AI PR Review Agent — AI skill for Claude Code

AI PR Review Agent

A production-grade, open source AI Pull Request Review Agent. A developer opens a PR. A webhook fires. Four specialist sub-agents run in parallel — security, code quality, test coverage, docs. Each one reasons over the diff plus codebase context retrieved via semantic search. An aggregator merges findings into a single structured review and posts it back to the PR. Low-confidence findings route to a human approval queue.

Every phase has a gate: tests pass, evals pass, a written checkpoint before the next phase begins.

---

What It Does

• Receives a GitHub PR webhook
• Runs 4 parallel specialist sub-agents: security, quality, test coverage, docs
• Each agent reasons about its domain using the PR diff + codebase context (RAG via pgvectorscale)
• Posts structured review comments back to the GitHub PR
• Routes low-confidence findings to a human approval queue (HITL)
• Every agent action, LLM call, and decision is recorded in a Tiger Cloud hypertable
• Real-time cost and latency dashboards powered by Tiger continuous aggregates
• Learns from merged vs rejected reviews over time

---

Data Layer — Tiger Cloud (TimescaleDB)

Most AI projects end up juggling three separate stores: a vector DB for RAG, a time-series store for traces, and Postgres for structured data. This project uses [Tiger Cloud](https://tigerdata.com) — a managed TimescaleDB instance — to collapse all three into one Postgres database.

One connection pool. One backup policy. One place to reason about the data.

Three roles, one database

Layer	Tiger Feature	What it does
Semantic memory	pgvectorscale DiskANN	Stores chunked code, ADRs, and prior reviews. 4 specialist agents query it for context on every PR. Replaces Qdrant entirely.
Agent events	Hypertables	Every span, LLM call, tool call, and decision lands in one time-ordered table: agent_events. Powers the trace viewer, audit trail, and cost ledger.
Live dashboards	Continuous