Symphony

raxol_symphony orchestrates coding agents against a ticket tracker. It's an Elixir/OTP port of OpenAI's Symphony. The orchestrator polls a tracker, claims eligible issues, isolates each one in a per-issue workspace, and runs a coding agent until the workflow hits a terminal state.

Status: pre-alpha. Not yet on Hex; use the path dep at packages/raxol_symphony/.

Quick Start

mix raxol.symphony --workflow ./WORKFLOW.md

WORKFLOW.md defines tracker source, eligibility rules, retry policy, and per-issue runner config. It's hot-reloaded via file_system, so editing it doesn't restart the orchestrator. Last-known-good is served if a save leaves the file in an invalid state.

Architecture

Tracker (Memory | Linear | GitHub Issues)
    |
    v
Orchestrator (BaseManager GenServer)
    |-- polls tracker, claims eligible issues
    |-- per-issue workspace under config.workspace.root
    |
    v
Runner (RaxolAgent | Codex)          PubSub
    |                                   |
    v                                   v
Coding agent run                    Six surfaces
                                    (terminal, LiveView,
                                     MCP, Telegram, Watch, JSON API)

Runners

Runner	What it wraps	Notes
`RaxolAgent`	`Raxol.Agent.Stream`	Default. Same stack as `raxol_agent`.
`Codex`	`codex app-server` via Port	JSON-RPC 2.0 over stdio. Three-step handshake (`initialize` -> `initialized` -> `thread/start`), per-turn `turn/start` cycles.

Pick a runner per workflow. Mix-and-match isn't supported in a single run.

Surfaces

Every surface subscribes to the same orchestrator snapshot via Phoenix.PubSub, so they stay consistent without per-surface state:

Terminal: TEA dashboard listing active runs and their state.
LiveView: /symphony mounts the same dashboard in the browser.
MCP: 7 tools (list_runs, get_run, resume_run, etc) plus symphony://runs as an MCP resource.
Telegram: per-issue session, inline keyboards, approval prompts.
Watch: debounced push to APNS/FCM, tap-to-approve actions.
JSON API: GET /api/v1/state, GET /api/v1/runs/:issue_id, POST /api/v1/refresh, POST /api/v1/runs/:issue_id/stop.

Evidence Collection

Raxol.Symphony.Evidence.collect/3 runs per dispatch. It pulls:

GitHub CI status and PR comments via the GitHub API
Code complexity via cloc (falls back to SLOC if cloc isn't installed)
Asciinema .cast recording of the agent's terminal session

Set recording.enabled: true in the workflow to capture casts. The Evidence.Capture GenServer writes one .cast per run under evidence.dir.

Retry Behaviour

Three retry classes, configured per workflow:

Class	Trigger	Backoff
Continuation	Agent yields, expects re-prompt	Fixed 1s
Failure	Run exits with error	Exponential, `10s * 2^n`, capped
Stall	No output for `read_timeout_ms`	Restart from snapshot, no backoff

turn_timeout_ms bounds each individual turn; exceeding it bumps the stall counter.

Configuration

WORKFLOW.md is parsed into Raxol.Symphony.Workflow at load time. Sample shape:

---
tracker:
  type: github
  owner: example
  repo: thing
  labels: [agent-eligible]

workspace:
  root: ./symphony-workspaces

runner:
  type: raxol_agent
  read_timeout_ms: 120000
  turn_timeout_ms: 300000

recording:
  enabled: true
  dir: ./evidence

retry:
  failure_cap_ms: 600000
---