Symposium

A Rust implementation of the OpenAI Symphony spec — a long-running orchestration service that polls an issue tracker, creates isolated per-issue workspaces, and runs coding agent sessions.

Key Differences from the Spec

1. Notion as tracker — issues live in a Notion database (queried via MCP server) instead of Linear
2. MCP proxy architecture — communicates with Notion through a spawned MCP server process (JSON-RPC over stdio) rather than calling the REST API directly

Architecture

CLI (clap)
  └─ Orchestrator (tokio event loop)
       ├─ Config Layer (WORKFLOW.md watcher + typed config)
       ├─ Notion Tracker (MCP client → notion MCP server)
       ├─ Workspace Manager (fs + hook subprocess)
       ├─ Agent Runner (app-server subprocess, JSON-RPC stdio)
       ├─ HTTP Server (axum: dashboard + /api/v1/*)
       └─ Structured Logging (tracing)

Getting Started

Prerequisites

• Rust 1.85+ (edition 2024)
• A Notion database with ID, Status, and Priority properties
• A Notion MCP server (e.g. npx -y @notionhq/notion-mcp-server)
• A coding agent that speaks the app-server JSON-RPC protocol

Build

cargo build --release

Configure

Create a WORKFLOW.md file with YAML front matter and a prompt template:

---
tracker:
  kind: notion
  mcp_command: "npx -y @notionhq/notion-mcp-server"
  database_id: "your-database-id"
  active_states: ["Todo", "In Progress"]
  terminal_states: ["Done", "Cancelled"]
  property_id: "ID"
  property_status: "Status"
  property_priority: "Priority"

polling:
  interval_ms: 30000

workspace:
  root: "~/symposium_workspaces"

hooks:
  after_create: |
    git clone https://github.com/your/repo.git .
  before_run: |
    git pull --rebase origin main

agent:
  max_concurrent_agents: 5
  max_turns: 20

codex:
  command: "claude-code app-server"
  turn_timeout_ms: 3600000
  stall_timeout_ms: 300000

server:
  port: 8080
---
You are working on issue {{ issue.identifier }}: {{ issue.title }}.

{{ issue.description }}

{% if attempt %}
This is retry attempt {{ attempt }}. Review what happened in the previous attempt and continue.
{% endif %}

Run

symposium WORKFLOW.md --port 8080

The orchestrator will:

1. Parse WORKFLOW.md for config and prompt template
2. Start an HTTP dashboard at http://localhost:8080
3. Poll Notion for issues in active states
4. For each eligible issue: create a workspace, run hooks, start an agent session
5. Retry failed issues with exponential backoff
6. Clean up workspaces when issues reach terminal states

HTTP API

Endpoint	Method	Description
/	GET	HTML dashboard with auto-refresh
/api/v1/state	GET	Full orchestrator state as JSON
/api/v1/issues/{id}	GET	Detail for a single running issue
/api/v1/refresh	POST	Trigger an immediate poll cycle

Config Hot-Reload

Edit WORKFLOW.md while the service is running — changes are picked up automatically (500ms debounce). Parse errors are logged and the previous config is retained.

License

MIT