Pitch

A user just had the TUI "basically die" — sustained lag / unresponsiveness during normal use. This is an audit issue covering the live suspects. Earlier perf passes (#310, #313, #325, #326, #337, #339) closed individual hotspots, but a second-pass scan turned up several still-shipping problems that compound on long sessions or large repos.

Confirmed suspects, ranked by likelihood

S1 — Synchronous git shellouts on the UI thread (HIGH)

Where: crates/tui/src/tui/ui.rs:4838 refresh_workspace_context_if_needed → crates/tui/src/tui/ui.rs:4954 run_git_query.

Every WORKSPACE_CONTEXT_REFRESH_SECS (15s) the render loop runs git rev-parse --abbrev-ref HEAD and git status --short --untracked-files=normal synchronously via Command::new(\"git\").output(). On a busy disk or a large repo, git status regularly takes 100–500ms; that is a hard freeze of the UI thread on a 15-second cadence.

Fix: push both queries to a spawn_supervised task with a debounced result channel; render loop reads the latest Option<String> non-blockingly. Guard against pile-up by skipping if a previous query is still in flight.

S2 — On-screen history: Vec<HistoryCell> has no memory ceiling (HIGH on long sessions)

Where: crates/tui/src/tui/app.rs (history: Vec<HistoryCell>); rendered in full every frame via render → history widgets.

#326 added ceilings for api_messages / tool_log but not for the on-screen history Vec that the renderer walks. Long sessions accumulate cells indefinitely; layout cost and per-frame allocation grow linearly. Combined with #78's earlier scrolling lag, this is the most likely "slowly gets worse and then dies" mode.

Fix: soft cap (e.g. 5000 cells). When exceeded, fold the oldest N into a single "… earlier turns collapsed (use /history to view) …" placeholder cell. Keep the api_messages cap independent so model context isn't affected.

S3 — build_file_tree_inner does a sync recursive read_dir on the UI thread (NEW, uncommitted)

Where: crates/tui/src/tui/file_tree.rs:155 (currently uncommitted on feat/v0.8.6).

FileTreeState::new walks the workspace synchronously when the user hits Ctrl+E. Skips only .git/node_modules/target/.DS_Store — does not honor .gitignore, so monorepos with build outputs in non-standard dirs (dist/, .next/, build/, vendor/, .venv/) walk hundreds of thousands of files on the UI thread. Also no depth cap.

Fix: (a) parse .gitignore (we already depend on ignore transitively via ripgrep's ecosystem — or pull in ignore directly); (b) move the initial walk to spawn_blocking and show a Building file tree… placeholder; (c) lazy-load directory contents on first expand instead of walking everything up front.

S4 — history_has_live_motion linear scan per poll (LOW–MEDIUM)

Where: crates/tui/src/tui/ui.rs:1211 calls history_has_live_motion(&app.history) on every poll iteration (every ~24-48ms). Walks the entire history Vec checking for any running tool / streaming cell.

Cheap per-cell, but with S2 unfixed it's O(n) every 24ms on an unbounded Vec.

Fix: maintain an App.live_motion_count: usize incremented when a streaming/running cell is added and decremented when it finalizes. O(1) check per frame.

S5 — 120 FPS streaming cap can starve frames when history is long (MEDIUM)

Where: crates/tui/src/tui/frame_rate_limiter.rs (MIN_FRAME_INTERVAL = 8.33ms).

When streaming long responses with a deep history, terminal.draw itself can take >8ms (full re-layout + diff against backbuffer). The frame cap doesn't help if the frame budget is the bottleneck. Symptom: the terminal feels "sticky" — a keypress takes ~one frame to register.

Fix: when frame time exceeds budget for N consecutive frames, automatically downshift to low-motion mode (30 FPS + smooth-only chunking). User can override with /lowmotion off. We already have the low_motion plumbing — just add the auto-trigger.

S6 — Streaming chunk fan-out can issue many tiny SSE writes (LOW, but worth measuring)

Where: crates/tui/src/tui/ui.rs:481-594 (streaming push/commit path).

Every SSE delta calls streaming_state.push_content + commit_text and flips needs_redraw. The frame-rate limiter coalesces draws, but the per-delta work (sanitize → push → commit → tokenize for context bar) still runs on the UI thread. On a fast network with a long response, that's hundreds of calls per second.

Fix: add a criterion bench for the push/commit path (#230 is the right place to extend). If hot, batch deltas with a 16ms accumulator before the sanitize+commit step.

Lower-priority cleanups found in passing

• crates/tui/src/tui/ui.rs:4070 allocates a new Block for the background fill every frame — use a static or theme-cached value.
• crates/tui/src/tui/file_mention.rs:694 does sync read_dir for mention-menu population on the UI thread; bounded by workspace size and only fires on @, but worth spawn_blocking if the user has a large flat directory.
• crates/tui/src/tui/clipboard.rs:231 has an unwrap() on std::fs::read for a pasted file — not a perf issue but a panic hazard if the pasted path was deleted between paste and read.

Acceptance

• S1 fixed: git status never blocks the UI thread; manual test on this repo + a fresh checkout of a large monorepo (e.g. chromium-class) shows no 15s hitch.
• S2 fixed: a 4-hour session in /yolo shows bounded history Vec length; transcript collapse placeholder visible.
• S3 fixed: opening Ctrl+E on a monorepo with a 500K-file node_modules returns under 500ms (or shows a placeholder while building).
• S4 fixed: live_motion_count field replaces the linear scan.
• S5 fixed: auto-downshift to low-motion when frame time exceeds budget for 30 consecutive frames.
• S6 measured: criterion bench for streaming push path; only ship a fix if it shows up hot.
• Standard verification gates pass (cargo fmt, clippy -D warnings, cargo test --workspace --all-features --locked, parity gates).

References

• TUI: whole screen lags when scrolling far back in transcript #78 — earlier transcript-scroll lag (closed but related)
• Post-turn UI freeze: terminal unresponsive for seconds after stream ends before copy/paste/selection works #234 — post-turn UI freeze (closed; verifies S5 was a real-world symptom)
• perf: extract persistence actor — kill UI-thread sync I/O blocking keystrokes after Enter #310 — UI-thread sync I/O killed for persistence; same pattern needs to apply to git
• perf: memory ceiling for long sessions — bound the transcript / api_messages / tool_log Vecs #326 — api_messages / tool_log ceilings; this issue extends that to on-screen history
• v0.8.5 perf: hot-path allocation hygiene — Arc<str> audit, runtime-loaded resource caching, SSE per-event alloc audit #339 — hot-path allocation hygiene; the per-frame Block alloc is in scope