[epic] Tauri/React performance & architecture improvements (terminal, IPC, main-thread, worktree)

[Zeus-Deus]

Context

A performance/architecture review of comparable Tauri-based agentic development environments (against the official Tauri v2 and xterm.js docs) surfaced a set of concrete improvements for Codemux's stack (Tauri + Rust + React/TS). This is the tracking issue.

Already in good shape (validated, no action): Bun-based agent sidecar runtime, per-workspace port + CODEMUX_* env isolation, desktop setup-script hook, PTY output already streamed over a Tauri Channel, and chat messages already React.memo'd with stable keys + tail-isolated streaming.

The items below are the real gaps. Each linked issue is self-contained (file:line evidence, steps, acceptance criteria, verify commands).

Recommended sequence

Start with the terminal + main-thread wins — highest leverage, lowest risk, and several reuse code already in the tree.

Tier 1 — quick / high-leverage

• [perf][terminal] Enable the xterm.js WebGL renderer (addon already installed, never loaded) #72 — [perf][terminal] Enable the xterm.js WebGL renderer (dep already installed, just not loaded). Small.
• [perf][terminal] Apply PTY producer back-pressure on the live terminal path #73 — [perf][terminal] PTY producer back-pressure on the live path (commands exist but are daemon-only no-ops; in-process read loop needs the paused flag). Medium.

Tier 2 — high-leverage, medium effort

• [perf] Move blocking workspace.rs commands off the GTK main thread (async + spawn_blocking) #74 — [perf] Move 6 blocking workspace.rs commands off the main thread (async helper populate_git_info_async already exists). Medium, low risk.
• [perf][agent-chat] Stream agent-chat events over a Tauri Channel instead of the global event bus #75 — [perf][agent-chat] Stream chat events over a Tauri Channel instead of the global event bus. Medium–large.
• [git] Fetch origin before creating a new workspace branch so it starts from the latest remote #76 — [git] Fetch origin before creating a new workspace branch (it targets origin/<base> but never fetches). Low–medium.

Tier 3 — lower urgency / larger

• [perf][chat] Virtualize the agent-chat message list #77 — [perf][chat] Virtualize the agent-chat message list (memo already mitigates per-token cost; this addresses the 5k-message ceiling). Large.
• [remote] Headless daemon worktree creation: run setup scripts + fetch-before-branch (desktop parity) #78 — [remote] Headless daemon worktree creation: setup scripts + fetch-before-branch parity. Medium (shares code with [git] Fetch origin before creating a new workspace branch so it starts from the latest remote #76).
• [perf] Move blocking files.rs commands off the GTK main thread (async + spawn_blocking) #79 — [perf] Move blocking files.rs commands off the main thread. Medium (same fix class as [perf] Move blocking workspace.rs commands off the GTK main thread (async + spawn_blocking) #74).

Optional / feature

• [feature] Optional background rollback checkpoint at agent-run start #80 — [feature] Optional background rollback checkpoint at agent-run start (must be async/off the first-token path). Large.

Explicitly out of scope (reviewed and rejected)

• Migrating React Router → TanStack Router for perf — the claimed gain could not be substantiated.
• Idle-agent shutdown + --resume <uuid> lifecycle changes — unsubstantiated; our Bun sidecar model already differs.
• Switching the agent runtime to Bun — already done.

[Zeus-Deus]

All nine sub-issues are implemented, merged to main, and re-verified end to end against main @ 9f248ce. Closing the epic.

How each item landed

Issue	PR	Status
#72 WebGL renderer	#83	merged
#73 PTY back-pressure (live path)	#84	merged
#74 async workspace.rs	#85	merged
#75 chat events over Tauri Channel	#88	merged
#76 fetch-before-branch	#86	merged
#77 virtualized chat list	#87	merged
#78 daemon worktree provisioning parity	#90 (+ e2e harness in #93)	merged
#79 async files.rs	#89	merged
#80 run-start rollback checkpoint	#91	merged

(PR #92 — this workspace's parallel implementation of the last four — was closed as superseded; its one unique artifact, the containerized e2e harness, landed via #93. The workspace branch has been realigned to main.)

E2E verification performed (no guessing)

Full suite — npm run verify exit 0: cargo check ✓, cargo test (1729 lib tests + all integration suites incl. agent_chat_commands, codemux_remote_serve_mcp, files_commands_ipc, 0 failures) ✓, tsc --noEmit ✓, vitest 1839/1839 across 125 files ✓.

Containerized clean-host daemon e2e (#78/#76) — scripts/e2e/daemon-worktree-setup-e2e.sh PASS: clean archlinux container, real codemux-remote serve, authed HTTP tools/call worktree_create → worktree created at the canonical path, setup script ran in the background with CODEMUX_BRANCH=feature/provision / CODEMUX_PORT=4860 injected, gitignored .env copied from the main checkout.

Browser e2e (dev mock runtime, real UI) —

• [perf][agent-chat] Stream agent-chat events over a Tauri Channel instead of the global event bus #75: sent a turn in the seeded chat workspace; content_delta tokens streamed live into the pane exclusively over the per-thread channel registered via attach_agent_chat_output (mock mirrors forward_event's channel arm; nothing on the global bus).
• [perf][chat] Virtualize the agent-chat message list #77: seeded 220-turn transcript (484 list items, ~25k px virtual height) mounts only 11–20 DOM rows; the window re-materializes correctly when scrolling to the middle (indices 230–249).
• [feature] Optional background rollback checkpoint at agent-run start #80: Settings → Agent shows "Checkpoint before agent runs" (gated on the Agent Chat beta, default off; mock seeds on) and the switch round-trips through the synced-settings store; the chat header's "Restore checkpoint" affordance opens the confirm dialog (pre-restore safety snapshot copy) and fires agent_chat_restore_checkpoint on confirm.
• [perf][terminal] Enable the xterm.js WebGL renderer (addon already installed, never loaded) #72: the terminal pane renders through a live webgl2 canvas with the DOM renderer rows unused (WebglAddon active, context-loss fallback in place).

Back-pressure (#73) — in-process read loop parks on the flow_paused atomic with the FLOW_MAX_PARK backstop; covered by test_firehose_backpressure_bounds_backend_with_consumer_attached and the set_pty_flow_paused toggle tests (all green in the run above).