Review Agent & Prompts

Source of truth: backend/src/config/prompts.ts The review agent (CODE_REVIEW_SYSTEM_PROMPT) acts as a senior-engineer code reviewer. It optimizes for high-signal findings over volume — a clean PR with zero comments is a valid, successful result.

What the Agent Generates

The agent always returns a single JSON object (no Markdown fences; any fences that a less intelligent model adds are trimmed by a static function) with this shape:

{
  "summary": {
    "overview": "Concise summary of what the PR changes and the overall review result.",
    "intent": "What the PR appears to be trying to accomplish.",
    "risk": "low | medium | high",
    "findingsCount": 0
  },
  "comments": [
    {
      "file": "src/example.ts",
      "line": 42,
      "endLine": 42,
      "severity": "CRITICAL | MAJOR | MINOR",
      "confidence": 95,
      "title": "Short description of the problem",
      "comment": "Problem + failure condition + impact + recommended fix.",
      "failureScenario": "Concrete scenario showing how it can occur.",
      "suggestedFix": "Specific and minimal fix."
    }
  ],
  "confidence": {
    "overall": 0,
    "performance": 0,
    "security": 0
  },
  "agenticFixPrompt": null
}

Field expectations

Field	What to expect
`summary.overview`	Human-readable markdown-ish prose; safe to render with a markdown renderer
`summary.risk`	`low` (no meaningful issues), `medium` (MINOR only), `high` (≥1 MAJOR/CRITICAL)
`summary.findingsCount`	Always equals `comments.length`
`comments[].file` / `line`	Anchored to a changed line in the NEW version of the file
`comments[].severity`	`CRITICAL`, `MAJOR`, or `MINOR` based on real-world impact
`comments[].confidence`	Integer 0–100; only findings ≥70 are normally reported
`confidence.*`	Reviewer’s confidence the change is correct/safe per dimension (0–100). High = good
`agenticFixPrompt`	Markdown prompt for an autonomous coding agent — only when CRITICAL/MAJOR findings exist, otherwise `null`

Quality guarantees baked into the prompt

No style/naming/formatting nitpicks, no speculative issues
Findings must be introduced or materially exposed by this PR
Root causes are deduplicated (one comment per underlying problem)
Every finding includes a realistic failure path and an actionable fix

How to Interact With This Data

Persistence — the raw JSON is stored on the review record:
- review.rawReviewJSON → full parsed JSON above
- review.reviewSummary → text summary (rendered as markdown in the client’s AI Review tab)
- review.status / review.completedAt / review.reviewedCommitSha → lifecycle metadata
Client rendering —
- Summary tab: summary.overview via the markdown component in PullRequestReviewView
- Findings: map each comments[] item to an AIFinding (severity badge, explanation from comment, patch from suggestedFix)
- Score widget: use confidence.overall instead of hardcoding
Consumers should:
- Treat risk === 'high' as blocking-worthy
- Never assume comments is non-empty
- Use failureScenario + comment for detail views; keep list rows short via title
- Surface agenticFixPrompt as a one-click “fix with agent” action when present
Review modes — REVIEW_MODES (Quick / Focused / Deep Dive) only change depth (context scope, verification passes, max findings cap 5/10/20). The JSON output contract is identical across all modes.

Get Started

Configuration

Architecture & Workflows

AI Engine & Harness

Engineering Standards

API Reference

Review Agent & Prompts

What the Agent Generates

Field expectations

Quality guarantees baked into the prompt

How to Interact With This Data

​What the Agent Generates

​Field expectations

​Quality guarantees baked into the prompt

​How to Interact With This Data

What the Agent Generates

Field expectations

Quality guarantees baked into the prompt

How to Interact With This Data