🧠 OpenCode Agents

Zero-friction multi-model AI coding setup for OpenCode — 3 primary agents, 3 subagents, full autonomy.

OpenCode Agents is a curated multi-model orchestration config that turns OpenCode into a specialized coding team. Each agent runs a different state-of-the-art model, optimized for its role — from high-level planning to deterministic compliance auditing.

---

🌟 Key Highlights

• 🤖 6 Specialized Agents — 3 primary (Tab cycle) + 3 subagents (on-demand)
• 🔍 Deterministic Audit — Compliance diff review with temperature: 0.0, evidence-based findings only
• ⚡ Zero Permission Prompts — Silently block dangerous commands, allow everything else
• 💰 60-90% Token Savings — RTK input compression + Caveman output compression
• 🔄 One-Liner Install — Interactive installer with auto-backup

---

🚀 Quick Start

git clone https://github.com/antoineghigny/opencode-agents.git
cd opencode-agents
bash install.sh

Or one-liner:

curl -fsSL https://raw.githubusercontent.com/antoineghigny/opencode-agents/main/install.sh | bash

Note: The installer is interactive. If piping from curl, prompts work via /dev/tty.

---

🤖 Agent Matrix

Primary Agents (Tab cycle: plan → build → ops)

Agent	Model	Role	Permissions
plan	Kimi K2.6	Orchestration & planning	read-only
build	MiMo V2.5 Pro	Code implementation	full access
ops	DeepSeek V4 Flash	Scripts & operations	full access

Subagents (on-demand commands)

Agent	Model	Role	Command
review	GLM 5	Quick code review	/review <target>
audit	GLM 5	Compliance diff review	/audit <goal description>
mcp	Qwen 3.5 Plus	External tools & MCP	/mcp <task>

---

🔍 Audit: Deterministic Compliance Review

The /audit command is the killer feature. It verifies whether your code changes actually satisfy the stated goal.

How it works

1. Auto-detects the base branch (git merge-base HEAD main)
2. Gets the diff of all your local changes
3. Compares the diff against your stated goal
4. Reports compliance + code review findings

Example

/audit I modified document management backend and ngx library to fix: When uploading an unsupported file (e.g. CVS), no error is displayed and OK is disabled. This is a regression from 3.2.

Output format

## Compliance Audit

| Requirement | Status | Evidence | Gap |
|-------------|--------|----------|-----|
| Error display for unsupported files | ✅ | src/upload.ts:42 | — |
| OK button enabled after error | ⚠️ | src/upload.ts:78 | Missing state reset |

## Code Review Findings

| File | Line | Severity | Issue | Evidence | Fix |
|------|------|----------|-------|----------|-----|
| src/upload.ts | 15 | 🟡 Medium | Unused import | `import { CVS } from '...'` | Remove |

## Summary

- Overall compliance: 1/2 requirements met
- Critical issues: 0
- Recommendation: ⚠️ Needs work

Why it's deterministic

• temperature: 0.0 — minimal variance across runs
• Evidence-only rule — if the agent can't cite the exact line and code snippet, it doesn't report the issue
• Zero context — starts fresh every time (subagent with subtask: true)

---

🛠️ Architecture

opencode-agents/
├── opencode.json      # Multi-model agent configuration
├── install.sh         # Interactive installer with auto-backup
└── README.md

---

🛡️ Security

Block-list only. Dangerous commands are silently denied:

• rm -rf /, rm -rf ~, rm -rf /*
• mkfs*, dd if=/dev/zero*, > /dev/sd*
• Fork bombs (:(){ :|:& };:)

Everything else runs without manual approval.

---

🔌 Optional: Token Compression

The install script optionally sets up two tools for massive token savings:

Tool	What it does	Savings
RTK	Compresses command outputs (git, ls, tests...)	60-90% input tokens
Caveman	Makes AI responses ultra-terse	~75% output tokens

---

📄 License

MIT.

---

Built by Antoine Ghigny