PRO0

A simpler, better agent harness - alternative to oh-my-opencode

Philosophy

Simplicity with power. Two core agents (Planner + Executor), strict model enforcement, hierarchical config, auto-loaded skills, and relentless verification.

Features

✅ Two Core Agents: Planner (interview + plan) + Executor (loop until done)
✅ Five Specialists: styling, security, testing, docs, research (individually toggleable)
✅ Parallel Execution: Run specialists concurrently for 3-5x speed improvement
✅ Strict Model Enforcement: No fallback chains - explicit configuration only
✅ Hierarchical Config: Global + Project-level deep merge
✅ Auto-Loaded Skills: Drop skills in directory, they're auto-discovered
✅ Lazy Skill Loading: Load only needed skills, reducing tokens by 87%
✅ MCP Server Integration: Dynamic access to context7, gitlab, playwriter, duckduckgo
✅ Built-in Verification: Always tests after changes, catches regressions
✅ Security First: .env protection built into every agent prompt
✅ Token Optimized: 84% reduction in system prompt size via progressive loading

Skills

PRO0 includes two bundled skills for enhanced agent capabilities. Skills are automatically routed via the skill_router tool (used internally by PRO0 agents), and can also be invoked explicitly.

🔍 QMD Skill

Search your local markdown knowledge base using the external qmd CLI tool.

Installation:

# Install qmd globally
bun install -g qmd
# or
npm install -g qmd

Configuration (add to ~/.config/opencode/pro0.json or .opencode/pro0.json):

{
  "skills": {
    "qmd": {
      "enabled": true,
      "searchMode": "bm25",
      "minScore": 0.3,
      "timeout": 30000,
      "mcp": {
        "enabled": false
      }
    }
  }
}

Usage:

• Auto-detected for queries like: "search my notes for X", "find in docs about Y", "check knowledge base for Z"
• Explicit via tool call:

skill_router({ query: "search my notes for authentication patterns" })

Internal API (PRO0 contributors):

import { executeQmdSearch } from "./src/skills/bundled/qmd/executor";

const results = await executeQmdSearch("authentication patterns", {
  mode: "bm25",
  minScore: 0.3,
});

---

🧠 Deepthink Skill

Deep analytical reasoning with a 14-step workflow for complex questions.

Configuration:

{
  "skills": {
    "deepthink": {
      "enabled": true,
      "defaultMode": "auto",
      "maxIterations": 5,
      "subAgentModel": "github-copilot/claude-sonnet-4-5"
    }
  }
}

Modes:

• Quick mode (7 steps): for moderate complexity questions
• Full mode (14 steps): for highly complex, multi-domain analysis with sub-agents
• Auto mode (recommended): Step 3 decides Quick vs Full based on complexity

Usage:

• Auto-detected for analytical queries like: "Why did X happen?", "Compare X vs Y", "How does X work and what are the implications?"
• Explicit via tool call:

skill_router({ query: "Why did the Roman Empire fall?", force_skill: "deepthink" })

Internal API (PRO0 contributors):

import { executeDeepthink } from "./src/skills/bundled/deepthink";

const analysis = await executeDeepthink("Why did the Roman Empire fall?", "auto");

Guides:

• docs/skills/qmd.md
• docs/skills/deepthink.md

See docs/configuration.md for configuration file locations and all options.

Installation

As OpenCode Plugin (Recommended)

1. Install the plugin:

npm install -g pro0
# or
bun add -g pro0

2. Configure OpenCode to use PRO0 agents (~/.config/opencode/opencode.json):

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["pro0"],
  "default_agent": "proPlanner"
}

This configuration:

• Loads the PRO0 plugin (adds 8 agents: proPlanner, proExecutor + 6 specialists)
• Sets proPlanner as the default agent

3. Start using OpenCode:

cd your-project
opencode

You should now see proPlanner as the active agent in the UI!

4. Switch between agents:
   • Tab key: Cycle through all primary agents (proPlanner ↔ proExecutor ↔ build ↔ plan)
   • @ mention: Invoke specialists (@styling, @security, @testing, @docs, @research, @self-review)

Note: PRO0 agents coexist with OpenCode's built-in agents. You can switch between them using Tab. If you only want PRO0 agents, you can disable the built-in ones:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["pro0"],
  "default_agent": "proPlanner",
  "agent": {
    "build": { "disable": true },
    "plan": { "disable": true }
  }
}

This disables build and plan while keeping PRO0's proPlanner and proExecutor.

As Standalone Package

npm install pro0
# or
bun add pro0

Initialize in your project:

npx pro0 init

This creates:

• ~/.config/opencode/pro0.json (global config, auto-created with defaults)
• .opencode/ (project config directory)
• .pro0/plans/ (where plans are saved)

Quick Start

1. Configure Models

Edit ~/.config/opencode/pro0.json:

{
  "$schema": "https://raw.githubusercontent.com/YOUR_REPO/main/pro0.schema.json",
  "proPlanner": {
    "model": "github-copilot/claude-opus-4-5",
    "temperature": 0.7
  },
  "proManager": {
    "model": "github-copilot/claude-sonnet-4-5",
    "temperature": 0.3
  }
}

2. Use in Your Code

import { loadConfig, PLANNER_PROMPT, EXECUTOR_PROMPT } from 'pro0';

const config = loadConfig(process.cwd());
console.log('Planner model:', config.proPlanner.model);
console.log('Manager model:', config.proManager.model);

3. CLI Commands

# Show current config
npx pro0 config

# List installed skills
npx pro0 skills

# List all plans
npx pro0 plans

Configuration

File Locations

• Global: ~/.config/opencode/pro0.json (auto-created on first run)
• Project: .opencode/pro0.json (optional, overrides global)

Deep Merge Behavior

Project config extends global config (doesn't replace):

Global:  { planner: { model: "gpt-4", temperature: 0.7 }, executor: { model: "claude" } }
Project: { planner: { temperature: 0.3 } }
Result:  { planner: { model: "gpt-4", temperature: 0.3 }, executor: { model: "claude" } }

Strict Model Enforcement

NO FALLBACKS. If a model is missing, you get a clear error:

Error: No model configured for agent 'proPlanner'.
Please set 'proPlanner.model' in ~/.config/opencode/pro0.json or .opencode/pro0.json

Custom Skills System

Auto-Loading

Skills are auto-discovered from:

• ~/.config/opencode/skills/
• .opencode/skills/

Skill Structure

~/.config/opencode/skills/
  └── my-skill/
      ├── skill.json
      └── prompts/
          └── prompt.md

skill.json:

{
  "name": "my-skill",
  "version": "1.0.0",
  "description": "Does something useful",
  "prompts": ["prompts/prompt.md"]
}

Disabling Skills

{
  "skills": {
    "auto_load": true,
    "disabled": ["old-skill", "deprecated-tool"]
  }
}

Agents

Planner

Interviews user, spawns research subagents, creates detailed plan in .pro0/plans/<timestamp>-<slug>.md.

Executor

Takes plan, executes tasks in loop, spawns specialists, runs verification.

Specialists

• styling: UI/UX, CSS, design
• security: Vulnerability checks, code review
• testing: Write/update tests
• docs: Documentation updates
• research: External docs, OSS examples

Verification

After execution completes:

1. Runs configured test command (npm test by default)
2. Checks that unchanged functions still pass (regression check)
3. Retries on failure (max 3 attempts by default)
4. Reports results

{
  "verification": {
    "run_tests_after_completion": true,
    "test_command": "npm test",
    "allow_partial_success": false,
    "regression_check": true
  }
}

Security

Every agent prompt includes a prominent .env safety warning:

⚠️ **SECURITY WARNING: NEVER READ .env FILES** ⚠️

NEVER use Read, Grep, or any other tool to access .env, .env.local,
.env.production, or any environment variable files.

LSP Tools (Code Intelligence)

PRO0 agents can use LSP (Language Server Protocol) for advanced code intelligence features like go-to-definition, find-references, and symbol search.

Setup

1. Enable experimental features:

   export OPENCODE_EXPERIMENTAL=true

2. LSP servers are auto-configured for 30+ languages in OpenCode (TypeScript, Python, Rust, Go, Java, etc.)

3. Optional: Configure custom LSP servers in ~/.config/opencode/opencode.json:

   {
     "lsp": {
       "typescript": {
         "disabled": false,
         "command": ["typescript-language-server", "--stdio"]
       }
     }
   }

Available Operations

Agents can use these LSP operations:

• goToDefinition - Jump to symbol definition
• findReferences - Find all usages across workspace
• hover - Get type information and documentation
• documentSymbol - Get file outline/structure
• workspaceSymbol - Search for symbols across workspace
• goToImplementation - Find implementations of interfaces

See OpenCode LSP docs for complete documentation.

Background Agents (Parallel Execution)

PRO0 supports running multiple specialists in parallel for faster execution.

Using Background Tasks

// Start background task
delegate_task({
  subagent_type: "styling",
  prompt: "Design login form",
  run_in_background: true
})
// Returns: { task_id: "bg_styling_123...", status: "pending" }

// Continue working while background task runs...

// Retrieve results later
background_output({ task_id: "bg_styling_123..." })

// Or get all background task results
background_output({ all: true })

// Cancel if needed
background_cancel({ task_id: "bg_styling_123..." })
background_cancel({ all: true })  // Cancel all

Benefits

• 3-5x faster execution for independent tasks
• Specialists work simultaneously
• Main agent continues while specialists process
• Retrieve results when ready

Why PRO0 vs oh-my-opencode?

Feature	oh-my-opencode	PRO0
Agents	10+ agents	2 core + 6 specialists
Config file	oh-my-opencode.json	pro0.json
Model fallback	3-step provider chain	Strict: error if missing
Skills	Explicit config	Auto-scan + blacklist
Verification	Optional hooks	Built-in, always runs
Background agents	✅ Complex system	✅ Simple, built-in
LSP tools	✅ Via OpenCode	✅ Via OpenCode
MCP integration	✅ Via skill_mcp	✅ Via skill_mcp
Ralph Loop	✅ With todo enforcer	✅ Max 5 iterations
Code review	✅ Via hooks	✅ Self-review specialist
Complexity	High	Low
.env protection	Not emphasized	Built-in warnings

Development

# Install dependencies
bun install

# Build
bun run build

# Watch mode
bun run dev

# Clean
bun run clean

License

MIT

Contributing

PRs welcome! See PRO0_SPEC.md for architecture details.