DrizzyAgent

DrizzyAgent is an AI agent harness that extends OpenCode (oh-my-opencode fork) with multi-agent orchestration. It provides a powerful system for coding tasks with specialized agents for different types of work.

Built for those who want the power of multi-model AI agents without the complexity of managing them manually.

What It Does

DrizzyAgent provides discipline agents that work together to complete coding tasks:

| Agent | Purpose | Default Model Chain | |-------|---------|---------------------| | Coder | Main orchestrator - plans, delegates, drives to completion | Claude Opus 4.6 / Kimi K2.5 / GPT-5.4 / GLM-5 | | GPTCoder | Deep autonomous worker - explores, researches, executes end-to-end | GPT-5.4 | | Planner | Strategic planner - interviews, identifies scope, builds plans | Claude Opus 4.6 / Kimi K2.5 / GPT-5.4 / Gemini 3.1 Pro | | Oracle | Architecture consultant - debugging, complex logic decisions | GPT-5.4 / Kimi K2.5 / Gemini 3.1 Pro / Claude Opus 4.6 | | Librarian | Documentation/code search - external references, OSS examples | Gemini 3 Flash / GLM-4.7 / Claude Sonnet 4.5 | | Explore | Fast codebase grep - pattern discovery, cross-layer search | Custom resolver (Claude / Copilot / OpenAI) | | Orchestrator | Runs Planner-generated plans step by step | Kimi K2.5 / Claude Sonnet 4.6 / GPT-5.4 / Gemini 3.1 Pro | | Researcher | Deep research orchestrator - multi-source investigation with cited reports | Gemini 3.1 Pro / Kimi K2.5 / Claude Sonnet 4.6 / GPT-5.4 / GLM-4.7 Free | | Coder Junior | Delegated sub-tasks from Coder via category system | Claude Sonnet 4.6 / GPT-5.4 / Gemini 3 Flash | | Researcher Junior | Focused web research subagent for parallel sub-topic investigation | Kimi K2.5 / Claude Sonnet 4.6 / GPT-5.4 / GLM-4.7 Free | | Plan Consultant | Pre-planning analysis - scope clarification, ambiguity detection | Claude Opus 4.6 / Kimi K2.5 / GPT-5.4 / Gemini 3.1 Pro | | Plan Reviewer | Plan quality review - verifiability, completeness checks | GPT-5.4 / Kimi K2.5 / Claude Opus 4.6 / Gemini 3.1 Pro | | Multimodal Looker | Image and visual analysis | GPT-5.4 / Kimi K2.5 / Gemini 3 Flash |

Main selectable agents: Coder, GPTCoder, Planner, Orchestrator, Researcher. Subagent-only specialists include Researcher Junior, Oracle, Librarian, Explore, and other focused helpers.

Key Features

ultrawork Mode

Type ultrawork (or ulw). Every agent activates. Doesn't stop until done.

Agent Categories

When Coder delegates, it picks a category, not a model:

| Category | For | Default Model Chain | |----------|-----|---------------------| | visual-engineering | Frontend, UI/UX, design | Gemini 3.1 Pro / GLM-5 / Claude Opus 4.6 / Kimi K2.5 | | ultrabrain | Hard logic, architecture decisions | GPT-5.4 / Gemini 3.1 Pro / Claude Opus 4.6 / Kimi K2.5 | | deep | Autonomous research + execution | GPT-5.4 / Claude Opus 4.6 / Gemini 3.1 Pro / Kimi K2.5 | | artistry | Creative, unconventional approaches | Gemini 3.1 Pro / Claude Opus 4.6 / GPT-5.4 / Kimi K2.5 | | quick | Single-file changes, typos | Claude Haiku 4.5 / Gemini 3 Flash / GPT-5.1 Codex Mini | | unspecified-low | General tasks, low effort | Claude Sonnet 4.6 / Kimi K2.5 / GPT-5.4 / Gemini 3 Flash | | unspecified-high | General tasks, high effort | Claude Opus 4.6 / GPT-5.4 / GLM-5 / Kimi K2.5 | | writing | Documentation, prose, technical writing | Gemini 3 Flash / Kimi K2.5 / Claude Sonnet 4.6 |

The harness maps categories to the right models automatically based on provider availability.

Built-in Tools

• LSP Tools: lsp_rename, lsp_goto_definition, lsp_find_references, lsp_diagnostics
• AST-Grep: Pattern-aware code search and rewriting across 25 languages
• Background Agents: Fire 5+ specialists in parallel
• Session Manager: List, read, search session history
• Hashline Edit: Content-hash validated edits (no stale-line errors)
• Tmux Integration: Full interactive terminal for REPLs, debuggers, TUIs

Built-in MCPs

• Web Search (Exa): Real-time web search
• Context7: Official documentation lookup
• Grep.app: GitHub code search

Commands

| Command | Description | |---------|-------------| | /ultrawork or /ulw | Activate all agents, work until done | | /init-deep | Generate hierarchical AGENTS.md files | | /start-work | Planner interview-mode planning | | /ralph-loop | Self-referential development loop | | /stop-continuation | Stop all continuation mechanisms |

Skills System

Skills bring domain expertise:

• playwright: Browser automation
• git-master: Atomic commits, rebase surgery
• frontend-ui-ux: Design-first UI generation
• Custom skills: Add your own in .opencode/skills/

Installation

To install drizzy-agent, run:

npx drizzy-agent install
# or
bunx drizzy-agent install

Or install as an OpenCode plugin:

// ~/.config/opencode/opencode.json
{
  "plugins": ["drizzy-agent"]
}

Configuration

Create .opencode/drizzy-agent.jsonc or ~/.config/opencode/drizzy-agent.jsonc:

{
  "agents": {
    "coder": {
      "model": "claude-opus-4-6"
    },
    "researcher": {
      "model": "gemini-3.1-pro"
    }
  },
  "categories": {
    "visual-engineering": {
      "model": "gemini-3.1-pro"
    }
  },
  "disabled_hooks": []
}

How Configuration Works (Snapshot-Based Defaults)

DrizzyAgent uses a snapshot-based configuration system. When you run drizzy-agent install, the tool creates a minimal config containing only your provider selections:

{
  "$schema": "...",
  "_install_defaults": {
    "snapshot_version": 1,
    "providers": {
      "claude": "yes",
      "openai": false,
      "gemini": false,
      // ... which providers you selected during install
    }
  }
}

The _install_defaults section is install-managed and read-only. It captures which providers you selected. The actual agents and categories configuration is computed at runtime from current fallback rules based on your available providers.

Benefits of this approach:

• Your config stays small and focused on your provider choices
• Model fallbacks automatically improve with each plugin update
• No manual updates needed when better models become available

Adding Custom Overrides

You can add explicit overrides on top of the computed defaults:

{
  "_install_defaults": { /* ... */ },
  "agents": {
    "coder": {
      "model": "claude-opus-4-6"  // This overrides the computed default
    },
    "researcher": {
      "model": "gemini-3.1-pro"
    }
  },
  "categories": {
    "deep": {
      "model": "o3-mini"  // Override for specific category
    }
  }
}

Explicit overrides take precedence over computed defaults. Omitting a field means "use the computed default."

Updating Provider Selections

Rerun the install command to update provider selections:

bunx drizzy-agent install

Or manually override by editing ~/.config/opencode/drizzy-agent.jsonc:

{
  "agents": {
    "coder": { "model": "claude-opus-4-6" }
  },
  "categories": {
    "deep": { "model": "gpt-5.4" }
  }
}

Development

# Install dependencies
bun install

# Build
bun run build

# Test
bun test

Architecture

src/
├── agents/          # 13 agents (Coder, GPTCoder, Oracle, Librarian, Explore, Orchestrator, Planner, Researcher, Researcher Junior, Plan Consultant, Plan Reviewer, Coder Junior, Multimodal Looker)
├── hooks/           # 46 lifecycle hooks
├── tools/           # 26 built-in tools
├── features/        # Core feature modules
├── cli/             # CLI implementation
└── shared/          # Utilities

License

SUL-1.0 (Source Available License)

Named after the author's nickname "Drizzy" — built for developers who want serious AI assistance without the overhead.