Symphony

Sternberg-maintained fork of openai/symphony with current Linear compatibility fixes, better defaults for production-style use, and a complete onboarding flow. Push tickets to a Linear board, agents ship the code.

Quick start

If you have an AI coding agent, one command:

npx skills add davidste/symphony -s symphony-setup -y

Then ask your agent to set up Symphony for your repo.

How it works

Symphony polls a Linear project for active tickets. Each ticket gets an isolated workspace clone and a Codex agent. The agent reads the ticket, writes a plan, implements, validates, and opens a PR. You review PRs and move tickets through states — the agents handle the rest.

The state machine lives in WORKFLOW.md — a markdown file with YAML frontmatter for config and a prompt body that defines agent behavior. Hot-reloads in under a second, no restart needed.

What's different from upstream

• Cheaper Linear calls — agents no longer burn tokens on schema introspection before every GraphQL call, and workpad sync is a single dynamic tool instead of a hand-rolled mutation
• Correct sandbox — the workflow is git + GitHub PR centric. Upstream's default sandbox blocks .git/ writes, which silently breaks the entire flow. Fixed.
• Media uploads via Linear — upstream references a GitHub media upload skill that doesn't ship. The workflow and Linear skill now use Linear's native fileUpload mutation for screenshots and recordings
• Setup skill — auto-detects your repo, installs worker skills, creates Linear workflow states, and verifies everything before launch
• Current Linear comment schema — the bundled linear skill uses resolvedAt for comment resolution state, which matches current Linear GraphQL behavior

Project-specific workflow overrides

The shared repo should stay generic. Put product- or repo-specific policy into the target repository's WORKFLOW.md, not into the shared Symphony fork.

Common examples:

• override hooks.after_create for the real source repo and setup commands
• lower concurrency or turn limits for a cautious first rollout
• explicitly override legacy repo-local instructions that conflict with Symphony's control plane (for example, another tracker or mandatory local automation)
• teach workers how to launch and validate your app via a repo-local skill such as launch-app

Manual setup

1. Build: git clone https://github.com/davidste/symphony && cd symphony/elixir && mise trust && mise install && mise exec -- mix setup && mise exec -- mix build
2. Install skills: npx skills add davidste/symphony -a codex -s linear land commit push pull debug --copy -y and copy elixir/WORKFLOW.md to your repo
3. In WORKFLOW.md, set tracker.project_slug and hooks.after_create (clone your repo + setup commands)
4. Add Rework, Human Review, Merging as custom states in Linear (Team Settings → Workflow)
5. Commit, push, then: mise exec -- ./bin/symphony /path/to/your-repo/WORKFLOW.md

For a fully owned setup, rely on this repo's README, elixir/README.md, and the bundled symphony-setup skill rather than any third-party walkthrough.

License

Apache License 2.0