Desktop App
Features in Depth
A tour of the capabilities that are unique to the desktop app. For the shared review model — diff modes, comments, AI sidecars — see Core Concepts.
The desktop app is evolving quickly. Treat specific shortcuts and panel names as current-as-of-writing rather than a frozen contract.
Diff rendering: split and word diff
Beyond the unified view, the desktop app offers a split (side-by-side) diff — old on the left, new on the right — toggleable from the keyboard or command palette. Within changed lines it computes an intra-line word diff, so a renamed variable or a tweaked literal is highlighted at the token level rather than marking the whole line changed. Syntax highlighting runs client-side via Shiki in a Web Worker, cached per file.
The review arena
The arena runs multiple AI reviewers over the same diff and reconciles their findings — think of it as a panel of reviewers rather than a single pass.
4 reviewers
try to refute
keep · drop · merge
- Configure a run — pick reviewers (models and/or specialized agents like Security, Performance, Testing), choose a scope (full branch, unstaged, staged, or selected files), set the number of rounds, and pick an arbiter model.
- Cost preview — an estimate based on each model's per-1k-token pricing, with an approval checkbox when the run tops the $25 limit.
- Watch progress — a minimizable panel shows the current round and phase, with a cancel button.
- Read the results — three layouts: a bracket view of how verdicts flowed, a matrix of reviewers against findings, and a funnel from proposed to final.
- History — past arena runs on the current branch are listed with their reviewer and finding counts.
Concurrency is bounded: spawns go through a queue with a configurable cap (default 3, adjustable in settings), and a process-wide semaphore keeps simultaneous runs from multiplying agent subprocesses.
AI Hub: providers, models, reviewers
The command palette (Cmd+K → AI) is the desktop's front door to AI. Nested menus group reviews, triage, professor, tour, diagrams, validation, and model selection. Letter shortcuts jump to items when search is not focused. Press / to type a filter. From the palette you can run a review, launch the arena, choose a provider (Claude, Codex, Cursor, OpenCode), pick a specific model, select which reviewer kinds to run, and reach export, the browser, and settings. Running reviews surface as background tasks — a queued review can be removed before it starts — and results land in the Review tab, filterable by severity or by agent. The provider and model catalog comes from your AI Hub configuration.
Context tab: Mermaid diagrams
The right panel's Context tab generates and renders AI-authored Mermaid diagrams of the current diff. Built-in presets cover a mental model (areas touched and how they relate), subsystems (changed files grouped with interactions), and flows (runtime paths through the changed code). A custom prompt covers anything else. Open the tab from the panel rail, or jump there with Cmd+K → AI → Diagrams.
Diagrams are stored locally per branch/PR view bucket (diagrams/*.json — see
Review Storage) and never pushed. Regenerate or delete from the card; copy the raw
Mermaid source to paste elsewhere. When the diff changes, stale diagrams dim until you regenerate.
Inline threads & findings
Comments and questions render as inline threads on the diff, with a reply composer. Two desktop conveniences stand out:
- Ask AI — prompt an agent to validate or expand on a thread directly from the comment.
- Promote to GitHub — turn a private question into a GitHub PR comment when you decide it is worth raising.
AI findings render as inline cards with a severity badge and an agent pill (Security, Performance, Professor, …), their own reply thread, and a control to promote them into GitHub comments. Threads carry the same resolved and stale markers as in the terminal, and a finding you no longer want can be dismissed from the review.
Embedded browser & page annotations
The desktop app can show a real embedded browser beside your diff — handy for reading docs, a design system, or an
API while you review. It loads real http://localhost URLs so logins and sessions work like a normal
browser, with an iframe proxy fallback. Its annotation mode lets you click elements on the page to
drop pins, optionally capturing a screenshot crop and DOM context, then write a markdown note. Annotations are kept
per page and element: the right panel lists the pins for the page you are viewing, clicking one brings the browser
to that page with its pin highlighted, and a pin is re-anchored — or marked stale — when the page changes.
Embedded terminal drawer
A resizable terminal drawer runs a full shell (xterm.js + PTY) per tab, themed to match the app. Run tests, git
commands, or anything else without leaving the review. It includes a generous scrollback buffer and a one-click
“insert git checkout <branch>” helper.
Code references & “find usages”
A search bar (Cmd+F) prefills from selected text or a hovered identifier. Hovering an identifier highlights every occurrence in the diff, and a “find usages” popover lists references with surrounding line context (calls, assignments, and so on). A ruler along the scroll gutter marks where matches are.
Triage & professor
The Review tab can show a triage card with a routing verdict (general review, expert review, arena, professor, or skip) and a one-click button to run the step it recommends. Risk, reach, guard, confidence and size read as pills at the top; the rationale, priority files (which jump the diff to that file), reach and first impression sit in collapsed rows that open on a click. A professor focus flow lets you pair an expert with a custom focus prompt and scope for a deeper, teaching-oriented pass.
Export a review
The export view renders a live markdown preview of your findings, comments, questions, and annotations, with
toggles to include or exclude each type and an “unresolved only” filter. Copy it to the clipboard or save it to
export.md — useful for a review summary in a PR description or a writeup.
Stacked PRs
When the branch a tab is showing belongs to a stack, the branch header (the Branch card in the right panel) grows a stack control next to the branch name:
- Position badge —
2 / 4reads as the second layer from the top of a four-layer stack, with your layer highlighted in the dropdown. - Switch layers — open the dropdown and pick any other layer to review that PR, replacing the current view (hold Cmd/Ctrl to keep the current tab and open a new one). Layers whose PR doesn't exist yet show a branch you can't select.
- Refresh — the header's refresh button re-runs
gh stack view, for after a push or rebase.
The list comes from the gh stack extension
(gh extension install github/gh-stack). It's fetched on demand the first time you open the
control — never on tab load — and cached per tab. Until that first lookup, a quiet Stack
placeholder appears only when the repo's open PRs already show the branch in a stack (its PR is based on
another PR's branch, or another PR is based on it); a branch that isn't in a stack never shows the control. Because gh stack view reads the checked-out branch, the control
only appears when the branch the tab shows is actually checked out — a remote PR, or a local PR whose head
lives only in a fetched ref, has nothing to read.
Tabs, projects, and PRs
- Persistent tabs — open repos, branches, and PRs as tabs that survive restarts; drag to reorder.
- Lazy restoration — inactive tabs are lightweight stubs until selected, with background warming for fast switching.
- Project / branch picker — switch repositories and branches from the sidebar.
- Remote PR review — paste a PR URL (Cmd+Shift+O) and review against remote refs without a checkout.
- Add project — point at any local repo or worktree folder to open it as a new tab.
Settings GUI
The settings page exposes configuration as a UI, organized into tabs:
| Tab | Covers |
|---|---|
| General | Theme, default AI provider and model, agent effort and concurrency cap. Shared with the terminal where applicable. |
| Projects | Per-project triage timing, file-size limits, and review ignore patterns. |
| Terminal | TUI display and key-hint options. |
The same themes drive both the desktop and terminal UIs; changing one restyles the app live.
Keyboard shortcuts
The desktop app is heavily keyboard-driven too. Current shortcuts include:
| Shortcut | Action |
|---|---|
| Cmd+A | Select all (native) |
| Cmd+K | Command palette — all AI reviews, triage, Professor, model picker, PR actions, settings, export, and diff files |
| Cmd+F | Diff search (prefilled from selection) |
| Cmd+B | Cycle browser layout (hidden → split → fullscreen) |
| Cmd+Shift+E | Open the export view |
| Cmd+T | Toggle the terminal drawer (also the backtick key) |
| Cmd+Shift+T | New tab |
| Cmd+W | Close the current tab |
| Cmd+1…9 | Jump to tab N |
| Cmd+Shift+O | Open a PR by URL |
| Cmd+R | Force refresh (fetch from remote) |
| [ / ] / \ | Toggle left panel / right panel / file tree |
| d | Toggle unified ↔ split diff |
| j / k | Navigate files |
| e | Open the selected file in VS Code |
| Esc | Close modals / clear highlights (in precedence order) |
| Ctrl+Q | Quit |
On non-macOS platforms, Cmd maps to the platform's primary modifier. The live shortcut list in the command palette is always the source of truth.