ? error.message : String(error);
              updateMemorySyncCommand(
                cmdId,
                `Failed to reset memfs: ${errorText}`,
                false,
                msg,
              );
            } finally {
              setCommandRunning(false);
            }

            return { submitted: true };
          }

          if (subcommand === "disable") {
            updateMemorySyncCommand(
              cmdId,
              "Disabling memory filesystem...",
              true,
              msg,
              true,
            );
            setCommandRunning(true);

            try {
              // 1. Re-attach memory tool
              const { reattachMemoryTool } = await import("../tools/toolset");
              const modelId = currentModelId || "anthropic/claude-sonnet-4";
              await reattachMemoryTool(agentId, modelId);

              // 2. Update system prompt to remove memfs section
              const { updateAgentSystemPromptMemfs } = await import(
                "../agent/modify"
              );
              await updateAgentSystemPromptMemfs(agentId, false);

              // 3. Update settings
              settingsManager.setMemfsEnabled(agentId, false);

              // 4. Remove git-memory-enabled tag from agent
              const { removeGitMemoryTag } = await import("../agent/memoryGit");
              await removeGitMemoryTag(agentId);

              // 5. Move local memory dir to /tmp (backup, not delete)
              let backupInfo = "";
              const memoryDir = getMemoryFilesystemRoot(agentId);
              if (existsSync(memoryDir)) {
                const backupDir = join(
                  tmpdir(),
                  `letta-memfs-disable-${agentId}-${Date.now()}`,
                );
                renameSync(memoryDir, backupDir);
                backupInfo = `\nLocal files backed up to ${backupDir}`;
              }

              updateMemorySyncCommand(
                cmdId,
                `Memory filesystem disabled. Memory tool re-attached.${backupInfo}`,
                true,
                msg,
              );
            } catch (error) {
              const errorText =
                error instanceof Error ? error.message : String(error);
              updateMemorySyncCommand(
                cmdId,
                `Failed to disable memfs: ${errorText}`,
                false,
                msg,
              );
            } finally {
              setCommandRunning(false);
            }

            return { submitted: true };
          }

          // Unknown subcommand
          cmd.fail(
            `Unknown subcommand: ${subcommand}. Use /memfs, /memfs enable, /memfs disable, /memfs sync, or /memfs reset.`,
          );
          return { submitted: true };
        }

        // Special handling for /skill command - enter skill creation mode
        if (trimmed.startsWith("/skill")) {
          // Extract optional description after `/skill`
          const [, ...rest] = trimmed.split(/\s+/);
          const description = rest.join(" ").trim();

          const initialOutput = description
            ? `Starting skill creation for: ${description}`
            : "Starting skill creation. I’ll load the creating-skills skill and ask a few questions about the skill you want to build...";

          const cmd = commandRunner.start(msg, initialOutput);

          // Check for pending approvals before sending
          const approvalCheck = await checkPendingApprovalsForSlashCommand();
          if (approvalCheck.blocked) {
            cmd.fail(
              "Pending approval(s). Resolve approvals before running /skill.",
            );
            return { submitted: false }; // Keep /skill in input box, user handles approval first
          }

          setCommandRunning(true);

          try {
            // Import the skill-creation prompt
            const { SKILL_CREATOR_PROMPT } = await import(
              "../agent/promptAssets.js"
            );

            // Build system-reminder content for skill creation
            const userDescriptionLine = description
              ? `\n\nUser-provided skill description:\n${description}`
              : "\n\nThe user did not provide a description with /skill. Ask what kind of skill they want to create before proceeding.";

            const skillMessage = `${SYSTEM_REMINDER_OPEN}\n${SKILL_CREATOR_PROMPT}${userDescriptionLine}\n${SYSTEM_REMINDER_CLOSE}`;

            // Mark command as finished before sending message
            cmd.finish(
              "Entered skill creation mode. Answer the assistant’s questions to design your new skill.",
              true,
            );

            // Process conversation with the skill-creation prompt
            await processConversation([
              {
                type: "message",
                role: "user",
                content: buildTextParts(skillMessage),
              },
            ]);
          } catch (error) {
            const errorDetails = formatErrorDetails(error, agentId);
            cmd.fail(`Failed: ${errorDetails}`);
          } finally {
            setCommandRunning(false);
          }

          return { submitted: true };
        }

        // Special handling for /remember command - remember something from conversation
        if (trimmed.startsWith("/remember")) {
          // Extract optional description after `/remember`
          const [, ...rest] = trimmed.split(/\s+/);
          const userText = rest.join(" ").trim();

          const initialOutput = userText
            ? "Storing to memory..."
            : "Processing memory request...";

          const cmd = commandRunner.start(msg, initialOutput);

          // Check for pending approvals before sending (mirrors regular message flow)
          const approvalCheck = await checkPendingApprovalsForSlashCommand();
          if (approvalCheck.blocked) {
            cmd.fail(
              "Pending approval(s). Resolve approvals before running /remember.",
            );
            return { submitted: false }; // Keep /remember in input box, user handles approval first
          }

          setCommandRunning(true);

          try {
            // Import the remember prompt
            const { REMEMBER_PROMPT } = await import(
              "../agent/promptAssets.js"
            );

            // Build system-reminder content for memory request
            const rememberReminder = userText
              ? `${SYSTEM_REMINDER_OPEN}\n${REMEMBER_PROMPT}\n${SYSTEM_REMINDER_CLOSE}`
              : `${SYSTEM_REMINDER_OPEN}\n${REMEMBER_PROMPT}\n\nThe user did not specify what to remember. Look at the recent conversation context to identify what they likely want you to remember, or ask them to clarify.\n${SYSTEM_REMINDER_CLOSE}`;
            const rememberParts = userText
              ? buildTextParts(rememberReminder, userText)
              : buildTextParts(rememberReminder);

            // Mark command as finished before sending message
            cmd.finish(
              userText
                ? "Storing to memory..."
                : "Processing memory request from conversation context...",
              true,
            );

            // Process conversation with the remember prompt
            await processConversation([
              {
                type: "message",
                role: "user",
                content: rememberParts,
              },
            ]);
          } catch (error) {
            const errorDetails = formatErrorDetails(error, agentId);
            cmd.fail(`Failed: ${errorDetails}`);
          } finally {
            setCommandRunning(false);
          }

          return { submitted: true };
        }

        // Special handling for /plan command - enter plan mode
        if (trimmed === "/plan") {
          // Generate plan file path and enter plan mode
          const planPath = generatePlanFilePath();
          permissionMode.setPlanFilePath(planPath);
          permissionMode.setMode("plan");
          setUiPermissionMode("plan");

          const cmd = commandRunner.start(
            "/plan",
            `Plan mode enabled. Plan file: ${planPath}`,
          );
          cmd.finish(`Plan mode enabled. Plan file: ${planPath}`, true);

          return { submitted: true };
        }

        // Special handling for /init command - initialize agent memory
        if (trimmed === "/init") {
          const cmd = commandRunner.start(msg, "Gathering project context...");

          // Check for pending approvals before sending
          const approvalCheck = await checkPendingApprovalsForSlashCommand();
          if (approvalCheck.blocked) {
            cmd.fail(
              "Pending approval(s). Resolve approvals before running /init.",
            );
            return { submitted: false }; // Keep /init in input box, user handles approval first
          }

          setCommandRunning(true);

          try {
            // Gather git context if available
            let gitContext = "";
            try {
              const { execSync } = await import("node:child_process");
              const cwd = process.cwd();

              // Check if we're in a git repo
              try {
                execSync("git rev-parse --git-dir", {
                  cwd,
                  stdio: "pipe",
                });

                // Gather git info
                const branch = execSync("git branch --show-current", {
                  cwd,
                  encoding: "utf-8",
                }).trim();
                const mainBranch = execSync(
                  "git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@' || echo 'main'",
                  { cwd, encoding: "utf-8", shell: "/bin/bash" },
                ).trim();
                const status = execSync("git status --short", {
                  cwd,
                  encoding: "utf-8",
                }).trim();
                const recentCommits = execSync(
                  "git log --oneline -10 2>/dev/null || echo 'No commits yet'",
                  { cwd, encoding: "utf-8" },
                ).trim();

                gitContext = `
## Current Project Context

**Working directory**: ${cwd}

### Git Status
- **Current branch**: ${branch}
- **Main branch**: ${mainBranch}
- **Status**:
${status || "(clean working tree)"}

### Recent Commits
${recentCommits}
`;
              } catch {
                // Not a git repo, just include working directory
                gitContext = `
## Current Project Context

**Working directory**: ${cwd}
**Git**: Not a git repository
`;
              }
            } catch {
              // execSync import failed, skip git context
            }

            // Mark command as finished before sending message
            cmd.finish(
              "Assimilating project context and defragmenting memories...",
              true,
            );

            // Send trigger message instructing agent to load the initializing-memory skill
            // Only include memfs path if memfs is enabled for this agent
            const memfsSection = settingsManager.isMemfsEnabled(agentId)
              ? `
## Memory Filesystem Location

Your memory blocks are synchronized with the filesystem at:
\`~/.letta/agents/${agentId}/memory/\`

Use this path when working with memory files during initialization.
`
              : "";

            const initMessage = `${SYSTEM_REMINDER_OPEN}
The user has requested memory initialization via /init.
${memfsSection}
## 1. Invoke the initializing-memory skill

Use the \`Skill\` tool with \`skill: "initializing-memory"\` to load the comprehensive instructions for memory initialization.

If the skill fails to invoke, proceed with your best judgment based on these guidelines:
- Ask upfront questions (research depth, identity, related repos, workflow style)
- Research the project based on chosen depth
- Create/update memory blocks incrementally
- Reflect and verify completeness

## 2. Follow the skill instructions

Once invoked, follow the instructions from the \`initializing-memory\` skill to complete the initialization.
${gitContext}
${SYSTEM_REMINDER_CLOSE}`;

            // Process conversation with the init prompt
            await processConversation([
              {
                type: "message",
                role: "user",
                content: buildTextParts(initMessage),
              },
            ]);
          } catch (error) {
            const errorDetails = formatErrorDetails(error, agentId);
            cmd.fail(`Failed: ${errorDetails}`);
          } finally {
            setCommandRunning(false);
          }
          return { submitted: true };
        }

        if (trimmed.startsWith("/feedback")) {
          const maybeMsg = msg.slice("/feedback".length).trim();
          setFeedbackPrefill(maybeMsg);
          startOverlayCommand(
            "feedback",
            "/feedback",
            "Opening feedback dialog...",
            "Feedback dialog dismissed",
          );
          setActiveOverlay("feedback");
          return { submitted: true };
        }

        // === Custom command handling ===
        // Check BEFORE falling through to executeCommand()
        const { findCustomCommand, substituteArguments, expandBashCommands } =
          await import("./commands/custom.js");
        const customCommandName = trimmed.split(/\s+/)[0]?.slice(1) || ""; // e.g., "review" from "/review arg"
        const matchedCustom = await findCustomCommand(customCommandName);

        if (matchedCustom) {
          const cmd = commandRunner.start(
            trimmed,
            `Running /${matchedCustom.id}...`,
          );

          // Check for pending approvals before sending
          const approvalCheck = await checkPendingApprovalsForSlashCommand();
          if (approvalCheck.blocked) {
            cmd.fail(
              `Pending approval(s). Resolve approvals before running /${matchedCustom.id}.`,
            );
            return { submitted: false }; // Keep custom command in input box, user handles approval first
          }

          // Extract arguments (everything after command name)
          const args = trimmed.slice(`/${matchedCustom.id}`.length).trim();

          // Build prompt: 1) substitute args, 2) expand bash commands
          let prompt = substituteArguments(matchedCustom.content, args);
          prompt = await expandBashCommands(prompt);

          // Show command in transcript (running phase for visual feedback)
          setCommandRunning(true);

          try {
            // Mark command as finished BEFORE sending to agent
            // (matches /remember pattern - command succeeded in triggering agent)
            cmd.finish("Running custom command...", true);

            // Send prompt to agent
            // NOTE: Unlike /remember, we DON'T append args separately because
            // they're already substituted into the prompt via $ARGUMENTS
            await processConversation([
              {
                type: "message",
                role: "user",
                content: buildTextParts(
                  `${SYSTEM_REMINDER_OPEN}\n${prompt}\n${SYSTEM_REMINDER_CLOSE}`,
                ),
              },
            ]);
          } catch (error) {
            // Only catch errors from processConversation setup, not agent execution
            const errorDetails = formatErrorDetails(error, agentId);
            cmd.fail(`Failed to run command: ${errorDetails}`);
          } finally {
            setCommandRunning(false);
          }

          return { submitted: true };
        }
        // === END custom command handling ===

        // Check if this is a known command before treating it as a slash command
        const { commands, executeCommand } = await import(
          "./commands/registry"
        );
        const registryCommandName = trimmed.split(/\s+/)[0] ?? "";
        const isRegistryCommand = Boolean(commands[registryCommandName]);
        const registryCmd = isRegistryCommand
          ? commandRunner.start(msg, `Running ${registryCommandName}...`)
          : null;
        const result = await executeCommand(aliasedMsg);

        // If command not found, fall through to send as regular message to agent
        if (result.notFound) {
          if (registryCmd) {
            registryCmd.fail(`Unknown command: ${registryCommandName}`);
          }
          // Don't treat as command - continue to regular message handling below
        } else {
          // Known command - show in transcript and handle result
          if (registryCmd) {
            registryCmd.finish(result.output, result.success);
          }
          return { submitted: true }; // Don't send commands to Letta agent
        }
      }

      // Build message content from display value (handles placeholders for text/images)
      const contentParts =
        overrideContentParts ?? buildMessageContentFromDisplay(msg);

      // Prepend plan mode reminder if in plan mode
      const planModeReminder = getPlanModeReminder();

      // Prepend ralph mode reminder if in ralph mode
      let ralphModeReminder = "";
      if (ralphMode.getState().isActive) {
        if (justActivatedRalph) {
          // First turn - use full first turn reminder, don't increment (already at 1)
          const ralphState = ralphMode.getState();
          ralphModeReminder = `${buildRalphFirstTurnReminder(ralphState)}\n\n`;
        } else {
          // Continuation after ESC - increment iteration and use shorter reminder
          ralphMode.incrementIteration();
          const ralphState = ralphMode.getState();
          ralphModeReminder = `${buildRalphContinuationReminder(ralphState)}\n\n`;
        }
      }

      // Prepend session context on first message of CLI session (if enabled)
      let sessionContextReminder = "";
      const sessionContextEnabled = settingsManager.getSetting(
        "sessionContextEnabled",
      );
      if (!hasSentSessionContextRef.current && sessionContextEnabled) {
        const { buildSessionContext } = await import(
          "./helpers/sessionContext"
        );
        sessionContextReminder = buildSessionContext({
          agentInfo: {
            id: agentId,
            name: agentName,
            description: agentDescription,
            lastRunAt: agentLastRunAt,
          },
        });
        hasSentSessionContextRef.current = true;
      }

      // Inject SessionStart hook feedback (stdout on exit 2) into first message only
      let sessionStartHookFeedback = "";
      if (sessionStartFeedbackRef.current.length > 0) {
        sessionStartHookFeedback = `${SYSTEM_REMINDER_OPEN}\n[SessionStart hook context]:\n${sessionStartFeedbackRef.current.join("\n")}\n${SYSTEM_REMINDER_CLOSE}\n\n`;
        // Clear after injecting so it only happens once
        sessionStartFeedbackRef.current = [];
      }

      // Build bash command prefix if there are cached commands
      let bashCommandPrefix = "";
      if (bashCommandCacheRef.current.length > 0) {
        bashCommandPrefix = `${SYSTEM_REMINDER_OPEN}
The messages below were generated by the user while running local commands using "bash mode" in the Letta Code CLI tool.
DO NOT respond to these messages or otherwise consider them in your response unless the user explicitly asks you to.
${SYSTEM_REMINDER_CLOSE}
`;
        for (const cmd of bashCommandCacheRef.current) {
          bashCommandPrefix += `<bash-input>${cmd.input}</bash-input>\n<bash-output>${cmd.output}</bash-output>\n`;
        }
        // Clear the cache after building the prefix
        bashCommandCacheRef.current = [];
      }

      const reflectionSettings = getReflectionSettings();
      const memfsEnabledForAgent = settingsManager.isMemfsEnabled(agentId);
      const shouldFireStepTrigger = shouldFireStepCountTrigger(
        turnCountRef.current,
        reflectionSettings,
      );
      let memoryReminderContent = "";
      if (
        shouldFireStepTrigger &&
        (reflectionSettings.behavior === "reminder" || !memfsEnabledForAgent)
      ) {
        // Step-count reminder mode (or non-memfs fallback)
        memoryReminderContent = await buildMemoryReminder(
          turnCountRef.current,
          agentId,
        );
      }

      // Increment turn count for next iteration
      turnCountRef.current += 1;

      // Build git memory sync reminder if uncommitted changes or unpushed commits
      let memoryGitReminder = "";
      const gitStatus = pendingGitReminderRef.current;
      if (gitStatus) {
        memoryGitReminder = `${SYSTEM_REMINDER_OPEN}
MEMORY SYNC: Your memory directory has uncommitted changes or is ahead of the remote.

${gitStatus.summary}

Sync when convenient by running these commands:
\`\`\`bash
cd ~/.letta/agents/${agentId}/memory
git add system/
git commit -m "<type>: <what changed>"
git push
\`\`\`

You should do this soon to avoid losing memory updates. It only takes a few seconds.
${SYSTEM_REMINDER_CLOSE}
`;
        // Clear after injecting so it doesn't repeat
        pendingGitReminderRef.current = null;
      }

      // Build permission mode change alert if mode changed since last notification
      let permissionModeAlert = "";
      const currentMode = permissionMode.getMode();
      if (currentMode !== lastNotifiedModeRef.current) {
        const modeDescriptions: Record<PermissionMode, string> = {
          default: "Normal approval flow.",
          acceptEdits: "File edits auto-approved.",
          plan: "Read-only mode. Focus on exploration and planning.",
          bypassPermissions: "All tools auto-approved. Bias toward action.",
        };
        permissionModeAlert = `${SYSTEM_REMINDER_OPEN}Permission mode changed to: ${currentMode}. ${modeDescriptions[currentMode]}${SYSTEM_REMINDER_CLOSE}\n\n`;
        lastNotifiedModeRef.current = currentMode;
      }

      // Combine reminders with content as separate text parts.
      // This preserves each reminder boundary in the API payload.
      // Note: Task notifications now come through messageQueue directly (added by messageQueueBridge)
      const reminderParts: Array<{ type: "text"; text: string }> = [];
      const pushReminder = (text: string) => {
        if (!text) return;
        reminderParts.push({ type: "text", text });
      };
      const maybeLaunchReflectionSubagent = async (
        triggerSource: "step-count" | "compaction-event",
      ) => {
        if (!memfsEnabledForAgent) {
          return false;
        }
        if (hasActiveReflectionSubagent()) {
          debugLog(
            "memory",
            `Skipping auto reflection launch (${triggerSource}) because one is already active`,
          );
          return false;
        }
        try {
          const { spawnBackgroundSubagentTask } = await import(
            "../tools/impl/Task"
          );
          spawnBackgroundSubagentTask({
            subagentType: "reflection",
            prompt: AUTO_REFLECTION_PROMPT,
            description: AUTO_REFLECTION_DESCRIPTION,
          });
          debugLog(
            "memory",
            `Auto-launched reflection subagent (${triggerSource})`,
          );
          return true;
        } catch (error) {
          debugWarn(
            "memory",
            `Failed to auto-launch reflection subagent (${triggerSource}): ${
              error instanceof Error ? error.message : String(error)
            }`,
          );
          return false;
        }
      };
      pushReminder(sessionContextReminder);

      // Inject available skills as system-reminder (LET-7353)
      // Lazy-discover on first message, reinject after compaction
      {
        if (!discoveredSkillsRef.current) {
          try {
            const { discoverSkills: discover, SKILLS_DIR: defaultDir } =
              await import("../agent/skills");
            const { getSkillsDirectory } = await import("../agent/context");
            const skillsDir =
              getSkillsDirectory() || join(process.cwd(), defaultDir);
            const { skills } = await discover(skillsDir, agentId);
            discoveredSkillsRef.current = skills;
          } catch {
            discoveredSkillsRef.current = [];
          }
        }

        const needsSkillsReinject =
          contextTrackerRef.current.pendingSkillsReinject;
        if (!hasInjectedSkillsRef.current || needsSkillsReinject) {
          const { formatSkillsAsSystemReminder } = await import(
            "../agent/skills"
          );
          const skillsReminder = formatSkillsAsSystemReminder(
            discoveredSkillsRef.current,
          );
          if (skillsReminder) {
            pushReminder(skillsReminder);
          }
          hasInjectedSkillsRef.current = true;
          contextTrackerRef.current.pendingSkillsReinject = false;
        }
      }

      // Build conversation switch alert if a switch is pending (behind feature flag)
      let conversationSwitchAlert = "";
      if (
        pendingConversationSwitchRef.current &&
        settingsManager.getSetting("conversationSwitchAlertEnabled")
      ) {
        const { buildConversationSwitchAlert } = await import(
          "./helpers/conversationSwitchAlert"
        );
        conversationSwitchAlert = buildConversationSwitchAlert(
          pendingConversationSwitchRef.current,
        );
      }
      pendingConversationSwitchRef.current = null;

      pushReminder(sessionStartHookFeedback);
      pushReminder(permissionModeAlert);
      pushReminder(conversationSwitchAlert);
      pushReminder(planModeReminder);
      pushReminder(ralphModeReminder);

      pushReminder(bashCommandPrefix);
      pushReminder(userPromptSubmitHookFeedback);
      pushReminder(memoryReminderContent);

      // Step-count auto-launch mode: fire reflection in background on interval.
      if (
        shouldFireStepTrigger &&
        reflectionSettings.trigger === "step-count" &&
        reflectionSettings.behavior === "auto-launch"
      ) {
        await maybeLaunchReflectionSubagent("step-count");
      }

      // Consume compaction-triggered reflection behavior on next user turn.
      if (contextTrackerRef.current.pendingReflectionTrigger) {
        contextTrackerRef.current.pendingReflectionTrigger = false;
        if (reflectionSettings.trigger === "compaction-event") {
          if (
            reflectionSettings.behavior === "auto-launch" &&
            memfsEnabledForAgent
          ) {
            await maybeLaunchReflectionSubagent("compaction-event");
          } else {
            const compactionReminderContent =
              await buildCompactionMemoryReminder(agentId);
            pushReminder(compactionReminderContent);
          }
        }
      }

      pushReminder(memoryGitReminder);
      const messageContent =
        reminderParts.length > 0
          ? [...reminderParts, ...contentParts]
          : contentParts;

      // Append task notifications (if any) as event lines before the user message
      appendTaskNotificationEvents(taskNotifications);

      // Append the user message to transcript IMMEDIATELY (optimistic update)
      const userId = uid("user");
      if (userTextForInput) {
        buffersRef.current.byId.set(userId, {
          kind: "user",
          id: userId,
          text: userTextForInput,
        });
        buffersRef.current.order.push(userId);
      }

      // Reset token counter for this turn (only count the agent's response)
      buffersRef.current.tokenCount = 0;
      // If the previous trajectory ended, ensure the live token display resets.
      if (!sessionStatsRef.current.getTrajectorySnapshot()) {
        trajectoryTokenDisplayRef.current = 0;
        setTrajectoryTokenBase(0);
        trajectoryRunTokenStartRef.current = 0;
      }
      // Clear interrupted flag from previous turn
      buffersRef.current.interrupted = false;
      // Rotate to a new thinking message for this turn
      setThinkingMessage(getRandomThinkingVerb());
      // Show streaming state immediately for responsiveness (pending approval check takes ~100ms)
      setStreaming(true);
      openTrajectorySegment();
      refreshDerived();

      // Check for pending approvals before sending message (skip if we already have
      // a queued approval response to send first).
      // Only do eager check when resuming a session (LET-7101) - otherwise lazy recovery handles it
      if (needsEagerApprovalCheck && !queuedApprovalResults) {
        // Log for debugging
        const eagerStatusId = uid("status");
        buffersRef.current.byId.set(eagerStatusId, {
          kind: "status",
          id: eagerStatusId,
          lines: [
            "[EAGER CHECK] Checking for pending approvals (resume mode)...",
          ],
        });
        buffersRef.current.order.push(eagerStatusId);
        refreshDerived();

        try {
          const client = await getClient();
          // Fetch fresh agent state to check for pending approvals with accurate in-context messages
          const agent = await client.agents.retrieve(agentId);
          const { pendingApprovals: existingApprovals } = await getResumeData(
            client,
            agent,
            conversationIdRef.current,
          );

          // Remove eager check status
          buffersRef.current.byId.delete(eagerStatusId);
          buffersRef.current.order = buffersRef.current.order.filter(
            (id) => id !== eagerStatusId,
          );

          // Check if user cancelled while we were fetching approval state
          if (
            userCancelledRef.current ||
            abortControllerRef.current?.signal.aborted
          ) {
            // User hit ESC during the check - abort and clean up
            buffersRef.current.byId.delete(userId);
            const orderIndex = buffersRef.current.order.indexOf(userId);
            if (orderIndex !== -1) {
              buffersRef.current.order.splice(orderIndex, 1);
            }
            setStreaming(false);
            refreshDerived();
            return { submitted: false };
          }

          if (existingApprovals && existingApprovals.length > 0) {
            // There are pending approvals - check permissions first (respects yolo mode)
            const { needsUserInput, autoAllowed, autoDenied } =
              await classifyApprovals(existingApprovals, {
                getContext: analyzeToolApproval,
                alwaysRequiresUserInput,
                missingNameReason: "Tool call incomplete - missing name",
              });

            // Check if user cancelled during permission check
            if (
              userCancelledRef.current ||
              abortControllerRef.current?.signal.aborted
            ) {
              buffersRef.current.byId.delete(userId);
              const orderIndex = buffersRef.current.order.indexOf(userId);
              if (orderIndex !== -1) {
                buffersRef.current.order.splice(orderIndex, 1);
              }
              setStreaming(false);
              refreshDerived();
              return { submitted: false };
            }

            // If all approvals can be auto-handled (yolo mode), process them immediately
            if (needsUserInput.length === 0) {
              // Precompute diffs for file edit tools before execution (both auto-allowed and needs-user-input)
              for (const ac of [...autoAllowed, ...needsUserInput]) {
                const toolName = ac.approval.toolName;
                const toolCallId = ac.approval.toolCallId;
                try {
                  const args = JSON.parse(ac.approval.toolArgs || "{}");

                  if (isFileWriteTool(toolName)) {
                    const filePath = args.file_path as string | undefined;
                    if (filePath) {
                      const result = computeAdvancedDiff({
                        kind: "write",
                        filePath,
                        content: (args.content as string) || "",
                      });
                      if (result.mode === "advanced") {
                        precomputedDiffsRef.current.set(toolCallId, result);
                      }
                    }
                  } else if (isFileEditTool(toolName)) {
                    const filePath = args.file_path as string | undefined;
                    if (filePath) {
                      // Check if it's a multi-edit (has edits array) or single edit
                      if (args.edits && Array.isArray(args.edits)) {
                        const result = computeAdvancedDiff({
                          kind: "multi_edit",
                          filePath,
                          edits: args.edits as Array<{
                            old_string: string;
                            new_string: string;
                            replace_all?: boolean;
                          }>,
                        });
                        if (result.mode === "advanced") {
                          precomputedDiffsRef.current.set(toolCallId, result);
                        }
                      } else {
                        const result = computeAdvancedDiff({
                          kind: "edit",
                          filePath,
                          oldString: (args.old_string as string) || "",
                          newString: (args.new_string as string) || "",
                          replaceAll: args.replace_all as boolean | undefined,
                        });
                        if (result.mode === "advanced") {
                          precomputedDiffsRef.current.set(toolCallId, result);
                        }
                      }
                    }
                  } else if (isPatchTool(toolName) && args.input) {
                    // Patch tools - parse hunks directly (patches ARE diffs)
                    const operations = parsePatchOperations(
                      args.input as string,
                    );
                    for (const op of operations) {
                      const key = `${toolCallId}:${op.path}`;
                      if (op.kind === "add" || op.kind === "update") {
                        const result = parsePatchToAdvancedDiff(
                          op.patchLines,
                          op.path,
                        );
                        if (result) {
                          precomputedDiffsRef.current.set(key, result);
                        }
                      }
                      // Delete operations don't need diffs
                    }
                  }
                } catch {
                  // Ignore errors in diff computation for auto-allowed tools
                }
              }

              const autoAllowedToolCallIds = autoAllowed.map(
                (ac) => ac.approval.toolCallId,
              );
              const autoAllowedAbortController =
                abortControllerRef.current ?? new AbortController();
              const shouldTrackAutoAllowed = autoAllowedToolCallIds.length > 0;
              let autoAllowedResults: Array<{
                toolCallId: string;
                result: ToolExecutionResult;
              }> = [];
              let autoDeniedResults: ApprovalResult[] = [];

              if (shouldTrackAutoAllowed) {
                setIsExecutingTool(true);
                executingToolCallIdsRef.current = autoAllowedToolCallIds;
                toolAbortControllerRef.current = autoAllowedAbortController;
                autoAllowedExecutionRef.current = {
                  toolCallIds: autoAllowedToolCallIds,
                  results: null,
                  conversationId: conversationIdRef.current,
                  generation: conversationGenerationRef.current,
                };
              }

              try {
                if (autoAllowedToolCallIds.length > 0) {
                  // Set phase to "running" for auto-allowed tools
                  setToolCallsRunning(
                    buffersRef.current,
                    autoAllowedToolCallIds,
                  );
                  refreshDerived();
                }

                // Execute auto-allowed tools (sequential for writes, parallel for reads)
                autoAllowedResults =
                  autoAllowed.length > 0
                    ? await executeAutoAllowedTools(
                        autoAllowed,
                        (chunk) => onChunk(buffersRef.current, chunk),
                        {
                          abortSignal: autoAllowedAbortController.signal,
                          onStreamingOutput: updateStreamingOutput,
                        },
                      )
                    : [];

                // Create denial results for auto-denied and update UI
                autoDeniedResults = autoDenied.map((ac) => {
                  // Prefer the detailed reason over the short matchedRule name
                  const reason = ac.permission.reason
                    ? `Permission denied: ${ac.permission.reason}`
                    : "matchedRule" in ac.permission &&
                        ac.permission.matchedRule
                      ? `Permission denied by rule: ${ac.permission.matchedRule}`
                      : "Permission denied: Unknown";

                  // Update buffers with denial for UI
                  onChunk(buffersRef.current, {
                    message_type: "tool_return_message",
                    id: "dummy",
                    date: new Date().toISOString(),
                    tool_call_id: ac.approval.toolCallId,
                    tool_return: `Error: request to call tool denied. User reason: ${reason}`,
                    status: "error",
                    stdout: null,
                    stderr: null,
                  });

                  return {
                    type: "approval" as const,
                    tool_call_id: ac.approval.toolCallId,
                    approve: false,
                    reason,
                  };
                });

                const queuedResults: ApprovalResult[] = [
                  ...autoAllowedResults.map((ar) => ({
                    type: "tool" as const,
                    tool_call_id: ar.toolCallId,
                    tool_return: ar.result.toolReturn,
                    status: ar.result.status,
                    stdout: ar.result.stdout,
                    stderr: ar.result.stderr,
                  })),
                  ...autoDeniedResults,
                ];

                if (autoAllowedExecutionRef.current) {
                  autoAllowedExecutionRef.current.results = queuedResults;
                }
                const autoAllowedMetadata = autoAllowedExecutionRef.current
                  ? {
                      conversationId:
                        autoAllowedExecutionRef.current.conversationId,
                      generation: conversationGenerationRef.current,
                    }
                  : undefined;

                if (
                  userCancelledRef.current ||
                  autoAllowedAbortController.signal.aborted ||
                  interruptQueuedRef.current
                ) {
                  if (queuedResults.length > 0) {
                    queueApprovalResults(queuedResults, autoAllowedMetadata);
                  }
                  setStreaming(false);
                  markIncompleteToolsAsCancelled(
                    buffersRef.current,
                    true,
                    "user_interrupt",
                  );
                  refreshDerived();
                  return { submitted: false };
                }

                refreshDerived();

                // Combine results and send directly with the user's message
                // (can't use state here as it won't be available until next render)
                const recoveryApprovalResults = [
                  ...autoAllowedResults.map((ar) => ({
                    type: "approval" as const,
                    tool_call_id: ar.toolCallId,
                    approve: true,
                    tool_return: ar.result.toolReturn,
                  })),
                  ...autoDeniedResults,
                ];

                // Build and send initialInput directly
                const initialInput: Array<MessageCreate | ApprovalCreate> = [
                  {
                    type: "approval",
                    approvals: recoveryApprovalResults,
                  },
                  {
                    type: "message",
                    role: "user",
                    content:
                      messageContent as unknown as MessageCreate["content"],
                  },
                ];

                toolResultsInFlightRef.current = true;
                await processConversation(initialInput);
                toolResultsInFlightRef.current = false;
                clearPlaceholdersInText(msg);
                return { submitted: true };
              } finally {
                if (shouldTrackAutoAllowed) {
                  setIsExecutingTool(false);
                  toolAbortControllerRef.current = null;
                  executingToolCallIdsRef.current = [];
                  autoAllowedExecutionRef.current = null;
                  toolResultsInFlightRef.current = false;
                }
              }
            } else {
              // Some approvals need user input - show dialog
              // Remove the optimistic user message from transcript
              buffersRef.current.byId.delete(userId);
              const orderIndex = buffersRef.current.order.indexOf(userId);
              if (orderIndex !== -1) {
                buffersRef.current.order.splice(orderIndex, 1);
              }

              setStreaming(false);
              setPendingApprovals(needsUserInput.map((ac) => ac.approval));
              setApprovalContexts(
                needsUserInput
                  .map((ac) => ac.context)
                  .filter(Boolean) as ApprovalContext[],
              );

              // Precompute diffs for file edit tools before execution (both auto-allowed and needs-user-input)
              for (const ac of [...autoAllowed, ...needsUserInput]) {
                const toolName = ac.approval.toolName;
                const toolCallId = ac.approval.toolCallId;
                try {
                  const args = JSON.parse(ac.approval.toolArgs || "{}");

                  if (isFileWriteTool(toolName)) {
                    const filePath = args.file_path as string | undefined;
                    if (filePath) {
                      const result = computeAdvancedDiff({
                        kind: "write",
                        filePath,
                        content: (args.content as string) || "",
                      });
                      if (result.mode === "advanced") {
                        precomputedDiffsRef.current.set(toolCallId, result);
                      }
                    }
                  } else if (isFileEditTool(toolName)) {
                    const filePath = args.file_path as string | undefined;
                    if (filePath) {
                      // Check if it's a multi-edit (has edits array) or single edit
                      if (args.edits && Array.isArray(args.edits)) {
                        const result = computeAdvancedDiff({
                          kind: "multi_edit",
                          filePath,
                          edits: args.edits as Array<{
                            old_string: string;
                            new_string: string;
                            replace_all?: boolean;
                          }>,
                        });
                        if (result.mode === "advanced") {
                          precomputedDiffsRef.current.set(toolCallId, result);
                        }
                      } else {
                        const result = computeAdvancedDiff({
                          kind: "edit",
                          filePath,
                          oldString: (args.old_string as string) || "",
                          newString: (args.new_string as string) || "",
                          replaceAll: args.replace_all as boolean | undefined,
                        });
                        if (result.mode === "advanced") {
                          precomputedDiffsRef.current.set(toolCallId, result);
                        }
                      }
                    }
                  } else if (isPatchTool(toolName) && args.input) {
                    // Patch tools - parse hunks directly (patches ARE diffs)
                    const operations = parsePatchOperations(
                      args.input as string,
                    );
                    for (const op of operations) {
                      const key = `${toolCallId}:${op.path}`;
                      if (op.kind === "add" || op.kind === "update") {
                        const result = parsePatchToAdvancedDiff(
                          op.patchLines,
                          op.path,
                        );
                        if (result) {
                          precomputedDiffsRef.current.set(key, result);
                        }
                      }
                      // Delete operations don't need diffs
                    }
                  }
                } catch {
                  // Ignore errors in diff computation for auto-allowed tools
                }
              }

              const autoAllowedToolCallIds = autoAllowed.map(
                (ac) => ac.approval.toolCallId,
              );
              const autoAllowedAbortController =
                abortControllerRef.current ?? new AbortController();
              const shouldTrackAutoAllowed = autoAllowedToolCallIds.length > 0;
              let autoAllowedWithResults: Array<{
                toolCallId: string;
                result: ToolExecutionResult;
              }> = [];
              let autoDeniedWithReasons: Array<{
                approval: ApprovalRequest;
                reason: string;
              }> = [];

              if (shouldTrackAutoAllowed) {
                setIsExecutingTool(true);
                executingToolCallIdsRef.current = autoAllowedToolCallIds;
                toolAbortControllerRef.current = autoAllowedAbortController;
                autoAllowedExecutionRef.current = {
                  toolCallIds: autoAllowedToolCallIds,
                  results: null,
                  conversationId: conversationIdRef.current,
                  generation: conversationGenerationRef.current,
                };
              }

              try {
                // Execute auto-allowed tools (sequential for writes, parallel for reads)
                autoAllowedWithResults =
                  autoAllowed.length > 0
                    ? await executeAutoAllowedTools(
                        autoAllowed,
                        (chunk) => onChunk(buffersRef.current, chunk),
                        {
                          abortSignal: autoAllowedAbortController.signal,
                          onStreamingOutput: updateStreamingOutput,
                        },
                      )
                    : [];

                // Create denial reasons for auto-denied and update UI
                autoDeniedWithReasons = autoDenied.map((ac) => {
                  // Prefer the detailed reason over the short matchedRule name
                  const reason = ac.permission.reason
                    ? `Permission denied: ${ac.permission.reason}`
                    : "matchedRule" in ac.permission &&
                        ac.permission.matchedRule
                      ? `Permission denied by rule: ${ac.permission.matchedRule}`
                      : "Permission denied: Unknown";

                  // Update buffers with denial for UI
                  onChunk(buffersRef.current, {
                    message_type: "tool_return_message",
                    id: "dummy",
                    date: new Date().toISOString(),
                    tool_call_id: ac.approval.toolCallId,
                    tool_return: `Error: request to call tool denied. User reason: ${reason}`,
                    status: "error",
                    stdout: null,
                    stderr: null,
                  });

                  return {
                    approval: ac.approval,
                    reason,
                  };
                });

                const queuedResults: ApprovalResult[] = [
                  ...autoAllowedWithResults.map((ar) => ({
                    type: "tool" as const,
                    tool_call_id: ar.toolCallId,
                    tool_return: ar.result.toolReturn,
                    status: ar.result.status,
                    stdout: ar.result.stdout,
                    stderr: ar.result.stderr,
                  })),
                  ...autoDeniedWithReasons.map((ad) => ({
                    type: "approval" as const,
                    tool_call_id: ad.approval.toolCallId,
                    approve: false,
                    reason: ad.reason,
                  })),
                ];

                if (autoAllowedExecutionRef.current) {
                  autoAllowedExecutionRef.current.results = queuedResults;
                }
                const autoAllowedMetadata = autoAllowedExecutionRef.current
                  ? {
                      conversationId:
                        autoAllowedExecutionRef.current.conversationId,
                      generation: conversationGenerationRef.current,
                    }
                  : undefined;

                if (
                  userCancelledRef.current ||
                  autoAllowedAbortController.signal.aborted ||
                  interruptQueuedRef.current
                ) {
                  if (queuedResults.length > 0) {
                    queueApprovalResults(queuedResults, autoAllowedMetadata);
                  }
                  setStreaming(false);
                  markIncompleteToolsAsCancelled(
                    buffersRef.current,
                    true,
                    "user_interrupt",
                  );
                  refreshDerived();
                  return { submitted: false };
                }

                // Store auto-handled results to send along with user decisions
                setAutoHandledResults(autoAllowedWithResults);
                setAutoDeniedApprovals(autoDeniedWithReasons);

                refreshDerived();
                return { submitted: false };
              } finally {
                if (shouldTrackAutoAllowed) {
                  setIsExecutingTool(false);
                  toolAbortControllerRef.current = null;
                  executingToolCallIdsRef.current = [];
                  autoAllowedExecutionRef.current = null;
                }
              }
            }
          }
        } catch (_error) {
          // If check fails, proceed anyway (don't block user)
        }
      }

      // Start the conversation loop. If we have queued approval results from an interrupted
      // client-side execution, send them first before the new user message.
      const initialInput: Array<MessageCreate | ApprovalCreate> = [];

      if (queuedApprovalResults) {
        const queuedMetadata = queuedApprovalMetadataRef.current;
        const isQueuedValid =
          queuedMetadata &&
          queuedMetadata.conversationId === conversationIdRef.current &&
          queuedMetadata.generation === conversationGenerationRef.current;

        if (isQueuedValid) {
          initialInput.push({
            type: "approval",
            approvals: queuedApprovalResults,
          });
        } else {
          debugWarn(
            "queue",
            "Dropping stale queued approval results for mismatched conversation or generation",
          );
        }
        queueApprovalResults(null);
        interruptQueuedRef.current = false;
      }

      initialInput.push({
        type: "message",
        role: "user",
        content: messageContent as unknown as MessageCreate["content"],
      });

      await processConversation(initialInput, { submissionGeneration });

      // Clean up placeholders after submission
      clearPlaceholdersInText(msg);

      return { submitted: true };
    },
    [
      streaming,
      commandRunning,
      processConversation,
      refreshDerived,
      agentId,
      agentName,
      agentDescription,
      agentLastRunAt,
      commandRunner,
      handleExit,
      isExecutingTool,
      queuedApprovalResults,
      queueApprovalResults,
      pendingApprovals,
      profileConfirmPending,
      handleAgentSelect,
      startOverlayCommand,
      tokenStreamingEnabled,
      isAgentBusy,
      setStreaming,
      setCommandRunning,
      pendingRalphConfig,
      openTrajectorySegment,
      resetTrajectoryBases,
      appendTaskNotificationEvents,
    ],
  );

  const onSubmitRef = useRef(onSubmit);
  useEffect(() => {
    onSubmitRef.current = onSubmit;
  }, [onSubmit]);

  // Process queued messages when streaming ends
  // Task notifications are now added directly to messageQueue via messageQueueBridge
  useEffect(() => {
    // Reference dequeueEpoch to satisfy exhaustive-deps - it's used to force
    // re-runs when userCancelledRef is reset (refs aren't in deps)
    // Also triggers when task notifications are added to queue
    void dequeueEpoch;

    const hasAnythingQueued = messageQueue.length > 0;

    if (
      !streaming &&
      hasAnythingQueued &&
      pendingApprovals.length === 0 &&
      !commandRunning &&
      !isExecutingTool &&
      !anySelectorOpen && // Don't dequeue while a selector/overlay is open
      !waitingForQueueCancelRef.current && // Don't dequeue while waiting for cancel
      !userCancelledRef.current && // Don't dequeue if user just cancelled
      !abortControllerRef.current // Don't dequeue while processConversation is still active
    ) {
      // Concatenate all queued messages into one (better UX when user types multiple
      // messages quickly - they get combined into one context for the agent)
      // Task notifications are already in the queue as XML strings
      const concatenatedMessage = messageQueue
        .map((item) => item.text)
        .join("\n");
      const queuedContentParts = buildQueuedContentParts(messageQueue);

      debugLog(
        "queue",
        `Dequeuing ${messageQueue.length} message(s): "${concatenatedMessage.slice(0, 50)}${concatenatedMessage.length > 50 ? "..." : ""}"`,
      );

      // Store the message before clearing queue - allows restoration on error
      lastDequeuedMessageRef.current = concatenatedMessage;
      setMessageQueue([]);

      // Submit the concatenated message using the normal submit flow
      // This ensures all setup (reminders, UI updates, etc.) happens correctly
      overrideContentPartsRef.current = queuedContentParts;
      onSubmitRef.current(concatenatedMessage);
    } else if (hasAnythingQueued) {
      // Log why dequeue was blocked (useful for debugging stuck queues)
      debugLog(
        "queue",
        `Dequeue blocked: streaming=${streaming}, pendingApprovals=${pendingApprovals.length}, commandRunning=${commandRunning}, isExecutingTool=${isExecutingTool}, anySelectorOpen=${anySelectorOpen}, waitingForQueueCancel=${waitingForQueueCancelRef.current}, userCancelled=${userCancelledRef.current}, abortController=${!!abortControllerRef.current}`,
      );
    }
  }, [
    streaming,
    messageQueue,
    pendingApprovals,
    commandRunning,
    isExecutingTool,
    anySelectorOpen,
    dequeueEpoch, // Triggered when userCancelledRef is reset OR task notifications added
  ]);

  // Helper to send all approval results when done
  const sendAllResults = useCallback(
    async (
      additionalDecision?:
        | { type: "approve"; approval: ApprovalRequest }
        | { type: "deny"; approval: ApprovalRequest; reason: string },
    ) => {
      try {
        // Don't send results if user has already cancelled
        if (
          userCancelledRef.current ||
          abortControllerRef.current?.signal.aborted
        ) {
          setStreaming(false);
          setIsExecutingTool(false);
          setPendingApprovals([]);
          setApprovalContexts([]);
          setApprovalResults([]);
          setAutoHandledResults([]);
          setAutoDeniedApprovals([]);
          return;
        }

        // Snapshot current state before clearing dialog
        const approvalResultsSnapshot = [...approvalResults];
        const autoHandledSnapshot = [...autoHandledResults];
        const autoDeniedSnapshot = [...autoDeniedApprovals];
        const pendingSnapshot = [...pendingApprovals];

        // Clear dialog state immediately so UI updates right away
        setPendingApprovals([]);
        setApprovalContexts([]);
        setApprovalResults([]);
        setAutoHandledResults([]);
        setAutoDeniedApprovals([]);

        // Show "thinking" state and lock input while executing approved tools client-side
        setStreaming(true);
        openTrajectorySegment();
        // Ensure interrupted flag is cleared for this execution
        buffersRef.current.interrupted = false;

        const approvalAbortController = new AbortController();
        toolAbortControllerRef.current = approvalAbortController;

        // Combine all decisions using snapshots
        const allDecisions = [
          ...approvalResultsSnapshot,
          ...(additionalDecision ? [additionalDecision] : []),
        ];

        const approvedDecisions = allDecisions.filter(
          (
            decision,
          ): decision is {
            type: "approve";
            approval: ApprovalRequest;
            precomputedResult?: ToolExecutionResult;
          } => decision.type === "approve",
        );
        const runningDecisions = approvedDecisions.filter(
          (decision) => !decision.precomputedResult,
        );

        executingToolCallIdsRef.current = runningDecisions.map(
          (decision) => decision.approval.toolCallId,
        );

        // Set phase to "running" for all approved tools
        if (runningDecisions.length > 0) {
          setToolCallsRunning(
            buffersRef.current,
            runningDecisions.map((d) => d.approval.toolCallId),
          );
        }
        refreshDerived();

        // Execute approved tools and format results using shared function
        const { executeApprovalBatch } = await import(
          "../agent/approval-execution"
        );
        sessionStatsRef.current.startTrajectory();
        const toolRunStart = performance.now();
        let executedResults: Awaited<ReturnType<typeof executeApprovalBatch>>;
        try {
          executedResults = await executeApprovalBatch(
            allDecisions,
            (chunk) => {
              onChunk(buffersRef.current, chunk);
              // Also log errors to the UI error display
              if (
                chunk.status === "error" &&
                chunk.message_type === "tool_return_message"
              ) {
                const isToolError = chunk.tool_return?.startsWith(
                  "Error executing tool:",
                );
                if (isToolError) {
                  appendError(chunk.tool_return);
                }
              }
              // Flush UI so completed tools show up while the batch continues
              refreshDerived();
            },
            {
              abortSignal: approvalAbortController.signal,
              onStreamingOutput: updateStreamingOutput,
            },
          );
        } finally {
          const toolRunMs = performance.now() - toolRunStart;
          sessionStatsRef.current.accumulateTrajectory({
            localToolMs: toolRunMs,
          });
        }

        // Combine with auto-handled and auto-denied results using snapshots
        const allResults = [
          ...autoHandledSnapshot.map((ar) => ({
            type: "tool" as const,
            tool_call_id: ar.toolCallId,
            tool_return: ar.result.toolReturn,
            status: ar.result.status,
            stdout: ar.result.stdout,
            stderr: ar.result.stderr,
          })),
          ...autoDeniedSnapshot.map((ad) => ({
            type: "approval" as const,
            tool_call_id: ad.approval.toolCallId,
            approve: false,
            reason: ad.reason,
          })),
          ...executedResults,
        ];

        // Dev-only validation: ensure outgoing IDs match expected IDs (using snapshots)
        if (process.env.NODE_ENV !== "production") {
          // Include ALL tool call IDs: auto-handled, auto-denied, and pending approvals
          const expectedIds = new Set([
            ...autoHandledSnapshot.map((ar) => ar.toolCallId),
            ...autoDeniedSnapshot.map((ad) => ad.approval.toolCallId),
            ...pendingSnapshot.map((a) => a.toolCallId),
          ]);
          const sendingIds = new Set(
            allResults.map((r) => r.tool_call_id).filter(Boolean),
          );

          const setsEqual = (a: Set<string>, b: Set<string>) =>
            a.size === b.size && [...a].every((id) => b.has(id));

          if (!setsEqual(expectedIds, sendingIds)) {
            console.error("[BUG] Approval ID mismatch detected");
            console.error("Expected IDs:", Array.from(expectedIds));
            console.error("Sending IDs:", Array.from(sendingIds));
            throw new Error(
              "Approval ID mismatch - refusing to send mismatched IDs",
            );
          }
        }

        // Rotate to a new thinking message
        setThinkingMessage(getRandomThinkingVerb());
        refreshDerived();

        const wasAborted = approvalAbortController.signal.aborted;
        // Check if user cancelled via ESC. We use wasAborted (toolAbortController was aborted)
        // as the primary signal, plus userCancelledRef for cancellations that happen just before
        // tools complete. Note: we can't use `abortControllerRef.current === null` because
        // abortControllerRef is also null in the normal approval flow (no stream running).
        const userCancelled = userCancelledRef.current;

        if (wasAborted || userCancelled) {
          // Queue results to send alongside the next user message so the backend
          // doesn't keep requesting the same approvals after an interrupt.
          if (!interruptQueuedRef.current) {
            queueApprovalResults(allResults as ApprovalResult[]);
          }
          setStreaming(false);
          closeTrajectorySegment();
          syncTrajectoryElapsedBase();

          // Reset queue-cancel flag so dequeue effect can fire
          waitingForQueueCancelRef.current = false;
          queueSnapshotRef.current = [];
        } else {
          const queuedItemsToAppend = consumeQueuedMessages();
          const queuedNotifications = queuedItemsToAppend
            ? getQueuedNotificationSummaries(queuedItemsToAppend)
            : [];
          const hadNotifications =
            appendTaskNotificationEvents(queuedNotifications);
          const input: Array<MessageCreate | ApprovalCreate> = [
            { type: "approval", approvals: allResults as ApprovalResult[] },
          ];
          if (queuedItemsToAppend && queuedItemsToAppend.length > 0) {
            const queuedUserText = buildQueuedUserText(queuedItemsToAppend);
            if (queuedUserText) {
              const userId = uid("user");
              buffersRef.current.byId.set(userId, {
                kind: "user",
                id: userId,
                text: queuedUserText,
              });
              buffersRef.current.order.push(userId);
            }
            input.push({
              type: "message",
              role: "user",
              content: buildQueuedContentParts(queuedItemsToAppend),
            });
            refreshDerived();
          } else if (hadNotifications) {
            refreshDerived();
          }
          // Flush finished items synchronously before reentry. This avoids a
          // race where deferred non-Task commits delay Task grouping while the
          // reentry path continues.
          flushEligibleLinesBeforeReentry(
            commitEligibleLines,
            buffersRef.current,
          );
          toolResultsInFlightRef.current = true;
          await processConversation(input, { allowReentry: true });
          toolResultsInFlightRef.current = false;

          // Clear any stale queued results from previous interrupts.
          // This approval flow supersedes any previously queued results - if we don't
          // clear them here, they persist with matching generation and get sent on the
          // next onSubmit, causing "Invalid tool call IDs" errors.
          queueApprovalResults(null);
        }
      } finally {
        // Always release the execution guard, even if an error occurred
        setIsExecutingTool(false);
        toolAbortControllerRef.current = null;
        executingToolCallIdsRef.current = [];
        interruptQueuedRef.current = false;
        toolResultsInFlightRef.current = false;
      }
    },
    [
      approvalResults,
      autoHandledResults,
      autoDeniedApprovals,
      pendingApprovals,
      processConversation,
      refreshDerived,
      appendError,
      setStreaming,
      updateStreamingOutput,
      queueApprovalResults,
      consumeQueuedMessages,
      appendTaskNotificationEvents,
      syncTrajectoryElapsedBase,
      closeTrajectorySegment,
      openTrajectorySegment,
      commitEligibleLines,
    ],
  );

  // Handle approval callbacks - sequential review
  const handleApproveCurrent = useCallback(
    async (diffs?: Map<string, AdvancedDiffSuccess>) => {
      if (isExecutingTool) return;

      const currentIndex = approvalResults.length;
      const currentApproval = pendingApprovals[currentIndex];

      if (!currentApproval) return;

      // Store precomputed diffs before execution
      if (diffs) {
        for (const [key, diff] of diffs) {
          precomputedDiffsRef.current.set(key, diff);
        }
      }

      setIsExecutingTool(true);

      try {
        // Store approval decision (don't execute yet - batch execute after all approvals)
        const decision = {
          type: "approve" as const,
          approval: currentApproval,
        };

        // Check if we're done with all approvals
        if (currentIndex + 1 >= pendingApprovals.length) {
          // All approvals collected, execute and send to backend
          // sendAllResults owns the lock release via its finally block
          await sendAllResults(decision);
        } else {
          // Not done yet, store decision and show next approval
          setApprovalResults((prev) => [...prev, decision]);
          setIsExecutingTool(false);
        }
      } catch (e) {
        const errorDetails = formatErrorDetails(e, agentId);
        appendError(errorDetails);
        setStreaming(false);
        setIsExecutingTool(false);
      }
    },
    [
      agentId,
      pendingApprovals,
      approvalResults,
      sendAllResults,
      appendError,
      isExecutingTool,
      setStreaming,
    ],
  );

  const handleApproveAlways = useCallback(
    async (
      scope?: "project" | "session",
      diffs?: Map<string, AdvancedDiffSuccess>,
    ) => {
      if (isExecutingTool) return;

      if (pendingApprovals.length === 0 || approvalContexts.length === 0)
        return;

      const currentIndex = approvalResults.length;
      const approvalContext = approvalContexts[currentIndex];
      if (!approvalContext) return;

      const rule = approvalContext.recommendedRule;
      const actualScope = scope || approvalContext.defaultScope;

      const cmd = commandRunner.start(
        "/approve-always",
        "Adding permission...",
      );

      // Save the permission rule
      try {
        await savePermissionRule(rule, "allow", actualScope);
      } catch (error) {
        const errorDetails = formatErrorDetails(error, agentId);
        cmd.fail(`Failed to add permission: ${errorDetails}`);
        return;
      }

      // Show confirmation in transcript
      const scopeText =
        actualScope === "session" ? " (session only)" : " (project)";
      cmd.finish(`Added permission: ${rule}${scopeText}`, true);

      // Re-check remaining approvals against the newly saved permission
      // This allows subsequent approvals that match the new rule to be auto-allowed
      const remainingApprovals = pendingApprovals.slice(currentIndex + 1);
      if (remainingApprovals.length > 0) {
        const recheckResults = await Promise.all(
          remainingApprovals.map(async (approval) => {
            const parsedArgs = safeJsonParseOr<Record<string, unknown>>(
              approval.toolArgs,
              {},
            );
            const permission = await checkToolPermission(
              approval.toolName,
              parsedArgs,
            );
            return { approval, permission };
          }),
        );

        const nowAutoAllowed = recheckResults.filter(
          (r) => r.permission.decision === "allow",
        );
        const stillNeedAsking = recheckResults.filter(
          (r) => r.permission.decision === "ask",
        );

        // Only auto-handle if ALL remaining are now allowed
        // (avoids complex state synchronization issues with partial batches)
        if (stillNeedAsking.length === 0 && nowAutoAllowed.length > 0) {
          const currentApproval = pendingApprovals[currentIndex];
          if (!currentApproval) return;

          // Store diffs before execution
          if (diffs) {
            for (const [key, diff] of diffs) {
              precomputedDiffsRef.current.set(key, diff);
            }
          }

          setIsExecutingTool(true);

          // Snapshot current state BEFORE clearing (critical for ID matching!)
          // This must include ALL previous decisions, auto-handled, and auto-denied
          const approvalResultsSnapshot = [...approvalResults];
          const autoHandledSnapshot = [...autoHandledResults];
          const autoDeniedSnapshot = [...autoDeniedApprovals];

          // Build ALL decisions: previous + current + auto-allowed remaining
          const allDecisions: Array<
            | { type: "approve"; approval: ApprovalRequest }
            | { type: "deny"; approval: ApprovalRequest; reason: string }
          > = [
            ...approvalResultsSnapshot, // Include decisions from previous rounds
            { type: "approve", approval: currentApproval },
            ...nowAutoAllowed.map((r) => ({
              type: "approve" as const,
              approval: r.approval,
            })),
          ];

          // Clear dialog state immediately
          setPendingApprovals([]);
          setApprovalContexts([]);
          setApprovalResults([]);
          setAutoHandledResults([]);
          setAutoDeniedApprovals([]);

          setStreaming(true);
          openTrajectorySegment();
          buffersRef.current.interrupted = false;

          // Set phase to "running" for all approved tools
          setToolCallsRunning(
            buffersRef.current,
            allDecisions
              .filter((d) => d.type === "approve")
              .map((d) => d.approval.toolCallId),
          );
          refreshDerived();

          try {
            // Execute ALL decisions together
            const { executeApprovalBatch } = await import(
              "../agent/approval-execution"
            );
            const executedResults = await executeApprovalBatch(
              allDecisions,
              (chunk) => {
                onChunk(buffersRef.current, chunk);
                refreshDerived();
              },
              { onStreamingOutput: updateStreamingOutput },
            );

            // Combine with auto-handled and auto-denied results (from initial check)
            const allResults = [
              ...autoHandledSnapshot.map((ar) => ({
                type: "tool" as const,
                tool_call_id: ar.toolCallId,
                tool_return: ar.result.toolReturn,
                status: ar.result.status,
                stdout: ar.result.stdout,
                stderr: ar.result.stderr,
              })),
              ...autoDeniedSnapshot.map((ad) => ({
                type: "approval" as const,
                tool_call_id: ad.approval.toolCallId,
                approve: false,
                reason: ad.reason,
              })),
              ...executedResults,
            ];

            setThinkingMessage(getRandomThinkingVerb());
            refreshDerived();

            // Continue conversation with all results
            await processConversation([
              {
                type: "approval",
                approvals: allResults as ApprovalResult[],
              },
            ]);
          } finally {
            setIsExecutingTool(false);
          }
          return; // Don't call handleApproveCurrent - we handled everything
        }
      }

      // Fallback: proceed with normal flow (will prompt for remaining approvals)
      await handleApproveCurrent(diffs);
    },
    [
      agentId,
      commandRunner,
      approvalResults,
      approvalContexts,
      pendingApprovals,
      autoHandledResults,
      autoDeniedApprovals,
      handleApproveCurrent,
      processConversation,
      refreshDerived,
      isExecutingTool,
      setStreaming,
      openTrajectorySegment,
      updateStreamingOutput,
    ],
  );

  const handleDenyCurrent = useCallback(
    async (reason: string) => {
      if (isExecutingTool) return;

      const currentIndex = approvalResults.length;
      const currentApproval = pendingApprovals[currentIndex];

      if (!currentApproval) return;

      setIsExecutingTool(true);

      try {
        // Store denial decision
        const decision = {
          type: "deny" as const,
          approval: currentApproval,
          reason: reason || "User denied the tool execution",
        };

        // Check if we're done with all approvals
        if (currentIndex + 1 >= pendingApprovals.length) {
          // All approvals collected, execute and send to backend
          // sendAllResults owns the lock release via its finally block
          setThinkingMessage(getRandomThinkingVerb());
          await sendAllResults(decision);
        } else {
          // Not done yet, store decision and show next approval
          setApprovalResults((prev) => [...prev, decision]);
          setIsExecutingTool(false);
        }
      } catch (e) {
        const errorDetails = formatErrorDetails(e, agentId);
        appendError(errorDetails);
        setStreaming(false);
        setIsExecutingTool(false);
      }
    },
    [
      agentId,
      pendingApprovals,
      approvalResults,
      sendAllResults,
      appendError,
      isExecutingTool,
      setStreaming,
    ],
  );

  // Cancel all pending approvals - queue denials to send with next message
  // Similar to interrupt flow during tool execution
  const handleCancelApprovals = useCallback(() => {
    if (pendingApprovals.length === 0) return;

    // Create denial results for all pending approvals and queue for next message
    const denialResults = pendingApprovals.map((approval) => ({
      type: "approval" as const,
      tool_call_id: approval.toolCallId,
      approve: false,
      reason: "User cancelled the approval",
    }));
    queueApprovalResults(denialResults);

    // Mark the pending approval tool calls as cancelled in the buffers
    markIncompleteToolsAsCancelled(buffersRef.current, true, "approval_cancel");
    refreshDerived();

    // Clear all approval state
    setPendingApprovals([]);
    setApprovalContexts([]);
    setApprovalResults([]);
    setAutoHandledResults([]);
    setAutoDeniedApprovals([]);
  }, [pendingApprovals, refreshDerived, queueApprovalResults]);

  const handleModelSelect = useCallback(
    async (modelId: string, commandId?: string | null) => {
      const overlayCommand = commandId
        ? commandRunner.getHandle(commandId, "/model")
        : consumeOverlayCommand("model");

      let selectedModel: {
        id: string;
        handle?: string;
        label: string;
        updateArgs?: { context_window?: number };
      } | null = null;

      try {
        const { models } = await import("../agent/model");
        selectedModel = models.find((m) => m.id === modelId) ?? null;

        if (!selectedModel && modelId.includes("/")) {
          const { getModelContextWindow } = await import(
            "../agent/available-models"
          );
          const apiContextWindow = await getModelContextWindow(modelId);

          selectedModel = {
            id: modelId,
            handle: modelId,
            label: modelId.split("/").pop() ?? modelId,
            description: "Custom model",
            updateArgs: apiContextWindow
              ? { context_window: apiContextWindow }
              : undefined,
          } as unknown as (typeof models)[number];
        }

        if (!selectedModel) {
          const output = `Model not found: ${modelId}. Run /model and press R to refresh available models.`;
          const cmd = overlayCommand ?? commandRunner.start("/model", output);
          cmd.fail(output);
          return;
        }
        const model = selectedModel;
        const modelHandle = model.handle ?? model.id;

        if (isAgentBusy()) {
          setActiveOverlay(null);
          const cmd =
            overlayCommand ??
            commandRunner.start(
              "/model",
              `Model switch queued – will switch after current task completes`,
            );
          cmd.update({
            output: `Model switch queued – will switch after current task completes`,
            phase: "running",
          });
          setQueuedOverlayAction({
            type: "switch_model",
            modelId,
            commandId: cmd.id,
          });
          return;
        }

        await withCommandLock(async () => {
          const cmd =
            overlayCommand ??
            commandRunner.start(
              "/model",
              `Switching model to ${model.label}...`,
            );
          cmd.update({
            output: `Switching model to ${model.label}...`,
            phase: "running",
          });

          const { updateAgentLLMConfig } = await import("../agent/modify");
          const updatedConfig = await updateAgentLLMConfig(
            agentId,
            modelHandle,
            model.updateArgs,
          );
          setLlmConfig(updatedConfig);
          setCurrentModelId(modelId);

          // Reset context token tracking since different models have different tokenizers
          resetContextHistory(contextTrackerRef.current);

          const { isOpenAIModel, isGeminiModel } = await import(
            "../tools/manager"
          );
          const targetToolset:
            | "codex"
            | "codex_snake"
            | "default"
            | "gemini"
            | "gemini_snake"
            | "none" = isOpenAIModel(modelHandle)
            ? "codex"
            : isGeminiModel(modelHandle)
              ? "gemini"
              : "default";

          let toolsetName:
            | "codex"
            | "codex_snake"
            | "default"
            | "gemini"
            | "gemini_snake"
            | "none"
            | null = null;
          if (currentToolset !== targetToolset) {
            const { switchToolsetForModel } = await import("../tools/toolset");
            toolsetName = await switchToolsetForModel(modelHandle, agentId);
            setCurrentToolset(toolsetName);
          }

          const autoToolsetLine = toolsetName
            ? `Automatically switched toolset to ${toolsetName}. Use /toolset to change back if desired.\nConsider switching to a different system prompt using /system to match.`
            : null;
          const outputLines = [
            `Switched to ${model.label}`,
            ...(autoToolsetLine ? [autoToolsetLine] : []),
          ].join("\n");

          cmd.finish(outputLines, true);
        });
      } catch (error) {
        const errorDetails = formatErrorDetails(error, agentId);
        const modelLabel = selectedModel?.label ?? modelId;
        const guidance =
          "Run /model and press R to refresh available models. If the model is still unavailable, choose another model or connect a provider with /connect.";
        const cmd =
          overlayCommand ??
          commandRunner.start(
            "/model",
            `Failed to switch model to ${modelLabel}.`,
          );
        cmd.fail(
          `Failed to switch model to ${modelLabel}: ${errorDetails}\n${guidance}`,
        );
      }
    },
    [
      agentId,
      commandRunner,
      consumeOverlayCommand,
      currentToolset,
      isAgentBusy,
      withCommandLock,
    ],
  );

  const handleSystemPromptSelect = useCallback(
    async (promptId: string, commandId?: string | null) => {
      const overlayCommand = commandId
        ? commandRunner.getHandle(commandId, "/system")
        : consumeOverlayCommand("system");

      let selectedPrompt:
        | { id: string; label: string; content: string }
        | undefined;

      try {
        const { SYSTEM_PROMPTS } = await import("../agent/promptAssets");
        selectedPrompt = SYSTEM_PROMPTS.find((p) => p.id === promptId);

        if (!selectedPrompt) {
          const cmd =
            overlayCommand ??
            commandRunner.start(
              "/system",
              `System prompt not found: ${promptId}`,
            );
          cmd.fail(`System prompt not found: ${promptId}`);
          return;
        }
        const prompt = selectedPrompt;

        if (isAgentBusy()) {
          setActiveOverlay(null);
          const cmd =
            overlayCommand ??
            commandRunner.start(
              "/system",
              "System prompt switch queued – will switch after current task completes",
            );
          cmd.update({
            output:
              "System prompt switch queued – will switch after current task completes",
            phase: "running",
          });
          setQueuedOverlayAction({
            type: "switch_system",
            promptId,
            commandId: cmd.id,
          });
          return;
        }

        await withCommandLock(async () => {
          const cmd =
            overlayCommand ??
            commandRunner.start(
              "/system",
              `Switching system prompt to ${prompt.label}...`,
            );
          cmd.update({
            output: `Switching system prompt to ${prompt.label}...`,
            phase: "running",
          });

          const { updateAgentSystemPromptRaw } = await import(
            "../agent/modify"
          );
          const result = await updateAgentSystemPromptRaw(
            agentId,
            prompt.content,
          );

          if (result.success) {
            setCurrentSystemPromptId(promptId);
            cmd.finish(`Switched system prompt to ${prompt.label}`, true);
          } else {
            cmd.fail(result.message);
          }
        });
      } catch (error) {
        const errorDetails = formatErrorDetails(error, agentId);
        const cmd =
          overlayCommand ??
          commandRunner.start("/system", "Failed to switch system prompt.");
        cmd.fail(`Failed to switch system prompt: ${errorDetails}`);
      }
    },
    [
      agentId,
      commandRunner,
      consumeOverlayCommand,
      isAgentBusy,
      withCommandLock,
    ],
  );

  const handleSleeptimeModeSelect = useCallback(
    async (
      reflectionSettings: ReflectionSettings,
      commandId?: string | null,
    ) => {
      const overlayCommand = commandId
        ? commandRunner.getHandle(commandId, "/sleeptime")
        : consumeOverlayCommand("sleeptime");

      if (isAgentBusy()) {
        setActiveOverlay(null);
        const cmd =
          overlayCommand ??
          commandRunner.start(
            "/sleeptime",
            "Sleeptime settings update queued – will apply after current task completes",
          );
        cmd.update({
          output:
            "Sleeptime settings update queued – will apply after current task completes",
          phase: "running",
        });
        setQueuedOverlayAction({
          type: "set_sleeptime",
          settings: reflectionSettings,
          commandId: cmd.id,
        });
        return;
      }

      await withCommandLock(async () => {
        const cmd =
          overlayCommand ??
          commandRunner.start("/sleeptime", "Saving sleeptime settings...");
        cmd.update({
          output: "Saving sleeptime settings...",
          phase: "running",
        });

        try {
          const legacyMode = reflectionSettingsToLegacyMode(reflectionSettings);
          settingsManager.updateLocalProjectSettings({
            memoryReminderInterval: legacyMode,
            reflectionTrigger: reflectionSettings.trigger,
            reflectionBehavior: reflectionSettings.behavior,
            reflectionStepCount: reflectionSettings.stepCount,
          });
          settingsManager.updateSettings({
            memoryReminderInterval: legacyMode,
            reflectionTrigger: reflectionSettings.trigger,
            reflectionBehavior: reflectionSettings.behavior,
            reflectionStepCount: reflectionSettings.stepCount,
          });

          cmd.finish(
            `Updated sleeptime settings to: ${formatReflectionSettings(reflectionSettings)}`,
            true,
          );
        } catch (error) {
          const errorDetails = formatErrorDetails(error, agentId);
          cmd.fail(`Failed to save sleeptime settings: ${errorDetails}`);
        }
      });
    },
    [
      agentId,
      commandRunner,
      consumeOverlayCommand,
      isAgentBusy,
      withCommandLock,
    ],
  );

  const handleToolsetSelect = useCallback(
    async (
      toolsetId:
        | "codex"
        | "codex_snake"
        | "default"
        | "gemini"
        | "gemini_snake"
        | "none",
      commandId?: string | null,
    ) => {
      const overlayCommand = commandId
        ? commandRunner.getHandle(commandId, "/toolset")
        : consumeOverlayCommand("toolset");

      if (isAgentBusy()) {
        setActiveOverlay(null);
        const cmd =
          overlayCommand ??
          commandRunner.start(
            "/toolset",
            "Toolset switch queued – will switch after current task completes",
          );
        cmd.update({
          output:
            "Toolset switch queued – will switch after current task completes",
          phase: "running",
        });
        setQueuedOverlayAction({
          type: "switch_toolset",
          toolsetId,
          commandId: cmd.id,
        });
        return;
      }

      await withCommandLock(async () => {
        const cmd =
          overlayCommand ??
          commandRunner.start(
            "/toolset",
            `Switching toolset to ${toolsetId}...`,
          );
        cmd.update({
          output: `Switching toolset to ${toolsetId}...`,
          phase: "running",
        });

        try {
          const { forceToolsetSwitch } = await import("../tools/toolset");
          await forceToolsetSwitch(toolsetId, agentId);
          setCurrentToolset(toolsetId);
          cmd.finish(`Switched toolset to ${toolsetId}`, true);
        } catch (error) {
          const errorDetails = formatErrorDetails(error, agentId);
          cmd.fail(`Failed to switch toolset: ${errorDetails}`);
        }
      });
    },
    [
      agentId,
      commandRunner,
      consumeOverlayCommand,
      isAgentBusy,
      withCommandLock,
    ],
  );

  // Process queued overlay actions when streaming ends
  // These are actions from interactive commands (like /agents, /model) that were
  // used while the agent was busy. The change is applied after end_turn.
  useEffect(() => {
    if (
      !streaming &&
      !commandRunning &&
      !isExecutingTool &&
      pendingApprovals.length === 0 &&
      queuedOverlayAction !== null
    ) {
      const action = queuedOverlayAction;
      setQueuedOverlayAction(null); // Clear immediately to prevent re-runs

      // Process the queued action
      if (action.type === "switch_agent") {
        // Call handleAgentSelect - it will see isAgentBusy() as false now
        handleAgentSelect(action.agentId, { commandId: action.commandId });
      } else if (action.type === "switch_model") {
        // Call handleModelSelect - it will see isAgentBusy() as false now
        handleModelSelect(action.modelId, action.commandId);
      } else if (action.type === "set_sleeptime") {
        handleSleeptimeModeSelect(action.settings, action.commandId);
      } else if (action.type === "switch_conversation") {
        const cmd = action.commandId
          ? commandRunner.getHandle(action.commandId, "/resume")
          : commandRunner.start(
              "/resume",
              "Processing queued conversation switch...",
            );
        cmd.update({
          output: "Processing queued conversation switch...",
          phase: "running",
        });

        // Execute the conversation switch asynchronously
        (async () => {
          setCommandRunning(true);
          try {
            if (action.conversationId === conversationId) {
              cmd.finish("Already on this conversation", true);
            } else {
              const client = await getClient();
              if (agentState) {
                const resumeData = await getResumeData(
                  client,
                  agentState,
                  action.conversationId,
                );

                setConversationId(action.conversationId);

                pendingConversationSwitchRef.current = {
                  origin: "resume-selector",
                  conversationId: action.conversationId,
                  isDefault: action.conversationId === "default",
                  messageCount: resumeData.messageHistory.length,
                  messageHistory: resumeData.messageHistory,
                };

                settingsManager.setLocalLastSession(
                  { agentId, conversationId: action.conversationId },
                  process.cwd(),
                );
                settingsManager.setGlobalLastSession({
                  agentId,
                  conversationId: action.conversationId,
                });

                // Reset context tokens for new conversation
                resetContextHistory(contextTrackerRef.current);

                cmd.finish(
                  `Switched to conversation (${resumeData.messageHistory.length} messages)`,
                  true,
                );
              }
            }
          } catch (error) {
            cmd.fail(
              `Failed to switch conversation: ${error instanceof Error ? error.message : String(error)}`,
            );
          } finally {
            setCommandRunning(false);
            refreshDerived();
          }
        })();
      } else if (action.type === "switch_toolset") {
        handleToolsetSelect(action.toolsetId, action.commandId);
      } else if (action.type === "switch_system") {
        handleSystemPromptSelect(action.promptId, action.commandId);
      }
    }
  }, [
    streaming,
    commandRunning,
    isExecutingTool,
    pendingApprovals,
    queuedOverlayAction,
    handleAgentSelect,
    handleModelSelect,
    handleSleeptimeModeSelect,
    handleToolsetSelect,
    handleSystemPromptSelect,
    agentId,
    agentState,
    conversationId,
    refreshDerived,
    setCommandRunning,
    commandRunner.getHandle,
    commandRunner.start,
  ]);

  // Handle escape when profile confirmation is pending
  const handleFeedbackSubmit = useCallback(
    async (message: string) => {
      // Consume command handle BEFORE closing overlay; otherwise closeOverlay()
      // finishes it as "Feedback dialog dismissed" and we emit a duplicate entry.
      const overlayCommand = consumeOverlayCommand("feedback");
      closeOverlay();

      await withCommandLock(async () => {
        const cmd =
          overlayCommand ??
          commandRunner.start("/feedback", "Sending feedback...");

        try {
          const resolvedMessage = resolvePlaceholders(message);

          cmd.update({
            output: "Sending feedback...",
            phase: "running",
          });

          const settings = settingsManager.getSettings();
          const apiKey =
            process.env.LETTA_API_KEY || settings.env?.LETTA_API_KEY;

          // Only send anonymized, safe settings for debugging
          const {
            env: _env,
            refreshToken: _refreshToken,
            ...safeSettings
          } = settings;

          const response = await fetch(
            "https://api.letta.com/v1/metadata/feedback",
            {
              method: "POST",
              headers: {
                ...getLettaCodeHeaders(apiKey),
                "X-Letta-Code-Device-ID": settingsManager.getOrCreateDeviceId(),
              },
              body: JSON.stringify({
                message: resolvedMessage,
                feature: "letta-code",
                agent_id: agentId,
                session_id: telemetry.getSessionId(),
                version: process.env.npm_package_version || "unknown",
                platform: process.platform,
                settings: JSON.stringify(safeSettings),
                // Additional context for debugging
                system_info: {
                  local_time: getLocalTime(),
                  device_type: getDeviceType(),
                  cwd: process.cwd(),
                },
                session_stats: (() => {
                  const stats = sessionStatsRef.current?.getSnapshot();
                  if (!stats) return undefined;
                  return {
                    total_api_ms: stats.totalApiMs,
                    total_wall_ms: stats.totalWallMs,
                    step_count: stats.usage.stepCount,
                    prompt_tokens: stats.usage.promptTokens,
                    completion_tokens: stats.usage.completionTokens,
                    total_tokens: stats.usage.totalTokens,
                    cached_input_tokens: stats.usage.cachedInputTokens,
                    cache_write_tokens: stats.usage.cacheWriteTokens,
                    reasoning_tokens: stats.usage.reasoningTokens,
                    context_tokens: stats.usage.contextTokens,
                  };
                })(),
                agent_info: {
                  agent_name: agentName,
                  agent_description: agentDescription,
                  model: currentModelId,
                },
                account_info: {
                  billing_tier: billingTier,
                },
              }),
            },
          );

          if (!response.ok) {
            const errorText = await response.text();
            throw new Error(
              `Failed to send feedback (${response.status}): ${errorText}`,
            );
          }

          cmd.finish(
            "Feedback submitted! To chat with the Letta dev team live, join our Discord (https://discord.gg/letta).",
            true,
          );
        } catch (error) {
          const errorDetails = formatErrorDetails(error, agentId);
          cmd.fail(`Failed to send feedback: ${errorDetails}`);
        }
      });
    },
    [
      agentId,
      agentName,
      agentDescription,
      currentModelId,
      billingTier,
      commandRunner,
      consumeOverlayCommand,
      withCommandLock,
      closeOverlay,
    ],
  );

  const handleProfileEscapeCancel = useCallback(() => {
    if (profileConfirmPending) {
      const { cmdId, name } = profileConfirmPending;
      const cmd = commandRunner.getHandle(cmdId, `/profile load ${name}`);
      cmd.fail("Cancelled");
      setProfileConfirmPending(null);
    }
  }, [commandRunner, profileConfirmPending]);

  // Handle ralph mode exit from Input component (shift+tab)
  const handleRalphExit = useCallback(() => {
    const ralph = ralphMode.getState();
    if (ralph.isActive) {
      const wasYolo = ralph.isYolo;
      ralphMode.deactivate();
      setUiRalphActive(false);
      if (wasYolo) {
        permissionMode.setMode("default");
        setUiPermissionMode("default");
      }
    }
  }, []);

  // Handle permission mode changes from the Input component (e.g., shift+tab cycling)
  const handlePermissionModeChange = useCallback(
    (mode: PermissionMode) => {
      // When entering plan mode via tab cycling, generate and set the plan file path
      if (mode === "plan") {
        const planPath = generatePlanFilePath();
        permissionMode.setPlanFilePath(planPath);
      }
      // permissionMode.setMode() is called in InputRich.tsx before this callback
      setUiPermissionMode(mode);
      triggerStatusLineRefresh();
    },
    [triggerStatusLineRefresh],
  );

  const handlePlanApprove = useCallback(
    async (acceptEdits: boolean = false) => {
      const currentIndex = approvalResults.length;
      const approval = pendingApprovals[currentIndex];
      if (!approval) return;

      const isLast = currentIndex + 1 >= pendingApprovals.length;

      // Capture plan file path BEFORE exiting plan mode (for post-approval rendering)
      const planFilePath = permissionMode.getPlanFilePath();
      lastPlanFilePathRef.current = planFilePath;

      // Exit plan mode
      const restoreMode = acceptEdits
        ? "acceptEdits"
        : (permissionMode.getModeBeforePlan() ?? "default");
      permissionMode.setMode(restoreMode);
      setUiPermissionMode(restoreMode);

      try {
        // Execute ExitPlanMode tool to get the result
        const parsedArgs = safeJsonParseOr<Record<string, unknown>>(
          approval.toolArgs,
          {},
        );
        const toolResult = await executeTool("ExitPlanMode", parsedArgs);

        // Update buffers with tool return
        onChunk(buffersRef.current, {
          message_type: "tool_return_message",
          id: "dummy",
          date: new Date().toISOString(),
          tool_call_id: approval.toolCallId,
          tool_return: getDisplayableToolReturn(toolResult.toolReturn),
          status: toolResult.status,
          stdout: toolResult.stdout,
          stderr: toolResult.stderr,
        });

        setThinkingMessage(getRandomThinkingVerb());
        refreshDerived();

        const decision = {
          type: "approve" as const,
          approval,
          precomputedResult: toolResult,
        };

        if (isLast) {
          setIsExecutingTool(true);
          await sendAllResults(decision);
        } else {
          setApprovalResults((prev) => [...prev, decision]);
        }
      } catch (e) {
        const errorDetails = formatErrorDetails(e, agentId);
        appendError(errorDetails);
        setStreaming(false);
      }
    },
    [
      agentId,
      pendingApprovals,
      approvalResults,
      sendAllResults,
      appendError,
      refreshDerived,
      setStreaming,
    ],
  );

  const handlePlanKeepPlanning = useCallback(
    async (reason: string) => {
      const currentIndex = approvalResults.length;
      const approval = pendingApprovals[currentIndex];
      if (!approval) return;

      const isLast = currentIndex + 1 >= pendingApprovals.length;

      // Stay in plan mode
      const denialReason =
        reason ||
        "The user doesn't want to proceed with this tool use. The tool use was rejected (eg. if it was a file edit, the new_string was NOT written to the file). STOP what you are doing and wait for the user to tell you how to proceed.";

      const decision = {
        type: "deny" as const,
        approval,
        reason: denialReason,
      };

      if (isLast) {
        setIsExecutingTool(true);
        await sendAllResults(decision);
      } else {
        setApprovalResults((prev) => [...prev, decision]);
      }
    },
    [pendingApprovals, approvalResults, sendAllResults],
  );

  // Auto-reject ExitPlanMode if plan mode is not enabled or plan file doesn't exist
  useEffect(() => {
    const currentIndex = approvalResults.length;
    const approval = pendingApprovals[currentIndex];
    if (approval?.toolName === "ExitPlanMode") {
      // First check if plan mode is enabled
      if (permissionMode.getMode() !== "plan") {
        // Plan mode state was lost (e.g., CLI restart) - queue rejection with helpful message
        // This is different from immediate rejection because we want the user to see what happened
        // and be able to type their next message

        // Add status message to explain what happened
        const statusId = uid("status");
        buffersRef.current.byId.set(statusId, {
          kind: "status",
          id: statusId,
          lines: ["⚠️ Plan mode session expired (use /plan to re-enter)"],
        });
        buffersRef.current.order.push(statusId);

        // Queue denial to send with next message (same pattern as handleCancelApprovals)
        const denialResults = [
          {
            type: "approval" as const,
            tool_call_id: approval.toolCallId,
            approve: false,
            reason:
              "Plan mode session expired (CLI restarted). Use EnterPlanMode to re-enter plan mode, or request the user to re-enter plan mode.",
          },
        ];
        queueApprovalResults(denialResults);

        // Mark tool as cancelled in buffers
        markIncompleteToolsAsCancelled(
          buffersRef.current,
          true,
          "internal_cancel",
        );
        refreshDerived();

        // Clear all approval state (same as handleCancelApprovals)
        setPendingApprovals([]);
        setApprovalContexts([]);
        setApprovalResults([]);
        setAutoHandledResults([]);
        setAutoDeniedApprovals([]);
        return;
      }
      // Then check if plan file exists (keep existing behavior - immediate rejection)
      // This case means plan mode IS active, but agent forgot to write the plan file
      if (!planFileExists()) {
        const planFilePath = permissionMode.getPlanFilePath();
        const plansDir = join(homedir(), ".letta", "plans");
        handlePlanKeepPlanning(
          `You must write your plan to a plan file before exiting plan mode.\n` +
            (planFilePath ? `Plan file path: ${planFilePath}\n` : "") +
            `Use a write tool to create your plan in ${plansDir}, then use ExitPlanMode to present the plan to the user.`,
        );
      }
    }
  }, [
    pendingApprovals,
    approvalResults.length,
    handlePlanKeepPlanning,
    refreshDerived,
    queueApprovalResults,
  ]);

  const handleQuestionSubmit = useCallback(
    async (answers: Record<string, string>) => {
      const currentIndex = approvalResults.length;
      const approval = pendingApprovals[currentIndex];
      if (!approval) return;

      const isLast = currentIndex + 1 >= pendingApprovals.length;

      // Get questions from approval args
      const questions = getQuestionsFromApproval(approval);

      // Check for memory preference question and update setting
      parseMemoryPreference(questions, answers);

      // Format the answer string like Claude Code does
      // Filter out malformed questions (LLM might send invalid data)
      const answerParts = questions
        .filter((q) => q.question)
        .map((q) => {
          const answer = answers[q.question] || "";
          return `"${q.question}"="${answer}"`;
        });
      const toolReturn = `User has answered your questions: ${answerParts.join(", ")}. You can now continue with the user's answers in mind.`;

      const precomputedResult: ToolExecutionResult = {
        toolReturn,
        status: "success",
      };

      // Update buffers with tool return
      onChunk(buffersRef.current, {
        message_type: "tool_return_message",
        id: "dummy",
        date: new Date().toISOString(),
        tool_call_id: approval.toolCallId,
        tool_return: toolReturn,
        status: "success",
        stdout: null,
        stderr: null,
      });

      setThinkingMessage(getRandomThinkingVerb());
      refreshDerived();

      const decision = {
        type: "approve" as const,
        approval,
        precomputedResult,
      };

      if (isLast) {
        setIsExecutingTool(true);
        await sendAllResults(decision);
      } else {
        setApprovalResults((prev) => [...prev, decision]);
      }
    },
    [pendingApprovals, approvalResults, sendAllResults, refreshDerived],
  );

  const handleEnterPlanModeApprove = useCallback(async () => {
    const currentIndex = approvalResults.length;
    const approval = pendingApprovals[currentIndex];
    if (!approval) return;

    const isLast = currentIndex + 1 >= pendingApprovals.length;

    // Generate plan file path
    const planFilePath = generatePlanFilePath();

    // Toggle plan mode on and store plan file path
    permissionMode.setMode("plan");
    permissionMode.setPlanFilePath(planFilePath);
    setUiPermissionMode("plan");

    // Get the tool return message from the implementation
    const toolReturn = `Entered plan mode. You should now focus on exploring the codebase and designing an implementation approach.

In plan mode, you should:
1. Thoroughly explore the codebase to understand existing patterns
2. Identify similar features and architectural approaches
3. Consider multiple approaches and their trade-offs
4. Use AskUserQuestion if you need to clarify the approach
5. Design a concrete implementation strategy
6. When ready, use ExitPlanMode to present your plan for approval

Remember: DO NOT write or edit any files yet. This is a read-only exploration and planning phase.

Plan file path: ${planFilePath}`;

    const precomputedResult: ToolExecutionResult = {
      toolReturn,
      status: "success",
    };

    // Update buffers with tool return
    onChunk(buffersRef.current, {
      message_type: "tool_return_message",
      id: "dummy",
      date: new Date().toISOString(),
      tool_call_id: approval.toolCallId,
      tool_return: toolReturn,
      status: "success",
      stdout: null,
      stderr: null,
    });

    setThinkingMessage(getRandomThinkingVerb());
    refreshDerived();

    const decision = {
      type: "approve" as const,
      approval,
      precomputedResult,
    };

    if (isLast) {
      setIsExecutingTool(true);
      await sendAllResults(decision);
    } else {
      setApprovalResults((prev) => [...prev, decision]);
    }
  }, [pendingApprovals, approvalResults, sendAllResults, refreshDerived]);

  const handleEnterPlanModeReject = useCallback(async () => {
    const currentIndex = approvalResults.length;
    const approval = pendingApprovals[currentIndex];
    if (!approval) return;

    const isLast = currentIndex + 1 >= pendingApprovals.length;

    const rejectionReason =
      "User chose to skip plan mode and start implementing directly.";

    const decision = {
      type: "deny" as const,
      approval,
      reason: rejectionReason,
    };

    if (isLast) {
      setIsExecutingTool(true);
      await sendAllResults(decision);
    } else {
      setApprovalResults((prev) => [...prev, decision]);
    }
  }, [pendingApprovals, approvalResults, sendAllResults]);

  // Live area shows only in-progress items
  // biome-ignore lint/correctness/useExhaustiveDependencies: staticItems.length and deferredCommitAt are intentional triggers to recompute when items are promoted to static or deferred commits complete
  const liveItems = useMemo(() => {
    return lines.filter((ln) => {
      if (!("phase" in ln)) return false;
      if (emittedIdsRef.current.has(ln.id)) return false;
      if (ln.kind === "command" || ln.kind === "bash_command") {
        return ln.phase === "running";
      }
      if (ln.kind === "tool_call") {
        // Task tool_calls need special handling:
        // - Only include if pending approval (phase: "ready" or "streaming")
        // - Running/finished Task tools are handled by SubagentGroupDisplay
        if (ln.name && isTaskTool(ln.name)) {
          // Only show Task tools that are awaiting approval (not running/finished)
          return ln.phase === "ready" || ln.phase === "streaming";
        }
        // Always show other tool calls in progress
        return (
          ln.phase !== "finished" ||
          deferredToolCallCommitsRef.current.has(ln.id)
        );
      }
      // Events (like compaction) show while running
      if (ln.kind === "event") {
        if (!showCompactionsEnabled && ln.eventType === "compaction")
          return false;
        return ln.phase === "running";
      }
      if (!tokenStreamingEnabled && ln.phase === "streaming") return false;
      return ln.phase === "streaming";
    });
  }, [
    lines,
    tokenStreamingEnabled,
    showCompactionsEnabled,
    staticItems.length,
    deferredCommitAt,
  ]);

  // Subscribe to subagent state for reactive overflow detection
  const { agents: subagents } = useSyncExternalStore(
    subscribeToSubagents,
    getSubagentSnapshot,
  );

  // Estimate live area height for overflow detection.
  const estimatedLiveHeight = useMemo(() => {
    // Count actual lines in live content by counting newlines
    const countLines = (text: string | undefined): number => {
      if (!text) return 0;
      return (text.match(/\n/g) || []).length + 1;
    };

    // Estimate height for each live item based on actual content
    let liveItemsHeight = 0;
    for (const item of liveItems) {
      // Base height for each item (header line, margins)
      let itemHeight = 2;

      if (item.kind === "bash_command" || item.kind === "command") {
        // Count lines in command input and output
        itemHeight += countLines(item.input);
        itemHeight += countLines(item.output);
      } else if (item.kind === "tool_call") {
        // Count lines in tool args and result
        itemHeight += Math.min(countLines(item.argsText), 5); // Cap args display
        itemHeight += countLines(item.resultText);
      } else if (
        item.kind === "assistant" ||
        item.kind === "reasoning" ||
        item.kind === "error"
      ) {
        itemHeight += countLines(item.text);
      }

      liveItemsHeight += itemHeight;
    }

    // Subagents: 4 lines each (description + URL + status + margin)
    const LINES_PER_SUBAGENT = 4;
    const subagentsHeight = subagents.length * LINES_PER_SUBAGENT;

    // Fixed buffer for header, input area, status bar, margins
    // Using larger buffer to catch edge cases and account for timing lag
    const FIXED_BUFFER = 20;

    const estimatedHeight = liveItemsHeight + subagentsHeight + FIXED_BUFFER;

    return estimatedHeight;
  }, [liveItems, subagents.length]);

  // Overflow detection with hysteresis: disable quickly on overflow, re-enable
  // only after we've recovered extra headroom to avoid flap near the boundary.
  const [shouldAnimate, setShouldAnimate] = useState(
    () => estimatedLiveHeight < terminalRows,
  );
  useEffect(() => {
    if (terminalRows <= 0) {
      setShouldAnimate(false);
      return;
    }

    const disableThreshold = terminalRows;
    const resumeThreshold = Math.max(
      0,
      terminalRows - ANIMATION_RESUME_HYSTERESIS_ROWS,
    );

    setShouldAnimate((prev) => {
      if (prev) {
        return estimatedLiveHeight < disableThreshold;
      }
      return estimatedLiveHeight < resumeThreshold;
    });
  }, [estimatedLiveHeight, terminalRows]);

  // Commit welcome snapshot once when ready for fresh sessions (no history)
  // Wait for agentProvenance to be available for new agents (continueSession=false)
  useEffect(() => {
    if (
      loadingState === "ready" &&
      !welcomeCommittedRef.current &&
      messageHistory.length === 0
    ) {
      // For new agents, wait until provenance is available
      // For resumed agents, provenance stays null (that's expected)
      if (!continueSession && !agentProvenance) {
        return; // Wait for provenance to be set
      }
      welcomeCommittedRef.current = true;
      setStaticItems((prev) => [
        ...prev,
        {
          kind: "welcome",
          id: `welcome-${Date.now().toString(36)}`,
          snapshot: {
            continueSession,
            agentState,
            agentProvenance,
            terminalWidth: columns,
          },
        },
      ]);

      // Add status line showing agent info
      const statusId = `status-agent-${Date.now().toString(36)}`;

      // Check if agent is pinned (locally or globally)
      const isPinned = agentState?.id
        ? settingsManager.getLocalPinnedAgents().includes(agentState.id) ||
          settingsManager.getGlobalPinnedAgents().includes(agentState.id)
        : false;

      // Build status message based on session type
      const agentName = agentState?.name || "Unnamed Agent";
      const headerMessage = resumedExistingConversation
        ? `Resuming (empty) conversation with **${agentName}**`
        : continueSession
          ? `Starting new conversation with **${agentName}**`
          : "Creating a new agent";

      // Command hints - for pinned agents show /memory, for unpinned show /pin
      const commandHints = isPinned
        ? [
            "→ **/agents**    list all agents",
            "→ **/resume**    resume a previous conversation",
            "→ **/memory**    view your agent's memory blocks",
            "→ **/init**      initialize your agent's memory",
            "→ **/remember**  teach your agent",
          ]
        : [
            "→ **/agents**    list all agents",
            "→ **/resume**    resume a previous conversation",
            "→ **/pin**       save + name your agent",
            "→ **/init**      initialize your agent's memory",
            "→ **/remember**  teach your agent",
          ];

      // Build status lines with optional release notes above header
      const statusLines: string[] = [];

      // Add release notes first (above everything) - same styling as rest of status block
      if (releaseNotes) {
        statusLines.push(releaseNotes);
        statusLines.push(""); // blank line separator
      }

      statusLines.push(headerMessage);
      statusLines.push(...commandHints);

      buffersRef.current.byId.set(statusId, {
        kind: "status",
        id: statusId,
        lines: statusLines,
      });
      buffersRef.current.order.push(statusId);
      refreshDerived();
      commitEligibleLines(buffersRef.current, { deferToolCalls: false });
    }
  }, [
    loadingState,
    continueSession,
    resumedExistingConversation,
    messageHistory.length,
    commitEligibleLines,
    columns,
    agentProvenance,
    agentState,
    refreshDerived,
    releaseNotes,
  ]);

  const liveTrajectorySnapshot =
    sessionStatsRef.current.getTrajectorySnapshot();
  const liveTrajectoryTokenBase =
    liveTrajectorySnapshot?.tokens ?? trajectoryTokenBase;
  const liveTrajectoryElapsedBaseMs =
    liveTrajectorySnapshot?.wallMs ?? trajectoryElapsedBaseMs;
  const runTokenDelta = Math.max(
    0,
    tokenCount - trajectoryRunTokenStartRef.current,
  );
  const trajectoryTokenDisplay = Math.max(
    liveTrajectoryTokenBase + runTokenDelta,
    trajectoryTokenDisplayRef.current,
  );
  const inputVisible = !showExitStats;
  const inputEnabled =
    !showExitStats && pendingApprovals.length === 0 && !anySelectorOpen;
  const currentApprovalPreviewCommitted = currentApproval?.toolCallId
    ? eagerCommittedPreviewsRef.current.has(currentApproval.toolCallId)
    : false;
  const showApprovalPreview =
    !currentApprovalShouldCommitPreview && !currentApprovalPreviewCommitted;

  useEffect(() => {
    trajectoryTokenDisplayRef.current = trajectoryTokenDisplay;
  }, [trajectoryTokenDisplay]);

  return (
    <Box key={resumeKey} flexDirection="column">
      <Static
        key={staticRenderEpoch}
        items={staticItems}
        style={{ flexDirection: "column" }}
      >
        {(item: StaticItem, index: number) => {
          return (
            <Box key={item.id} marginTop={index > 0 ? 1 : 0}>
              {item.kind === "welcome" ? (
                <WelcomeScreen loadingState="ready" {...item.snapshot} />
              ) : item.kind === "user" ? (
                <UserMessage line={item} />
              ) : item.kind === "reasoning" ? (
                <ReasoningMessage line={item} />
              ) : item.kind === "assistant" ? (
                <AssistantMessage line={item} />
              ) : item.kind === "tool_call" ? (
                <ToolCallMessage
                  line={item}
                  precomputedDiffs={precomputedDiffsRef.current}
                  lastPlanFilePath={lastPlanFilePathRef.current}
                />
              ) : item.kind === "subagent_group" ? (
                <SubagentGroupStatic agents={item.agents} />
              ) : item.kind === "error" ? (
                <ErrorMessage line={item} />
              ) : item.kind === "status" ? (
                <StatusMessage line={item} />
              ) : item.kind === "event" ? (
                !showCompactionsEnabled &&
                item.eventType === "compaction" ? null : (
                  <EventMessage line={item} />
                )
              ) : item.kind === "separator" ? (
                <Box marginTop={1}>
                  <Text dimColor>{"─".repeat(columns)}</Text>
                </Box>
              ) : item.kind === "command" ? (
                <CommandMessage line={item} />
              ) : item.kind === "bash_command" ? (
                <BashCommandMessage line={item} />
              ) : item.kind === "trajectory_summary" ? (
                <TrajectorySummary line={item} />
              ) : item.kind === "approval_preview" ? (
                <ApprovalPreview
                  toolName={item.toolName}
                  toolArgs={item.toolArgs}
                  precomputedDiff={item.precomputedDiff}
                  allDiffs={precomputedDiffsRef.current}
                  planContent={item.planContent}
                  planFilePath={item.planFilePath}
                  toolCallId={item.toolCallId}
                />
              ) : null}
            </Box>
          );
        }}
      </Static>

      <Box flexDirection="column">
        {/* Loading screen / intro text */}
        {loadingState !== "ready" && (
          <WelcomeScreen
            loadingState={loadingState}
            continueSession={continueSession}
            agentState={agentState}
          />
        )}

        {loadingState === "ready" && (
          <>
            {/* Transcript - wrapped in AnimationProvider for overflow-based animation control */}
            <AnimationProvider shouldAnimate={shouldAnimate}>
              {/* Show liveItems always - all approvals now render inline */}
              {liveItems.length > 0 && (
                <Box flexDirection="column">
                  {liveItems.map((ln) => {
                    const isFileTool =
                      ln.kind === "tool_call" &&
                      ln.name &&
                      (isFileEditTool(ln.name) ||
                        isFileWriteTool(ln.name) ||
                        isPatchTool(ln.name));
                    const isApprovalTracked =
                      ln.kind === "tool_call" &&
                      ln.toolCallId &&
                      (ln.toolCallId === currentApproval?.toolCallId ||
                        pendingIds.has(ln.toolCallId) ||
                        queuedIds.has(ln.toolCallId));
                    if (isFileTool && !isApprovalTracked) {
                      return null;
                    }
                    // Skip Task tools that don't have a pending approval
                    // They render as empty Boxes (ToolCallMessage returns null for non-finished Task tools)
                    // which causes N blank lines when N Task tools are called in parallel
                    // Note: pendingIds doesn't include the ACTIVE approval (currentApproval),
                    // so we must also check if this is the active approval
                    if (
                      ln.kind === "tool_call" &&
                      ln.name &&
                      isTaskTool(ln.name) &&
                      ln.toolCallId &&
                      !pendingIds.has(ln.toolCallId) &&
                      ln.toolCallId !== currentApproval?.toolCallId
                    ) {
                      return null;
                    }

                    // Check if this tool call matches the current approval awaiting user input
                    const matchesCurrentApproval =
                      ln.kind === "tool_call" &&
                      currentApproval &&
                      ln.toolCallId === currentApproval.toolCallId;

                    return (
                      <Box key={ln.id} flexDirection="column" marginTop={1}>
                        {matchesCurrentApproval ? (
                          <ApprovalSwitch
                            approval={currentApproval}
                            onApprove={handleApproveCurrent}
                            onApproveAlways={handleApproveAlways}
                            onDeny={handleDenyCurrent}
                            onCancel={handleCancelApprovals}
                            onPlanApprove={handlePlanApprove}
                            onPlanKeepPlanning={handlePlanKeepPlanning}
                            onQuestionSubmit={handleQuestionSubmit}
                            onEnterPlanModeApprove={handleEnterPlanModeApprove}
                            onEnterPlanModeReject={handleEnterPlanModeReject}
                            precomputedDiff={
                              ln.toolCallId
                                ? precomputedDiffsRef.current.get(ln.toolCallId)
                                : undefined
                            }
                            allDiffs={precomputedDiffsRef.current}
                            isFocused={true}
                            approveAlwaysText={
                              currentApprovalContext?.approveAlwaysText
                            }
                            allowPersistence={
                              currentApprovalContext?.allowPersistence ?? true
                            }
                            showPreview={showApprovalPreview}
                          />
                        ) : ln.kind === "user" ? (
                          <UserMessage line={ln} />
                        ) : ln.kind === "reasoning" ? (
                          <ReasoningMessage line={ln} />
                        ) : ln.kind === "assistant" ? (
                          <AssistantMessage line={ln} />
                        ) : ln.kind === "tool_call" &&
                          ln.toolCallId &&
                          queuedIds.has(ln.toolCallId) ? (
                          // Render stub for queued (decided but not executed) approval
                          <PendingApprovalStub
                            toolName={
                              approvalMap.get(ln.toolCallId)?.toolName ||
                              ln.name ||
                              "Unknown"
                            }
                            description={stubDescriptions.get(ln.toolCallId)}
                            decision={queuedDecisions.get(ln.toolCallId)}
                          />
                        ) : ln.kind === "tool_call" &&
                          ln.toolCallId &&
                          pendingIds.has(ln.toolCallId) ? (
                          // Render stub for pending (undecided) approval
                          <PendingApprovalStub
                            toolName={
                              approvalMap.get(ln.toolCallId)?.toolName ||
                              ln.name ||
                              "Unknown"
                            }
                            description={stubDescriptions.get(ln.toolCallId)}
                          />
                        ) : ln.kind === "tool_call" ? (
                          <ToolCallMessage
                            line={ln}
                            precomputedDiffs={precomputedDiffsRef.current}
                            lastPlanFilePath={lastPlanFilePathRef.current}
                            isStreaming={streaming}
                          />
                        ) : ln.kind === "error" ? (
                          <ErrorMessage line={ln} />
                        ) : ln.kind === "status" ? (
                          <StatusMessage line={ln} />
                        ) : ln.kind === "event" ? (
                          <EventMessage line={ln} />
                        ) : ln.kind === "command" ? (
                          <CommandMessage line={ln} />
                        ) : ln.kind === "bash_command" ? (
                          <BashCommandMessage line={ln} />
                        ) : null}
                      </Box>
                    );
                  })}
                </Box>
              )}

              {/* Fallback approval UI when backfill is disabled (no liveItems) */}
              {liveItems.length === 0 && currentApproval && (
                <Box flexDirection="column">
                  <ApprovalSwitch
                    approval={currentApproval}
                    onApprove={handleApproveCurrent}
                    onApproveAlways={handleApproveAlways}
                    onDeny={handleDenyCurrent}
                    onCancel={handleCancelApprovals}
                    onPlanApprove={handlePlanApprove}
                    onPlanKeepPlanning={handlePlanKeepPlanning}
                    onQuestionSubmit={handleQuestionSubmit}
                    onEnterPlanModeApprove={handleEnterPlanModeApprove}
                    onEnterPlanModeReject={handleEnterPlanModeReject}
                    allDiffs={precomputedDiffsRef.current}
                    isFocused={true}
                    approveAlwaysText={
                      currentApprovalContext?.approveAlwaysText
                    }
                    allowPersistence={
                      currentApprovalContext?.allowPersistence ?? true
                    }
                    showPreview={showApprovalPreview}
                  />
                </Box>
              )}

              {/* Subagent group display - shows running/completed subagents */}
              <SubagentGroupDisplay />
            </AnimationProvider>

            {/* Exit stats - shown when exiting via double Ctrl+C */}
            {showExitStats &&
              (() => {
                const stats = sessionStatsRef.current.getSnapshot();
                return (
                  <Box flexDirection="column" marginTop={1}>
                    {/* Alien + Stats (3 lines) */}
                    <Box>
                      <Text color={colors.footer.agentName}>{" ▗▖▗▖   "}</Text>
                      <Text dimColor>
                        Total duration (API): {formatDuration(stats.totalApiMs)}
                      </Text>
                    </Box>
                    <Box>
                      <Text color={colors.footer.agentName}>{"▙█▜▛█▟  "}</Text>
                      <Text dimColor>
                        Total duration (wall):{" "}
                        {formatDuration(stats.totalWallMs)}
                      </Text>
                    </Box>
                    <Box>
                      <Text color={colors.footer.agentName}>{"▝▜▛▜▛▘  "}</Text>
                      <Text dimColor>
                        Session usage: {stats.usage.stepCount} steps,{" "}
                        {formatCompact(stats.usage.promptTokens)} input,{" "}
                        {formatCompact(stats.usage.completionTokens)} output
                      </Text>
                    </Box>
                    {/* Resume commands (no alien) */}
                    <Box height={1} />
                    <Text dimColor>Resume this agent with:</Text>
                    <Text color={colors.link.url}>
                      {/* Show -n "name" if agent has name and is pinned, otherwise --agent */}
                      {agentName &&
                      (settingsManager
                        .getLocalPinnedAgents()
                        .includes(agentId) ||
                        settingsManager
                          .getGlobalPinnedAgents()
                          .includes(agentId))
                        ? `letta -n "${agentName}"`
                        : `letta --agent ${agentId}`}
                    </Text>
                    {/* Only show conversation hint if not on default (default is resumed automatically) */}
                    {conversationId !== "default" && (
                      <>
                        <Box height={1} />
                        <Text dimColor>Resume this conversation with:</Text>
                        <Text color={colors.link.url}>
                          {`letta --conv ${conversationId}`}
                        </Text>
                      </>
                    )}
                  </Box>
                );
              })()}

            {/* Input row - always mounted to preserve state */}
            <Box marginTop={1}>
              <Input
                visible={inputVisible}
                streaming={streaming}
                tokenCount={trajectoryTokenDisplay}
                elapsedBaseMs={liveTrajectoryElapsedBaseMs}
                thinkingMessage={thinkingMessage}
                onSubmit={onSubmit}
                onBashSubmit={handleBashSubmit}
                bashRunning={bashRunning}
                onBashInterrupt={handleBashInterrupt}
                inputEnabled={inputEnabled}
                collapseInputWhenDisabled={
                  pendingApprovals.length > 0 || anySelectorOpen
                }
                permissionMode={uiPermissionMode}
                onPermissionModeChange={handlePermissionModeChange}
                onExit={handleExit}
                onInterrupt={handleInterrupt}
                interruptRequested={interruptRequested}
                agentId={agentId}
                agentName={agentName}
                currentModel={currentModelDisplay}
                currentModelProvider={currentModelProvider}
                messageQueue={messageQueue}
                onEnterQueueEditMode={handleEnterQueueEditMode}
                onEscapeCancel={
                  profileConfirmPending ? handleProfileEscapeCancel : undefined
                }
                ralphActive={uiRalphActive}
                ralphPending={pendingRalphConfig !== null}
                ralphPendingYolo={pendingRalphConfig?.isYolo ?? false}
                onRalphExit={handleRalphExit}
                conversationId={conversationId}
                onPasteError={handlePasteError}
                restoredInput={restoredInput}
                onRestoredInputConsumed={() => setRestoredInput(null)}
                networkPhase={networkPhase}
                terminalWidth={columns}
                shouldAnimate={shouldAnimate}
                statusLineText={statusLine.text || undefined}
                statusLineRight={statusLine.rightText || undefined}
                statusLinePadding={statusLine.padding || 0}
              />
            </Box>

            {/* Model Selector - conditionally mounted as overlay */}
            {activeOverlay === "model" && (
              <ModelSelector
                currentModelId={currentModelId ?? undefined}
                onSelect={handleModelSelect}
                onCancel={closeOverlay}
                filterProvider={modelSelectorOptions.filterProvider}
                forceRefresh={modelSelectorOptions.forceRefresh}
                billingTier={billingTier ?? undefined}
                isSelfHosted={(() => {
                  const settings = settingsManager.getSettings();
                  const baseURL =
                    process.env.LETTA_BASE_URL ||
                    settings.env?.LETTA_BASE_URL ||
                    "https://api.letta.com";
                  return !baseURL.includes("api.letta.com");
                })()}
              />
            )}

            {activeOverlay === "sleeptime" && (
              <SleeptimeSelector
                initialSettings={getReflectionSettings()}
                memfsEnabled={settingsManager.isMemfsEnabled(agentId)}
                onSave={handleSleeptimeModeSelect}
                onCancel={closeOverlay}
              />
            )}

            {/* Provider Selector - for connecting BYOK providers */}
            {activeOverlay === "connect" && (
              <ProviderSelector
                onCancel={closeOverlay}
                onStartOAuth={async () => {
                  const overlayCommand = consumeOverlayCommand("connect");
                  // Close selector and start OAuth flow
                  closeOverlay();
                  const cmd =
                    overlayCommand ??
                    commandRunner.start("/connect", "Starting connection...");
                  const {
                    handleConnect,
                    setActiveCommandId: setActiveConnectCommandId,
                  } = await import("./commands/connect");
                  setActiveConnectCommandId(cmd.id);
                  try {
                    await handleConnect(
                      {
                        buffersRef,
                        refreshDerived,
                        setCommandRunning,
                        onCodexConnected: () => {
                          setModelSelectorOptions({
                            filterProvider: "chatgpt-plus-pro",
                            forceRefresh: true,
                          });
                          startOverlayCommand(
                            "model",
                            "/model",
                            "Opening model selector...",
                            "Models dialog dismissed",
                          );
                          setActiveOverlay("model");
                        },
                      },
                      "/connect codex",
                    );
                  } finally {
                    setActiveConnectCommandId(null);
                  }
                }}
              />
            )}

            {/* Toolset Selector - conditionally mounted as overlay */}
            {activeOverlay === "toolset" && (
              <ToolsetSelector
                currentToolset={currentToolset ?? undefined}
                onSelect={handleToolsetSelect}
                onCancel={closeOverlay}
              />
            )}

            {/* System Prompt Selector - conditionally mounted as overlay */}
            {activeOverlay === "system" && (
              <SystemPromptSelector
                currentPromptId={currentSystemPromptId ?? undefined}
                onSelect={handleSystemPromptSelect}
                onCancel={closeOverlay}
              />
            )}

            {/* Subagent Manager - for managing custom subagents */}
            {activeOverlay === "subagent" && (
              <SubagentManager onClose={closeOverlay} />
            )}

            {/* Agent Selector - for browsing/selecting agents */}
            {activeOverlay === "resume" && (
              <AgentSelector
                currentAgentId={agentId}
                onSelect={async (id) => {
                  const overlayCommand = consumeOverlayCommand("resume");
                  closeOverlay();
                  await handleAgentSelect(id, {
                    commandId: overlayCommand?.id,
                  });
                }}
                onCancel={closeOverlay}
                onCreateNewAgent={() => {
                  closeOverlay();
                  setActiveOverlay("new");
                }}
              />
            )}

            {/* Conversation Selector - for resuming conversations */}
            {activeOverlay === "conversations" && (
              <ConversationSelector
                agentId={agentId}
                agentName={agentName ?? undefined}
                currentConversationId={conversationId}
                onSelect={async (convId, selectorContext) => {
                  const overlayCommand = consumeOverlayCommand("conversations");
                  closeOverlay();

                  // Skip if already on this conversation
                  if (convId === conversationId) {
                    const cmd =
                      overlayCommand ??
                      commandRunner.start(
                        "/resume",
                        "Already on this conversation",
                      );
                    cmd.finish("Already on this conversation", true);
                    return;
                  }

                  // If agent is busy, queue the switch for after end_turn
                  if (isAgentBusy()) {
                    const cmd =
                      overlayCommand ??
                      commandRunner.start(
                        "/resume",
                        "Conversation switch queued – will switch after current task completes",
                      );
                    cmd.update({
                      output:
                        "Conversation switch queued – will switch after current task completes",
                      phase: "running",
                    });
                    setQueuedOverlayAction({
                      type: "switch_conversation",
                      conversationId: convId,
                      commandId: cmd.id,
                    });
                    return;
                  }

                  // Lock input for async operation
                  setCommandRunning(true);

                  const inputCmd = "/resume";
                  const cmd =
                    overlayCommand ??
                    commandRunner.start(inputCmd, "Switching conversation...");
                  cmd.update({
                    output: "Switching conversation...",
                    phase: "running",
                  });

                  try {
                    // Validate conversation exists BEFORE updating state
                    // (getResumeData throws 404/422 for non-existent conversations)
                    if (agentState) {
                      const client = await getClient();
                      const resumeData = await getResumeData(
                        client,
                        agentState,
                        convId,
                      );

                      // Only update state after validation succeeds
                      setConversationId(convId);

                      pendingConversationSwitchRef.current = {
                        origin: "resume-selector",
                        conversationId: convId,
                        isDefault: convId === "default",
                        messageCount:
                          selectorContext?.messageCount ??
                          resumeData.messageHistory.length,
                        summary: selectorContext?.summary,
                        messageHistory: resumeData.messageHistory,
                      };

                      settingsManager.setLocalLastSession(
                        { agentId, conversationId: convId },
                        process.cwd(),
                      );
                      settingsManager.setGlobalLastSession({
                        agentId,
                        conversationId: convId,
                      });

                      // Build success command with agent + conversation info
                      const currentAgentName =
                        agentState.name || "Unnamed Agent";
                      const successLines =
                        resumeData.messageHistory.length > 0
                          ? [
                              `Resumed conversation with "${currentAgentName}"`,
                              `⎿  Agent: ${agentId}`,
                              `⎿  Conversation: ${convId}`,
                            ]
                          : [
                              `Switched to conversation with "${currentAgentName}"`,
                              `⎿  Agent: ${agentId}`,
                              `⎿  Conversation: ${convId} (empty)`,
                            ];
                      const successOutput = successLines.join("\n");
                      cmd.finish(successOutput, true);
                      const successItem: StaticItem = {
                        kind: "command",
                        id: cmd.id,
                        input: cmd.input,
                        output: successOutput,
                        phase: "finished",
                        success: true,
                      };

                      // Clear current transcript and static items
                      buffersRef.current.byId.clear();
                      buffersRef.current.order = [];
                      buffersRef.current.tokenCount = 0;
                      resetContextHistory(contextTrackerRef.current);
                      emittedIdsRef.current.clear();
                      resetDeferredToolCallCommits();
                      setStaticItems([]);
                      setStaticRenderEpoch((e) => e + 1);
                      resetTrajectoryBases();

                      // Backfill message history with visual separator
                      if (resumeData.messageHistory.length > 0) {
                        hasBackfilledRef.current = false;
                        backfillBuffers(
                          buffersRef.current,
                          resumeData.messageHistory,
                        );
                        // Collect backfilled items
                        const backfilledItems: StaticItem[] = [];
                        for (const id of buffersRef.current.order) {
                          const ln = buffersRef.current.byId.get(id);
                          if (!ln) continue;
                          emittedIdsRef.current.add(id);
                          backfilledItems.push({ ...ln } as StaticItem);
                        }
                        // Add separator before backfilled messages, then success at end
                        const separator = {
                          kind: "separator" as const,
                          id: uid("sep"),
                        };
                        setStaticItems([
                          separator,
                          ...backfilledItems,
                          successItem,
                        ]);
                        setLines(toLines(buffersRef.current));
                        hasBackfilledRef.current = true;
                      } else {
                        // Add separator for visual spacing even without backfill
                        const separator = {
                          kind: "separator" as const,
                          id: uid("sep"),
                        };
                        setStaticItems([separator, successItem]);
                        setLines(toLines(buffersRef.current));
                      }

                      // Restore pending approvals if any (fixes #540 for ConversationSelector)
                      if (resumeData.pendingApprovals.length > 0) {
                        setPendingApprovals(resumeData.pendingApprovals);

                        // Analyze approval contexts (same logic as startup)
                        try {
                          const contexts = await Promise.all(
                            resumeData.pendingApprovals.map(
                              async (approval) => {
                                const parsedArgs = safeJsonParseOr<
                                  Record<string, unknown>
                                >(approval.toolArgs, {});
                                return await analyzeToolApproval(
                                  approval.toolName,
                                  parsedArgs,
                                );
                              },
                            ),
                          );
                          setApprovalContexts(contexts);
                        } catch (approvalError) {
                          // If analysis fails, leave context as null (will show basic options)
                          console.error(
                            "Failed to analyze resume approvals:",
                            approvalError,
                          );
                        }
                      }
                    }
                  } catch (error) {
                    // Update existing loading message instead of creating new one
                    // Format error message to be user-friendly (avoid raw JSON/internal details)
                    let errorMsg = "Unknown error";
                    if (error instanceof APIError) {
                      if (error.status === 404) {
                        errorMsg = "Conversation not found";
                      } else if (error.status === 422) {
                        errorMsg = "Invalid conversation ID";
                      } else {
                        errorMsg = error.message;
                      }
                    } else if (error instanceof Error) {
                      errorMsg = error.message;
                    }
                    cmd.fail(`Failed to switch conversation: ${errorMsg}`);
                  } finally {
                    setCommandRunning(false);
                  }
                }}
                onNewConversation={async () => {
                  const overlayCommand = consumeOverlayCommand("conversations");
                  closeOverlay();

                  // Lock input for async operation
                  setCommandRunning(true);

                  const cmd =
                    overlayCommand ??
                    commandRunner.start(
                      "/resume",
                      "Creating new conversation...",
                    );
                  cmd.update({
                    output: "Creating new conversation...",
                    phase: "running",
                  });

                  try {
                    // Create a new conversation
                    const client = await getClient();
                    const conversation = await client.conversations.create({
                      agent_id: agentId,
                      isolated_block_labels: [...ISOLATED_BLOCK_LABELS],
                    });
                    setConversationId(conversation.id);
                    settingsManager.setLocalLastSession(
                      { agentId, conversationId: conversation.id },
                      process.cwd(),
                    );
                    settingsManager.setGlobalLastSession({
                      agentId,
                      conversationId: conversation.id,
                    });

                    // Build success command with agent + conversation info
                    const currentAgentName =
                      agentState?.name || "Unnamed Agent";
                    const shortConvId = conversation.id.slice(0, 20);
                    const successLines = [
                      `Started new conversation with "${currentAgentName}"`,
                      `⎿  Agent: ${agentId}`,
                      `⎿  Conversation: ${shortConvId}... (new)`,
                    ];
                    const successOutput = successLines.join("\n");
                    cmd.finish(successOutput, true);
                    const successItem: StaticItem = {
                      kind: "command",
                      id: cmd.id,
                      input: cmd.input,
                      output: successOutput,
                      phase: "finished",
                      success: true,
                    };

                    // Clear current transcript and static items
                    buffersRef.current.byId.clear();
                    buffersRef.current.order = [];
                    buffersRef.current.tokenCount = 0;
                    resetContextHistory(contextTrackerRef.current);
                    emittedIdsRef.current.clear();
                    resetDeferredToolCallCommits();
                    setStaticItems([]);
                    setStaticRenderEpoch((e) => e + 1);
                    resetTrajectoryBases();
                    setStaticItems([successItem]);
                    setLines(toLines(buffersRef.current));
                  } catch (error) {
                    cmd.fail(
                      `Failed to create conversation: ${error instanceof Error ? error.message : String(error)}`,
                    );
                  } finally {
                    setCommandRunning(false);
                  }
                }}
                onCancel={closeOverlay}
              />
            )}

            {/* Message Search - conditionally mounted as overlay */}
            {activeOverlay === "search" && (
              <MessageSearch
                onClose={closeOverlay}
                initialQuery={searchQuery || undefined}
                agentId={agentId}
                conversationId={conversationId}
                onOpenConversation={async (
                  targetAgentId,
                  targetConvId,
                  searchContext,
                ) => {
                  const overlayCommand = consumeOverlayCommand("search");
                  closeOverlay();

                  // Different agent: use handleAgentSelect (which supports optional conversationId)
                  if (targetAgentId !== agentId) {
                    await handleAgentSelect(targetAgentId, {
                      conversationId: targetConvId,
                      commandId: overlayCommand?.id,
                    });
                    return;
                  }

                  // Normalize undefined/null to "default"
                  const actualTargetConv = targetConvId || "default";

                  // Same agent, same conversation: nothing to do
                  if (actualTargetConv === conversationId) {
                    const cmd =
                      overlayCommand ??
                      commandRunner.start(
                        "/search",
                        "Already on this conversation",
                      );
                    cmd.finish("Already on this conversation", true);
                    return;
                  }

                  // Same agent, different conversation: switch conversation
                  // (Reuses ConversationSelector's onSelect logic pattern)
                  if (isAgentBusy()) {
                    const cmd =
                      overlayCommand ??
                      commandRunner.start(
                        "/search",
                        "Conversation switch queued – will switch after current task completes",
                      );
                    cmd.update({
                      output:
                        "Conversation switch queued – will switch after current task completes",
                      phase: "running",
                    });
                    setQueuedOverlayAction({
                      type: "switch_conversation",
                      conversationId: actualTargetConv,
                      commandId: cmd.id,
                    });
                    return;
                  }

                  setCommandRunning(true);
                  const cmd =
                    overlayCommand ??
                    commandRunner.start("/search", "Switching conversation...");
                  cmd.update({
                    output: "Switching conversation...",
                    phase: "running",
                  });

                  try {
                    if (agentState) {
                      const client = await getClient();
                      const resumeData = await getResumeData(
                        client,
                        agentState,
                        actualTargetConv,
                      );

                      setConversationId(actualTargetConv);

                      pendingConversationSwitchRef.current = {
                        origin: "search",
                        conversationId: actualTargetConv,
                        isDefault: actualTargetConv === "default",
                        messageCount: resumeData.messageHistory.length,
                        messageHistory: resumeData.messageHistory,
                        searchQuery: searchContext?.query,
                        searchMessage: searchContext?.message,
                      };

                      settingsManager.setLocalLastSession(
                        { agentId, conversationId: actualTargetConv },
                        process.cwd(),
                      );
                      settingsManager.setGlobalLastSession({
                        agentId,
                        conversationId: actualTargetConv,
                      });

                      const currentAgentName =
                        agentState.name || "Unnamed Agent";
                      const successOutput = [
                        `Switched to conversation with "${currentAgentName}"`,
                        `⎿  Conversation: ${actualTargetConv}`,
                      ].join("\n");
                      cmd.finish(successOutput, true);
                      const successItem: StaticItem = {
                        kind: "command",
                        id: cmd.id,
                        input: cmd.input,
                        output: successOutput,
                        phase: "finished",
                        success: true,
                      };

                      // Clear current transcript and static items
                      buffersRef.current.byId.clear();
                      buffersRef.current.order = [];
                      buffersRef.current.tokenCount = 0;
                      resetContextHistory(contextTrackerRef.current);
                      emittedIdsRef.current.clear();
                      resetDeferredToolCallCommits();
                      setStaticItems([]);
                      setStaticRenderEpoch((e) => e + 1);
                      resetTrajectoryBases();

                      // Backfill message history
                      if (resumeData.messageHistory.length > 0) {
                        hasBackfilledRef.current = false;
                        backfillBuffers(
                          buffersRef.current,
                          resumeData.messageHistory,
                        );
                        const backfilledItems: StaticItem[] = [];
                        for (const id of buffersRef.current.order) {
                          const ln = buffersRef.current.byId.get(id);
                          if (!ln) continue;
                          emittedIdsRef.current.add(id);
                          backfilledItems.push({ ...ln } as StaticItem);
                        }
                        const separator = {
                          kind: "separator" as const,
                          id: uid("sep"),
                        };
                        setStaticItems([
                          separator,
                          ...backfilledItems,
                          successItem,
                        ]);
                        setLines(toLines(buffersRef.current));
                        hasBackfilledRef.current = true;
                      } else {
                        const separator = {
                          kind: "separator" as const,
                          id: uid("sep"),
                        };
                        setStaticItems([separator, successItem]);
                        setLines(toLines(buffersRef.current));
                      }

                      // Restore pending approvals if any
                      if (resumeData.pendingApprovals.length > 0) {
                        setPendingApprovals(resumeData.pendingApprovals);
                        try {
                          const contexts = await Promise.all(
                            resumeData.pendingApprovals.map(
                              async (approval) => {
                                const parsedArgs = safeJsonParseOr<
                                  Record<string, unknown>
                                >(approval.toolArgs, {});
                                return await analyzeToolApproval(
                                  approval.toolName,
                                  parsedArgs,
                                );
                              },
                            ),
                          );
                          setApprovalContexts(contexts);
                        } catch {
                          // If analysis fails, leave context as null
                        }
                      }
                    }
                  } catch (error) {
                    let errorMsg = "Unknown error";
                    if (error instanceof APIError) {
                      if (error.status === 404) {
                        errorMsg = "Conversation not found";
                      } else if (error.status === 422) {
                        errorMsg = "Invalid conversation ID";
                      } else {
                        errorMsg = error.message;
                      }
                    } else if (error instanceof Error) {
                      errorMsg = error.message;
                    }
                    cmd.fail(`Failed: ${errorMsg}`);
                  } finally {
                    setCommandRunning(false);
                  }
                }}
              />
            )}

            {/* Feedback Dialog - conditionally mounted as overlay */}
            {activeOverlay === "feedback" && (
              <FeedbackDialog
                onSubmit={handleFeedbackSubmit}
                onCancel={closeOverlay}
                initialValue={feedbackPrefill}
              />
            )}

            {/* Memory Viewer - conditionally mounted as overlay */}
            {/* Use tree view for memfs-enabled agents, tab view otherwise */}
            {activeOverlay === "memory" &&
              (settingsManager.isMemfsEnabled(agentId) ? (
                <MemfsTreeViewer
                  agentId={agentId}
                  onClose={closeOverlay}
                  conversationId={conversationId}
                />
              ) : (
                <MemoryTabViewer
                  blocks={agentState?.memory?.blocks || []}
                  agentId={agentId}
                  onClose={closeOverlay}
                  conversationId={conversationId}
                />
              ))}

            {/* Memory sync conflict overlay removed - git-backed memory
                uses standard git merge conflicts resolved by the agent */}

            {/* MCP Server Selector - conditionally mounted as overlay */}
            {activeOverlay === "mcp" && (
              <McpSelector
                agentId={agentId}
                onAdd={() => {
                  // Switch to the MCP connect flow
                  setActiveOverlay("mcp-connect");
                }}
                onCancel={closeOverlay}
              />
            )}

            {/* MCP Connect Flow - interactive TUI for OAuth connection */}
            {activeOverlay === "mcp-connect" && (
              <McpConnectFlow
                onComplete={(serverName, serverId, toolCount) => {
                  const overlayCommand = consumeOverlayCommand("mcp-connect");
                  closeOverlay();
                  const cmd =
                    overlayCommand ??
                    commandRunner.start(
                      "/mcp connect",
                      "Connecting MCP server...",
                    );
                  cmd.finish(
                    `Successfully created MCP server "${serverName}"\n` +
                      `ID: ${serverId}\n` +
                      `Discovered ${toolCount} tool${toolCount === 1 ? "" : "s"}\n` +
                      "Open /mcp to attach or detach tools for this server.",
                    true,
                  );
                }}
                onCancel={closeOverlay}
              />
            )}

            {/* Help Dialog - conditionally mounted as overlay */}
            {activeOverlay === "help" && <HelpDialog onClose={closeOverlay} />}

            {/* Hooks Manager - for managing hooks configuration */}
            {activeOverlay === "hooks" && (
              <HooksManager onClose={closeOverlay} agentId={agentId} />
            )}

            {/* New Agent Dialog - for naming new agent before creation */}
            {activeOverlay === "new" && (
              <NewAgentDialog
                onSubmit={handleCreateNewAgent}
                onCancel={closeOverlay}
              />
            )}

            {/* Pin Dialog - for naming agent before pinning */}
            {activeOverlay === "pin" && (
              <PinDialog
                currentName={agentName || ""}
                local={pinDialogLocal}
                onSubmit={async (newName) => {
                  const overlayCommand = consumeOverlayCommand("pin");
                  closeOverlay();
                  setCommandRunning(true);

                  const cmd =
                    overlayCommand ??
                    commandRunner.start("/pin", "Pinning agent...");
                  const scopeText = pinDialogLocal
                    ? "to this project"
                    : "globally";
                  const displayName =
                    newName || agentName || agentId.slice(0, 12);

                  cmd.update({
                    output: `Pinning "${displayName}" ${scopeText}...`,
                    phase: "running",
                  });

                  try {
                    const client = await getClient();

                    // Rename if new name provided
                    if (newName && newName !== agentName) {
                      await client.agents.update(agentId, { name: newName });
                      updateAgentName(newName);
                    }

                    // Pin the agent
                    if (pinDialogLocal) {
                      settingsManager.pinLocal(agentId);
                    } else {
                      settingsManager.pinGlobal(agentId);
                    }

                    cmd.finish(
                      `Pinned "${newName || agentName || agentId.slice(0, 12)}" ${scopeText}.`,
                      true,
                    );
                  } catch (error) {
                    cmd.fail(`Failed to pin: ${error}`);
                  } finally {
                    setCommandRunning(false);
                    refreshDerived();
                  }
                }}
                onCancel={closeOverlay}
              />
            )}

            {/* Plan Mode Dialog - NOW RENDERED INLINE with tool call (see liveItems above) */}
            {/* ExitPlanMode approval is handled by InlinePlanApproval component */}

            {/* AskUserQuestion now rendered inline via InlineQuestionApproval */}
            {/* EnterPlanMode now rendered inline in liveItems above */}
            {/* ApprovalDialog removed - all approvals now render inline via InlineGenericApproval fallback */}
          </>
        )}
      </Box>
    </Box>
  );
}