mahbot 0.5.2

//! "Running Agents" dashboard page — a live view of every currently-running
//! agent and in-flight non-agent LLM work, grouped by WORKSPACE, then by
//! DIRECT PARENT INVOCATION (ticket / analyze round / research run / workspace
//! singleton / unattributable orchestrator calls).
//!
//! The view is read-only: at render time it reads the in-memory registries
//! (`AGENT_REGISTRY` and `NON_AGENT_CALLS`) plus the live transcript snapshots
//! (`TRANSCRIPT_REGISTRY`), so it reflects the live in-memory conversation
//! including the unpersisted tail. No database reads, no schema changes, no
//! history retained between ticks. Re-renders are driven by coalesced runtime
//! change events (agent/non-agent registries, live transcript content, voice
//! status) — the GUI refreshes as activity happens, not on a fixed cadence.
//!
//! Truthfulness rules:
//! - The card's first row is the role icon followed inline by the token/elapsed
//!   metrics (collapsed and expanded alike), with no fill spacers; the rest of
//!   the card content (thinking fallback, trace groups, tool blocks, activity
//!   line) starts on the second row at the card's left edge. Narration (the
//!   assistant's short reasoning text) is shown exactly once — as the current
//!   (latest) trace-group label, which shows a neutral "clanking" placeholder
//!   while its narration hasn't committed; a brand-new agent — or a current
//!   user turn with no tool rounds yet — shows a card-level "clanking"
//!   placeholder. The narration renders at 14px italic, never truncated — the
//!   full reasoning text stays a hover tooltip — while a silent round (empty
//!   narration) with a decoded Reasoning block promotes that reasoning into
//!   the label: capped, plain, no tooltip. The live activity phase label (a
//!   non-tool LLM call) renders as a separate line at the very bottom of the
//!   card in both the collapsed and expanded state.
//!   In-flight tool state is derived from the live transcript history: a tool
//!   call whose result has not been committed yet renders with a live
//!   in-flight marker. Only non-tool LLM phases remain invisible in the
//!   transcript and are still shown solely by the registry activity label.
//! - Trace groups (below the LIVE line) are a projection of the shared
//!   session ledger (the live transcript snapshot decoded via
//!   `session_view::build_ledger`): tool-call assistant rounds grouped by
//!   shared narration, bounded by real user turns. Content-only final answers
//!   are omitted.
//! - The trace groups shown on a card are scoped to the CURRENT user turn:
//!   groups from before the last real user message are never shown.
//! - Parallel tool execution is a history-derived over-approximation: a call
//!   shows as in-flight while its result is uncommitted (its tool-call round
//!   is already in the ledger) — finished-but-uncommitted calls in a parallel
//!   batch, unknown/pre-flight-cancelled calls before their round commits,
//!   and drain-suspended calls all appear in-flight. This is accepted for an
//!   observability view; tools that never produce a committed round simply
//!   never appear.
//! - The instrumentation is purely observational — it never affects
//!   shutdown/drain logic and gains no cancellation semantics.
//! - The live view shows tool arguments EXACTLY as the agent passed them —
//!   including any secrets. This is a deliberate, documented divergence from
//!   the durable logs (tool-call stats, failure feedback), which remain
//!   credential-scrubbed: the live view is transient and exists for full
//!   visibility of what the agent is doing.
//!
//! The page is DB-free by design: everything it shows is derived from the
//! live in-memory registries and the live transcript snapshots, plus the
//! dashboard's registered-workspace map (only for the "(external)" marker on
//! unregistered/ephemeral workspace sections). All header labels (ticket
//! titles, analyze/research questions) are captured observationally at spawn —
//! never read from the DB.

use std::collections::HashMap;
use std::collections::HashSet;

use crate::ChatRole;
use crate::agent::registry::{AgentHandle, NonAgentCallHandle, ParentKey};
use crate::gui::dialog;
use crate::gui::session_view::{
    MAX_TOOL_TOOLTIP_WIDTH, SessionEntry, ToolBlockView, ToolCallEntry, build_ledger,
    collapse_control_chars, promoted_reasoning, tool_block, truncate_at_boundary,
};
use crate::gui::theme;
use crate::gui::widgets;
use chrono::{DateTime, Utc};

use iced::widget::{
    Column, Row, Space, button, column, container, mouse_area, row, stack, text, tooltip,
};
use iced::{Alignment, Element, Length};

use iced_fonts::lucide;

use super::Message;

use crate::Workspace;

/// Maximum display length (Unicode chars) of an analyze/research group header
/// label (the question/task text). Truncated at a word/path-delimiter boundary
/// with "…" (see [`truncate_at_boundary`]).
const MAX_GROUP_LABEL_CHARS: usize = 104;

/// Maximum width (px) of a metric hover tooltip body. Much narrower than
/// [`MAX_TOOL_TOOLTIP_WIDTH`] (tuned for multi-arg tool calls) since a metric
/// tooltip holds one short line of explanatory text.
const MAX_METRIC_TOOLTIP_WIDTH: f32 = 360.0;

/// Messages emitted by the Running Agents page.
///
/// The page itself stays stateless — the pending research-cancel
/// confirmation lives on the [`Dashboard`](super::Dashboard) so the dialog
/// requires no per-run page state beyond the pending confirmation itself.
#[derive(Debug, Clone)]
pub(crate) enum RunningMessage {
    /// Cancel button pressed on a research-run group header. Carries the
    /// run's durable job id — NEVER rendered as text; the button action
    /// alone carries it.
    CancelRequest(String),
    /// The confirmation dialog's danger button — start the async cancel.
    CancelConfirmed,
    /// Keep/Dismiss (or Escape/backdrop) — close the dialog with no effect.
    CancelDismissed,
    /// The async cancel finished: the run stopped and the cleanup agent is
    /// removing its traces (cancel_research_run is infallible).
    CancelFinished,
    /// Toggle the expanded/collapsed state of a running-agent card. Keyed by
    /// (agent_id, generation) so a recycled agent_id never inherits a stale
    /// expansion.
    ToggleAgentExpanded { agent_id: String, generation: u64 },
}

/// Render the live running-agents page.
///
/// The page is observational — it reads the live in-memory registries and
/// renders directly into the dashboard's [`Message`] type; its only
/// self-emitted messages are the research-run manual-cancel flow
/// ([`RunningMessage`]) and the card expand/collapse toggle. `workspaces` is
/// the dashboard's registered-workspace map (name → info) — used only to mark
/// unregistered/ephemeral workspace sections with the "(external)" suffix.
/// Everything else comes from the live in-memory registries.
///
/// `expanded` is the dashboard's set of expanded agent cards, keyed by
/// (agent_id, generation). Stale keys are pruned by the dashboard's
/// `RuntimeChanged` handler against the freshly-listed agents; rendering here
/// only reads the set.
pub(crate) fn view(
    workspaces: &HashMap<String, Workspace>,
    pending_cancel: Option<&str>,
    expanded: &HashSet<(String, u64)>,
) -> Element<'static, Message> {
    let agents = crate::agent::registry::AGENT_REGISTRY.list();
    let calls = crate::agent::registry::NON_AGENT_CALLS.list();

    // Workspace-first sections: each section holds its groups, groups
    // sorted by kind (tickets → analyze rounds → research runs →
    // singletons → unattributed), sections alphabetically by name.
    let sections = build_sections(build_groups(&agents, &calls), workspaces);

    let body: Element<'_, RunningMessage> = if sections.is_empty() {
        // Shared empty-state pattern: large radar glyph (the page's nav
        // icon) + label, centered.
        widgets::empty_state_placeholder(
            lucide::radar::<iced::Theme, iced::Renderer>(),
            "Nothing is currently running.",
            theme::TEXT_MUTED,
        )
    } else {
        let mut content = Column::new().spacing(theme::SPACE_20);
        for section in &sections {
            content = content.push(render_section(section, expanded));
        }
        widgets::vscroll(content)
    };

    // Uniform page chrome with the rest of the dashboard: base Flexoki
    // fill + 24px padding, matching the other pages.
    let page: Element<'_, RunningMessage> = container(body)
        .width(Length::Fill)
        .height(Length::Fill)
        .padding(theme::PAGE_PADDING)
        .style(theme::base_container_style)
        .into();
    // Confirm-dialog overlay: the page content is stack child 0, the dialog
    // (or a type-stable placeholder) child 1 — the widget shapes never
    // change, so no page state is lost when the dialog opens/closes.
    let confirm_layer: Element<'_, RunningMessage> = if let Some(run_key) = pending_cancel {
        cancel_confirm_dialog(run_key)
    } else {
        stack([widgets::empty_stack_placeholder()]).into()
    };
    let stacked: Element<'_, RunningMessage> = stack([page, confirm_layer]).into();
    stacked.map(Message::RunningAgents)
}

// ── Display model ─────────────────────────────────────────────────────────

/// One running work item: either an agent card or a non-agent LLM call row.
/// The agent card is boxed — `AgentHandle` is large and the variant would
/// otherwise dominate the enum's size (clippy::large_enum_variant).
#[derive(Debug, Clone)]
enum DisplayItem {
    Agent(Box<AgentCard>),
    Call(CallRow),
}

#[derive(Debug, Clone)]
struct AgentCard {
    handle: AgentHandle,
}

#[derive(Debug, Clone)]
struct CallRow {
    handle: NonAgentCallHandle,
}

/// Ordered group kinds — the page's display order.
#[derive(Debug, Clone, PartialEq, Eq)]
enum GroupKind {
    Ticket,
    AnalyzeRound,
    Research,
    Singleton,
    Unattributed,
}

#[derive(Debug, Clone)]
struct DisplayGroup {
    kind: GroupKind,
    /// Group key (ticket id / round key / run key / workspace name).
    key: String,
    /// Workspace name for the group header.
    workspace: String,
    /// Human-readable header label for the group (ticket title / analyze
    /// question / research question), captured observationally from any
    /// member's `parent_label`. `None` for singletons/unattributed and for
    /// parent-keyed groups whose members carried no label (defensive fallback
    /// — the header then degrades to a generic label, never a raw key).
    label: Option<String>,
    items: Vec<DisplayItem>,
    /// True when this research group carries a run-lifetime orchestrator
    /// marker (rendered as a run-lifetime indicator, not a call row).
    run_lifetime: bool,
}

/// One workspace section on the page: the groups running in one workspace,
/// keyed by the raw workspace name and ordered by its resolved label.
#[derive(Debug, Clone)]
struct WorkspaceSection {
    /// Raw workspace name — the grouping key.
    workspace: String,
    /// Resolved display label: the bare name for registered workspaces, the
    /// "(external)"-suffixed form for unregistered/ephemeral ones, and
    /// "workspace" for an empty name.
    label: String,
    groups: Vec<DisplayGroup>,
}

impl DisplayGroup {
    /// (order, key) — groups sort by kind order first, then by key within
    /// the kind (stable, deterministic).
    fn sort_key(&self) -> (u8, String) {
        let order = match self.kind {
            GroupKind::Ticket => 0,
            GroupKind::AnalyzeRound => 1,
            GroupKind::Research => 2,
            GroupKind::Singleton => 3,
            GroupKind::Unattributed => 4,
        };
        (order, self.key.clone())
    }
}

/// Map a parent key to its display group kind — the DIRECT PARENT INVOCATION
/// the work item belongs to.
fn parent_group_kind(parent: &ParentKey) -> GroupKind {
    match parent {
        ParentKey::Ticket(_) => GroupKind::Ticket,
        ParentKey::AnalyzeRound(_) => GroupKind::AnalyzeRound,
        ParentKey::Research(_) => GroupKind::Research,
    }
}

/// The group key payload for a parent key (ticket id / round key / run key).
fn parent_group_key(parent: &ParentKey) -> &str {
    match parent {
        ParentKey::Ticket(id) => id,
        ParentKey::AnalyzeRound(key) | ParentKey::Research(key) => key,
    }
}

/// Group the running agents and non-agent calls by DIRECT PARENT INVOCATION.
///
/// - Tickets: agents with a [`ParentKey::Ticket`] plus calls with the same
///   ticket parent.
/// - Analyze rounds: agents + calls sharing an [`ParentKey::AnalyzeRound`] key.
/// - Research runs: agents + calls sharing a [`ParentKey::Research`] key.
/// - Workspace singletons: agents with no parent key (manager / maintainer /
///   discovery / direct chat).
/// - Unattributed: non-agent calls with no parent key (workspace-scoped).
///
/// Group identity is (kind, key) plus — for everything except research — the
/// workspace: ticket ids already embed the workspace name and analyze-round keys
/// are scoped to one workspace, so the extra component closes any
/// cross-workspace key collision (e.g. two concurrent sync analyze rounds in
/// different workspaces drawing the same short NanoID suffix). Research groups
/// are matched on (kind, key) alone because a run's coder round executes in an
/// EPHEMERAL per-run workspace whose name IS the run key (job_id) while its
/// analysts and orchestrator calls run in the real workspace — a workspace
/// component would split the group. The run key is a durable 10-char NanoID
/// (~60 bits entropy), so (kind, key) stays unique across workspaces; the
/// group's displayed workspace is resolved to the REAL workspace whenever a
/// non-ephemeral member is seen.
fn build_groups(agents: &[AgentHandle], calls: &[NonAgentCallHandle]) -> Vec<DisplayGroup> {
    let mut groups: Vec<DisplayGroup> = Vec::new();

    let find_group =
        |kind: GroupKind, key: &str, workspace: &str, groups: &mut Vec<DisplayGroup>| -> usize {
            let research = kind == GroupKind::Research;
            if let Some(idx) = groups.iter().position(|g| {
                g.kind == kind && g.key == key && (research || g.workspace == workspace)
            }) {
                if research && groups[idx].workspace == key && workspace != key {
                    // The group currently shows the ephemeral per-run
                    // workspace; adopt the joining member's real workspace.
                    groups[idx].workspace = workspace.to_string();
                }
                return idx;
            }
            groups.push(DisplayGroup {
                kind,
                key: key.to_string(),
                workspace: workspace.to_string(),
                label: None,
                items: Vec::new(),
                run_lifetime: false,
            });
            groups.len() - 1
        };

    // Adopt a member's parent label into the group (first non-None wins —
    // all members of one invocation carry the same label; the fallback keeps
    // a group whose only remaining member is label-less from degrading).
    let adopt_label = |label: &Option<String>, group: &mut DisplayGroup| {
        if group.label.is_none() && label.is_some() {
            group.label.clone_from(label);
        }
    };

    for agent in agents {
        let workspace = agent.workspace_name.clone();
        if let Some(parent) = &agent.parent_key {
            let idx = find_group(
                parent_group_kind(parent),
                parent_group_key(parent),
                &workspace,
                &mut groups,
            );
            adopt_label(&agent.parent_label, &mut groups[idx]);
            groups[idx]
                .items
                .push(DisplayItem::Agent(Box::new(AgentCard {
                    handle: agent.clone(),
                })));
        } else {
            // Workspace singleton. An empty workspace name groups under the
            // empty key; the section label resolver renders it as "workspace".
            let idx = find_group(GroupKind::Singleton, &workspace, &workspace, &mut groups);
            groups[idx]
                .items
                .push(DisplayItem::Agent(Box::new(AgentCard {
                    handle: agent.clone(),
                })));
        }
    }

    for call in calls {
        let workspace = call.workspace.clone();
        if let Some(parent) = &call.parent_key {
            let idx = find_group(
                parent_group_kind(parent),
                parent_group_key(parent),
                &workspace,
                &mut groups,
            );
            adopt_label(&call.parent_label, &mut groups[idx]);
            // Run-lifetime orchestrator markers render as a run-lifetime
            // indicator on the group, not as a transient call card.
            if call.run_lifetime {
                groups[idx].run_lifetime = true;
            } else {
                groups[idx].items.push(DisplayItem::Call(CallRow {
                    handle: call.clone(),
                }));
            }
        } else {
            let idx = find_group(GroupKind::Unattributed, &workspace, &workspace, &mut groups);
            groups[idx].items.push(DisplayItem::Call(CallRow {
                handle: call.clone(),
            }));
        }
    }

    groups
}

/// Group the display groups into WORKSPACE SECTIONS — the page is organized
/// by workspace, whose resolved label is used only for deterministic ordering
/// (the label no longer renders as a per-section heading).
///
/// Sections are ordered alphabetically by their resolved label (deterministic
/// — activity-based ordering would flicker as agents come and go); within a
/// section, groups keep the canonical kind order (tickets → analyze rounds →
/// research runs → singletons → unattributed) via [`DisplayGroup::sort_key`].
fn build_sections(
    groups: Vec<DisplayGroup>,
    workspaces: &std::collections::HashMap<String, Workspace>,
) -> Vec<WorkspaceSection> {
    let mut sections: Vec<WorkspaceSection> = Vec::new();
    for group in groups {
        let label = workspace_label_for(&group.workspace, workspaces);
        if let Some(section) = sections.iter_mut().find(|s| s.workspace == group.workspace) {
            section.groups.push(group);
        } else {
            sections.push(WorkspaceSection {
                workspace: group.workspace.clone(),
                label,
                groups: vec![group],
            });
        }
    }
    sections.sort_by(|a, b| a.label.cmp(&b.label));
    for section in &mut sections {
        section.groups.sort_by_key(DisplayGroup::sort_key);
    }
    sections
}

// ── Rendering ─────────────────────────────────────────────────────────────

/// Render one workspace section: just the section's groups. The workspace
/// name no longer gets a visual heading — `section.label` exists only for the
/// deterministic section ordering in [`build_sections`].
///
/// The returned element owns all rendered content (text widgets take owned
/// Strings), so its lifetime is independent of the `section` borrow.
fn render_section(
    section: &WorkspaceSection,
    expanded: &HashSet<(String, u64)>,
) -> Element<'static, RunningMessage> {
    let mut groups = Column::new().spacing(theme::SPACE_10);
    for group in &section.groups {
        groups = groups.push(render_group(group, expanded));
    }
    groups.into()
}

/// Resolve a group's header parts: the primary label plus an optional ticket
/// id. `render_group` composes a single flat header — for ticket groups
/// `[{id}] {title}`, otherwise just the label.
///
/// - Ticket groups: the ticket NAME plus the ticket ID. When no title was
///   captured (defensive — every live ticket group carries one via its agents
///   or the synthesis call), the key becomes the sole label.
/// - Analyze/research groups: the truncated question/task text; a generic
///   fallback when no label was captured. The raw NanoID key is NEVER shown.
/// - Singleton/unattributed groups: generic labels — their workspace name is
///   already the section grouping, so repeating it would be redundant.
fn group_title(group: &DisplayGroup) -> (String, Option<String>) {
    match &group.kind {
        GroupKind::Ticket => match &group.label {
            Some(title) => (title.clone(), Some(group.key.clone())),
            None => (group.key.clone(), None),
        },
        GroupKind::AnalyzeRound => (
            group.label.as_deref().map_or_else(
                || "Analyze round".to_string(),
                |l| truncate_at_boundary(l, MAX_GROUP_LABEL_CHARS),
            ),
            None,
        ),
        GroupKind::Research => (
            group.label.as_deref().map_or_else(
                || "Research run".to_string(),
                |l| truncate_at_boundary(l, MAX_GROUP_LABEL_CHARS),
            ),
            None,
        ),
        GroupKind::Singleton => ("Standalone".to_string(), None),
        GroupKind::Unattributed => ("Other LLM work".to_string(), None),
    }
}

/// Render one group: header (flat label + run-lifetime marker) then its
/// cards/rows directly on the page background — each agent card carries its
/// own card style; call rows are unadorned by design.
fn render_group(
    group: &DisplayGroup,
    expanded: &HashSet<(String, u64)>,
) -> Element<'static, RunningMessage> {
    let (title, id) = group_title(group);
    let mut header_parts: Vec<Element<'_, RunningMessage>> = Vec::new();
    // Ticket groups render a single flat `[{id}] {title}` header — the id is
    // composed into the text, never styled as a separate ACCENT element. When
    // no title was captured (defensive), the key is the sole label. All other
    // groups render their label as the same flat line.
    let header = if group.kind == GroupKind::Ticket {
        match id {
            Some(id) => format!("[{id}] {title}"),
            None => title.clone(),
        }
    } else {
        title.clone()
    };
    header_parts.push(
        text(header)
            .size(theme::TEXT_12)
            .color(theme::TEXT_SECONDARY)
            .into(),
    );
    if group.run_lifetime {
        header_parts.push(
            text("run active")
                .size(theme::TEXT_11)
                .color(theme::TEXT_SECONDARY)
                .into(),
        );
    }
    // Manual cancel: danger-styled button on RESEARCH-run group headers only.
    // The run key is carried by the message, never rendered as text.
    if group.kind == GroupKind::Research {
        header_parts.push(Space::new().width(Length::Fill).into());
        header_parts.push(
            button(text("Cancel run").size(theme::TEXT_11))
                .style(theme::button_danger)
                .on_press(RunningMessage::CancelRequest(group.key.clone()))
                .into(),
        );
    }

    let mut items = Column::new().spacing(theme::SPACE_6);
    for item in &group.items {
        match item {
            DisplayItem::Agent(card) => {
                let is_expanded =
                    expanded.contains(&(card.handle.agent_id.clone(), card.handle.generation));
                items = items.push(render_agent_card(card, is_expanded));
            }
            DisplayItem::Call(call) => {
                items = items.push(render_call_row(call));
            }
        }
    }

    column![
        Row::with_children(header_parts)
            .spacing(theme::SPACE_8)
            .align_y(Alignment::Center),
        items,
    ]
    .spacing(theme::SPACE_6)
    .into()
}

/// Build the research-run manual-cancel confirmation dialog, mirroring the
/// board's ticket-cancel pattern: consequences listed, danger confirm
/// button, Keep/Dismiss (and Escape/backdrop) to close with no effect.
/// `run_key` is the run's durable job id — never displayed as text.
fn cancel_confirm_dialog(run_key: &str) -> Element<'static, RunningMessage> {
    // Deliberately unused: the run key must never appear in the dialog's
    // rendered text — the button action alone carries it (see the enum docs).
    let _ = run_key;
    widgets::modal_backdrop(
        dialog::confirm_dialog(
            dialog::dialog_title("Cancel this research run?"),
            dialog::dialog_body([
                "Confirming will:",
                "• stop all agents of this run — an in-flight tool or LLM call \
                 may finish, but no further work happens;\n\
                 • stop the orchestrator — no more rounds, no report, and \
                 nothing is delivered to the Manager;\n\
                 • remove the run's temporary traces and its run folder — \
                 dispatching the cleanup agent while the orchestrator is still \
                 running, a direct sweep otherwise;\n\
                 • remove the run permanently — it can never resume.",
            ]),
            [
                dialog::DialogAction::secondary("Keep run", RunningMessage::CancelDismissed),
                dialog::DialogAction::danger("Cancel run", RunningMessage::CancelConfirmed),
            ],
        ),
        RunningMessage::CancelDismissed,
        0.5,
    )
}

/// Fallback label for workspaces outside the dashboard's registered set
/// (personal user spaces, ephemeral run-scoped workspaces).
fn fallback_workspace_label(workspace: &str) -> String {
    if workspace.is_empty() {
        "workspace".to_string()
    } else {
        format!("{workspace} (external)")
    }
}

/// Resolve a workspace's display label, used for deterministic section
/// ordering: the bare name for registered workspaces, the fallback label
/// (with the "(external)" marker) for unregistered/ephemeral ones, and
/// "workspace" for an empty name.
fn workspace_label_for(
    name: &str,
    workspaces: &std::collections::HashMap<String, Workspace>,
) -> String {
    if name.is_empty() {
        "workspace".to_string()
    } else if workspaces.contains_key(name) {
        name.to_string()
    } else {
        fallback_workspace_label(name)
    }
}

/// Render one running-agent card: the role icon and the token/elapsed metrics
/// inline in the first row, then the current user turn's trace groups. The
/// whole card is clickable to expand/collapse the trace groups.
///
/// Narration (the assistant's short reasoning text) lives solely in the
/// current trace-group label, which shows a "clanking" placeholder while its
/// narration hasn't committed; a brand-new agent — or a current user turn
/// with no tool rounds yet — shows a card-level "clanking" placeholder. The
/// live activity phase label (an
/// in-flight non-tool LLM call, e.g. extracting/summarizing/transcribing)
/// renders as a separate accent line at the very bottom of the card in both
/// the collapsed and expanded state. Trace groups come from the live
/// transcript snapshot (the unpersisted tail included): the current (latest)
/// group, plus up to 5 previous groups when expanded.
fn render_agent_card(card: &AgentCard, expanded: bool) -> Element<'static, RunningMessage> {
    let h = &card.handle;
    // Color resolution via the canonical string helper (handles derivative
    // names and falls back to muted grey for unknown roles); the icon still
    // needs a typed Role, so parse with the same Engineer fallback as before.
    let (fg, _bg) = theme::role_badge_color(&h.role);
    let role: crate::Role = h.role.parse().unwrap_or(crate::Role::Engineer);
    let icon = theme::role_icon(&role).size(theme::TEXT_24).color(fg);
    let elapsed = format_elapsed(h.started_at);

    // The live transcript snapshot feeds the trace groups below (and the
    // current group's narration label), scoped to the current user turn, so
    // fetch it once up front.
    let snapshot = crate::session::TRANSCRIPT_REGISTRY.snapshot(&h.agent_id);
    let groups = snapshot
        .as_ref()
        .map(|s| derive_trace_groups(current_turn_entries(&build_ledger(&s.history))))
        .unwrap_or_default();

    let token_count = snapshot.as_ref().and_then(|s| s.token_count);

    let mut content = Column::new()
        .spacing(theme::SPACE_6)
        .align_x(Alignment::Start)
        .width(Length::Fill);
    // The card's first row is the role icon followed inline by the token /
    // elapsed metrics, sitting at the left edge of the card (no fill spacers).
    let first_row = Row::new()
        .spacing(theme::SPACE_12)
        .align_y(Alignment::Center)
        .push(icon)
        .push(render_metrics(token_count, &elapsed));
    content = content.push(first_row);
    // A brand-new agent — or the current user turn with no tool rounds yet —
    // shows a card-level "clanking" placeholder. Once the current turn has any
    // group, the current group's own label handles narration / thinking, so
    // this fallback must not render at card level.
    if groups.is_empty() {
        content = content.push(clanking_placeholder());
    }
    // Trace groups from the live snapshot, scoped to the current user turn.
    // Content-only assistant messages (final answers) are deliberately
    // omitted — only tool-call assistant messages form groups.
    for group in render_visible_groups(&groups, expanded) {
        content = content.push(group);
    }
    // The live activity phase label (an in-flight non-tool LLM call) renders
    // as a separate accent line at the very bottom of the card, below the
    // trace groups, in both the collapsed and expanded view.
    if let Some(activity) = &h.activity {
        content = content.push(
            text(activity.to_owned())
                .size(theme::TEXT_13)
                .color(theme::ACCENT),
        );
    }

    let on_press = RunningMessage::ToggleAgentExpanded {
        agent_id: h.agent_id.clone(),
        generation: h.generation,
    };
    // The whole card is a click target (expand/collapse), but it stays a
    // `container` under a transparent `mouse_area` so the inner tool blocks
    // keep their hover tooltips (a `button` wrapper would swallow them).
    mouse_area(
        container(content)
            .width(Length::Fill)
            .padding(theme::PAD_8)
            .style(theme::surface_card_style),
    )
    .on_press(on_press)
    .into()
}

/// The narration typography of the trace-group labels: italic at
/// [`theme::NARRATION_TEXT_SIZE`] with wrapping within the row. A narration
/// label is never truncated and keeps the full reasoning as a hover tooltip;
/// a promoted reasoning label is pre-capped by
/// [`promoted_reasoning_label`] and carries no tooltip.
fn narration_text(narration: &str) -> iced::widget::Text<'static, iced::Theme, iced::Renderer> {
    text(narration.to_owned())
        .size(theme::NARRATION_TEXT_SIZE)
        .font(theme::FONT_ITALIC)
        .color(theme::TEXT_SECONDARY)
        .wrapping(iced::widget::text::Wrapping::WordOrGlyph)
}

/// Max display length (Unicode chars) of a promoted reasoning label. The
/// narration itself never truncates, but a promoted long CoT fallback is
/// capped so it cannot dominate the card.
const MAX_PROMOTED_LABEL_CHARS: usize = 200;

/// Single-line display form of a promoted reasoning label: control
/// characters collapsed to spaces ([`collapse_control_chars`]), then capped
/// at [`MAX_PROMOTED_LABEL_CHARS`] chars at a word/path-delimiter boundary
/// with "…" ([`truncate_at_boundary`]).
fn promoted_reasoning_label(reasoning: &str) -> String {
    truncate_at_boundary(
        &collapse_control_chars(reasoning, MAX_PROMOTED_LABEL_CHARS),
        MAX_PROMOTED_LABEL_CHARS,
    )
}

/// The neutral "clanking" placeholder shown while a narration has not
/// committed yet (a brand-new agent with no trace groups, or a current
/// group awaiting its first narration).
fn clanking_placeholder() -> iced::widget::Text<'static, iced::Theme, iced::Renderer> {
    text("clanking")
        .size(theme::TEXT_13)
        .color(theme::TEXT_SECONDARY)
}

/// Render the card's metrics: the live session-token count (when known) and
/// the elapsed run time, each as a compact labeled tooltip. Renders inline in
/// the card's first row, right after the role icon (see
/// [`render_agent_card`]).
fn render_metrics(token_count: Option<u64>, elapsed: &str) -> Element<'static, RunningMessage> {
    let mut metrics = Row::new()
        .spacing(theme::SPACE_12)
        .align_y(Alignment::Center);
    // The live token count comes from the published transcript snapshot (the
    // registry no longer carries a session_tokens mirror). It renders as a
    // compact count; hovering reveals the label and the exact unrounded value.
    if let Some(token_count) = token_count {
        metrics = metrics.push(
            tooltip(
                text(theme::format_compact_tokens(token_count))
                    .size(theme::TEXT_11)
                    .color(theme::TEXT_SECONDARY),
                render_metric_tooltip(format!("Session tokens: {token_count}")),
                tooltip::Position::Top,
            )
            .gap(theme::SPACE_4)
            .style(theme::tooltip_style),
        );
    }
    // Elapsed time since the agent started; hovering explains the meaning.
    metrics = metrics.push(
        tooltip(
            text(elapsed.to_owned())
                .size(theme::TEXT_11)
                .color(theme::TEXT_SECONDARY),
            render_metric_tooltip("Elapsed run time since the agent started".to_string()),
            tooltip::Position::Top,
        )
        .gap(theme::SPACE_4)
        .style(theme::tooltip_style),
    );
    metrics.into()
}

/// Render the visible trace groups for a card: the current (latest) group
/// always (last, chronologically, newest round expanded), plus up to 5
/// previous groups (collapsed) when the card is expanded. Every group renders
/// its narration label (see [`render_trace_group`]). Empty input renders
/// nothing.
fn render_visible_groups(
    groups: &[TraceGroup],
    expanded: bool,
) -> Vec<Element<'static, RunningMessage>> {
    let mut rendered = Vec::new();
    let Some((current, previous)) = groups.split_last() else {
        return rendered;
    };
    if expanded {
        // The 5 most-recent previous groups, oldest first, then current.
        for group in previous.iter().rev().take(5).rev() {
            rendered.push(render_trace_group(group, false));
        }
    }
    // Current group: newest round expanded, earlier rounds collapsed.
    rendered.push(render_trace_group(current, true));
    rendered
}

/// The text a trace group renders as its narration label: the group's
/// narration (wrapped in the reasoning hover tooltip), or — when the
/// narration is empty (whitespace-only counts as empty) — the group's
/// promoted reasoning as a capped label (no tooltip: the label already
/// shows the text), or the neutral "clanking" placeholder for a silent
/// group (no narration and no reasoning).
enum GroupLabelText<'a> {
    Narration(&'a str),
    Promoted(String),
    Placeholder,
}

fn group_label_text(group: &TraceGroup) -> GroupLabelText<'_> {
    if let Some(reasoning) = promoted_reasoning(Some(&group.narration), group.reasoning.as_deref())
    {
        GroupLabelText::Promoted(promoted_reasoning_label(reasoning))
    } else if !group.narration.trim().is_empty() {
        GroupLabelText::Narration(&group.narration)
    } else {
        GroupLabelText::Placeholder
    }
}

/// Render one trace group. `expand_current` renders the newest round's calls
/// expanded (current group); otherwise every call collapses to a name-only
/// line (previous groups). The group's narration label always renders (the
/// short reasoning text, its promoted capped reasoning, or the "clanking"
/// placeholder); hovering the narration label — not the promoted one, which
/// already shows the text — reveals the group's full decoded Reasoning when
/// it carried one.
fn render_trace_group(
    group: &TraceGroup,
    expand_current: bool,
) -> Element<'static, RunningMessage> {
    let mut column = Column::new()
        .spacing(theme::SPACE_4)
        .align_x(Alignment::Start);
    let label: Element<'static, RunningMessage> = match group_label_text(group) {
        GroupLabelText::Narration(narration) => {
            reasoning_tooltip(group.reasoning.as_deref(), narration_text(narration).into())
        }
        GroupLabelText::Promoted(promoted) => narration_text(&promoted).into(),
        GroupLabelText::Placeholder => clanking_placeholder().into(),
    };
    column = column.push(label);

    if expand_current {
        let (newest, earlier) = group
            .rounds
            .split_last()
            .expect("a trace group always carries at least one round");
        if !earlier.is_empty() {
            column = column.push(
                text(collapsed_calls_line(earlier))
                    .size(theme::TEXT_10)
                    .color(theme::TEXT_SECONDARY),
            );
        }
        for entry in newest {
            column = column.push(tool_row(entry));
        }
    } else {
        column = column.push(
            text(collapsed_calls_line(&group.rounds))
                .size(theme::TEXT_10)
                .color(theme::TEXT_SECONDARY),
        );
    }
    column.into()
}

/// A single expanded tool-call row in the current group. Settled calls render
/// exactly as [`tool_block`]; an in-flight call (result not yet committed)
/// gains a small accent `loader_circle` marker after the block.
fn tool_row(entry: &ToolCallEntry) -> Element<'static, RunningMessage> {
    let block = tool_block(&entry.tool, ToolBlockView::Compact);
    if entry.result.is_none() {
        Row::new()
            .spacing(theme::SPACE_6)
            .align_y(Alignment::Center)
            .push(block)
            .push(
                lucide::loader_circle::<iced::Theme, iced::Renderer>()
                    .size(theme::TEXT_14)
                    .color(theme::ACCENT),
            )
            .into()
    } else {
        block
    }
}

/// Wrap a narration label in a hover tooltip revealing the full decoded
/// Reasoning behind it (see [`TraceGroup::reasoning`]). When there is no
/// reasoning the label renders bare.
fn reasoning_tooltip<'a>(
    reasoning: Option<&str>,
    label: Element<'a, RunningMessage>,
) -> Element<'a, RunningMessage> {
    let Some(reasoning) = reasoning.filter(|r| !r.is_empty()) else {
        return label;
    };
    let tooltip_content: Element<'static, RunningMessage> = container(
        text(reasoning.to_string())
            .size(theme::TEXT_11)
            .color(theme::TEXT_SECONDARY)
            .wrapping(iced::widget::text::Wrapping::WordOrGlyph),
    )
    .max_width(MAX_TOOL_TOOLTIP_WIDTH)
    .into();
    tooltip(label, tooltip_content, tooltip::Position::Top)
        .gap(theme::SPACE_4)
        .style(theme::tooltip_style)
        .into()
}

/// Compose the collapsed name-only line for a set of tool-call rounds: tool
/// names with underscores replaced by spaces, in first-appearance order, each
/// unique name suffixed with `xN` when it appears more than once within the
/// collapsed set (e.g. `read file x2, list files`). An in-flight call
/// (result not yet committed) gets its display name suffixed with a running
/// marker (`…`) before the xN counting, so it never looks settled.
fn collapsed_calls_line(rounds: &[Vec<ToolCallEntry>]) -> String {
    let mut order: Vec<String> = Vec::new();
    let mut counts: HashMap<String, usize> = HashMap::new();
    for round in rounds {
        for call in round {
            let base = call.tool.name.replace('_', " ");
            let display = if call.result.is_none() {
                format!("{base} …")
            } else {
                base
            };
            if !counts.contains_key(&display) {
                order.push(display.clone());
            }
            *counts.entry(display).or_default() += 1;
        }
    }
    order
        .iter()
        .map(|name| {
            let c = counts[name];
            if c > 1 {
                format!("{name} x{c}")
            } else {
                name.clone()
            }
        })
        .collect::<Vec<_>>()
        .join(", ")
}

/// Prune stale expanded-state keys: an agent that finished (or was replaced by
/// a new generation of the same agent_id) must not keep a stale expansion.
pub(crate) fn prune_expanded(expanded: &mut HashSet<(String, u64)>, agents: &[AgentHandle]) {
    let live: HashSet<(String, u64)> = agents
        .iter()
        .map(|h| (h.agent_id.clone(), h.generation))
        .collect();
    expanded.retain(|key| live.contains(key));
}

/// One trace group in a running agent's live transcript: a maximal run of
/// consecutive tool-call assistant rounds sharing a single (possibly empty)
/// visible narration, bounded by a real user turn or a new narration.
///
/// Derived from the shared session ledger: the group's narration comes from
/// the decoded assistant `content` (the short visible reasoning text); its
/// `reasoning` is the long decoded `Reasoning` block of the group's rounds,
/// surfaced on hover or promoted into the label. A group always carries at
/// least one round.
struct TraceGroup {
    /// The assistant's short visible narration (empty → renders "clanking",
    /// or the promoted capped reasoning label when the group carries one).
    /// This is the decoded assistant `content`, NEVER the long `Reasoning`
    /// / `[thinking]` block.
    narration: String,
    /// The full decoded Reasoning of the group: the first non-empty reasoning
    /// among its rounds. Surfaced on hover behind a narration label, or
    /// promoted into the label itself when the narration is empty. `None` when
    /// the rounds carried none (a silent group then renders "clanking").
    reasoning: Option<String>,
    /// Tool-call rounds in first-appearance order; the LAST round is the newest
    /// (expanded in the current group), earlier rounds collapse. Each round
    /// carries the full ledger tool-call entries (name/args pairs plus the
    /// tool_call_id-matched result), so the display can distinguish an in-flight
    /// call (result not yet committed) from a settled one.
    rounds: Vec<Vec<ToolCallEntry>>,
}

/// A real user turn boundary in the ledger: a User message that is neither
/// an injected tool-result-image message (content starting with
/// [`crate::util::INJECTED_IMAGE_TAG`]) nor an empty resume signal (no/empty
/// content). Only such a message starts a new user turn.
fn is_user_turn_boundary(entry: &SessionEntry) -> bool {
    matches!(entry, SessionEntry::Message { role: ChatRole::User, content, .. }
    if content.as_deref().is_some_and(|c| {
        !c.trim().is_empty() && !c.starts_with(crate::util::INJECTED_IMAGE_TAG)
    }))
}

/// The slice of the ledger belonging to the CURRENT user turn: everything
/// after the last real user turn boundary ([`is_user_turn_boundary`]). The
/// running card shows only the current turn, so groups from previous turns
/// never leak onto it. When no boundary exists the whole ledger is current.
fn current_turn_entries(entries: &[SessionEntry]) -> &[SessionEntry] {
    let start = entries
        .iter()
        .rposition(is_user_turn_boundary)
        .map_or(0, |i| i + 1);
    &entries[start..]
}

/// Project the shared session ledger into a running agent's trace groups, in
/// chronological order.
///
/// Boundary rules:
/// - A new (non-empty) narration in a tool-call assistant round starts a new
///   group; that narration becomes the group's `narration`.
/// - Consecutive tool-call rounds with no narration accumulate into the current
///   group.
/// - A real user turn closes the current group; the next tool-call round then
///   starts a fresh one even with no narration (its `narration` slots to
///   "clanking"). An empty (whitespace) User message or a synthetic
///   tool-injected one (content starting with the
///   [`crate::util::INJECTED_IMAGE_TAG`] marker) is NOT a boundary.
fn derive_trace_groups(entries: &[SessionEntry]) -> Vec<TraceGroup> {
    let mut groups: Vec<TraceGroup> = Vec::new();
    let mut current: Option<TraceGroup> = None;
    for entry in entries {
        match entry {
            SessionEntry::ToolRound {
                narration,
                reasoning,
                calls,
            } => {
                let narration = narration.clone().unwrap_or_default();
                let new_group = !narration.is_empty() || current.is_none();
                if new_group {
                    if let Some(group) = current.take() {
                        groups.push(group);
                    }
                }
                let group = current.get_or_insert_with(|| TraceGroup {
                    narration,
                    reasoning: None,
                    rounds: Vec::new(),
                });
                if group.reasoning.is_none()
                    && let Some(reasoning) = reasoning
                    && !reasoning.is_empty()
                {
                    group.reasoning = Some(reasoning.clone());
                }
                group.rounds.push(calls.clone());
            }
            SessionEntry::Message { .. } => {
                if is_user_turn_boundary(entry)
                    && let Some(group) = current.take()
                {
                    groups.push(group);
                }
            }
        }
    }
    if let Some(group) = current.take() {
        groups.push(group);
    }
    groups
}

/// Render a one-line metric tooltip body in a narrow, styled container: the
/// explanatory text (metric label plus, where useful, the precise value). Kept
/// narrower than [`MAX_TOOL_TOOLTIP_WIDTH`] because a metric tooltip holds a
/// single short line, not multi-arg tool calls.
fn render_metric_tooltip(label: String) -> Element<'static, RunningMessage> {
    container(
        text(label)
            .size(theme::TEXT_11)
            .color(theme::TEXT_PRIMARY)
            .wrapping(iced::widget::text::Wrapping::WordOrGlyph),
    )
    .max_width(MAX_METRIC_TOOLTIP_WIDTH)
    .into()
}

/// Render a compact non-agent LLM call row: zap marker + purpose + elapsed.
/// No workspace name (the workspace grouping already names it); the elapsed
/// time uses the same brighter tone as agent cards. The purpose is the static
/// human-readable label — raw kind names never leak.
fn render_call_row(call: &CallRow) -> Element<'static, RunningMessage> {
    let h = &call.handle;
    let elapsed = format_elapsed(h.started_at);
    let purpose = crate::agent::registry::call_kind_label(h.kind);

    row![
        lucide::zap::<iced::Theme, iced::Renderer>()
            .size(theme::TEXT_16)
            .color(theme::ACCENT),
        text(purpose)
            .size(theme::TEXT_12)
            .color(theme::TEXT_SECONDARY),
        Space::new().width(Length::Fill),
        text(elapsed)
            .size(theme::TEXT_11)
            .color(theme::TEXT_SECONDARY),
    ]
    .spacing(theme::SPACE_6)
    .align_y(Alignment::Center)
    .into()
}

/// Format the elapsed time since `started_at` at render time.
fn format_elapsed(started_at: DateTime<Utc>) -> String {
    let now = Utc::now();
    let secs = now.signed_duration_since(started_at).num_seconds().max(0);
    if secs < 60 {
        format!("{secs}s")
    } else if secs < 3600 {
        format!("{}m {:02}s", secs / 60, secs % 60)
    } else {
        format!("{}h {:02}m", secs / 3600, (secs % 3600) / 60)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::ChatMessage;
    use crate::agent::registry::{AgentHandle, RunningTool};
    use crate::gui::session_view::build_ledger;

    fn agent_handle(
        id: &str,
        role: &str,
        ticket_id: Option<String>,
        workspace: &str,
        parent: Option<ParentKey>,
    ) -> AgentHandle {
        AgentHandle {
            agent_id: id.to_string(),
            role: role.to_string(),
            ticket_id,
            workspace_path: format!("/ws/{workspace}"),
            workspace_name: workspace.to_string(),
            parent_key: parent,
            parent_label: None,
            started_at: Utc::now(),
            label: role.to_string(),
            generation: 0,
            activity: None,
        }
    }

    fn call_handle(
        kind: &'static str,
        workspace: &str,
        parent: Option<ParentKey>,
        run_lifetime: bool,
    ) -> NonAgentCallHandle {
        NonAgentCallHandle {
            kind,
            workspace: workspace.to_string(),
            started_at: Utc::now(),
            parent_key: parent,
            parent_label: None,
            run_lifetime,
        }
    }

    #[test]
    fn groups_ticket_agents_and_synthesis_together() {
        let agents = vec![
            agent_handle(
                "a1",
                "analyst",
                Some("T1".to_string()),
                "ws1",
                Some(ParentKey::Ticket("T1".to_string())),
            ),
            agent_handle("mgr", "manager", None, "ws1", None),
        ];
        let calls = vec![call_handle(
            "synthesis",
            "ws1",
            Some(ParentKey::Ticket("T1".to_string())),
            false,
        )];
        let groups = build_groups(&agents, &calls);
        assert_eq!(groups.len(), 2, "ticket group + singleton group");
        let ticket = &groups[0];
        assert_eq!(ticket.kind, GroupKind::Ticket);
        assert_eq!(ticket.key, "T1");
        assert_eq!(ticket.items.len(), 2, "analyst agent + synthesis call");
    }

    #[test]
    fn two_analyze_rounds_never_mix_members() {
        let agents = vec![
            agent_handle(
                "analyze_ws_AAA_0_analyst",
                "analyst",
                None,
                "ws1",
                Some(ParentKey::AnalyzeRound("roundA".to_string())),
            ),
            agent_handle(
                "analyze_ws_AAA_1_analyst",
                "analyst",
                None,
                "ws1",
                Some(ParentKey::AnalyzeRound("roundA".to_string())),
            ),
            agent_handle(
                "analyze_ws_BBB_0_analyst",
                "analyst",
                None,
                "ws1",
                Some(ParentKey::AnalyzeRound("roundB".to_string())),
            ),
        ];
        let calls = vec![call_handle(
            "consolidate",
            "ws1",
            Some(ParentKey::AnalyzeRound("roundA".to_string())),
            false,
        )];
        let groups = build_groups(&agents, &calls);
        let analyze_groups: Vec<_> = groups
            .iter()
            .filter(|g| g.kind == GroupKind::AnalyzeRound)
            .collect();
        assert_eq!(analyze_groups.len(), 2, "two distinct analyze round groups");
        let round_a = analyze_groups
            .iter()
            .find(|g| g.key == "roundA")
            .expect("round A exists");
        assert_eq!(round_a.items.len(), 3, "2 analysts + consolidation call");
        let round_b = analyze_groups
            .iter()
            .find(|g| g.key == "roundB")
            .expect("round B exists");
        assert_eq!(round_b.items.len(), 1, "only its own analyst");
    }

    #[test]
    fn research_run_members_share_one_key_across_phases() {
        // Two concurrent research runs in one workspace must never mix
        // members even though every spawn carries a unique agent-id suffix.
        let agents = vec![
            agent_handle(
                "research_ws_x1_decompose_0",
                "analyst",
                None,
                "ws1",
                Some(ParentKey::Research("run1".to_string())),
            ),
            agent_handle(
                "research_ws_x2_r1_0",
                "analyst",
                None,
                "ws1",
                Some(ParentKey::Research("run1".to_string())),
            ),
            agent_handle(
                "research_ws_y1_decompose_0",
                "analyst",
                None,
                "ws1",
                Some(ParentKey::Research("run2".to_string())),
            ),
        ];
        let calls = vec![
            call_handle(
                "synthesize",
                "ws1",
                Some(ParentKey::Research("run1".to_string())),
                false,
            ),
            call_handle(
                "research_orchestrator",
                "ws1",
                Some(ParentKey::Research("run1".to_string())),
                true,
            ),
        ];
        let groups = build_groups(&agents, &calls);
        let research_groups: Vec<_> = groups
            .iter()
            .filter(|g| g.kind == GroupKind::Research)
            .collect();
        assert_eq!(research_groups.len(), 2);
        let run1 = research_groups
            .iter()
            .find(|g| g.key == "run1")
            .expect("run 1 exists");
        assert_eq!(run1.items.len(), 3, "2 analysts + synthesize call");
        assert!(
            run1.run_lifetime,
            "run-lifetime orchestrator marker attached"
        );
        let run2 = research_groups
            .iter()
            .find(|g| g.key == "run2")
            .expect("run 2 exists");
        assert_eq!(run2.items.len(), 1, "only its own member");
        assert!(!run2.run_lifetime);
    }

    #[test]
    fn singletons_and_unattributed_calls_group_by_workspace() {
        let agents = vec![
            agent_handle("manager_ws1", "manager", None, "ws1", None),
            agent_handle("maintainer_ws2_abc", "maintainer", None, "ws2", None),
        ];
        let calls = vec![call_handle("some_orchestrator_call", "ws1", None, false)];
        let groups = build_groups(&agents, &calls);
        let singleton: Vec<_> = groups
            .iter()
            .filter(|g| g.kind == GroupKind::Singleton)
            .collect();
        assert_eq!(singleton.len(), 2, "one per workspace");
        let unattributed: Vec<_> = groups
            .iter()
            .filter(|g| g.kind == GroupKind::Unattributed)
            .collect();
        assert_eq!(unattributed.len(), 1);
        assert_eq!(unattributed[0].workspace, "ws1");
    }

    #[test]
    fn sort_key_orders_groups_by_kind() {
        let mut groups = [
            DisplayGroup {
                kind: GroupKind::Unattributed,
                key: "ws".to_string(),
                workspace: "ws".to_string(),
                label: None,
                items: Vec::new(),
                run_lifetime: false,
            },
            DisplayGroup {
                kind: GroupKind::Ticket,
                key: "T1".to_string(),
                workspace: "ws".to_string(),
                label: None,
                items: Vec::new(),
                run_lifetime: false,
            },
            DisplayGroup {
                kind: GroupKind::Research,
                key: "r".to_string(),
                workspace: "ws".to_string(),
                label: None,
                items: Vec::new(),
                run_lifetime: false,
            },
        ];
        groups.sort_by_key(DisplayGroup::sort_key);
        assert_eq!(groups[0].kind, GroupKind::Ticket);
        assert_eq!(groups[1].kind, GroupKind::Research);
        assert_eq!(groups[2].kind, GroupKind::Unattributed);
    }

    #[test]
    fn research_groups_resolve_workspace_across_ephemeral_coder_members() {
        // A run's coder round executes in an ephemeral per-run workspace whose
        // name IS the run key (job_id), while its analysts and orchestrator
        // calls run in the real workspace. The group must resolve to the REAL
        // workspace whether the coder registers before or after the run's
        // real-workspace members, and two concurrent runs must never mix.
        let agents = vec![
            // run1: ephemeral coder registers first, then a real-workspace
            // analyst.
            agent_handle(
                "research_ws1_run1_coder0",
                "coder",
                None,
                "run1",
                Some(ParentKey::Research("run1".to_string())),
            ),
            agent_handle(
                "research_ws1_run1_r1_0",
                "analyst",
                None,
                "ws1",
                Some(ParentKey::Research("run1".to_string())),
            ),
            // run2: real-workspace analyst registers first, then the
            // ephemeral coder.
            agent_handle(
                "research_ws1_run2_r1_0",
                "analyst",
                None,
                "ws1",
                Some(ParentKey::Research("run2".to_string())),
            ),
            agent_handle(
                "research_ws1_run2_coder0",
                "coder",
                None,
                "run2",
                Some(ParentKey::Research("run2".to_string())),
            ),
        ];
        let groups = build_groups(&agents, &[]);
        let runs: Vec<_> = groups
            .iter()
            .filter(|g| g.kind == GroupKind::Research)
            .collect();
        assert_eq!(runs.len(), 2, "two concurrent runs never mix members");
        for run in runs {
            assert_eq!(
                run.workspace, "ws1",
                "the real workspace wins over the ephemeral per-run workspace"
            );
            assert_eq!(run.items.len(), 2, "coder + analyst in one group");
        }
    }

    #[test]
    fn cross_workspace_round_keys_never_merge_groups() {
        // Two concurrent sync analyze rounds in DIFFERENT workspaces could draw
        // the same short NanoID suffix — the workspace component of the group
        // identity must keep their groups separate.
        let agents = vec![
            agent_handle(
                "analyze_ws_AAA_0_analyst",
                "analyst",
                None,
                "ws1",
                Some(ParentKey::AnalyzeRound("abc123".to_string())),
            ),
            agent_handle(
                "analyze_ws_BBB_0_analyst",
                "analyst",
                None,
                "ws2",
                Some(ParentKey::AnalyzeRound("abc123".to_string())),
            ),
        ];
        let groups = build_groups(&agents, &[]);
        let analyze_groups: Vec<_> = groups
            .iter()
            .filter(|g| g.kind == GroupKind::AnalyzeRound)
            .collect();
        assert_eq!(
            analyze_groups.len(),
            2,
            "same round key in two workspaces stays separate"
        );
    }

    /// Build a handle with a parent label (the display-level label captured
    /// at spawn — ticket title / question text).
    fn agent_handle_labeled(
        id: &str,
        role: &str,
        workspace: &str,
        parent: Option<ParentKey>,
        parent_label: Option<&str>,
    ) -> AgentHandle {
        let mut h = agent_handle(id, role, None, workspace, parent);
        h.parent_label = parent_label.map(ToString::to_string);
        h
    }

    fn call_handle_labeled(
        kind: &'static str,
        workspace: &str,
        parent: Option<ParentKey>,
        parent_label: Option<&str>,
    ) -> NonAgentCallHandle {
        let mut h = call_handle(kind, workspace, parent, false);
        h.parent_label = parent_label.map(ToString::to_string);
        h
    }

    #[test]
    fn ticket_group_label_comes_from_members_not_the_key() {
        // Ticket title is captured observationally at spawn (agents via
        // Agent::new's ticket, the synthesis call via the joint-verdict
        // threading). A group must carry the title even when only the
        // synthesis call remains — never degrade to an ID-only header.
        let agents = vec![agent_handle_labeled(
            "a1",
            "analyst",
            "ws1",
            Some(ParentKey::Ticket("T1".to_string())),
            Some("Fix the login flow"),
        )];
        let calls = vec![call_handle_labeled(
            "synthesis",
            "ws1",
            Some(ParentKey::Ticket("T1".to_string())),
            Some("Fix the login flow"),
        )];
        let groups = build_groups(&agents, &calls);
        let ticket = groups
            .iter()
            .find(|g| g.kind == GroupKind::Ticket)
            .expect("ticket group exists");
        assert_eq!(
            ticket.label.as_deref(),
            Some("Fix the login flow"),
            "group label adopted from a member's parent label"
        );
        // group_title preserves (title, id); render_group composes `[id] title`.
        let (title, secondary) = group_title(ticket);
        assert_eq!(title, "Fix the login flow");
        assert_eq!(secondary.as_deref(), Some("T1"));
    }

    #[test]
    fn analyze_and_research_groups_render_question_not_raw_key() {
        // Analyze/research headers show the question text; the raw NanoID key
        // is never rendered. A missing label degrades to a generic label,
        // never the key.
        let agents = vec![agent_handle_labeled(
            "a1",
            "analyst",
            "ws1",
            Some(ParentKey::AnalyzeRound("job_abc".to_string())),
            Some("Why is CI flaky?"),
        )];
        let groups = build_groups(&agents, &[]);
        let analyze = groups
            .iter()
            .find(|g| g.kind == GroupKind::AnalyzeRound)
            .expect("analyze group exists");
        let (title, secondary) = group_title(analyze);
        assert_eq!(title, "Why is CI flaky?");
        assert!(secondary.is_none(), "no id secondary for analyze groups");
        assert!(!title.contains("job_abc"), "raw key never leaks");

        let mut no_label = analyze.clone();
        no_label.label = None;
        let (fallback, _) = group_title(&no_label);
        assert_eq!(fallback, "Analyze round");
        assert!(!fallback.contains("job_abc"), "generic fallback, no key");
    }

    #[test]
    fn sections_group_by_workspace_and_keep_kind_order() {
        // Workspace-first layout. Sections are alphabetical; within a section,
        // groups keep the canonical kind order. Singleton groups are keyed by
        // workspace, so they land in their own workspace's section.
        let mut ws_map = std::collections::HashMap::new();
        ws_map.insert(
            "ws2".to_string(),
            Workspace {
                path: "/ws/ws2".to_string(),
                paused: false,
                maintenance_enabled: false,
                ..Default::default()
            },
        );
        let agents = vec![
            agent_handle(
                "t_agent",
                "engineer",
                Some("T1".to_string()),
                "ws1",
                Some(ParentKey::Ticket("T1".to_string())),
            ),
            agent_handle("mgr_ws2", "manager", None, "ws2", None),
            agent_handle("mgr_ws1", "manager", None, "ws1", None),
        ];
        let groups = build_groups(&agents, &[]);
        let sections = build_sections(groups, &ws_map);
        assert_eq!(sections.len(), 2, "one section per workspace");
        assert_eq!(sections[0].workspace, "ws1");
        assert_eq!(sections[1].workspace, "ws2");
        // ws1: ticket group before singleton (kind order).
        assert_eq!(sections[0].groups[0].kind, GroupKind::Ticket);
        assert_eq!(sections[0].groups[1].kind, GroupKind::Singleton);
        // ws2 is registered → bare name; ws1 is unregistered → "(external)".
        assert_eq!(sections[1].label, "ws2");
        assert_eq!(sections[0].label, "ws1 (external)");
    }

    #[test]
    fn singleton_and_unattributed_headers_are_generic_not_workspace_titled() {
        // Inside a workspace section the singleton/unattributed group headers
        // must NOT repeat the workspace name — the workspace grouping already
        // names it.
        let agents = vec![agent_handle("mgr", "manager", None, "ws1", None)];
        let calls = vec![call_handle("some_orchestrator_call", "ws1", None, false)];
        let groups = build_groups(&agents, &calls);
        for group in &groups {
            let (title, _) = group_title(group);
            assert_ne!(title, "ws1", "group header must not duplicate the section");
            assert!(
                !title.contains("ws1"),
                "no workspace name in group headers: {title}"
            );
        }
    }

    // ── Trace-group derivation ──────────────────────────────────────

    fn assistant_tool_call(narration: &str, calls: &[(&str, serde_json::Value)]) -> ChatMessage {
        assistant_tool_call_reasoning(narration, calls, None)
    }

    /// Like [`assistant_tool_call`] but with an optional decoded `reasoning`
    /// field, so the ledger's `ToolRound.reasoning` is populated.
    fn assistant_tool_call_reasoning(
        narration: &str,
        calls: &[(&str, serde_json::Value)],
        reasoning: Option<&str>,
    ) -> ChatMessage {
        let calls_json: Vec<serde_json::Value> = calls
            .iter()
            .map(|(name, args)| {
                serde_json::json!({
                    "id": "call_1",
                    "name": name,
                    "arguments": serde_json::to_string(args).unwrap_or_default(),
                })
            })
            .collect();
        let content = if narration.is_empty() {
            serde_json::Value::Null
        } else {
            serde_json::Value::String(narration.to_string())
        };
        let mut body = serde_json::json!({ "content": content, "tool_calls": calls_json });
        if let Some(reasoning) = reasoning {
            body["reasoning"] = serde_json::json!(reasoning);
        }
        ChatMessage::assistant(body.to_string())
    }

    #[test]
    fn empty_and_non_tool_history_produce_no_groups() {
        assert!(derive_trace_groups(&[]).is_empty());
        // A content-only assistant message (final answer) is NOT a group.
        let history = vec![
            ChatMessage::user("hello\n\n<timestamp>2026-01-01 00:00:00 (UTC)</timestamp>"),
            ChatMessage::assistant("Final answer here."),
        ];
        assert!(derive_trace_groups(&build_ledger(&history)).is_empty());
    }

    #[test]
    fn narration_starts_and_boundary_groups() {
        let history = vec![
            assistant_tool_call(
                "Reading the file",
                &[("read", serde_json::json!({"path": "a.rs"}))],
            ),
            assistant_tool_call(
                "Editing it",
                &[("edit", serde_json::json!({"path": "b.rs"}))],
            ),
        ];
        let groups = derive_trace_groups(&build_ledger(&history));
        assert_eq!(groups.len(), 2, "new narration starts a new group");
        assert_eq!(groups[0].narration, "Reading the file");
        assert_eq!(groups[1].narration, "Editing it");
    }

    #[test]
    fn consecutive_no_narration_rounds_accumulate() {
        let history = vec![
            assistant_tool_call("", &[("read", serde_json::json!({"path": "a.rs"}))]),
            assistant_tool_call("", &[("list", serde_json::json!({"path": "."}))]),
        ];
        let groups = derive_trace_groups(&build_ledger(&history));
        assert_eq!(groups.len(), 1, "no narration → same group");
        assert_eq!(groups[0].rounds.len(), 2);
        assert!(groups[0].narration.is_empty());
    }

    #[test]
    fn real_user_turn_resets_group() {
        let history = vec![
            assistant_tool_call(
                "Narration",
                &[("read", serde_json::json!({"path": "a.rs"}))],
            ),
            ChatMessage::user("ok now fix it\n\n<timestamp>2026-01-01 00:00:00 (UTC)</timestamp>"),
            assistant_tool_call("", &[("edit", serde_json::json!({"path": "b.rs"}))]),
        ];
        let groups = derive_trace_groups(&build_ledger(&history));
        assert_eq!(groups.len(), 2, "user turn closes the group");
        assert_eq!(groups[0].narration, "Narration");
        // The post-reset empty-narration round starts a fresh group.
        assert_eq!(groups[1].rounds.len(), 1);
    }

    #[test]
    fn injected_image_user_message_does_not_reset_group() {
        let history = vec![
            assistant_tool_call("", &[("read", serde_json::json!({"path": "a.rs"}))]),
            ChatMessage::user(crate::util::injected_image_user_message(
                "data:image/png;base64,xxx",
            )),
            assistant_tool_call("", &[("read", serde_json::json!({"path": "a.rs"}))]),
        ];
        let groups = derive_trace_groups(&build_ledger(&history));
        assert_eq!(groups.len(), 1, "synthetic image stays in the same group");
        assert_eq!(groups[0].rounds.len(), 2);
    }

    #[test]
    fn empty_user_message_is_not_a_turn_boundary() {
        // An empty User message (resume signal) is NOT a real turn boundary:
        // tool rounds on both sides still form ONE group.
        let history = vec![
            assistant_tool_call("", &[("read", serde_json::json!({"path": "a.rs"}))]),
            ChatMessage::user(""),
            assistant_tool_call("", &[("edit", serde_json::json!({"path": "b.rs"}))]),
        ];
        let groups = derive_trace_groups(&build_ledger(&history));
        assert_eq!(groups.len(), 1, "empty user message is not a boundary");
        assert_eq!(groups[0].rounds.len(), 2);
    }

    #[test]
    fn trace_group_reasoning_is_first_non_empty_round() {
        // Reasoning is captured from the first non-empty reasoning among the
        // group's rounds; rounds with no reasoning are skipped, and a group
        // with none at all carries None (the narration label renders bare).
        let history = vec![
            assistant_tool_call_reasoning(
                "",
                &[("read", serde_json::json!({"path": "a.rs"}))],
                None,
            ),
            assistant_tool_call_reasoning(
                "",
                &[("edit", serde_json::json!({"path": "b.rs"}))],
                Some("first"),
            ),
            assistant_tool_call_reasoning(
                "",
                &[("list", serde_json::json!({"path": "."}))],
                Some("second"),
            ),
        ];
        let groups = derive_trace_groups(&build_ledger(&history));
        assert_eq!(groups.len(), 1, "no narration → same group");
        assert_eq!(groups[0].reasoning.as_deref(), Some("first"));
    }

    #[test]
    fn trace_group_without_reasoning_is_none() {
        let history = vec![assistant_tool_call_reasoning(
            "Narration",
            &[("read", serde_json::json!({"path": "a.rs"}))],
            None,
        )];
        let groups = derive_trace_groups(&build_ledger(&history));
        assert_eq!(groups[0].reasoning, None);
    }

    /// Build a ledger tool-call entry for the collapsed/in-flight projection
    /// tests: `result: None` marks the call in-flight.
    fn entry(name: &str, args: serde_json::Value, result: Option<&str>) -> ToolCallEntry {
        ToolCallEntry {
            tool_call_id: String::new(),
            tool: RunningTool::from_tool_call(&crate::ToolCall {
                id: String::new(),
                name: name.to_string(),
                arguments: args,
            }),
            result: result.map(ToString::to_string),
        }
    }

    #[test]
    fn collapsed_calls_line_counts_and_orders() {
        let rounds = vec![
            vec![
                entry("read_file", serde_json::json!({"path": "a.rs"}), Some("ok")),
                entry("list_files", serde_json::json!({"path": "."}), Some("ok")),
            ],
            vec![
                entry("read_file", serde_json::json!({"path": "b.rs"}), Some("ok")),
                entry("read_file", serde_json::json!({"path": "c.rs"}), None),
            ],
        ];
        assert_eq!(
            collapsed_calls_line(&rounds),
            "read file x2, list files, read file …",
            "underscores become spaces; settled names stay exact, an in-flight call gets a running suffix; `xN` counts collapsed repetitions; first-appearance order"
        );
    }

    #[test]
    fn card_shows_only_current_turn_groups() {
        // The card shows only the CURRENT user turn: the pre-turn group is gone.
        let history = vec![
            assistant_tool_call("First", &[("read", serde_json::json!({"path": "a.rs"}))]),
            ChatMessage::user("ok now do it\n\n<timestamp>2026-01-01 00:00:00 (UTC)</timestamp>"),
            assistant_tool_call("Second", &[("edit", serde_json::json!({"path": "b.rs"}))]),
        ];
        let ledger = build_ledger(&history);
        let groups = derive_trace_groups(current_turn_entries(&ledger));
        assert_eq!(groups.len(), 1, "only the post-turn group is shown");
        assert_eq!(groups[0].narration, "Second");
    }

    #[test]
    fn no_user_turn_shows_whole_history() {
        let history = vec![
            assistant_tool_call("A", &[("read", serde_json::json!({"path": "a.rs"}))]),
            assistant_tool_call("B", &[("edit", serde_json::json!({"path": "b.rs"}))]),
        ];
        let ledger = build_ledger(&history);
        let groups = derive_trace_groups(current_turn_entries(&ledger));
        assert_eq!(groups.len(), 2, "no boundary → whole ledger is current");
        assert_eq!(groups[0].narration, "A");
        assert_eq!(groups[1].narration, "B");
    }

    #[test]
    fn current_turn_without_tool_rounds_is_empty() {
        let history = vec![
            assistant_tool_call("A", &[("read", serde_json::json!({"path": "a.rs"}))]),
            ChatMessage::user("stop now\n\n<timestamp>2026-01-01 00:00:00 (UTC)</timestamp>"),
        ];
        let ledger = build_ledger(&history);
        let groups = derive_trace_groups(current_turn_entries(&ledger));
        assert!(
            groups.is_empty(),
            "no tool rounds after the last user turn → empty (clanking placeholder)"
        );
    }

    #[test]
    fn injected_image_and_empty_user_do_not_start_a_turn() {
        let history = vec![
            assistant_tool_call("", &[("read", serde_json::json!({"path": "a.rs"}))]),
            ChatMessage::user(crate::util::injected_image_user_message(
                "data:image/png;base64,xxx",
            )),
            ChatMessage::user(""),
            assistant_tool_call("", &[("edit", serde_json::json!({"path": "b.rs"}))]),
        ];
        let ledger = build_ledger(&history);
        let groups = derive_trace_groups(current_turn_entries(&ledger));
        assert_eq!(
            groups.len(),
            1,
            "injected + empty user messages are not boundaries"
        );
        assert_eq!(groups[0].rounds.len(), 2);
    }

    #[test]
    fn rounds_carry_results_for_in_flight_distinction() {
        // One assistant tool-call round with two calls; a tool result arrives
        // for only one of the two ids. After the ledger, one call is settled
        // and the other in-flight — both preserved in the group's round.
        let calls_json = serde_json::json!([
            {"id": "call_a", "name": "read", "arguments": "{\"path\": \"a.rs\"}"},
            {"id": "call_b", "name": "edit", "arguments": "{\"path\": \"b.rs\"}"},
        ]);
        let body = serde_json::json!({ "content": null, "tool_calls": calls_json });
        let history = vec![
            ChatMessage::assistant(body.to_string()),
            ChatMessage::tool_result("call_a", "contents of a.rs"),
        ];
        let ledger = build_ledger(&history);
        let groups = derive_trace_groups(&ledger);
        assert_eq!(groups.len(), 1);
        let round = &groups[0].rounds[0];
        assert_eq!(round.len(), 2, "both calls preserved in the round");
        let settled = round
            .iter()
            .find(|c| c.tool_call_id == "call_a")
            .expect("call_a present");
        assert_eq!(
            settled.result.as_deref(),
            Some("contents of a.rs"),
            "matched tool result attaches to call_a"
        );
        let in_flight = round
            .iter()
            .find(|c| c.tool_call_id == "call_b")
            .expect("call_b present");
        assert_eq!(
            in_flight.result, None,
            "call_b has no result yet → in-flight"
        );
    }

    #[test]
    fn promoted_reasoning_is_group_label() {
        // A silent round (empty narration) that carries a decoded Reasoning
        // block still forms exactly one group (boundary unchanged) and its
        // label is the promoted reasoning text — not a placeholder. A later
        // narrated round starts a second group (fragmentation did not grow).
        let history = vec![
            assistant_tool_call_reasoning(
                "",
                &[("read", serde_json::json!({"path": "a.rs"}))],
                Some("Checking the file first"),
            ),
            assistant_tool_call_reasoning(
                "Now editing",
                &[("edit", serde_json::json!({"path": "b.rs"}))],
                None,
            ),
        ];
        let groups = derive_trace_groups(&build_ledger(&history));
        assert_eq!(groups.len(), 2, "new narration starts a second group");
        assert!(
            groups[0].narration.is_empty(),
            "silent round keeps empty narration"
        );
        assert!(
            matches!(
                group_label_text(&groups[0]),
                GroupLabelText::Promoted(s) if s == "Checking the file first"
            ),
            "silent round with reasoning labels as its promoted reasoning"
        );
        assert_eq!(groups[1].narration, "Now editing");
    }

    #[test]
    fn silent_round_without_reasoning_is_placeholder() {
        let history = vec![assistant_tool_call(
            "",
            &[("read", serde_json::json!({"path": "a.rs"}))],
        )];
        let groups = derive_trace_groups(&build_ledger(&history));
        assert_eq!(groups.len(), 1);
        assert!(
            matches!(group_label_text(&groups[0]), GroupLabelText::Placeholder),
            "no narration and no reasoning → neutral placeholder"
        );
    }

    #[test]
    fn promoted_label_caps_at_limit() {
        // A long multi-word reasoning is collapsed to one line and capped at
        // the word/path-delimiter boundary with "…" (the narration itself is
        // never truncated, but a promoted fallback must not dominate the card).
        let long = "word ".repeat(80); // 400 chars
        let capped = promoted_reasoning_label(&long);
        assert!(capped.ends_with('…'));
        assert!(
            capped.chars().count() <= MAX_PROMOTED_LABEL_CHARS + 1,
            "cap adds at most the trailing ellipsis"
        );
        // Short text passes through unchanged, with control chars collapsed.
        assert_eq!(promoted_reasoning_label("do it now"), "do it now");
        assert_eq!(promoted_reasoning_label("line1\nline2"), "line1 line2");
    }

    #[test]
    fn prune_expanded_drops_stale_keys() {
        let mut expanded: HashSet<(String, u64)> =
            HashSet::from([("a".to_string(), 0), ("b".to_string(), 2)]);
        let agents = vec![agent_handle("a", "analyst", None, "ws1", None)];
        prune_expanded(&mut expanded, &agents);
        assert_eq!(expanded.len(), 1);
        assert!(
            expanded.contains(&("a".to_string(), 0)),
            "live agent keeps its expanded key"
        );
    }
}