Symphony

Why pick it

You want to manage work, not babysit coding agents: a long-running service polls an issue tracker, spins one isolated workspace per issue, and runs Codex (app-server) until the ticket hits a handoff or terminal state. OpenAI’s low-key engineering preview (Apache-2.0) — trusted environments only.

Policy lives in-repo as WORKFLOW.md (YAML front matter + Markdown prompt). The language-agnostic contract is SPEC.md; there is an experimental Elixir/OTP reference under elixir/.

What you get

• Tracker adapters — Linear, GitHub Issues, Jira Cloud, Asana, GitLab (poll + optional provider-native tools run host-side so tokens stay out of the Codex child).
• Per-issue workspaces — create/reuse dirs under workspace.root; hooks (after_create, before_run, …) for clone/bootstrap.
• Codex app-server runner — bounded concurrency, retries/backoff, stall detection, multi-turn continuation while the issue stays active.
• Observability — structured logs; optional Phoenix dashboard + JSON API (--port / server.port).
• Burrito binaries — self-contained macOS/Linux releases that still expect codex, git, and tracker creds on the host.

Demo shape: monitor Linear → agents implement → CI/PR/review signals → human accepts → agent lands the PR.

Why it over the others

vs running Codex/Claude by hand — Symphony is a daemon scheduler: poll, claim, isolate, retry, reconcile. You operate tickets and handoff states, not a grid of terminal panes.

vs Herdr — Herdr is a terminal multiplexer for watching live agent panes. Symphony is an issue→workspace→agent orchestrator; different layer.

vs CI-only agent jobs — CI is push-triggered and short-lived. Symphony continuously dispatches from the tracker with workspace reuse and workflow-defined handoffs (e.g. Human Review), not only “green build on PR”.

When it's not the right pick

• Production-hardened multi-tenant control plane — this is a preview / evaluate in trusted envs.
• No harness-ready repo (tests, conventions, agent-friendly workflow) — OpenAI’s harness engineering is assumed.
• You need a polished product UI for non-engineers — optional dashboard is ops-focused, not a full PM suite.

Quick start

Option A — implement the spec (any language): point an agent at SPEC.md and build for your stack.

Option B — Elixir reference:

git clone https://github.com/openai/symphony
cd symphony/elixir
mise trust && mise install
mise exec -- mix setup && mise exec -- mix build
# copy WORKFLOW.md + optional skills into your product repo; set tracker secrets
mise exec -- ./bin/symphony ./WORKFLOW.md

Or download a release binary and run ./symphony-… ./WORKFLOW.md. Customize Linear/GitHub/… states, project slug, and Codex sandbox in the workflow front matter.

Warning: evaluation-only posture. Prefer a hardened implementation of SPEC.md for serious production. Repo: openai/symphony.