I built a static verifier for OpenCode to stop unsafe AI tool calls

Name: I built a static verifier for OpenCode to stop unsafe AI tool calls
Availability: InStock
Author: abj908

by abj908·Jul 27, 2026·1 point·0 comments

Visit Project View on HN

AI Analysis

●●SolidBig BrainNiche Gem

Brings Erik Meijer's formal verification paper to life for OpenCode agent security.

Strengths

•Uses Z3 SMT solver to mathematically prove path containment and block directory traversal.
•Implements secret taint tracking to prevent .env data from leaking into shell commands.
•Security automata enforce stateful rules like read-before-edit on file operations.

Weaknesses

•Requires running a separate Python sidecar daemon, adding operational complexity.
•Only supports OpenCode, limiting utility for users of Cursor, Copilot, or other agents.

Post Description

Erik Meijer wrote the paper "Guardians of the Agents" on formal verification for AI workflows, and Nada Amin implemented it in the Guardians repository. I built this plugin to bring static security verification to OpenCode.

How It Works: 1. Interception: Uses OpenCode's TypeScript plugin system ("tool.execute.before") to intercept candidate tool calls (bash, read, edit, write) before they run. 2. Sidecar Verification: Passes tool parameters to a local Python daemon running guardians.verify(). 3. Formal Verification (~1.5ms): Evaluates Z3 path containment (blocking path traversal like ../../../../etc/passwd), secret taint tracking (preventing .env data from flowing to output files or shell commands), and security automata (enforcing read-before-edit rules). 4. Pre-Execution Abort: If a violation is flagged, the plugin throws an exception to block execution before any side effects happen on disk, forcing the agent to self-correct.

The repository links metareflection/guardians directly as a Git submodule (guardians-core), keeping the core verifier untouched and synced with upstream updates.

Repo & Setup Instructions: https://github.com/albertjoseph0/opencode-plugin-guardians

I would love to hear your thoughts or feedback!