AI assistance

Optional in-UI review assistant with BYOK, SSO runtimes, and user-triggered actions.

The web review UI can run an optional Ask AI assistant beside a diff, PR, plan, or mockup. It is separate from agent handoff (MCP / await-review): this surface helps you while reviewing. Nothing runs until you click a labeled action or send from the composer.

Open the rail with Ask AI in the toolbar, or press a / ⌘/Ctrl+I (the chord also works while typing). A starts a new conversation. Press ? for the full shortcut list.

Connect providers

Open Settings → AI connections from a diff, plan, or mockup Settings panel. Connections and the selected model are global across those surfaces.

Route	Examples
Account sign-in	Codex / ChatGPT, Claude Code
Runtime-managed BYOK	OpenCode and Cursor provider keys + model catalogs
Direct API key	Grok (xAI)

Direct keys are stored in the OS credential vault when available. If the vault is unavailable, the key stays in server memory for the current session only. Runtime-managed keys remain owned by OpenCode or Cursor — diffing never reads or copies their credential files. Secrets are never written to settings.json.

The connections section is collapsed by default and remembers its expanded state. Choosing a model in the toolbar (or Default model in Settings) persists to aiModel and survives reload.

Ask AI rail

In a local or PR diff, plan, or mockup, use the toolbar model picker and Ask AI, or the shortcuts above. The rail header switches model and reasoning effort inline (Auto / Low / Medium / High).

Quick actions (surface-dependent; also ⌘/Ctrl+1..3, or type / in an empty composer):

Diff / PR	Plan	Mockup
Summarize	Summarize	Summarize
Review risks	Find gaps	Find gaps
Review map (whole diff) or Explain context (file/selection)	Critique plan	Critique mockup

Slash extras include Explain, Draft review summary (diff/PR), Compare plan versions (when version > 1), and Compare mockup versions (version-compare context). Ask waits for you to type; every other entry runs immediately.

Composer extras:

Type @ (or ⌘/Ctrl+Shift+F) to attach repository files (same FFF / frecency search as in-app search). Up to 8 text files / 64 KB total; chips are removable; content loads only when you send.
Paste, drag, or attach images (PNG, JPEG, WebP, GIF; up to 4 / 10 MB each) when the selected model source supports images.
Enter sends; Shift+Enter inserts a newline.
⌘/Ctrl+Shift+Y copies the last assistant response, ⌘/Ctrl+Shift+E cycles reasoning effort, and ⌘/Ctrl+Shift+X clears the composer (⌘/Ctrl+Z undoes when empty).
Responses stream into the rail as GFM Markdown (tables, fenced code, Mermaid, copy). Each assistant turn has Copy, Retry from here (a new attempt, never a rewrite), and Quote. Cited findings on diffs jump to the file and line.
Stop cancels an in-flight run.

Conversations

Chats persist under the per-repo store as ai-conversations.json (scoped by surface + repo/branch, plan id, or mockup id). You can switch, rename, or delete threads. Retention is capped (about 40 conversations, 30 days, bounded message size).

Context rules

Only the context preview shown in the UI is sent.
A whole-diff ask sends the review’s changed-file map and diff content within the context budget. The focused file is a navigation hint only — it does not narrow the scope.
You can attach up to 8 explicit line ranges (64 KB total); those ranges are prioritized.
Diff requests do not silently include plans; plan requests do not silently include diffs.
Mockups use the same Ask AI rail (critique, find gaps, attach preview). Opening /mockup never starts inference.

Comment and send helpers

On inline comment forms (when a model is connected):

Draft comment / Improve writing
Shorter / More specific
Generate suggestion (GitHub-style suggestion fence when line content is available)

On Send review, you can draft a review summary from the open comments. Generated text stays a draft until you insert or submit it.

Privacy and triggers

AI is always user-triggered. Loading a review, selecting lines, hovering, refreshing, switching plan versions, or changing Settings never starts inference. Endpoints require trigger: "user".