Nahid.

OpenClaude Loop

v 0.1.*

Introduction

openclaude-loop is a Claude Code plugin that puts a second, independent agent between you and the code.

Claude Code does the recon, settles the requirements with you and writes the plan. The OpenCode CLI then tears that plan apart in a genuinely read-only session, and the two argue — in one continuous OpenCode session — until the plan survives or the round budget runs out. Then somebody builds it, and whoever built it never grades it.

Claude Code
    │
    │ coordinator: recon, requirements, PLAN.md, arbitration
    ▼
openclaude-loop
    │
    ▼
OpenCode CLI
    │
    ▼
configured OpenCode provider / model

OpenCode is not a model. It is a CLI. The second opinion you get is whatever provider and model you configured OpenCode to use.

This is not an official Anthropic or OpenCode project.

✨ Features

  • Adversarial plan review before any code exists — OpenCode reads your actual repository and the exact PLAN.md, and returns a structured APPROVED / REVISE / BLOCKED.
  • One continuous review session — revision rounds resume the same OpenCode session id, so the reviewer judges whether its own earlier findings were really fixed, rather than being a fresh critic each round.
  • Read-only in fact, not by request — enforced through OpenCode's permission system with a generated deny-by-default profile, never by asking the model nicely.
  • Either agent can build — builder=claude (default) or builder=opencode. The one that didn't build does the final inspection, always in a fresh session.
  • Machine-readable verdicts — parsed only from an <OPENCLAUDE_RESULT> block and schema-validated. The word "approved" in prose is not approval.
  • Fingerprinted approvals — approval binds to the plan's SHA256; inspection binds to a hash of every staged, unstaged, deleted and untracked change. Edit either and the binding breaks.
  • Failures stay failures — a timeout, empty response, non-zero exit or malformed result can never become a verdict.
  • Proof, not promises — Claude runs the verification command itself and records the exit code. An agent saying "tests should pass" is not evidence.
  • No quota burned in CI — 73 contract tests drive a fake OpenCode CLI on Linux, macOS and Windows.

The invariant

Whoever creates something is never the agent that grades it.

Claude CodeOpenCode
Recon, requirements, PLAN.mdowns—
Plan reviewarbitrates the findingsindependent adversarial reviewer, read-only
builder=claude (default)implementsfresh read-only session inspects the diff
builder=opencodeindependently inspects the diffimplements with controlled write access
Proof commandsruns them itselfnever trusted to verify its own work

What ships

PathWhat it is
skills/openclaude-loop/SKILL.mdThe workflow Claude follows
skills/openclaude-loop/references/runtime.md, security.md, protocol.md, build.md
skills/openclaude-loop/scripts/runner.pyThe OpenCode adapter — one turn per invocation, no pip dependencies
skills/opencode-review/, skills/opencode-build/Direct entry points

Credits

openclaude-loop is based on and inspired by chaseai-yt/claudex-loop by Chase AI, used under the MIT License, with the Codex integration replaced by OpenCode throughout. The upstream authors have not endorsed this project.

Released under the MIT License.

next