Struct App 

pub struct App {
    pub state: UiState,
    pub buf: Buffer,
    pub menu: MenuState,
    pub active_modal: Option<Box<dyn Modal>>,
    pub message_queue: VecDeque<QueuedMessage>,
    pub think: ThinkStripper,
    pub pending_tools: HashMap<String, (String, String, bool)>,
    pub exit_pending: Option<Instant>,
    pub fixissue_pending: Option<IssueRef>,
    pub fixissue_buffer: String,
    pub reasoning_buffer: String,
    pub setup_hint_shown: bool,
}

All the per-session UI state that flows through key/event handlers.

Before this aggregation, handlers took 7–9 &mut parameters each and the call sites filled a paragraph. Now the handlers take (&mut App, &mut LoopCtx, &mut dyn Renderer, …event) — the LoopCtx stays separate because the tokio select! in run_loop needs to borrow ctx.input_rx, ctx.runtime_event_rx, ctx.wake_rx independently, and bundling them into App would fight the borrow checker on every arm.

Fields

state: UiState

buf: Buffer

menu: MenuState

active_modal: Option<Box<dyn Modal>>

Exactly one overlay at a time — /model, /provider, /resume all push into the same slot. The Modal trait owns draw + key handling so adding a fourth overlay is Some(Box::new(X)), not a new field + new dispatch branch.

message_queue: VecDeque<QueuedMessage>

Messages the user submitted while a turn was already running. Drained one-at-a-time from the head whenever the current turn finishes. Matches CC’s “type-ahead” UX — queue the next prompt while the model is still thinking and it fires automatically.

think: ThinkStripper

Streaming-state <think>…</think> stripper. Kept on App (not a local in the streaming arm) because it carries state across agent events — a tag straddling two chunks would break if the stripper were re-constructed each event.

pending_tools: HashMap<String, (String, String, bool)>

call_id → (tool_name, detail, call_rendered). Populated on ToolCallStarted, read by ApprovalNeeded (which renders the ▸ Tool(detail) line eagerly so the user sees what they’re being asked to approve), and consumed on ToolCallResult. The call_rendered flag prevents rendering the tool-call line twice when ApprovalNeeded fired first.

exit_pending: Option<Instant>

Timestamp of the first Ctrl+C press on an empty idle buffer. Requires a second press within CTRL_C_EXIT_WINDOW to actually exit — protects against accidental single-tap exits.

fixissue_pending: Option<IssueRef>

Set by /fixissue <url> while the agent is resolving that issue. On TurnComplete the text buffered in fixissue_buffer is posted back as an issue comment + the fixed label is applied. Cleared on TurnComplete / TurnCancelled / Error so a subsequent normal message doesn’t accidentally trigger a post-back.

fixissue_buffer: String

Accumulates every visible AssistantText delta produced during a fixissue turn, verbatim. Sent as the AtomGit comment body on successful completion.

reasoning_buffer: String

Accumulates reasoning/thinking content for display in verbose mode. Flushed on newline or when buffer exceeds threshold.

setup_hint_shown: bool

Guards the one-shot /setup hint so it fires at most once per session. Flipped to true after the first render; subsequent redraws skip the check entirely.