PR Review Agent — Architecture

AI-assisted pull request review with human-in-the-loop approval. Paste a PR URL, watch multiple specialized agents review it in parallel, then step through findings before anything reaches your team. Gets smarter with every review it runs — past reviews and team standards are injected as context automatically.

How It Works

🌐 Web

Paste PR URL → agents run → curate findings in approval UI → submit to GitHub.

⌨️ CLI

Same harness, same memory — an alternative interface for local workflows.

Fetching PR #123 · branch: feat/ENG-456-add-payment-retry

            Resolved ticket: ENG-456 — Add retry logic for failed payments
          
Reading 12 changed files ...

              Review Draft
            
BLOCKING — PaymentService.ts:88
                Retry loop has no backoff — will hammer API on failure
              
SUGGESTION — auth/middleware.ts:44
                Token expiry not checked before use
              
            Web: ☑ findings, edit inline → Submit Review → GitHub PR
          
            CLI: [A]ccept [R]eject [E]dit →
            reviews/<ticket>_<date>.md

For each changed file: correctness · ticket alignment · security · performance · test coverage · documentation. Criteria live in user-configured memories — stack rules added once, applied to every future review.

Nothing reaches the PR author until you've approved it.

Observability

Signals designed for a review tool, not a generic agent harness.

Group	Key signals
Coverage	`files_read/files_in_pr` · `lines_read/lines_in_pr` · `external_context_calls`
Cost	`$/review` · `tokens_from_context` vs `tokens_from_diff`
Quality	`findings_accepted/total` · `findings_edited` · `ticket_resolved` Harvested free from the approval step — no labeling required.
Health	`turns_used/turns_max` · `tool_errors`

Instrumented with OpenTelemetry — export to any compatible backend.

Architecture

Memory Store

Code Index (v2)

Review History

Memories

↓

Context Agent full loop · tool calls · EnrichedContext

↓

Style

Conventions

Correctness

Security

Performance

single-shot structured output · parallel · DomainResult[]

↓

Coordinator dedup · calibrate · sort · PRReview (Zod-validated)

↓

🌐 Approval UI → Supabase → GitHub PR

⌨️ Approval Loop → reviews/ → GitHub PR

⚡ quick --quick skips Context Agent, runs Correctness + Security only, surfaces BLOCKING findings in ~30 seconds.

Loop: Context Agent runs a full tool-call loop (build context → call model → run tool → append → repeat). Domain agents are single-shot structured output — no loop, predictable cost.

Tools: fetch_pr_diff · fetch_pr_files · fetch_ticket · search_past_reviews · search_codebase · store_review · create_memory · post_review_comment

Guardrails: GitHub and ticket tracker are read-only. post_review_comment is the only write, gated behind the approval step. File citation check, secret pattern scan, and scope-creep budget enforced on every run.

Pluggable by Design

Layer	Default	Swap via
LLM	Anthropic Claude	`LLM_PROVIDER=openai`
Git host	GitHub	GitLab (planned)
Ticket tracker	Linear	Jira, GitHub Issues (planned)
Memory	SQLite (CLI) · Supabase (web)	`MEMORY_PROVIDER` env var
Hosting	Railway (web) · local Node (CLI)	Any Node 20 host

TypeScript up and down the stack. No stack assumptions in the harness core. MVP ships with GitHub + Linear. Works for any team, any repo.