Advanced Multi-Agent Orchestration for OpenCode (Claude Code Edition)
"Po po!" (Greek: Πω πω!) — An expression of astonishment or surprise.
A highly specialized fork of oh-my-opencode-slim, reverse-engineering the internal architecture of Anthropic's Claude Code CLI to solve LLM context flooding.
Status: Alpha — core architecture (context firewalls, advisor pattern, monitor tool) implemented. Phases D–E remaining.
Planned npm package name: po-po-code (not published yet)
Open-source coding agents typically dump every available tool and context file into a single model's prompt. If you are using a brilliant reasoning model like Codex 5.3 as your Orchestrator, feeding it 50 Chrome DevTools MCP schemas, 2MB screenshots, and raw server logs will cause Context Flooding. The model forgets instructions, slows down, and burns your token budget.
Po-po-code aims to restructure how agents communicate using the architectural patterns reflected in Anthropic's Claude Code (including leaked internals and the Monitor & Advisor tool direction):
🛡️ Context Firewalls (Domain Agents) — Implemented (Phase A).
Heavy MCPs stay off the Orchestrator. Instead of giving Codex the chrome-devtools MCP directly, a dedicated @browser agent runs on a large-context model (Gemini 3.1 Pro), absorbs visual/DOM noise internally, and returns only a dense text summary to the Orchestrator. The Orchestrator's MCP list defaults to [] — strict isolation by design.
🤝 The Advisor Pattern (Synchronous Delegation) — Implemented (Phase B).
The unified delegate_task tool replaces the old async-only background tasks. The Orchestrator delegates to a sub-agent and, when run_in_background is false (Advisor mode), receives the result inline without leaving the conversational loop. Set run_in_background: true for long-running tasks like builds.
👀 The Monitor Tool (Event-Driven Wakeups) — Implemented (Phase C).
No more token-heavy polling loops. The create_monitor tool lets the Orchestrator attach a detached script (e.g. tail -f | grep 'Error'); on match, OpenCode injects a <system-reminder> to wake the Orchestrator.
| Agent | Model (copilot preset) | Role & isolated MCPs |
|---|---|---|
| @orchestrator | gpt-5.3-codex |
The Brain. Answers simple queries; delegates heavy work. No heavy MCPs — pristine context. |
| @browser | gemini-3.1-pro-preview |
Visual / UI path. Holds chrome-devtools behind a context firewall. |
| @ops | gemini-3-flash-preview |
Execution / ops path. Builds, logs, bash, monitor. |
| @explorer | gemini-3-flash-preview |
Codebase scout. Holds serena, morph-mcp. |
| @designer | gemini-3.1-pro-preview |
UI/UX specialist. Design frameworks and implementation. |
| @oracle | claude-opus-4.6 |
Deep reasoning for hard bugs and architecture. |
bunx po-po-code@latest installgit clone https://github.com/tinof/po-po-code.git
cd po-po-code
bun install
bun run build
bun run devThe installer defaults to the copilot preset (Codex orchestrator + Gemini sub-agents via GitHub Copilot). After npm publish, you can use a different provider with:
bunx po-po-code@latest install --preset=openai
bunx po-po-code@latest install --preset=copilot
bunx po-po-code@latest install --preset=kimi
bunx po-po-code@latest install --preset=zai-planConfig files (JSONC supported) are written to:
- User:
~/.config/opencode/po-po-code.jsonc - Project:
.opencode/po-po-code.jsonc
Project config overrides user config. See docs/configuration.md for the full layering story.
The copilot preset keeps the Orchestrator lean while sub-agents use large-context Gemini models for context firewalls:
To override individual agents without changing presets, add entries under "agents" — they merge on top of the active preset.
Detailed tasks live in opencode-parity-plan.md. Summary:
| Phase | Focus | Status |
|---|---|---|
| A | Domain agents & context firewalls | ✅ @browser, @ops, @designer roles; strict orchestrator MCP firewall. |
| B | Advisor pattern | ✅ delegate_task with run_in_background; sync and async delegation. |
| C | Monitor tool | ✅ create_monitor — detached scripts, stdout triggers, <system-reminder> wakeups. |
| D | System reminders & caching | Planned — SYSTEM_PROMPT_DYNAMIC_BOUNDARY; hooks use <system-reminder> consistently. |
| E | Memory & cost | Planned — tiered CLAUDE.md hierarchy; cost visibility for large-context sub-agents. |
"Start the NextJS dev server. Monitor the output, and wake up to fix any TypeScript errors that appear in the logs."
The Orchestrator delegates to @ops to run the dev server, then uses create_monitor to attach a background watcher on stdout. When a TypeScript error appears in the logs, a <system-reminder> wakes the Orchestrator — no polling loops, no wasted tokens.
"Check why the login button isn't working on localhost:3000."
The Orchestrator delegates to @browser, which uses Chrome DevTools to capture screenshots, inspect the DOM, and check network requests. It returns a short text diagnosis (e.g. "CORS error on /api/auth") so the Orchestrator never ingests a multi-megabyte screenshot.
"What's the current directory structure of src/agents?"
The Orchestrator delegates to @explorer with run_in_background: false (Advisor mode). The explorer runs the lookup synchronously and returns the result inline — the Orchestrator continues its thought process without interruption.
Po-po-code is a specialized, heavily opinionated fork of oh-my-opencode-slim, which itself descends from oh-my-opencode created by Alvin and the Boring Dystopia Development team.
While the upstream project focuses on a broad, highly customizable agent suite with TUI multiplexing, Po-po-code strips away generic roles to focus on replicating the tight, autonomous, CLI-native developer experience of Anthropic's Claude Code using context firewalls and the parity roadmap above.
Huge thanks to the original contributors for building the foundation and hook system that makes this architectural direction possible.
"To make every token valuable, to make every cache hit, and to make every layer of extension a seamless integration."
{ "preset": "copilot", "agents": { "orchestrator": { "model": "github-copilot/gpt-5.3-codex", "temperature": 1, "variant": "high", "mcps": [] // Pristine context }, "browser": { "model": "google/gemini-3.1-pro-preview", "temperature": 1, "mcps": ["chrome-devtools"] // Context firewall }, "ops": { "model": "github-copilot/gemini-3-flash-preview", "temperature": 1, "mcps": [] }, "explorer": { "model": "github-copilot/gemini-3-flash-preview", "temperature": 1, "mcps": ["serena", "morph-mcp"] } } }