symphony-setup（交响乐设置）

Symphony Setup

Set up Symphony — OpenAI's orchestrator that turns Linear tickets into pull requests via autonomous Codex agents.

Preflight checks

Run these checks first and stop if any fail — resolve before continuing:

1. codex — run codex --version. Must be installed and authenticated.
2. mise — run mise --version. Needed for Elixir/Erlang version management.
3. gh — run gh auth status. Must be installed AND authenticated. Agents use gh to create PRs and close orphaned PRs. Silent failure without it.
4. LINEAR_API_KEY — run test -n "$LINEAR_API_KEY" && echo "set" || echo "missing". Must persist across sessions (shell config, not just export).
5. Linear MCP — verify Linear MCP is available. If not, set it up:

- Claude Code: claude mcp add --transport http linear https://mcp.linear.app/mcp - Codex: codex mcp add linear --url https://mcp.linear.app/mcp - Other clients: see Linear MCP docs

1. Git clone auth — the after_create hook runs git clone unattended. Verify the user's repo clone URL works non-interactively: git clone --depth 1 <url> /tmp/test-clone && rm -rf /tmp/test-clone. HTTPS with password prompts will silently fail. Use SSH keys (no passphrase) or HTTPS with credential helper / token.

Report results to the user before proceeding.

Build Symphony

Use the fork — easier to get started with:

git clone https://github.com/odysseus0/symphony
cd symphony/elixir
mise trust && mise install
mise exec -- mix setup
mise exec -- mix build

Note: mise install downloads precompiled Erlang/Elixir if available for the platform. If not, it compiles from source — this can take 10-20 minutes. Let the user know before starting.

Prepare the user's repo

Auto-detect as much as possible. Only ask the user to confirm or fill gaps.

Auto-detect repo info

• Repo path — git rev-parse --show-toplevel from the current directory. If not in a git repo, ask.
• Clone URL — git remote get-url origin. Verify it works non-interactively: git clone --depth 1 <url> /tmp/test-clone && rm -rf /tmp/test-clone.
• Setup commands — infer from lockfiles/manifests. Confirm with the user.

Auto-discover Linear project

Use Linear MCP to list projects. Present the list and let the user pick. The slugId is what goes in WORKFLOW.md's tracker.project_slug.

Auto-check and create workflow states

After the user picks a project, use Linear MCP to check the team's workflow states. Three custom states are required. If any are missing, create them via curl:

curl -s -X POST https://api.linear.app/graphql \
  -H "Content-Type: application/json" \
  -H "Authorization: $LINEAR_API_KEY" \
  -d '{"query": "mutation($input: WorkflowStateCreateInput!) { workflowStateCreate(input: $input) { success workflowState { id name } } }", "variables": {"input": {"teamId": "<team-id>", "name": "<name>", "type": "started", "color": "<color>"}}}'

Confirm with the user before creating.

Auto-detect app/UI

Check whether the project has a launchable UI before asking:

• electron or electron-builder in package.json dependencies → Electron app
• react-scripts, next, vite, nuxt in dependencies → web app with dev server
• start or dev script in package.json → likely has a dev server
• Procfile, docker-compose.yml → service with runtime

If detected, propose a launch-app skill based on what you find (framework, start script, default port). Confirm with the user and adjust. If nothing detected, ask whether there's a UI — for pure libraries/CLIs/APIs, skip the launch skill.

Install skills and workflow

Install two things from Symphony into the user's repo:

1. Skills — install via skills.sh (agents need these in their workspace clone): cd <user's repo> npx skills add odysseus0/symphony -a codex -s linear land commit push pull debug --copy -y The --copy flag is required — symlinks would break in workspace clones. The -s flag excludes symphony-setup (meta-skill, not needed by workers).
2. elixir/WORKFLOW.md — copy the entire file including the markdown body. The prompt body contains the state machine, planning protocol, and validation strategy that makes agents effective.

Patch WORKFLOW.md frontmatter

Two changes:

1. Project slug

tracker:
  project_slug: "<user's project slug>"

2. after_create hook

Replace entirely — the default clones the Symphony repo itself:

hooks:
  after_create: |
    git clone --depth 1 <user's repo clone URL> .
    <user's setup commands, if any>

Leave everything else as-is. Sandbox, approval_policy, polling interval, and concurrency settings all have good defaults in the fork.

App launch skill (if applicable)

If the user's project has a UI or app that needs runtime testing, create .agents/skills/launch-app/SKILL.md in their repo:

---
name: launch-app
description: Launch the app for runtime validation and testing.
---

# Launch App

<launch command and any setup steps specific to the user's project>
<how to verify the app is running>
<how to connect for testing — e.g., agent-browser URL, localhost port>

The WORKFLOW.md prompt tells agents to "run runtime validation" for app-touching changes. Without this skill, agents won't know how to launch the app. For non-app repos (libraries, CLIs, APIs), skip this.

Commit and push

Commit .agents/skills/, WORKFLOW.md, and launch-app skill (if created) to the user's repo and push. Push is critical — agents clone from the remote, so unpushed changes are invisible to workers.

After pushing, verify: git log origin/$(git branch --show-current) --oneline -1 should show your commit.

Pre-launch: check active tickets

Before starting Symphony, use Linear MCP to list all tickets in active states (Todo, In Progress, Rework). Symphony will immediately dispatch agents for every active ticket — not just new ones.

Show the list to the user and ask if they're comfortable with all of these being worked on. Move anything they're not ready to hand off back to Backlog.

Run

cd <symphony-path>/elixir
mise exec -- ./bin/symphony <repo-path>/WORKFLOW.md \
  --i-understand-that-this-will-be-running-without-the-usual-guardrails

The guardrails flag is required — Symphony runs Codex agents with danger-full-access sandboxing.

Add --port <port> to enable the Phoenix web dashboard.

Verify

Have the user push a test ticket to Todo in Linear. Watch for the first worker to claim it. If it fails, run this checklist:

• LINEAR_API_KEY available in the shell running Symphony?
• codex authenticated?
• gh auth status passing?
• Repo clone URL works non-interactively?
• .agents/skills/ and WORKFLOW.md pushed to remote?
• Custom Linear states (Rework, Human Review, Merging) added?

Getting started after setup

Once Symphony is running, help the user with their first workflows:

Break down a feature into tickets

The user has a big feature idea. Use Linear MCP to break it into tickets. For each ticket:

• Clear title and description with acceptance criteria
• Set blocking relationships where order matters
• Assign to the Symphony project so agents can pick them up
• Start with tickets that have no blockers in Todo

First run

Push a few tickets to Todo and watch. Walk the user through what to expect:

• Idle agents claim tickets within seconds
• Each agent writes a plan as a Linear comment before implementing
• PRs appear on GitHub with the symphony label
• The Linear board updates as agents move tickets through states

Tune on the fly

WORKFLOW.md hot-reloads within ~1 second — no restart needed. Common adjustments:

• agent.max_concurrent_agents — scale up/down based on API limits or repo complexity
• agent.max_turns — increase for complex tickets, decrease to limit token spend
• polling.interval_ms — how often Symphony checks for new/changed tickets