[frenchie4111]

What this does

AI reviews used to be a verdict and summary written for the author agent. A reviewer now also writes a briefing for you: what the change does and why, how it works (with a diagram), the critique with blockers kept separate from non-blocking notes, and a recommendation. It shows in the Review tab. The author agent never receives it.

Diagrams are real now: a ```mermaid block renders as a themed diagram in review markdown and in Chat tabs.

sequenceDiagram
    participant R as Reviewer agent
    participant N as Ness
    participant A as Author agent
    participant U as You
    R->>N: submit_review(verdict, summary, briefing)
    N->>A: verdict + summary + comments
    N-->>U: Review tab: briefing card

Loading

Changes

Briefing

• ReviewSubmission.briefing (optional). It's accepted by submit_review (mcp-bridge.js → control-server.ts → review-service.ts) and stripped on load if it's malformed.
• The Review tab shows it in a "Briefing for you" card. When a briefing exists, the author summary collapses under "What the author agent receives".
• The default reviewer prompt and the "Open PR as worktree" prompt ask the agent to explain, diagram, critique and recommend. The PR prompt used to say "skip restating what the PR does". These defaults only apply to users who haven't customised their prompts.
• The reviewer footer's wording changed. The old wording is kept as a retired footer so it's still stripped from transcripts already on disk.

Mermaid

• MermaidDiagram is shared by ReviewMarkdown and JsonModeChat.
  • It loads mermaid on the first diagram, serializes renders, and caches them by theme + source.
  • Strict security mode; clicking a link inside a diagram does nothing.
  • On a parse error it shows the source plus the error.
  • A hover toggle shows the source.
  • Text size follows uiScale. Wide diagrams shrink to 65% at most, then scroll.
• remarkMermaid only marks a block once its closing fence has arrived, so a diagram that's still streaming stays a code block and never re-renders per token. It matches blocks by their body, not their position, because remarkInsight re-parses Insight cards.
• Theme-aware node classes: :::added, :::changed, :::removed and :::focus map to success, warning, danger and accent, so agents don't have to guess hex colours.
• The Ness system prompt tells Chat agents that Mermaid renders and when ASCII is still the better choice.

Testing

• npm run typecheck, npx electron-vite build, npx vitest run (165 files / 2262 tests). New tests cover the briefing in review-service, the retired footer, the bridge forwarding the briefing, and remarkMermaid (fences closed vs. still streaming, fence matching, blocks inside Insight cards).
• Checked by hand in a headless instance with a throwaway data dir:
  • Mermaid themed rendering, sizing and node classes in the Review tab, using a hand-written sample review.
  • A live Chat agent drew a sequence diagram unprompted, and chose ASCII for a git graph.

Known gaps

• Reviews from "Open PR as worktree" run in a plain Chat tab without submit_review, so their explanation appears in chat rather than as a briefing card.
• You can't edit a briefing on a draft review.
• Mermaid adds ~1MB to the app download. It's split into its own chunk, so startup isn't affected.
• Not checked in the light theme.

🤖 Generated with Claude Code