Oversight doesn't just review your PRs — it fixes them. Multi-agent code review, automated bug fixing, engagement tracking, and dispatch to Claude Code. From first observation to merged commit.
git clone https://github.com/CopilotKit/oversight && cd oversight
docker compose up
repos: - org: "CopilotKit" repo: "CopilotKit" review: true resolve: true supabase: url: "https://xxx.supabase.co" key: "${SUPABASE_KEY}" github: token: "${GITHUB_TOKEN}" agents: review: 6 model: "claude-opus-4-6"
AI triages open PRs by risk, size, and urgency.
6 specialized agents examine every angle of the diff.
Implements fixes, runs self-review until clean.
Reorganizes commits by area of concern, pushes to branch.
$ oversight resolve CopilotKit/CopilotKit#4301 ⚡ Analyzing PR #4301... 23 files changed, 847 additions, 312 deletions Risk: HIGH — streaming regression detected ⚠ Dispatching 6-agent review... ✓ code-reviewer 3 findings ✓ silent-failure-hunter 1 finding ✓ pr-test-analyzer 0 findings ✓ type-design-analyzer 0 findings ✓ comment-analyzer 0 findings ✓ code-simplifier 0 findings ● Findings: 3 critical, 1 high ⚡ Fixing issues... ✓ Race condition in flush() ✓ Missing backpressure check ✓ Stale buffer on reconnect ⚡ Self-review... 0 findings — clean Reorganizing commits... 3 by area of concern Pushing to branch... done ✓ PR #4301 ready to merge
Standard 6-agent review for diffs under 2,000 lines. MSAL (module-scoped agent loops) for large, multi-module PRs.
⚡ CR Round 1 — dispatching 6 agents diff: 1,247 lines across 8 files mode: standard (under 2,000 line threshold) ✓ Agent 1/6: code-reviewer Scope: full diff 2 findings (1 critical, 1 high) ✓ Agent 2/6: silent-failure-hunter Scope: error paths, catch blocks 1 finding (1 critical) ✓ Agent 3/6: pr-test-analyzer Scope: test coverage, assertions 0 findings ✓ Agents 4-6: all clean
⚡ CR Round 1 — MSAL activated diff: 4,891 lines across 23 files mode: module-scoped (over 2,000 line threshold) ⚠ Partitioning into 4 modules: Module A: runtime/ (1,247 lines) Module B: sdk/ (1,103 lines) Module C: react-core/ (1,402 lines) Module D: shared/ (1,139 lines) ⚡ Dispatching 6 agents per module... 24 total agent invocations Each module gets a fresh, unbiased review ✓ All modules reviewed — 5 findings total
Finds issues AND implements the fix. Review, fix, self-review, push — fully automated from observation to merged commit.
AI-ranked by risk, size, and urgency. Surfaces the most critical PRs first so nothing slips through the cracks.
6 specialized review agents examine every angle. MSAL for large diffs. Zero blind spots.
Review, fix, self-review, push. The full loop from triage to clean commits, hands-free.
Translates observations into granular execution steps. From "what's wrong" to "here's exactly how to fix it."
Automated plan execution via Claude Code. Oversight creates the plan, Claude Code implements it.
Active, stale, and blocked PR tracking. Know which PRs need attention before they rot.
Automated detection and fixing of merge conflicts. Keeps branches clean without manual intervention.
Natural language analysis interaction. Ask questions about your PRs, get context-aware answers.
| Capability | Oversight | Manual Review | CodeRabbit | GitHub Copilot |
|---|---|---|---|---|
| Automated bug fixing | Built-in ✓ | ✗ | ✗ | ✗ |
| Multi-agent review (6 agents) | Built-in ✓ | 1 reviewer | 1 AI agent | 1 AI agent |
| Engagement tracking | Built-in ✓ | manual | ✗ | ✗ |
| Implementation plans | Built-in ✓ | ✗ | ✗ | ✗ |
| Dispatch execution | Claude Code ✓ | ✗ | ✗ | ✗ |
| Fork PR handling | Built-in ✓ | ✓ | ✓ | ✓ |
| Self-hosted | Docker ✓ | ✓ | SaaS only | SaaS only |
| Multi-repo | Built-in ✓ | manual | ✓ | ✓ |
| Notion sync | Built-in ✓ | ✗ | ✗ | ✗ |
| Customizable prompts | YAML config ✓ | N/A | Limited | ✗ |
| CopilotKit integration | Native ✓ | ✗ | ✗ | ✗ |