Skip to content

Worktrunk is a CLI for Git worktree management, designed for parallel AI agent workflows

86.4 KB
Raw Blame
//! Progressive-rendering glue between `collect::collect` and the skim picker.
//!
//! Rows funnel into skim's item stream (`tx`, consumed by the skeleton
//! batch — the handler's only send) and each item's shared `rendered` mutex
//! (in-place redraws). The skeleton also publishes session-shared state that
//! outlives this spawn — `shared_items`, the shortcut table, and the layout
//! `PickerCollector` uses for alt-x — and only the live spawn writes it, so a
//! superseded skeleton landing after a rapid second alt-r leaves it alone.
//!
//! # Why updates poke a render
//!
//! skim 4.x's ratatui backend renders **on demand** — it repaints on a key
//! press, when new items land on the channel, or when its matcher flags
//! `needs_render`. It does *not* unconditionally repaint on a timer (the
//! periodic `Heartbeat` only renders when `needs_render` is already set).
//! The old skim 0.20 tuikit backend *did* repaint every 100ms, which is what
//! made in-place `rendered`-mutex mutations surface for free.
//!
//! So after the one-shot skeleton batch, every later mutation is invisible to
//! skim unless we wake it. `request_render` pushes an `Event::Render` through
//! the sender skim hands us at startup (`render_tx`, filled once the TUI is
//! initialized). Intermediate updates coalesce to [`RENDER_THROTTLE`]; the
//! terminal `on_collect_complete` forces one final unthrottled frame.
//!
//! Preview work is staged in two tiers:
//! - `on_skeleton` fires the first item's 5 modes + first-item summary,
//!   plus the default-tab mode for branch-only items 1..N (their committed
//!   diff is cheap from the SHA-keyed cache; off-screen worktree rows are
//!   demand-loaded when selected). It also fills the static Summary hint
//!   for every row when summaries are disabled.
//! - `on_collect_complete` fires configured summaries for items 1..N once
//!   the row pipeline has torn down. Secondary local-git modes on those rows
//!   are computed only when selected by the dedicated demand worker, avoiding
//!   speculative worktree I/O for panes the user may never open.

use std::collections::{HashMap, HashSet};
use std::path::Path;
use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
use std::sync::{Arc, Mutex, OnceLock};
use std::time::{Duration, Instant};

use color_print::cformat;
use skim::prelude::*;
use worktrunk::git::Repository;
use worktrunk::styling::{HINT_SYMBOL, StyledLine};

/// Minimum gap between repaint pokes during collection. Holds the redraw rate
/// at ~60fps so a burst of task results can't flood skim's (effectively
/// unbounded) event channel ahead of the user's key presses. A directly-sent
/// `Event::Render` bypasses skim's own frame-rate cap, so we cap it here.
const RENDER_THROTTLE: Duration = Duration::from_millis(16);

use super::items::{
    HeaderFlash, HeaderLoading, HeaderSkimItem, LayoutSlot, LocalCheckout, LocalContent,
    LocalContentSlot, MorphHandle, PickerRow, PickerRowId, PickerRowSubject, PrStatusSlot,
    PreviewCache, RowShortcutData, RowUrl, ShortcutTable, pr_presence, pr_status_pane_eq,
    worktree_output_token,
};
use super::preview::PreviewMode;
use super::preview_notify::PrStatusDelta;
use super::preview_orchestrator::{PreviewOrchestrator, SpawnGeneration};
use crate::commands::list::collect::PickerProgressHandler;
use crate::commands::list::model::ListItem;

/// Handler owned by the background collect thread. Implements the
/// `PickerProgressHandler` trait that `collect` drives.
pub(super) struct PickerHandler {
    /// skim's item channel, consumed by `on_skeleton`'s single send. Sending
    /// the skeleton batch takes the sender out and drops it, so the channel
    /// closes as soon as the last live sender (the `--prs` thread's clone, if
    /// any) finishes — not when the handler itself drops. Row updates after the
    /// skeleton are in-place (`rendered` mutexes + `request_render`), never
    /// resent, so nothing needs the channel past that send, and holding it open
    /// is expensive: skim's reader polls an open-but-empty channel with kanal's
    /// `recv_timeout`, whose wait yields instead of parking — a full core spun
    /// for as long as collect still grinds (a slow CI fetch, an LLM summary).
    pub(super) tx: Mutex<Option<SkimItemSender>>,
    /// skim's event sender, published by the picker once `Skim::init_tui` has
    /// run. `None` until then — early skeleton sends drive the first paint via
    /// the item channel, so a missed poke before the TUI exists is harmless.
    /// Used to push `Event::Render` when a row mutates in place; see the module
    /// docstring.
    pub(super) render_tx: Arc<OnceLock<tokio::sync::mpsc::Sender<Event>>>,
    /// Last time `request_render` actually poked, for the [`RENDER_THROTTLE`]
    /// coalescing gate.
    pub(super) last_render_poke: Mutex<Instant>,
    /// Mirror of the skim item vec visible to `PickerCollector`. Populated
    /// atomically in `on_skeleton`.
    pub(super) shared_items: Arc<Mutex<Vec<Arc<dyn SkimItem>>>>,
    /// The `alt-y` / `alt-o` lookup table (token → branch + URL). Replaced
    /// atomically in `on_skeleton` with this skeleton's worktree/branch rows;
    /// the `--prs` thread extends it with PR/MR rows. See [`ShortcutTable`].
    pub(super) shortcut_table: ShortcutTable,
    /// One `Arc<Mutex<String>>` per data row — same Arcs `PickerRow`
    /// holds. Set once in `on_skeleton`, read lock-free thereafter.
    pub(super) rendered_slots: OnceLock<Box<[Arc<Mutex<String>>]>>,
    /// One live `pr_status` slot per data row — same `PrStatusSlot` Arcs the
    /// `PickerRow`s hold. Set once in `on_skeleton` (primed from the
    /// cache-filled snapshot), then written by `on_update` as the `CiStatus`
    /// task reports, so the `pr` tab reflects the live fetch.
    pub(super) pr_status_slots: OnceLock<Box<[PrStatusSlot]>>,
    /// Per data row, the PR/MR number whose `comments` thread was last fetched
    /// (`0` = none yet). Set once in `on_skeleton`. The `comments` tab's
    /// background forge fetch must fire at most once per PR, but the PR surfaces
    /// asynchronously (a cache prime at skeleton, or the live `CiStatus` task via
    /// `on_update`) and `on_update` fires repeatedly — so a fetch fires only when
    /// the observed number differs from this slot, and a corrected number (a
    /// reused branch, a stale prime) re-fetches and keeps the `comments` tab
    /// consistent with the `pr` tab. See [`Self::maybe_spawn_comments`].
    pub(super) comments_fetched: OnceLock<Box<[Arc<AtomicU64>]>>,
    /// One live `LocalContent` slot per data row — same Arcs the
    /// `PickerRow`s hold. Set once in `on_skeleton` (all-loading), then
    /// overwritten by `on_update` from the row's `ListItem` as the list pipeline
    /// lands, so the `working_tree` / `branch_diff` / `upstream` tabs dim once
    /// their diff is known empty.
    pub(super) local_content_slots: OnceLock<Box<[LocalContentSlot]>>,
    pub(super) preview_cache: PreviewCache,
    /// Fresh `Repository` for this spawn, used for the mutation-sensitive
    /// `on_skeleton` reads (`list_worktrees`, `local_branches`) — a
    /// `RepoCache` is a set of `OnceCell`s that are never invalidated, so
    /// these reads must go through a repo rebuilt after any in-picker
    /// removal, or they'd yield the stale pre-removal inventory. `spawn`
    /// rebuilds this repo per pass: the same `spawn_repo` the collect/prs
    /// threads use and `PreviewOrchestrator::refresh` rebinds for preview
    /// compute; this field is the handler's direct handle to it.
    pub(super) repo: Repository,
    pub(super) orchestrator: Arc<PreviewOrchestrator>,
    /// This spawn's producer token (see [`SpawnGeneration`]): every
    /// precompute spawn, comments fetch, and skim row this handler starts
    /// carries it, so a refresh that supersedes this spawn also supersedes
    /// everything it still has in flight — including this handler's own
    /// `on_collect_complete` firing after the refresh.
    pub(super) spawn_gen: SpawnGeneration,
    pub(super) preview_dims: (usize, usize),
    pub(super) llm_command: Option<String>,
    /// Filled into the Summary preview cache for every item when summaries
    /// are disabled — gives the Summary tab something useful instead of a
    /// perpetual "Generating…" placeholder.
    pub(super) summary_hint: Option<String>,
    /// Pre-formatted warning lines stashed by `collect::collect` while skim
    /// owns the terminal. The picker drains and emits these to stderr after
    /// `Skim::run_with` returns. Lines are kept in arrival order.
    pub(super) stashed_warnings: Arc<Mutex<Vec<String>>>,
    /// Items captured at `on_skeleton` and consumed at `on_collect_complete`
    /// to fan out the bulk preview pre-compute for items 1..N. Set once
    /// (`OnceLock`) because skeletons fire exactly once per collect.
    pub(super) deferred_items: OnceLock<Vec<Arc<ListItem>>>,
    /// Handoff to the `--prs` thread: the layout's column geometry (PR rows
    /// align to the worktree grid) plus the branches already shown (so `--prs`
    /// skips PRs already represented). Filled in `on_skeleton`. See
    /// [`super::prs::Skeleton`].
    pub(super) grid_slot: Arc<super::prs::GridSlot>,
    /// Handoff of the full collect layout to `PickerCollector`, filled in
    /// `provide_layout` and read at `alt-x` time to render a `/ branch` row on
    /// the same grid (the in-place morph). See [`LayoutSlot`].
    pub(super) layout_slot: LayoutSlot,
    /// Shared with the header: `Some(true)` while the `--prs` forge call is in
    /// flight, so the header shows a "loading…" marker. The `--prs` thread
    /// clears it when the fetch resolves. `None` on non-`--prs` pickers.
    pub(super) prs_loading: Option<Arc<AtomicBool>>,
    /// Shared picker-lifetime with the header and [`AltXRemover`](super::AltXRemover):
    /// a declined `alt-x` sets a transient "couldn't remove this row" message
    /// here, shown in place of the column labels for a beat (see [`HeaderFlash`]).
    pub(super) header_flash: Arc<HeaderFlash>,
}

impl PickerHandler {
    /// Wake skim to repaint after an in-place row mutation. No-op until the
    /// picker has published `render_tx` (TUI initialized). `force` skips the
    /// throttle for the final post-collection frame; otherwise pokes are
    /// coalesced to [`RENDER_THROTTLE`]. Best-effort: a full or closed channel
    /// just drops the poke (the next update, or `on_collect_complete`, catches
    /// the row up).
    fn request_render(&self, force: bool) {
        let Some(tx) = self.render_tx.get() else {
            return;
        };
        if !force {
            let mut last = self.last_render_poke.lock().unwrap();
            if last.elapsed() < RENDER_THROTTLE {
                return;
            }
            *last = Instant::now();
        }
        let _ = tx.try_send(Event::Render);
    }

    /// Spawn row `idx`'s `comments` background fetch if its branch has an open
    /// PR and that PR's thread hasn't been fetched yet. The `comments` tab on a
    /// worktree row shows the same forge-fetched thread a `--prs` row's tab does
    /// (see `items::PickerRow::render_comments_pane`); the PR number comes
    /// from the row's live status, which arrives asynchronously, so this fires
    /// from both the skeleton prime and `on_update`. The per-row
    /// [`Self::comments_fetched`] slot records which PR number was fetched, so
    /// repeated `on_update`s short-circuit — and a *changed* number (a reused
    /// branch, a stale prime corrected by the live fetch) drops the now-wrong
    /// cached thread and re-fetches, keeping the `comments` tab consistent with
    /// the `pr` tab. Both local and `--prs` rows use their structured row ID.
    fn maybe_spawn_comments(&self, idx: usize, item: &ListItem) {
        // A superseded handler (its collect thread draining past an `alt-r`)
        // must not act at all: its corrected-number path below *removes* the
        // shared `(row, Comments)` entry — possibly the one the live spawn
        // just fetched — and its replacement fetch would then drop at `fill`
        // on the stale token, stranding the tab on its loading placeholder
        // (nothing else refills Comments; the live handler's dedup slot
        // already records the number). Gate the removal with the fetch.
        if !self.spawn_gen.is_current() {
            return;
        }
        let Some(Some(status)) = &item.pr_status else {
            return;
        };
        let Some(pr_ref) = status.number else { return };
        // Skip the expired-cache prime: `populate_from_cache` carries an
        // unbounded-stale `updated_at` on a `is_priming` placeholder, and since
        // the fetch is deduped by PR number, keying off it here would lock the
        // comments cache to a wrong signature for the whole session (a stale
        // disk hit, or a mis-keyed write) — the live `CiStatus` fetch
        // (`is_priming: false`, within-TTL/fresh `updated_at`) re-runs this and
        // spawns with the right key. A valid (within-TTL) prime is not priming,
        // so it still warms the tab at skeleton.
        if status.is_priming {
            return;
        }
        let Some(slots) = self.comments_fetched.get() else {
            return;
        };
        let Some(slot) = slots.get(idx) else { return };
        let number = pr_ref.number;
        let previous = slot.swap(number, Ordering::Relaxed);
        if previous == number {
            return; // this PR's comments are already fetched (or in flight)
        }
        let row_id = PickerRowId::local(item);
        if previous != 0 {
            // The live fetch corrected the PR number — drop the stale thread so
            // the re-fetch (below) replaces it rather than serving the old PR.
            self.preview_cache
                .remove(&(row_id.clone(), PreviewMode::Comments));
        }
        super::prs::spawn_worktree_comments_fetch(
            &self.orchestrator,
            &self.spawn_gen,
            row_id,
            number as u32,
            status.updated_at.clone(),
            self.preview_dims.0,
        );
    }
}

/// Branch names the skeleton shows, for the `--prs` thread's dedup: a PR whose
/// head branch is in this set is already on screen, so its row is dropped. A
/// remote row ("origin/foo") also contributes its bare "foo", so a PR for
/// "foo" dedups against a shown "origin/foo". Detached rows (no branch)
/// contribute nothing.
fn collect_shown_branches(items: &[ListItem]) -> HashSet<String> {
    let mut shown = HashSet::new();
    for item in items {
        let Some(name) = item.branch() else {
            continue;
        };
        shown.insert(name.to_string());
        if let Some((_, bare)) = item.branch_ref().remote_parts() {
            shown.insert(bare.to_string());
        }
    }
    shown
}

impl PickerProgressHandler for PickerHandler {
    fn on_skeleton(
        &self,
        items: Vec<ListItem>,
        rendered: Vec<String>,
        header: StyledLine,
        grid: crate::commands::list::layout::ColumnGrid,
    ) {
        debug_assert_eq!(items.len(), rendered.len());

        // The branches already shown, so the `--prs` thread can skip PRs already
        // represented by a worktree/branch row (see `prs::Skeleton`). Computed here,
        // before the row loop consumes `items` — but the handoff that *wakes* the
        // `--prs` thread (`grid_slot.set`, with this set plus the column geometry) is
        // deferred until after the skeleton batch is sent (below). Setting it here
        // instead let the `--prs` thread race the skeleton: with an instant forge
        // call its rows could reach skim's channel first and a PR row would take the
        // reserved header slot (`header_lines(1)`), displacing the real header. The
        // grid is width-stable, so the brief extra wait costs nothing.
        let shown_branches = collect_shown_branches(&items);

        let mut slots: Vec<Arc<Mutex<String>>> = Vec::with_capacity(items.len());
        let mut pr_slots: Vec<PrStatusSlot> = Vec::with_capacity(items.len());
        let mut comments_slots: Vec<Arc<AtomicU64>> = Vec::with_capacity(items.len());
        let mut local_content_slots: Vec<LocalContentSlot> = Vec::with_capacity(items.len());
        let mut skim_items: Vec<Arc<dyn SkimItem>> = Vec::with_capacity(items.len() + 1);
        let mut list_items: Vec<Arc<ListItem>> = Vec::with_capacity(items.len());
        // `alt-y` / `alt-o` lookup entries for this skeleton's rows, keyed by the
        // same `output()` token the rows carry. Replaces the table wholesale below
        // (a refresh re-collect rebuilds it); the `--prs` thread extends it later.
        let mut shortcut_map: HashMap<String, RowShortcutData> =
            HashMap::with_capacity(items.len());

        // Synchronous skeleton-time tab-availability facts (see `TabAvailability`).
        // Branches with an upstream tracking ref drive the tab-5 (remote⇅) empty
        // state, read from the pre-skeleton `for-each-ref` inventory — never the
        // async `item.upstream`. A `local_branches()` failure yields the empty set
        // (every branch reads as no-upstream); preview rendering must not error.
        let upstream_branches: HashSet<String> = self
            .repo
            .local_branches()
            .map(|branches| {
                branches
                    .iter()
                    .filter(|b| b.upstream_short.is_some())
                    .map(|b| b.name.clone())
                    .collect()
            })
            .unwrap_or_default();
        let summaries_enabled = self.llm_command.is_some();

        // Parent of the main worktree — stripped from each row's matcher path
        // (below) so the fuzzy matcher indexes only the distinguishing tail
        // (`worktrunk.skim-features`), not the `~/workspace/` prefix every sibling
        // worktree shares. The base comes from `list_worktrees()` — the same
        // source the row paths come from — so it shares their exact
        // canonicalization (the strip can't miss on a `/private/var`-vs-`/var`
        // mismatch) and adds no network or uncached work (the skeleton was just
        // built from this list). `[0]` is the main worktree for normal repos and
        // the first linked worktree for bare ones; either way its parent is the
        // shared sibling parent. Computed once; a worktree outside that parent
        // keeps its full path via the `strip_prefix` fallback.
        let path_base = self
            .repo
            .list_worktrees()
            .ok()
            .and_then(|worktrees| worktrees.first())
            .and_then(|wt| wt.path.parent().map(Path::to_path_buf));

        // Header row — non-selectable via `header_lines(1)` on the options.
        // In `--prs` mode it shows a dim "↳ Loading open PRs…" line (in place of
        // the column labels) until the forge call's rows land — wording mirrors
        // the empty-list "No open PRs found", styled as a hint (the dim `↳` sits
        // in the pointer gutter, the text aligned with the row content at col 2).
        let loading = self.prs_loading.as_ref().map(|pending| {
            let noun = super::prs::forge_noun(&self.repo);
            HeaderLoading {
                pending: Arc::clone(pending),
                marker_ansi: cformat!("{HINT_SYMBOL} <dim>Loading open {noun}…</>"),
            }
        });
        skim_items.push(Arc::new(HeaderSkimItem {
            display_text: header.plain_text(),
            display_text_with_ansi: header.render(),
            loading,
            flash: Arc::clone(&self.header_flash),
        }) as Arc<dyn SkimItem>);

        for (item, rendered_line) in items.into_iter().zip(rendered) {
            let branch_name = item.branch_name().to_string();
            let has_upstream = upstream_branches.contains(&branch_name);
            // The *distinct* path (leaf relative to the shared worktree parent),
            // not the absolute path — see `path_base`. This feeds only the
            // matcher's `search_text`; the rendered Path column is a separate
            // field and is untouched.
            let path_str = item
                .worktree_path()
                .map(|p| {
                    path_base
                        .as_deref()
                        .and_then(|base| p.strip_prefix(base).ok())
                        .unwrap_or(p)
                        .to_string_lossy()
                        .into_owned()
                })
                .unwrap_or_default();
            // `search_base` is the stable head of the matcher text: branch +
            // distinct path. The PR/MR tokens (reference, title, author) and the
            // trailing gutter glyph are appended live in
            // `PickerRow::text()`, which reads the row's `pr_status` slot
            // each time skim matches — so a PR filters by number/title/author as
            // soon as the CI fetch lands them, the same fields a `--prs` row
            // carries. The gutter glyph stays last so a typed sigil filters by
            // row kind: `+` linked worktrees, `@` the current one, `/`/`|`
            // local/remote branches (`gutter_glyph` is the same skeleton-time
            // fact the rendered Gutter column uses). The folded reference also
            // lets `#` isolate PR-bearing rows (a plain fuzzy char); the GitLab
            // `!` is skim's inverse-match operator, so only the bare number
            // filters those (see
            // `folded_pr_reference_filters_under_skims_default_engine`).
            let gutter = item.kind().gutter_glyph();
            // `display_name`, not `branch_name`: a detached row shows its
            // abbreviated HEAD in the Branch column, so that's what the user
            // reads and types. `branch_name`'s `"(detached)"` matches nothing
            // on screen, and collapses every detached row onto one token.
            let mut search_base = item.display_name().to_string();
            if !path_str.is_empty() {
                search_base.push(' ');
                search_base.push_str(&path_str);
            }

            let rendered_arc = Arc::new(Mutex::new(rendered_line));
            slots.push(Arc::clone(&rendered_arc));

            let item_arc = Arc::new(item);
            list_items.push(Arc::clone(&item_arc));

            // Prime the live slot from the (cache-filled) snapshot so the `pr`
            // tab paints instantly; `on_update` overwrites it as the CiStatus
            // task reports.
            let pr_status_arc: PrStatusSlot = Arc::new(Mutex::new(item_arc.pr_status.clone()));
            pr_slots.push(Arc::clone(&pr_status_arc));
            comments_slots.push(Arc::new(AtomicU64::new(0)));

            // Prime the diff-content slot from the skeleton snapshot. Its async
            // fields are still `None` here (every diff tab reads as loading →
            // available), except a branch-only row, which has no working tree and
            // so resolves `working_tree` to empty immediately. `on_update`
            // overwrites it as the pipeline lands.
            let local_content_arc: LocalContentSlot =
                Arc::new(Mutex::new(LocalContent::from_item(&item_arc)));
            local_content_slots.push(Arc::clone(&local_content_arc));

            // `alt-x` morph handles: a non-primary worktree row with a branch
            // can have that branch outlive a removal, so the row morphs to
            // `/ branch` in place. The primary worktree (can't be removed),
            // detached worktrees (no branch), and branch-only rows (nothing to
            // morph from) carry no handle. Shares the row's live `Arc`s so the
            // collector's mutation surfaces on the same item skim renders.
            let morphed = Arc::new(AtomicBool::new(false));
            let morph = (item_arc.worktree_data().is_some()
                && item_arc.branch().is_some()
                && !item_arc.is_main())
            .then(|| MorphHandle {
                item: Arc::clone(&item_arc),
                rendered: Arc::clone(&rendered_arc),
                local_content: Arc::clone(&local_content_arc),
                morphed: Arc::clone(&morphed),
            });

            // Shortcut lookup for this row: `alt-y` copies `branch`, `alt-o`
            // opens the URL once the live `pr_status` slot reports one. Carry the
            // raw `Option` (not `branch_name()`'s `"(detached)"` fallback) so
            // `alt-y` no-ops on a detached worktree instead of copying the label.
            let output_token = worktree_output_token(&item_arc, &branch_name);
            shortcut_map.insert(
                output_token.clone(),
                RowShortcutData {
                    branch: item_arc.branch().map(str::to_string),
                    url: RowUrl::Live(Arc::clone(&pr_status_arc)),
                    morph,
                },
            );

            skim_items.push(Arc::new(PickerRow {
                search_base,
                gutter,
                rendered: rendered_arc,
                subject: PickerRowSubject::Local(LocalCheckout {
                    item: Arc::clone(&item_arc),
                    demand: Arc::clone(self.orchestrator.demand()),
                    spawn_gen: self.spawn_gen.clone(),
                    has_upstream,
                    summaries_enabled,
                    local_content: local_content_arc,
                    morphed,
                }),
                branch_name,
                output_token,
                preview_cache: Arc::clone(&self.preview_cache),
                pr_status: pr_status_arc,
                notifier: Arc::clone(self.orchestrator.notifier()),
            }) as Arc<dyn SkimItem>);
        }

        // Publish slots + skim items before sending to skim so alt-x reload
        // (which reads `shared_items`) sees a populated list the moment
        // skim calls `CommandCollector::invoke`.
        let _ = self.rendered_slots.set(slots.into_boxed_slice());
        let _ = self.pr_status_slots.set(pr_slots.into_boxed_slice());
        let _ = self.comments_fetched.set(comments_slots.into_boxed_slice());
        let _ = self
            .local_content_slots
            .set(local_content_slots.into_boxed_slice());
        // The session-shared list and shortcut table take only the live
        // spawn's rows: a superseded skeleton landing late (its collect
        // thread scheduled after a rapid second `alt-r`'s spawn already
        // published) would otherwise overwrite them with pre-refresh rows,
        // and a later `alt-x` resync or `alt-y`/`alt-o` would act on those.
        // Checked inside each lock so the check pairs with the live spawn's
        // own overwrite — its generation bump precedes its publish.
        {
            let mut list = self.shared_items.lock().unwrap();
            if self.spawn_gen.is_current() {
                *list = skim_items.clone();
            }
        }
        {
            let mut table = self.shortcut_table.lock().unwrap();
            if self.spawn_gen.is_current() {
                *table = shortcut_map;
            }
        }

        // skim 4.x's item channel carries Vec batches; the skeleton is a single
        // batch. This append wakes skim's reader (`items_available`) and drives
        // the first paint. Later field updates happen in place through each
        // item's shared `rendered` mutex and are *not* resent — they surface via
        // `request_render` (see module docstring), since skim won't repaint a
        // silent in-place mutation on its own. Sending consumes the sender
        // (see the `tx` field doc): this is the handler's only send, and
        // releasing it here lets skim's reader stop polling the channel while
        // collect's remaining tasks grind on.
        if let Some(tx) = self.tx.lock().unwrap().take() {
            let _ = tx.send(skim_items);
        }

        // Skeleton is in skim's channel; now wake the `--prs` thread (see the
        // `shown_branches` note above). Its rows append after the skeleton, so a PR
        // row can never land in the reserved header slot.
        self.grid_slot.set(super::prs::Skeleton {
            grid,
            shown_branches,
        });

        // Tier 1: warm the user's landing row (all modes) and the default tab
        // of branch-only rows. Configured summaries for items 1..N fire from
        // `on_collect_complete` after the row pipeline tears down. Secondary
        // local modes stay demand-loaded.
        self.orchestrator.spawn_initial_precompute(
            &self.spawn_gen,
            &list_items,
            self.preview_dims,
            self.llm_command.as_deref(),
        );
        // A row whose CI status was primed from cache already knows its PR at
        // skeleton time — kick off its `comments` fetch now so the tab is warm.
        // Rows whose PR only surfaces later get theirs from `on_update`.
        for (idx, item) in list_items.iter().enumerate() {
            self.maybe_spawn_comments(idx, item);
        }
        // Static Summary hint is a synchronous in-memory insert, no
        // contention concern. Pre-fill every row at skeleton time so the
        // Summary tab is usable for any selection immediately. Gated like
        // the shared publish above: the seeding bypasses `fill` (static
        // content, documented exception), so a superseded skeleton would
        // otherwise write its stale rows' keys into the refreshed cache.
        if self.llm_command.is_none()
            && self.spawn_gen.is_current()
            && let Some(hint) = self.summary_hint.as_deref()
        {
            self.orchestrator.seed_summary_hints(&list_items, hint);
        }
        let _ = self.deferred_items.set(list_items);
    }

    fn on_update(&self, idx: usize, rendered: String, item: &ListItem) {
        let row_id = PickerRowId::local(item);
        if let Some(slots) = self.rendered_slots.get()
            && let Some(slot) = slots.get(idx)
        {
            *slot.lock().unwrap() = rendered;
        }
        // Mirror the row's current CI status into its live slot so the `pr` /
        // `comments` tabs reflect the fetch as it lands. Track change at the
        // granularity each tab's body reads (see `PrStatusDelta`): `pane_changed`
        // (any field the `pr` pane draws, ignoring the `is_priming` dim hint it
        // doesn't) gates the slot write, the `pr` cache eviction, and the `pr`
        // tab re-render; `presence_changed` (Loading/NoPr/HasPr) gates only the
        // `comments` re-render. An unchanged-pane update touches nothing — a
        // re-render would just reset the preview scroll.
        let mut delta = PrStatusDelta {
            pane_changed: false,
            presence_changed: false,
        };
        if let Some(slots) = self.pr_status_slots.get()
            && let Some(slot) = slots.get(idx)
        {
            let mut slot = slot.lock().unwrap();
            delta.pane_changed = !pr_status_pane_eq(&slot, &item.pr_status);
            delta.presence_changed = pr_presence(&slot) != pr_presence(&item.pr_status);
            if delta.pane_changed {
                *slot = item.pr_status.clone();
            }
        }
        if delta.pane_changed {
            // Drop the memoized `pr` pane for this row so the next `preview()`
            // re-renders from the status just mirrored — see
            // `PickerRow::render_pr_pane_cached`.
            self.preview_cache
                .remove(&(row_id.clone(), PreviewMode::Pr));
        }
        // Mirror the row's live diff-content signals so the `working_tree` /
        // `branch_diff` / `upstream` tabs dim once their diff is known empty.
        // Cheap copy; the snapshot starts all-loading and resolves field-by-field
        // as the pipeline lands.
        if let Some(slots) = self.local_content_slots.get()
            && let Some(slot) = slots.get(idx)
        {
            *slot.lock().unwrap() = LocalContent::from_item(item);
        }
        // If this update is where the row's PR first surfaced, kick off its
        // `comments` background fetch (once per row) so the `comments` tab loads
        // the thread — the same fetch a `--prs` row makes.
        self.maybe_spawn_comments(idx, item);
        // `request_render` sends `Event::Render`, which repaints the *list* row
        // (its CI/status cells just changed) but does NOT re-run the preview.
        // When the `pr` pane changed, poke a `RunPreview` so the selected row's
        // `pr` / `comments` tab flips from "Fetching PR status…" to the resolved
        // PR without a keystroke. The notifier scopes the poke to the selected
        // row and to the visible tab's own change signal, so neither an
        // off-screen update nor a CI fetch landing while the user scrolls a diff
        // (or an unchanged `comments` thread) thrashes the preview — each would
        // reset its scroll (see `PreviewNotifier`). The `local_content` mirror
        // above feeds only the diff tabs' tab-bar dim, which refreshes on the
        // next selection change or tab switch rather than re-running here.
        if delta.pane_changed {
            self.orchestrator
                .notifier()
                .notify_pr_status_changed(&row_id, delta);
        }
        self.request_render(false);
    }

    fn repaint_rows(&self, rendered: Vec<String>) {
        let Some(slots) = self.rendered_slots.get() else {
            return;
        };
        for (slot, line) in slots.iter().zip(rendered) {
            *slot.lock().unwrap() = line;
        }
        self.request_render(false);
    }

    fn stash_warning(&self, line: String) {
        let mut warnings = self.stashed_warnings.lock().unwrap();
        if !warnings.contains(&line) {
            warnings.push(line);
        }
    }

    fn provide_layout(&self, layout: &crate::commands::list::layout::LayoutConfig) {
        // Stow a clone for `PickerCollector` to render the `alt-x` `/ branch`
        // row on this grid. Overwritten each skeleton (a refresh re-collects),
        // but only by the live spawn — a superseded skeleton's grid would
        // misalign the morph row, as in `on_skeleton`'s shared-list guard.
        let mut slot = self.layout_slot.lock().unwrap();
        if self.spawn_gen.is_current() {
            *slot = Some(layout.clone());
        }
    }

    fn on_collect_complete(&self) {
        // Collection is done — `on_update` may have throttled away the last
        // row's repaint, so force one final frame that reflects every slot.
        self.request_render(true);

        let Some(items) = self.deferred_items.get() else {
            return;
        };
        if items.len() <= 1 {
            return;
        }
        self.orchestrator.spawn_deferred_summaries(
            &self.spawn_gen,
            &items[1..],
            self.llm_command.as_deref(),
        );
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::commands::list::model::{ListItem, WorktreeData};
    use worktrunk::testing::TestRepo;

    fn as_worktree(item: ListItem, path: &std::path::Path, data: WorktreeData) -> ListItem {
        let subject = worktrunk::git::WorktreeRef::new(path, item.branch(), item.head());
        ListItem::new_worktree(subject, data)
    }

    /// Build a handler with explicit `repo` (the per-spawn inventory source)
    /// and `orchestrator` (preview compute). Diverging the two lets a test
    /// prove which one `on_skeleton`'s inventory reads consult. `render_tx` is
    /// shared with the orchestrator's notifier (as `handle_picker` shares it),
    /// so a test can publish a channel into it and observe both the list redraw
    /// (`Event::Render`, via `request_render`) and the notifier's
    /// `Event::RunPreview`.
    fn handler_with(
        repo: Repository,
        orchestrator: Arc<PreviewOrchestrator>,
        tx: SkimItemSender,
        render_tx: Arc<OnceLock<tokio::sync::mpsc::Sender<Event>>>,
    ) -> PickerHandler {
        let preview_cache: PreviewCache = Arc::clone(&orchestrator.cache);
        let spawn_gen = orchestrator.generation();
        PickerHandler {
            tx: Mutex::new(Some(tx)),
            render_tx,
            last_render_poke: Mutex::new(Instant::now()),
            shared_items: Arc::new(Mutex::new(Vec::new())),
            shortcut_table: Arc::new(Mutex::new(std::collections::HashMap::new())),
            rendered_slots: OnceLock::new(),
            pr_status_slots: OnceLock::new(),
            comments_fetched: OnceLock::new(),
            local_content_slots: OnceLock::new(),
            preview_cache,
            repo,
            orchestrator,
            spawn_gen,
            preview_dims: (80, 24),
            llm_command: None,
            summary_hint: Some("disabled".to_string()),
            stashed_warnings: Arc::new(Mutex::new(Vec::new())),
            deferred_items: OnceLock::new(),
            grid_slot: Arc::new(super::super::prs::GridSlot::new()),
            layout_slot: Arc::new(Mutex::new(None)),
            prs_loading: None,
            header_flash: Arc::new(HeaderFlash::default()),
        }
    }

    fn make_handler() -> (PickerHandler, TestRepo, SkimItemReceiver) {
        let test = TestRepo::with_initial_commit();
        let (tx, rx): (SkimItemSender, SkimItemReceiver) = unbounded();
        // Share one `render_tx` between the orchestrator's notifier and the
        // handler, as `handle_picker` does — so a test can publish a channel into
        // `handler.render_tx` and observe both the list redraw and the notifier's
        // `Event::RunPreview` on it.
        let render_tx: Arc<OnceLock<tokio::sync::mpsc::Sender<Event>>> = Arc::new(OnceLock::new());
        let orchestrator = Arc::new(PreviewOrchestrator::new(
            test.repo.clone(),
            Arc::clone(&render_tx),
        ));
        let handler = handler_with(test.repo.clone(), orchestrator, tx, render_tx);
        (handler, test, rx)
    }

    fn branch_key(branch: &str) -> PickerRowId {
        PickerRowId::local(&ListItem::new_branch(
            "0000000".to_string(),
            branch.to_string(),
        ))
    }

    fn header(text: &str) -> StyledLine {
        let mut line = StyledLine::new();
        line.push_raw(text);
        line
    }

    fn grid() -> crate::commands::list::layout::ColumnGrid {
        crate::commands::list::layout::ColumnGrid::default()
    }

    fn layout(items: &[ListItem]) -> crate::commands::list::layout::LayoutConfig {
        use crate::commands::list::layout;
        layout::calculate_layout_with_width(
            items,
            &crate::commands::list::columns::all_tasks(),
            layout::Destination {
                width: 80,
                link_style: layout::LinkStyle::Expanded,
            },
            std::path::Path::new("/test"),
            layout::ColumnSelection {
                custom: &[],
                selected: None,
            },
            layout::RepoFacts {
                has_remote: true,
                url_template: None,
                max_pr_number: None,
            },
        )
    }

    /// `on_skeleton` reads the worktree/branch inventory through the handler's
    /// own per-spawn `repo`, NOT the orchestrator's startup-cloned repo. `spawn`
    /// rebuilds `repo` fresh each refresh; reading inventory through the
    /// orchestrator's never-invalidated `RepoCache` would serve the stale
    /// pre-removal snapshot after an in-picker removal. Here the two repos live
    /// under different temp parents, so the matcher path-base
    /// (`list_worktrees().first()`) differs — a row under the handler-repo's tree
    /// strips to a relative tail only if `on_skeleton` consulted `self.repo`.
    #[test]
    fn on_skeleton_reads_inventory_from_handler_repo_not_orchestrator() {
        use crate::commands::list::model::WorktreeData;

        let self_repo = TestRepo::with_initial_commit();
        let orchestrator_repo = TestRepo::with_initial_commit();
        let (tx, rx): (SkimItemSender, SkimItemReceiver) = unbounded();
        let orchestrator = Arc::new(PreviewOrchestrator::new(
            orchestrator_repo.repo.clone(),
            Arc::new(OnceLock::new()),
        ));
        let handler = handler_with(
            self_repo.repo.clone(),
            orchestrator,
            tx,
            Arc::new(OnceLock::new()),
        );

        // A worktree row under the handler-repo's tree. The shared parent strips
        // only if path_base comes from self_repo (parent of `<temp>/repo`), not
        // from orchestrator_repo (a different temp parent).
        let item = as_worktree(
            ListItem::new_branch("abc".into(), "feature".into()),
            &self_repo.path().join("feature"),
            WorktreeData::default(),
        );
        handler.on_skeleton(vec![item], vec!["skel".into()], header("hdr"), grid());

        let received = rx.recv().expect("skeleton batch");
        // received[0] is the header; received[1] is the worktree row.
        let matcher_text = received[1].text().into_owned();
        // The stripped tail renders with the OS separator (`to_string_lossy`),
        // so build the expected tail the same way rather than hardcoding `/`
        // (the row path is `repo\feature` on Windows).
        let expected_tail = Path::new("repo").join("feature");
        assert!(
            matcher_text.contains(expected_tail.to_string_lossy().as_ref()),
            "row path should appear in the matcher text: {matcher_text:?}"
        );
        // The buggy orchestrator-repo path_base can't strip a self_repo path, so
        // it would leave the absolute `<temp>/repo` prefix in the matcher text.
        assert!(
            !matcher_text.contains(self_repo.path().to_string_lossy().as_ref()),
            "path_base from self.repo must strip the absolute prefix: {matcher_text:?}"
        );
    }

    /// Skeleton → update → reveal: verifies that each event writes through
    /// to the shared `rendered` string the `PickerRow` holds. Skim
    /// reads these strings each time it repaints (which `request_render`
    /// triggers); the matcher-stable search text (branch + path) never changes.
    #[test]
    fn handler_updates_render_strings_in_place() {
        let (handler, _test, rx) = make_handler();
        let items = vec![
            ListItem::new_branch("aaa".into(), "one".into()),
            ListItem::new_branch("bbb".into(), "two".into()),
        ];
        let rendered = vec!["skel-one".to_string(), "skel-two".to_string()];

        handler.on_skeleton(items, rendered, header("hdr"), grid());

        // Header + 2 items sent to skim as one batch.
        let received = rx.recv().expect("skeleton batch");
        assert_eq!(received.len(), 3, "expected header + 2 items");

        let slots = handler.rendered_slots.get().unwrap();
        assert_eq!(slots.len(), 2);
        assert_eq!(*slots[0].lock().unwrap(), "skel-one");
        assert_eq!(*slots[1].lock().unwrap(), "skel-two");

        // pr_status slots start primed from the (empty) snapshots.
        let pr_slots = handler.pr_status_slots.get().unwrap();
        assert!(pr_slots[0].lock().unwrap().is_none());
        assert!(pr_slots[1].lock().unwrap().is_none());

        // local_content slots start primed from the skeleton snapshots — these are
        // branch-only rows (no worktree), so `working_tree` already resolves empty
        // while the rest stays loading.
        let lc_slots = handler.local_content_slots.get().unwrap();
        let skeleton_lc =
            LocalContent::from_item(&ListItem::new_branch("aaa".into(), "one".into()));
        assert_eq!(*lc_slots[0].lock().unwrap(), skeleton_lc);
        assert_eq!(*lc_slots[1].lock().unwrap(), skeleton_lc);

        // on_update rewrites a single render slot (the second item here) and
        // mirrors that item's CI status and diff-content into its live slots.
        let mut updated = ListItem::new_branch("bbb".into(), "two".into());
        updated.pr_status = Some(None);
        updated.counts = Some(crate::commands::list::model::AheadBehind {
            ahead: 1,
            behind: 0,
        });
        handler.on_update(1, "updated-two".into(), &updated);
        assert_eq!(*slots[0].lock().unwrap(), "skel-one", "row 0 untouched");
        assert_eq!(*slots[1].lock().unwrap(), "updated-two");
        assert!(
            matches!(&*pr_slots[1].lock().unwrap(), Some(None)),
            "row 1 pr status mirrored from the updated item"
        );
        assert!(
            pr_slots[0].lock().unwrap().is_none(),
            "row 0 pr status untouched"
        );
        assert_eq!(
            *lc_slots[1].lock().unwrap(),
            LocalContent::from_item(&updated),
            "row 1 diff-content mirrored from the updated item"
        );
        assert_eq!(
            *lc_slots[0].lock().unwrap(),
            skeleton_lc,
            "row 0 diff-content untouched"
        );

        // repaint_rows rewrites every slot — slot writes are idempotent
        // through `Mutex<String>`, so unconditional updates are safe.
        handler.repaint_rows(vec!["rev-one".into(), "rev-two".into()]);
        assert_eq!(*slots[0].lock().unwrap(), "rev-one");
        assert_eq!(*slots[1].lock().unwrap(), "rev-two");
    }

    /// The skeleton send consumes the handler's sender (see the `tx` field
    /// doc): with no `--prs` clone in play, the channel closes at the skeleton
    /// batch even though the handler — and the collect thread driving it —
    /// lives on. Held open instead, skim's reader would spin a full core on the
    /// empty channel (kanal's `recv_timeout` yields rather than parking) for as
    /// long as collect's slower tasks (a CI fetch, an LLM summary) keep running.
    #[test]
    fn skeleton_send_consumes_sender_and_closes_channel() {
        let (handler, _test, rx) = make_handler();
        handler.on_skeleton(
            vec![ListItem::new_branch("abc".into(), "one".into())],
            vec!["skel".into()],
            header("hdr"),
            grid(),
        );

        let received = rx.recv().expect("skeleton batch");
        assert_eq!(received.len(), 2, "expected header + 1 item");
        assert!(
            rx.recv().is_err(),
            "channel should be closed after the skeleton send"
        );
        assert!(
            handler.tx.lock().unwrap().is_none(),
            "the sender should be consumed, not merely unused"
        );
    }

    /// `on_update` mirrors a row's live CI status into the preview-feeding slots,
    /// then pokes `Event::RunPreview` so the selected row's `pr` / `comments` tab
    /// re-renders from the new status without a keystroke. This is the loop the
    /// orchestrator-fill path can't close — `on_update` isn't a cache fill, so it
    /// can't ride `PreviewOrchestrator::fill`'s notify; the poke is wired
    /// separately here. It re-runs only when the *visible* tab's body would
    /// differ: an update for a row the cursor isn't on, while a diff tab shows,
    /// for an unchanged status, or for an `is_priming`-only flip (which the
    /// `pr` / `comments` panes don't draw) repaints the list (`Event::Render`)
    /// but must not re-run the preview — a re-run resets its scroll.
    #[test]
    fn on_update_pokes_run_preview_only_when_the_visible_pane_changes() {
        use crate::commands::list::ci_status::{CiSource, CiStatus, PrRef, PrStatus};

        let (handler, _test, rx) = make_handler();
        // Publish skim's event sender (shared with the orchestrator's notifier,
        // as in production) so both the list redraw and the preview re-run land
        // on this channel.
        let (tx, mut render_rx) = tokio::sync::mpsc::channel::<Event>(8);
        handler.render_tx.set(tx).unwrap();

        handler.on_skeleton(
            vec![ListItem::new_branch("aaa".into(), "feature".into())],
            vec!["s".into()],
            header("hdr"),
            grid(),
        );
        let _ = rx.recv(); // drain the skeleton batch

        // A worktree row carrying PR #7, with `is_priming` and `ci_status`
        // controllable so the test can vary just one field at a time.
        let row = |is_priming: bool, ci_status: CiStatus| {
            let mut item = ListItem::new_branch("aaa".into(), "feature".into());
            item.pr_status = Some(Some(PrStatus {
                ci_status,
                source: CiSource::PullRequest,
                is_stale: false,
                is_priming,
                url: None,
                number: Some(PrRef::pr(7)),
                review_state: None,
                title: None,
                body: None,
                author: None,
                comment_count: None,
                updated_at: None,
            }));
            item
        };
        let no_pr = || {
            let mut item = ListItem::new_branch("aaa".into(), "feature".into());
            item.pr_status = Some(None);
            item
        };

        let run_previews = |rx: &mut tokio::sync::mpsc::Receiver<Event>| {
            let mut n = 0;
            while let Ok(ev) = rx.try_recv() {
                if matches!(ev, Event::RunPreview) {
                    n += 1;
                }
            }
            n
        };
        // Arm the awaited `(row, mode)` — but first let every background fill
        // still in flight land, while `awaiting` is a key none of them match.
        // `on_skeleton` spawns a `LOCAL_GIT_MODES` precompute for this row,
        // which ends at `PreviewOrchestrator::fill` → `notify_filled` and pokes
        // the same unlabelled `Event::RunPreview` this oracle counts, as does
        // the `comments` fetch the first live CI report spawns. Left in
        // flight, a fill for the one
        // precompute key a later step also awaits (`WorkingTree`) pokes
        // legitimately but lands after that step's drain — charged to the *next*
        // step, which is the flake in #3725. Quiescing before each
        // `note_awaiting` keeps every counted poke attributable to the
        // `on_update` under test.
        let await_tab = |mode| {
            handler.orchestrator.wait_for_idle();
            handler
                .orchestrator
                .notifier()
                .note_awaiting(&branch_key("feature"), mode)
        };

        // Cursor on `feature`'s `pr` tab (still loading). The CI fetch lands as a
        // cache-prime placeholder → the pane resolves, so it re-runs.
        await_tab(PreviewMode::Pr);
        handler.on_update(0, "r".into(), &row(true, CiStatus::Passed));
        assert!(
            run_previews(&mut render_rx) >= 1,
            "the selected row's first CI report re-runs its pr tab"
        );

        // The live fetch overwrites the prime, clearing only `is_priming`. The
        // pr pane never draws that hint, so the body is byte-identical — it must
        // NOT re-run and reset the scroll of the PR body the user is reading.
        await_tab(PreviewMode::Pr);
        handler.on_update(0, "r".into(), &row(false, CiStatus::Passed));
        assert_eq!(
            run_previews(&mut render_rx),
            0,
            "an is_priming-only flip leaves the pr pane unchanged — no scroll reset"
        );

        // A real field the pane shows (the CI badge) changes → re-run.
        await_tab(PreviewMode::Pr);
        handler.on_update(0, "r".into(), &row(false, CiStatus::Failed));
        assert_eq!(
            run_previews(&mut render_rx),
            1,
            "a changed CI badge re-runs the pr tab"
        );

        // Cursor on the `comments` tab: a pr-field change that keeps the PR
        // present (#7, badge flips back) leaves the row-keyed thread
        // unchanged, so it must NOT re-run a scrolled thread.
        await_tab(PreviewMode::Comments);
        handler.on_update(0, "r".into(), &row(false, CiStatus::Passed));
        assert_eq!(
            run_previews(&mut render_rx),
            0,
            "a pane-only change must not reset a scrolled comments thread"
        );

        // A presence change (#7 → no PR) flips the comments body, so it re-runs.
        await_tab(PreviewMode::Comments);
        handler.on_update(0, "r".into(), &no_pr());
        assert_eq!(
            run_previews(&mut render_rx),
            1,
            "a presence change re-runs the comments tab"
        );

        // Cursor on a diff tab: a real status change (a PR surfaces again) feeds
        // the tab bar's dim but not the diff body, so it must NOT re-run.
        await_tab(PreviewMode::WorkingTree);
        handler.on_update(0, "r".into(), &row(false, CiStatus::Passed));
        assert_eq!(
            run_previews(&mut render_rx),
            0,
            "a CI update while a diff tab is showing must not reset its scroll"
        );

        // Cursor on a different row → `feature`'s update must not re-run the
        // preview the cursor is showing. `request_render`'s throttle may swallow
        // the `Event::Render` too, but the `RunPreview` poke is unthrottled, so
        // its absence is the meaningful signal.
        handler
            .orchestrator
            .notifier()
            .note_awaiting(&branch_key("other"), PreviewMode::Pr);
        handler.on_update(0, "r".into(), &no_pr());
        assert_eq!(
            run_previews(&mut render_rx),
            0,
            "an update to a row the cursor isn't on doesn't re-run the visible preview"
        );
    }

    /// A row's `comments` fetch is once-per-PR, keyed by row identity — but if the
    /// live CI fetch corrects the PR number (a stale prime, a reused branch), the
    /// now-wrong cached thread is dropped and re-fetched, so the `comments` tab
    /// can't keep serving the old PR's thread under the new number. The
    /// `make_handler` repo has no forge, so the fetch caches a terminal pane;
    /// the test watches the cache key being invalidated and repopulated across
    /// the number change.
    #[test]
    fn comments_refetch_on_pr_number_change() {
        use crate::commands::list::ci_status::{CiSource, CiStatus, PrRef, PrStatus};

        let (handler, _test, rx) = make_handler();
        handler.on_skeleton(
            vec![ListItem::new_branch("aaa".into(), "feature".into())],
            vec!["s".into()],
            header("hdr"),
            grid(),
        );
        let _ = rx.recv();

        let with_pr = |number: u64| {
            let mut item = ListItem::new_branch("aaa".into(), "feature".into());
            item.pr_status = Some(Some(PrStatus {
                ci_status: CiStatus::Passed,
                source: CiSource::PullRequest,
                is_stale: false,
                is_priming: false,
                url: None,
                number: Some(PrRef::pr(number)),
                review_state: None,
                title: None,
                body: None,
                author: None,
                comment_count: None,
                updated_at: None,
            }));
            item
        };
        let key = (branch_key("feature"), PreviewMode::Comments);

        // First the row resolves to PR #41 → its comments fetch fires and caches.
        handler.on_update(0, "r".into(), &with_pr(41));
        handler.orchestrator.wait_for_idle();
        assert!(
            handler.preview_cache.contains_key(&key),
            "PR #41's comments fetch populated the row-keyed cache"
        );

        // A repeat update with the SAME number must not re-fetch: overwrite with a
        // sentinel and confirm it survives.
        handler
            .preview_cache
            .insert(key.clone(), "SENTINEL-41".into());
        handler.on_update(0, "r".into(), &with_pr(41));
        handler.orchestrator.wait_for_idle();
        assert_eq!(
            handler
                .preview_cache
                .get(&key)
                .map(|v| v.clone())
                .as_deref(),
            Some("SENTINEL-41"),
            "same PR number short-circuits — no re-fetch"
        );

        // The live fetch corrects the number to #42 → the stale #41 thread is
        // dropped and a fresh fetch repopulates the key.
        handler.on_update(0, "r".into(), &with_pr(42));
        handler.orchestrator.wait_for_idle();
        assert_ne!(
            handler
                .preview_cache
                .get(&key)
                .map(|v| v.clone())
                .as_deref(),
            Some("SENTINEL-41"),
            "a corrected PR number drops the stale thread and re-fetches"
        );
        assert!(
            handler.preview_cache.contains_key(&key),
            "the re-fetch repopulated the comments cache for #42"
        );
    }

    /// An expired-cache prime (`is_priming: true`) carries an unbounded-stale
    /// `updated_at`; spawning the comments fetch off it would lock the disk cache
    /// to a wrong signature for the whole session. The prime is skipped, so the
    /// comments cache stays empty until the live (`is_priming: false`) fetch
    /// lands and spawns with a fresh timestamp.
    #[test]
    fn comments_skip_priming_prime_until_live() {
        use crate::commands::list::ci_status::{CiSource, CiStatus, PrRef, PrStatus};

        let (handler, _test, rx) = make_handler();
        handler.on_skeleton(
            vec![ListItem::new_branch("aaa".into(), "feature".into())],
            vec!["s".into()],
            header("hdr"),
            grid(),
        );
        let _ = rx.recv();

        let status = |is_priming: bool, updated_at: &str| {
            let mut item = ListItem::new_branch("aaa".into(), "feature".into());
            item.pr_status = Some(Some(PrStatus {
                ci_status: CiStatus::Passed,
                source: CiSource::PullRequest,
                is_stale: false,
                is_priming,
                url: None,
                number: Some(PrRef::pr(7)),
                review_state: None,
                title: None,
                body: None,
                author: None,
                comment_count: None,
                updated_at: Some(updated_at.to_string()),
            }));
            item
        };
        let key = (branch_key("feature"), PreviewMode::Comments);

        // Expired prime → skipped: no comments fetch, cache stays empty. (The
        // prime also must not consume the dedup slot, or the live update below
        // would short-circuit.)
        handler.on_update(0, "r".into(), &status(true, "2020-01-01T00:00:00Z"));
        handler.orchestrator.wait_for_idle();
        assert!(
            !handler.preview_cache.contains_key(&key),
            "an is_priming prime must not spawn the comments fetch"
        );

        // Live fetch (not priming) with a fresh timestamp → spawns and caches.
        handler.on_update(0, "r".into(), &status(false, "2026-06-28T00:00:00Z"));
        handler.orchestrator.wait_for_idle();
        assert!(
            handler.preview_cache.contains_key(&key),
            "the live fetch spawns the comments fetch with a fresh signature"
        );
    }

    /// Header + items get published in order. `output()` of the
    /// PickerRow is the branch name so skim returns the correct
    /// identifier when the user hits Enter.
    #[test]
    fn skeleton_publishes_header_then_items() {
        let (handler, _test, rx) = make_handler();
        let items = vec![
            ListItem::new_branch("aaa".into(), "feat-a".into()),
            ListItem::new_branch("bbb".into(), "feat-b".into()),
        ];

        handler.on_skeleton(
            items,
            vec!["skel-a".into(), "skel-b".into()],
            header("Branch Status"),
            grid(),
        );

        let received = rx.recv().expect("skeleton batch");
        assert_eq!(received.len(), 3);
        // Header emits empty output (not selectable).
        assert_eq!(received[0].output().as_ref(), "");
        // Branch items emit the branch name.
        assert_eq!(received[1].output().as_ref(), "feat-a");
        assert_eq!(received[2].output().as_ref(), "feat-b");

        // Shared state matches what was sent.
        let shared = handler.shared_items.lock().unwrap();
        assert_eq!(shared.len(), 3);
        assert_eq!(shared[1].output().as_ref(), "feat-a");
        assert_eq!(shared[2].output().as_ref(), "feat-b");

        // The shortcut table is keyed by each row's `output()` token (the branch
        // name for these branch-only rows) and carries the branch for `alt-y`.
        let table = handler.shortcut_table.lock().unwrap();
        assert_eq!(table.len(), 2);
        assert_eq!(
            table.get("feat-a").and_then(|d| d.branch.as_deref()),
            Some("feat-a")
        );
        assert_eq!(
            table.get("feat-b").and_then(|d| d.branch.as_deref()),
            Some("feat-b")
        );
    }

    /// A detached worktree row has no branch, so its shortcut entry carries
    /// `branch: None` — `alt-y` then no-ops rather than copying the
    /// `"(detached)"` label `branch_name()` falls back to.
    #[test]
    fn shortcut_table_branch_is_none_for_detached_worktree() {
        use crate::commands::list::model::WorktreeData;

        let (handler, _test, _rx) = make_handler();
        let path = std::path::PathBuf::from("/tmp/wt-detached");
        let item = ListItem::new_worktree(
            worktrunk::git::WorktreeRef::new(&path, None, "abc123"),
            WorktreeData {
                detached: true,
                ..Default::default()
            },
        );

        handler.on_skeleton(vec![item], vec!["skel".into()], header("hdr"), grid());

        let table = handler.shortcut_table.lock().unwrap();
        assert_eq!(table.len(), 1);
        let entry = table.values().next().expect("one row in the table");
        assert_eq!(
            entry.branch, None,
            "a detached worktree carries no branch, so alt-y no-ops"
        );
    }

    /// `on_skeleton` folds each row's gutter glyph onto the end of the
    /// matcher's `search_text`, so typing a sigil filters by row kind. The
    /// glyph comes from `ItemKind::gutter_glyph` (which handles every kind and
    /// is unit-tested in `item.rs`), so branch rows here prove the wiring.
    #[test]
    fn search_text_folds_in_gutter_glyph() {
        let (handler, _test, rx) = make_handler();
        let items = vec![
            ListItem::new_branch("abc".into(), "localbr".into()),
            ListItem::new_remote_branch("abc".into(), "origin".into(), "remotebr".into()),
        ];
        handler.on_skeleton(
            items,
            vec!["skel-local".into(), "skel-remote".into()],
            header("hdr"),
            grid(),
        );

        let received = rx.recv().expect("skeleton batch");
        // received[0] is the non-selectable header; data rows follow in order.
        let text = |i: usize| received[i].text().into_owned();
        assert!(text(1).ends_with('/'), "local branch: {:?}", text(1));
        assert!(text(2).ends_with('|'), "remote branch: {:?}", text(2));
        // Branch name stays searchable alongside the folded-in glyph.
        assert!(text(1).contains("localbr"));
        assert!(text(2).contains("origin/remotebr"));
    }

    /// A worktree row's matcher text folds in its PR's reference, title, and
    /// author — the same fields a `--prs` row carries — so a PR filters
    /// identically however it's shown. Here the PR is primed at skeleton (the
    /// cache case); the gutter glyph stays trailing for sigil filtering.
    #[test]
    fn worktree_row_matcher_text_folds_in_pr_number_title_author() {
        use crate::commands::list::ci_status::{CiSource, CiStatus, PrRef, PrStatus};

        let (handler, _test, rx) = make_handler();
        let mut item = ListItem::new_branch("abc".into(), "localbr".into());
        item.pr_status = Some(Some(PrStatus {
            ci_status: CiStatus::Passed,
            source: CiSource::PullRequest,
            is_stale: false,
            is_priming: false,
            url: None,
            number: Some(PrRef::pr(123)),
            review_state: None,
            title: Some("Speed up startup".into()),
            body: None,
            author: Some("alice".into()),
            comment_count: None,
            updated_at: None,
        }));
        handler.on_skeleton(vec![item], vec!["skel".into()], header("hdr"), grid());

        let received = rx.recv().expect("skeleton batch");
        let text = received[1].text().into_owned();
        // The CI column shows `#123`; typing the number (or title/author) filters.
        assert!(text.contains("#123"), "PR reference folded in: {text:?}");
        assert!(
            text.contains("Speed up startup"),
            "title folded in: {text:?}"
        );
        assert!(text.contains("alice"), "author folded in: {text:?}");
        assert!(
            text.contains("localbr"),
            "branch still searchable: {text:?}"
        );
        assert!(text.ends_with('/'), "gutter glyph stays trailing: {text:?}");
    }

    /// The freeze is relaxed: a worktree row's matcher text is read live from its
    /// `pr_status` slot, so a PR the live CI fetch discovers (cold cache, nothing
    /// primed at skeleton) becomes filterable by number/title/author once
    /// `on_update` lands it — not just on a later run.
    #[test]
    fn worktree_row_matcher_text_updates_when_live_fetch_lands_pr() {
        use crate::commands::list::ci_status::{CiSource, CiStatus, PrRef, PrStatus};

        let (handler, _test, rx) = make_handler();
        // Skeleton with no PR (cold cache).
        handler.on_skeleton(
            vec![ListItem::new_branch("abc".into(), "localbr".into())],
            vec!["skel".into()],
            header("hdr"),
            grid(),
        );
        let received = rx.recv().expect("skeleton batch");
        let row = Arc::clone(&received[1]);
        assert!(
            !row.text().contains("#7"),
            "no PR before the fetch: {:?}",
            row.text()
        );

        // The live fetch reports a PR for the row.
        let mut updated = ListItem::new_branch("abc".into(), "localbr".into());
        updated.pr_status = Some(Some(PrStatus {
            ci_status: CiStatus::Passed,
            source: CiSource::PullRequest,
            is_stale: false,
            is_priming: false,
            url: None,
            number: Some(PrRef::pr(7)),
            review_state: None,
            title: Some("Add caching".into()),
            body: None,
            author: Some("bob".into()),
            comment_count: None,
            updated_at: None,
        }));
        handler.on_update(0, "rendered".into(), &updated);

        // Same row item; its matcher text now reflects the landed PR.
        let text = row.text().into_owned();
        assert!(text.contains("#7"), "number now filters: {text:?}");
        assert!(text.contains("Add caching"), "title now filters: {text:?}");
        assert!(text.contains("bob"), "author now filters: {text:?}");
        assert!(text.ends_with('/'), "gutter glyph stays trailing: {text:?}");
    }

    /// `collect_shown_branches` gathers branch names for the `--prs` dedup, and
    /// adds a remote row's bare name so a PR for "foo" dedups against a shown
    /// "origin/foo".
    #[test]
    fn collect_shown_branches_adds_remote_bare_names() {
        let items = vec![
            ListItem::new_branch("a".into(), "local-feat".into()),
            ListItem::new_remote_branch("b".into(), "origin".into(), "remote-feat".into()),
            ListItem::new_remote_branch("c".into(), "team/fork".into(), "slash-feat".into()),
        ];
        let shown = collect_shown_branches(&items);
        assert!(shown.contains("local-feat"), "local branch name");
        assert!(shown.contains("origin/remote-feat"), "full remote ref");
        assert!(
            shown.contains("remote-feat"),
            "bare name so a PR head dedups against the remote row"
        );
        assert!(shown.contains("team/fork/slash-feat"), "slash remote ref");
        assert!(
            shown.contains("slash-feat"),
            "bare name from a slash-named remote"
        );
    }

    /// `on_skeleton` strips the shared main-worktree parent from each row's
    /// matcher `search_text`, so the fuzzy matcher indexes the distinguishing
    /// leaf (`repo.feat`) rather than the absolute prefix every sibling worktree
    /// carries. The inside row's path is read back from `worktree_for_branch`, so
    /// the strip runs against the real `list_worktrees()` path the rows carry, not
    /// `add_worktree`'s separately canonicalized return. A worktree outside the
    /// shared parent keeps its full path via the `strip_prefix` fallback.
    #[test]
    fn search_text_strips_shared_worktree_parent() {
        use crate::commands::list::model::WorktreeData;

        let (handler, mut test, rx) = make_handler();
        // A genuine sibling worktree at `{temp}/repo.feat`; read its path back from
        // the worktree list (the production path source — both the row path and
        // the stripped base come from `list_worktrees()`, so they share one
        // canonicalization).
        test.add_worktree("feat");
        let inside_path = test
            .repo
            .worktree_for_branch("feat")
            .unwrap()
            .expect("feat worktree is registered");
        let outside_path = std::path::PathBuf::from("/nonexistent-root/external-wt");

        let worktree_item = |branch: &str, path: &Path| {
            as_worktree(
                ListItem::new_branch("abc".into(), branch.into()),
                path,
                WorktreeData::default(),
            )
        };

        handler.on_skeleton(
            vec![
                worktree_item("inside", &inside_path),
                worktree_item("outside", &outside_path),
            ],
            vec!["skel-inside".into(), "skel-outside".into()],
            header("hdr"),
            grid(),
        );

        let received = rx.recv().expect("skeleton batch");
        let text = |i: usize| received[i].text().into_owned();
        let (inside, outside) = (text(1), text(2));

        // Inside the shared parent: only the leaf is indexed (gutter `+` for a
        // non-current linked worktree). The absolute prefix is gone.
        assert_eq!(inside, "inside repo.feat +", "leaf only: {inside:?}");

        // Outside the shared parent: the full path is retained (fallback).
        assert_eq!(
            outside, "outside /nonexistent-root/external-wt +",
            "out-of-tree path kept whole: {outside:?}"
        );
    }

    /// Locks the gutter-sigil filtering contract against skim's default engine —
    /// an `AndOrEngineFactory` wrapping an `ExactOrFuzzyEngineFactory`, the
    /// factory `Matcher::create_engine_factory` builds when the picker sets
    /// neither `--exact` nor `--regex` (skim 4.8 `src/matcher.rs`). `+`/`@` are
    /// literals that filter to their row kind; `^` and `|` collide with skim's
    /// prefix-anchor and OR operators and don't.
    #[test]
    fn gutter_glyphs_filter_under_skims_default_engine() {
        struct TextItem(String);
        impl SkimItem for TextItem {
            fn text(&self) -> std::borrow::Cow<'_, str> {
                std::borrow::Cow::Borrowed(&self.0)
            }
        }

        // One representative folded `search_text` per gutter kind.
        let current = "cur /tmp/cur @";
        let primary = "primary /tmp/primary ^";
        let linked = "linked /tmp/linked +";
        let local = "localbr /";
        let remote = "origin/remotebr |";

        let matches = |query: &str, haystack: &str| -> bool {
            let factory = AndOrEngineFactory::new(ExactOrFuzzyEngineFactory::builder().build());
            factory
                .create_engine(query)
                .match_item(&TextItem(haystack.to_string()))
                .is_some()
        };

        // `+`/`@` are literals — each filters to exactly its row kind.
        assert!(matches("+", linked), "+ selects the linked worktree");
        for other in [current, primary, local, remote] {
            assert!(!matches("+", other), "+ must not match {other:?}");
        }
        assert!(matches("@", current), "@ selects the current worktree");
        for other in [primary, linked, local, remote] {
            assert!(!matches("@", other), "@ must not match {other:?}");
        }

        // `^` is skim's prefix anchor: a bare `^` is an empty prefix pattern
        // that matches every row, so it can't isolate the primary worktree.
        // Proven by matching rows that contain no literal `^`.
        for row in [current, linked, local, remote] {
            assert!(
                matches("^", row),
                "^ (prefix anchor) matches all rows — not a filter: {row:?}"
            );
        }
        // A bare `|` is skim's OR separator with empty operands, which matches
        // nothing — so it can't isolate remote-branch rows (which carry `|`).
        for row in [remote, current, linked, local] {
            assert!(
                !matches("|", row),
                "| (OR operator) matches nothing — not a filter: {row:?}"
            );
        }
    }

    /// Locks the PR/MR filtering contract: a worktree row folds its PR's
    /// reference, title, and author into the matcher text, so under skim's
    /// default engine all three filter the row regardless of forge sigil. `#`
    /// (GitHub/Gitea/Azure) is a plain fuzzy char that also isolates PR-bearing
    /// rows, while the GitLab `!` is skim's inverse-match operator, so typing the
    /// literal `!7` *excludes* its own MR row — only the number filters there.
    /// Companion to `gutter_glyphs_filter_under_skims_default_engine`.
    #[test]
    fn folded_pr_reference_filters_under_skims_default_engine() {
        struct TextItem(String);
        impl SkimItem for TextItem {
            fn text(&self) -> std::borrow::Cow<'_, str> {
                std::borrow::Cow::Borrowed(&self.0)
            }
        }
        let matches = |query: &str, haystack: &str| -> bool {
            AndOrEngineFactory::new(ExactOrFuzzyEngineFactory::builder().build())
                .create_engine(query)
                .match_item(&TextItem(haystack.to_string()))
                .is_some()
        };

        // Folded matcher text for a worktree row carrying a PR (`#123`) / MR
        // (`!7`) — branch, path, reference, title, author, gutter — and one with
        // no PR. The PR fields are exactly what a `--prs` row also folds in.
        let github = "feature feature.wt #123 Add caching alice +";
        let gitlab = "feature feature.wt !7 Speed up startup bob +";
        let no_pr = "plain plain.wt +";

        // The number, title, and author all filter — the same fields in both
        // pickers, regardless of forge sigil.
        assert!(matches("123", github), "PR number filters the GitHub row");
        assert!(matches("7", gitlab), "MR number filters the GitLab row");
        assert!(matches("caching", github), "title filters");
        assert!(matches("alice", github), "author filters");
        assert!(matches("bob", gitlab), "MR author filters");
        // `#` is a plain fuzzy char: it isolates PR-bearing rows and skips
        // PR-less ones (same as the `--prs` gutter sigil).
        assert!(matches("#", github), "# selects a PR-bearing row");
        assert!(!matches("#", no_pr), "# skips a PR-less row");
        // `!7` is skim's inverse-match operator — it excludes rows containing
        // `7`, hiding the very MR row it names. The bare number is reliable.
        assert!(!matches("!7", gitlab), "!7 inverse-excludes its own MR row");
    }

    /// `stash_warning` accumulates distinct lines in arrival order so refreshes
    /// and concurrent producers cannot repeat a diagnostic after skim exits.
    #[test]
    fn stash_warning_preserves_order() {
        let (handler, _test, _rx) = make_handler();
        handler.stash_warning("first".into());
        handler.stash_warning("second".into());
        handler.stash_warning("first".into());
        handler.stash_warning("third".into());
        let stash = handler.stashed_warnings.lock().unwrap();
        assert_eq!(stash.as_slice(), &["first", "second", "third"]);
    }

    /// Preview pre-compute is tiered. After `on_skeleton`:
    /// - First item gets all 5 modes (the user's landing row).
    /// - Branch-only items 1..N get only `UnifiedDiff` (the picker's initial
    ///   tab) so quick j/k navigation hits warm content. Worktree-backed rows
    ///   are demand-loaded instead; that split is pinned in
    ///   `preview_orchestrator`, not here.
    /// - Secondary modes for items 1..N stay demand-loaded rather than
    ///   consuming worktree I/O speculatively.
    ///
    /// Summary hint is filled for every item synchronously at skeleton
    /// time so the Summary tab is usable for any selection immediately.
    #[test]
    fn precompute_staging_tiers_match_design() {
        use super::super::preview::PreviewMode;

        let (handler, _test, _rx) = make_handler();
        let items = vec![
            ListItem::new_branch("aaa".into(), "alpha".into()),
            ListItem::new_branch("bbb".into(), "beta".into()),
            ListItem::new_branch("ccc".into(), "gamma".into()),
        ];
        let rendered = vec!["s1".into(), "s2".into(), "s3".into()];

        handler.on_skeleton(items, rendered, header("hdr"), grid());
        handler.orchestrator.wait_for_idle();

        // Static Summary hint primed for every item at skeleton time.
        for branch in ["alpha", "beta", "gamma"] {
            let key = (branch_key(branch), PreviewMode::Summary);
            let hint = handler.preview_cache.get(&key).unwrap_or_else(|| {
                panic!("Summary hint should be filled for {branch} at skeleton time")
            });
            assert_eq!(
                hint.value(),
                "disabled",
                "Summary cache should contain the selected hint for {branch}"
            );
        }

        // First item: all 5 modes spawned at skeleton time.
        for mode in [
            PreviewMode::UnifiedDiff,
            PreviewMode::WorkingTree,
            PreviewMode::BranchDiff,
            PreviewMode::Log,
            PreviewMode::UpstreamDiff,
        ] {
            assert!(
                handler
                    .preview_cache
                    .contains_key(&(branch_key("alpha"), mode)),
                "first item should have {mode:?} cached after on_skeleton"
            );
        }

        // Branch-only items 1..N: UnifiedDiff cached at skeleton time.
        for branch in ["beta", "gamma"] {
            assert!(
                handler
                    .preview_cache
                    .contains_key(&(branch_key(branch), PreviewMode::UnifiedDiff)),
                "{branch}.UnifiedDiff should be cached after on_skeleton (initial tier)"
            );
        }

        // Items 1..N: secondary modes are not spawned speculatively.
        for branch in ["beta", "gamma"] {
            for mode in [
                PreviewMode::WorkingTree,
                PreviewMode::Log,
                PreviewMode::BranchDiff,
                PreviewMode::UpstreamDiff,
            ] {
                assert!(
                    !handler
                        .preview_cache
                        .contains_key(&(branch_key(branch), mode)),
                    "{branch}.{mode:?} should remain demand-loaded"
                );
            }
        }

        handler.on_collect_complete();
        handler.orchestrator.wait_for_idle();

        // Collection completion must not bulk-compute subsidiary local modes.
        // (Configured summaries use this hook, but this fixture has them off.)
        for branch in ["beta", "gamma"] {
            for mode in [
                PreviewMode::WorkingTree,
                PreviewMode::Log,
                PreviewMode::BranchDiff,
                PreviewMode::UpstreamDiff,
            ] {
                assert!(
                    !handler
                        .preview_cache
                        .contains_key(&(branch_key(branch), mode)),
                    "{branch}.{mode:?} should remain demand-loaded after collection"
                );
            }
        }
    }

    /// `on_collect_complete` is safe to call when no skeleton ever fired
    /// (e.g. zero-worktree early return in `collect`) and when only one
    /// item exists (nothing to defer). The post-conditions assert that no
    /// extra spawns leak into the cache — a regression that introduces
    /// work for items 1..N from this hook would surface here.
    #[test]
    fn on_collect_complete_is_no_op_when_no_rest_items() {
        use super::super::preview::PreviewMode;

        // Case 1: never called on_skeleton — cache must remain empty.
        let (handler, _test, _rx) = make_handler();
        handler.on_collect_complete();
        handler.orchestrator.wait_for_idle();
        assert_eq!(
            handler.preview_cache.iter().count(),
            0,
            "no work should be spawned when on_skeleton never fired"
        );

        // Case 2: single-item skeleton — first-item phase covered the 5
        // modes plus the static Summary hint (6 entries total). Nothing
        // left to defer; on_collect_complete must not add any entries.
        let (handler, _test, _rx) = make_handler();
        let items = vec![ListItem::new_branch("aaa".into(), "solo".into())];
        handler.on_skeleton(items, vec!["s1".into()], header("hdr"), grid());
        handler.orchestrator.wait_for_idle();
        let before = handler.preview_cache.iter().count();
        for mode in [
            PreviewMode::UnifiedDiff,
            PreviewMode::WorkingTree,
            PreviewMode::Log,
            PreviewMode::BranchDiff,
            PreviewMode::UpstreamDiff,
            PreviewMode::Summary,
        ] {
            assert!(
                handler
                    .preview_cache
                    .contains_key(&(branch_key("solo"), mode)),
                "first-item phase should have cached {mode:?}"
            );
        }
        assert_eq!(before, 6, "first-item phase populates exactly 6 entries");

        handler.on_collect_complete();
        handler.orchestrator.wait_for_idle();
        assert_eq!(
            handler.preview_cache.iter().count(),
            before,
            "on_collect_complete must not spawn additional work for a single-item skeleton"
        );
    }

    /// A handler superseded by a refresh must not publish its skeleton into
    /// the session-shared row list / shortcut table / alt-x layout, nor seed
    /// summary hints:
    /// a rapid second `alt-r` can schedule the new spawn's skeleton before a
    /// slow prior spawn's, and the stale skeleton landing second would
    /// otherwise overwrite the live rows (feeding a later `alt-x` resync and
    /// the `alt-y`/`alt-o` lookups) and write stale-row hint keys into the
    /// refreshed cache.
    #[test]
    fn superseded_skeleton_leaves_shared_state_alone() {
        // The live spawn publishes its layout for `alt-x`'s morph row.
        let (live, _live_test, _live_rx) = make_handler();
        let live_items = vec![ListItem::new_branch("abc".into(), "live".into())];
        live.provide_layout(&layout(&live_items));
        assert!(
            live.layout_slot.lock().unwrap().is_some(),
            "live skeleton publishes its layout"
        );

        let (handler, test, rx) = make_handler();
        handler.orchestrator.refresh(test.repo.clone());

        let stale_items = vec![ListItem::new_branch("abc".into(), "stale".into())];
        handler.on_skeleton(
            stale_items.clone(),
            vec!["skel".into()],
            header("hdr"),
            grid(),
        );
        // `collect` hands over the layout right after the skeleton.
        handler.provide_layout(&layout(&stale_items));
        handler.orchestrator.wait_for_idle();

        // The rows still stream to this spawn's (dead) skim channel...
        let received = rx.recv().expect("skeleton batch");
        assert_eq!(received.len(), 2, "header + row still sent");
        // ...but nothing session-shared takes them.
        assert!(
            handler.shared_items.lock().unwrap().is_empty(),
            "superseded skeleton must not overwrite the shared row list"
        );
        assert!(
            handler.shortcut_table.lock().unwrap().is_empty(),
            "superseded skeleton must not overwrite the shortcut table"
        );
        assert!(
            handler.preview_cache.is_empty(),
            "superseded skeleton must not seed hints into the refreshed cache"
        );
        assert!(
            handler.layout_slot.lock().unwrap().is_none(),
            "superseded skeleton must not overwrite the alt-x layout"
        );
    }

    /// A superseded handler's `maybe_spawn_comments` is fully inert. Its
    /// corrected-number path would otherwise evict the live spawn's
    /// `(row, Comments)` entry from the shared cache while its own refetch
    /// drops at `fill` on the stale token — stranding the tab on its loading
    /// placeholder, since the live handler's dedup slot already records the
    /// number and nothing else refills Comments.
    #[test]
    fn superseded_handler_does_not_evict_live_comments() {
        use crate::commands::list::ci_status::{CiSource, CiStatus, PrRef, PrStatus};

        let status = |n: u64| {
            Some(Some(PrStatus {
                ci_status: CiStatus::Passed,
                source: CiSource::PullRequest,
                is_stale: false,
                is_priming: false,
                url: None,
                number: Some(PrRef::pr(n)),
                review_state: None,
                title: None,
                body: None,
                author: None,
                comment_count: None,
                updated_at: None,
            }))
        };

        let (handler, test, rx) = make_handler();
        handler.on_skeleton(
            vec![ListItem::new_branch("abc".into(), "b".into())],
            vec!["skel".into()],
            header("hdr"),
            grid(),
        );
        let _ = rx.recv();
        let mut item = ListItem::new_branch("abc".into(), "b".into());
        let row_id = PickerRowId::local(&item);
        // Record PR #5 in this handler's dedup slot (the fetch resolves to
        // the "unsupported forge" pane — the test repo has no forge remote —
        // only the recorded number matters here).
        item.pr_status = status(5);
        handler.maybe_spawn_comments(0, &item);

        // A refresh supersedes this handler; the live spawn fetches the thread.
        handler.orchestrator.refresh(test.repo.clone());
        let key = (row_id.clone(), PreviewMode::Comments);
        PreviewOrchestrator::fill(
            &handler.orchestrator.cache,
            handler.orchestrator.notifier(),
            &handler.orchestrator.generation(),
            key.clone(),
            "live thread".to_string(),
        );

        // The stale handler observing a corrected number must not touch the
        // live spawn's entry.
        item.pr_status = status(6);
        handler.maybe_spawn_comments(0, &item);
        handler.orchestrator.wait_for_idle();
        assert_eq!(
            handler.preview_cache.get(&key).map(|v| v.clone()),
            Some("live thread".to_string()),
            "a superseded handler must not evict the live spawn's Comments entry"
        );
    }
}