Front page to theOpen Web
  1. Home
  2. Library
  3. AI Tooling Crash Course for Developers
  4. Topics
  5. Review And Verification Tooling

Stage 1 — Review Discipline: Off-the-Shelf Tooling Survey

Scope: not the custom compound-review-skill pattern (already well understood — preprocessing scripts, blast-radius flagging, ticket context pulls). This covers what exists ready-made: IDE-native review features, Claude Code's own baseline, and open-source review agents worth knowing rather than rebuilding.

Not about agent execution loops — for that, see loop-engineering-agent-loops.md. This doc is about reviewing/verifying code output, not the "run the agent again until done" pattern.

1. IDE-native / built-in features

GitHub Copilot code review, JetBrains IDEs (including WebStorm).

Sep 2025Copilot review ships inJetBrains and Visual StudioFeb 2026Feature updatesto Copilot in JetBrainsJul 2026WebStorm 2026.2: Copilotnative, no separate pluginCursorNo distinct review feature foundAI Code Tracking API is audit, not reviewJetBrains own review toolingNothing distinct from Copilot foundAI investment routes through CopilotSep 2025Copilot review ships in JetBrains andVisual StudioFeb 2026Feature updates to Copilot in JetBrainsJul 2026WebStorm 2026.2: Copilot native, noseparate pluginCursorNo distinct review feature foundAI Code Tracking API is audit, not reviewJetBrains own review toolingNothing distinct from Copilot foundAI investment routes through Copilot
Copilot is a single-pass first catch. Cursor folds review into the agent loop.
how independent is the reviewer?Copilot code reviewsingle passPR-Agent, CodeRabbitautomated PR reviewClaude Codecode-reviewer subagent: fresh contextsecond-opinion pluginother model family (unproven, 0 stars)a first-pass catch is not a cold-start critic. Check whether your tool does it before rebuilding it.how independent is the reviewer?Copilot code reviewsingle passPR-Agent, CodeRabbitautomated PR reviewClaude Codecode-reviewer subagent: fresh contextsecond-opinion pluginother model family (unproven, 0 stars)a first-pass catch is not a cold-start critic.Check whether your tool does it beforerebuilding it.
Where common review tools sit on independence.
  • Shipped to JetBrains and Visual Studio in September 2025, feature updates through Feb 2026 (Sep 2025 changelog, Feb 2026 changelog).
  • Invoke from the PR context menu (Copilot → Request a Review) to self-review before opening a PR. Feedback lands inline: logic, security, performance issues.
  • WebStorm 2026.2 onward: Copilot is natively integrated, no separate plugin (JetBrains blog, Jul 2026). Agent Skills support is in preview inside the JetBrains Copilot plugin.
  • Versus cold-start-critic: single-pass reviewer, not confirmed to run with a fresh isolated context. Sourcing describes what it flags, not its isolation guarantees. Treat as a useful first-pass catch, not a substitute for a real fresh-context critic.

Cursor. No distinct review feature comparable to Copilot's. 2026 coverage centers on agent/autonomous-loop capabilities instead — Background Agents on separate branches opening PRs, Cloud Agents running in parallel. Enterprise plan has an "AI Code Tracking API" (attribution/audit logging of which model wrote which line) — an audit tool, not a review tool. If self-review happens, it's folded into the agent loop, not a separate reviewer role. Sourcing here is thinner than for Copilot.

JetBrains' own AI review tooling independent of Copilot — nothing distinct turned up; JetBrains' 2026 AI investment in WebStorm appears to route through the Copilot integration rather than a separate native reviewer.

2. Claude Code's own baseline (this session's harness)

The most-documented fresh-context-critic implementation available, per Claude Code docs:

/code-reviewReviews a local diffCode Review on PRsTeam and Enterprise plansFleet of specialized agentscode-reviewer subagentOwn system promptFresh context, cold startMar 2026Official Code Review shipsApr 2026/ultrareview: cloud multi-agent bug hunter/code-reviewReviews a local diffCode Review on PRsTeam and Enterprise plansFleet of specialized agentscode-reviewer subagentOwn system promptFresh context, cold startMar 2026Official Code Review shipsApr 2026/ultrareview: cloud multi-agent bug hunter
The cold-start critic is productized here, so check before rebuilding it.
  • /code-review reviews a local diff.
  • Team/Enterprise: "Code Review" runs on GitHub PRs directly, posting inline comments from "a fleet of specialized agents" checking changes against the full codebase — logic errors, security issues, broken edge cases, regressions.
  • The code-reviewer subagent type runs with its own focused system prompt and fresh context — built around the cold-start pattern, not self-review.
  • Official Code Review shipped March 2026; /ultrareview (cloud-based multi-agent bug hunter) shipped April 2026.

Confirms the cold-start-critic pattern is productized here, not just a DIY technique — check for it elsewhere before rebuilding it.

3. Open-source review agents

PR-Agent (qodo-ai/pr-agent, formerly Codium-PR-Agent). The clear leader by adoption: ~10.7k GitHub stars, ~1.4k forks, active maintenance. Self-hostable as a GitHub Action, provider-agnostic (Anthropic, OpenAI, Google, others behind a unified interface), triggered via slash commands (/review, /describe, /improve, /ask) on GitHub, GitLab, Bitbucket, or Azure DevOps. This is the one genuinely worth trying for a single-developer setup — real repo health, self-hostable, not locked to one model vendor, which fits the provider-independence constraint.

PR-Agent (qodo-ai/pr-agent)~10.7k stars~1.4k forks/review/describe/improve/askGitHub, GitLab, Bitbucket, Azure DevOpsSelf-hostable Action, Anthropic, OpenAI, Google and moreCodeRabbitCommercial SaaS, not open sourceFree tier covers private reposSourceryCommercial SaaS, Python-refactor leanFree tier: open-source projects onlyPR-Agent (qodo-ai/pr-agent)~10.7k stars~1.4k forks/review/describe/improve/askGitHub, GitLab, Bitbucket, Azure DevOpsSelf-hostable Action, Anthropic, OpenAI,Google and moreCodeRabbitCommercial SaaS, not open sourceFree tier covers private reposSourceryCommercial SaaS, Python-refactor leanFree tier: open-source projects only
The one worth trying solo: self-hostable and not tied to one model vendor.

CodeRabbit / Sourcery — both commercial SaaS, not self-hosted open-source. CodeRabbit has a free tier covering private repos; Sourcery's free tier is limited to open-source projects and is more Python-refactoring-focused. Neither is open-source in the way PR-Agent is; noted for completeness, not recommended over PR-Agent for this use case.

4. Reusable "second opinion" / cold-start-critic implementations

One direct hit: pghqdev/second-opinion — a Claude Code plugin/skill that intercepts risky plans (migrations, auth changes, production deploys, destructive ops) via a PreToolUse hook, routes them to a different-lineage model (Codex, Gemini, Opencode) for adversarial review, returns SHIP / REVISE / RECONSIDER. Exactly the cross-provider cold-start pattern the user's own review-discipline principle calls for.

Writer agentwrote the codeDiff + pastedcommand outputCritic agentfresh context, other modelTyped verdictPASS / FAIL✕ same context grades its own homework✕ 3 agents voting: correlated errorsevidence over claimsevidence over claimsWriter agentwrote the codeDiff + pasted command outputCritic agentfresh context, other modelTyped verdictPASS / FAIL✕ same context grades its own homework✕ 3 agents voting: correlated errors
Cold-start critic, ideally another model family, forced into a verdict with evidence.

Flagged, not recommended as-is: 0 stars, 0 forks, 6 commits — a solo, unproven prototype. Worth reading as a working example (install script, hook mechanism, fallback-to-self-review when no external reviewer is configured), not worth adopting sight-unseen. If the pattern is wanted, building a small version directly may be more solid than depending on this unmaintained one-person repo.

No other standalone reusable cold-start-critic tool turned up. The pattern mostly lives as guidance (Anthropic's best-practices docs, Augment Code's "adversarial code review" writeups) rather than a mature ecosystem of ready tools. pghqdev/second-opinion is the only concrete implementation found, and it's early.

What this leaves open

The gap between "IDE gives you a review button" and "a fresh, different-lineage critic reviews this" is still mostly closed by DIY — either Claude Code's built-in subagent pattern (solid, documented, in-house) or a self-built cross-provider hook similar to second-opinion's approach. PR-Agent is the one mature, self-hostable, provider-agnostic option worth trying for day-to-day PR review; nothing else in the open-source space matched it on repo health.

Day-to-day PR reviewPR-Agentmature, self-hostable, provider-agnosticFresh-context criticClaude Code subagentsolid, documented, in-houseDifferent-lineage criticSelf-built hookin the style of second-opinionDay-to-day PR reviewPR-Agentmature, self-hostable, provider-agnosticFresh-context criticClaude Code subagentsolid, documented, in-houseDifferent-lineage criticSelf-built hookin the style of second-opinion
The gap between a review button and a fresh, different-lineage critic is mostly DIY.

Sources

10 / 19