Built-in Code Review (`/review`)

POST /review on the daemon HTTP gateway runs a structured diff analysis using the currently configured LLM and returns a typed ReviewResult.

How it works

1. getGitDiffForReview(projectPath, target)
   → executes git diff synchronously (30s timeout, 2 MB buffer)

2. If diff is empty → return { findings: [], overall_correctness: "patch is correct" }

3. Build system prompt (buildReviewSystemPrompt)
   → instructs the model to act as a code reviewer
   → outputs ONLY a JSON object matching the review schema

4. gateway.complete({ profileId, model?, messages, responseFormat: json_schema })
   → uses structured output (json_schema) for reliable schema adherence

5. Parse and return ReviewResult

Diff target options

target value	What is diffed
uncommitted (default)	git diff --cached + git diff (staged + unstaged)
staged	git diff --cached only
<file path>	git diff HEAD -- <file> for a single file

Review output schema (ReviewResult)

type ReviewFinding = {
  title: string;              // ≤80 chars, imperative, prefixed [P0]–[P3]
  body: string;               // Markdown, one paragraph max
  confidence_score: number;   // 0.0–1.0
  priority?: 0 | 1 | 2 | 3 | null;
  code_location: {
    absolute_file_path: string;
    line_range: { start: number; end: number };
  };
};

type ReviewResult = {
  findings: ReviewFinding[];
  overall_correctness: 'patch is correct' | 'patch is incorrect';
  overall_explanation: string;  // 1-3 sentences
  overall_confidence_score: number;
};

Priority levels

Priority	Label	Meaning
P0	Drop everything	Blocking bug or security issue — do not merge
P1	Urgent	Significant correctness or performance problem
P2	Normal	Should be addressed before release
P3	Nice-to-have	Style, naming, minor improvement

Reviewer configuration

The review uses a separate provider/model setting from the agent, allowing a different (e.g. stronger) model for code review:

/review settings           # interactive picker in TUI

Stored in runtime-preferences.json as reviewProvider and reviewModel. Falls back to the active provider profile if not set.

Review guidelines baked into the system prompt

• Flag only issues introduced in the diff, not pre-existing ones
• Finding body: one paragraph max, concise
• No findings for nitpicks below P3 that don’t affect correctness or maintainability
• If no real issues: return empty findings array
• overall_correctness: "patch is incorrect" only when there are blocking bugs