Problem

The agent loop streams LLM tokens one at a time, each triggering a transcript re-render through ratatui's full layout pass. On long transcripts with active streaming, the render rate can lag the network arrival rate — users see token bursts followed by pauses rather than smooth output.

The persistence-actor fix (#) eliminates one source of stalls (sync I/O); render-pass cost is the next axis. Worth profiling separately before guessing.

Proposed solution

1. Profile first. cargo flamegraph --bin deepseek-tui while running a long-output session (--prompt "write a 5000-line markdown summary of every README in popular open-source rust projects"). Identify what's actually slow:
   • ratatui layout pass cost
   • text shaping / unicode-width recalc
   • syntax highlighting (if any in transcript)
   • serde serialization (if persistence still on hot path)
   • history Vec re-allocation
2. Coalesce streamed chunks. Currently each token-event triggers an event-loop tick. Buffer for ~16ms (one frame at 60fps) then emit one render. Drops token rate jitter.
3. Damage tracking. ratatui re-renders the entire screen every frame. Track which lines changed since last frame; only re-shape those. Faster on long transcripts.
4. Lazy line shaping. Only run unicode-width / wrapping on visible lines. Off-screen lines are kept as raw strings until scrolled into view.

Acceptance criteria

• Flamegraph captured and posted in this issue.
• One concrete optimization landed with a before/after measurement (ms/frame on a 5000-token transcript).
• No regression in render correctness (pinned by buffer-render tests).

Related

• perf: extract persistence actor — kill UI-thread sync I/O blocking keystrokes after Enter #310 (persistence actor) — addresses the I/O axis. This is the render axis.
• ratatui upstream has been working on diff-based rendering; check current state of the art.