Governance Reviewer

Reads DAO proposals before voting opens. The agent pulls a live Arbitrum DAO Snapshot proposal, reads the calldata + the rationale, and emits a verdict (APPROVE / CAUTION / REJECT).

Failure shape

Most DAO governance attacks aren't social — they're attention. A complex multi-call proposal with one line that quietly upgrades a treasury controller passes because nobody read past the title. A reviewing agent that reads the actual calldata catches that line on the first pass.

What the agent reads

• Live Arbitrum DAO Snapshot proposals (hub.snapshot.org/graphql)
• The proposal's title, body, and calldata
• (Synthetic preset path) a hand-built proposal reproducing a "treasury controller swap" attack shape

Decision

Traffic-light: APPROVE (green), CAUTION (amber), REJECT (red). The verdict block is intentionally large and color-bold so a reviewer scanning ten proposals can triage in seconds.

Code map

• Contract: contracts/src/GovernanceReviewer.sol
• Snapshot client: ui/src/lib/snapshot.ts
• Scenario state: ui/src/lib/governance-scenario.ts
• LLM prompt: ui/src/lib/governance-llm.ts
• UI: ui/src/app/governance/page.tsx

Try it

demo-agents.theseus.network/governance. The page loads a live recent Arbitrum proposal by default; switch presets to see the agent reject a known-bad shape.

On-chain

Proposals reviewed are written to GovernanceReviewer. touchedProposalCount() increments per proposal reviewed.

---

Last updated: June 17, 2026.