@rhythmshandilya/opencode-conductor

An opencode plugin that answers each request with an orchestrated team of models: a conductor plans a workflow of (subtask, worker, access-list) steps, isolated worker agents execute them, and the final step's output is the answer. Zero config — it auto-selects your best 3 connected models.

Install

// opencode.json
{ "plugin": ["@rhythmshandilya/opencode-conductor"] }

Restart opencode, switch to the conductor agent, and ask. Two commands come with it:

• /conductor-models — the resolved model pool
• /conductor-last — the last run's plan, per-step status, and cost

How it works

• Plan — the conductor emits up to maxSteps steps from an anonymized pool of your models tagged by capability. A single successful step is treated as trivial and skips replanning.
• Isolate — each step runs in a fresh child session and sees prior steps only per its access list, preventing orchestration collapse.
• Parallelize — independent steps run concurrently; dependent steps wait only for what they reference.
• Recurse — the conductor reviews the final answer and may revise or terminate.
• Remember — completed workflows persist per project for follow-up turns.

Configuration

All optional. Set via plugin options in opencode.json, or a conductor.config.json in the project root (which wins):

{
  "$schema": "https://unpkg.com/@rhythmshandilya/opencode-conductor@latest/conductor.schema.json",
  "workers": ["anthropic/claude-sonnet-4-6", "openrouter/openai/gpt-oss-120b"],
  "conductor": "anthropic/claude-opus-4-8"
}

Model refs are "providerID/modelID" strings or { providerID, modelID, tag } objects.

| setting | default | meaning | | -------------------- | --------------- | --------------------------------------------------------- | | workers | best 3 auto | worker pool (1–10 models) | | conductor | qwen/qwen3-8b | model that plans workflows (falls back to your selection) | | maxSteps | 5 | max workflow steps per plan | | maxConcurrentSteps | 3 | max concurrently ready steps | | stepTimeoutMs | 300000 | per-step watchdog before a worker is aborted | | memory | true | project-scoped memory in .opencode/conductor/ | | budgetUsd | unset | optional USD spend ceiling |

See conductor.schema.json for the full set.

Research

Independent reimplementation of the orchestration architecture described below — driven by a prompted off-the-shelf LLM rather than a trained orchestrator. Reuses none of the authors' code or model weights. Not affiliated with or endorsed by Sakana AI.

• Learning to Orchestrate Agents in Natural Language with the Conductor — Nielsen et al., ICLR 2026, arXiv:2512.04388
• TRINITY: An Evolved LLM Coordinator — ICLR 2026, arXiv:2512.04695
• Sakana Fugu Technical Report — Sakana AI, 2026, arXiv:2606.21228

License

MIT