opencode-orchestrate

name	opencode-orchestrate
description	Delegate one or more tasks to OpenCode workers (Haiku subagents). Single worker for focused tasks, multiple workers in parallel for complex ones. No tmux required.
user-invocable	false
experimental	["agent-teams"]

OpenCode Orchestration

Spawn Haiku subagents that call OpenCode via bash. Claude's agent panel shows progress — no custom UI needed.

---

Decision: single vs parallel workers

Situation	Pattern
One focused task (review, plan, Q&A)	Single worker
2+ independent tasks (review + plan)	Parallel workers
Same task from multiple angles	Parallel workers

---

Single worker

Agent(
  subagent_type="swarm-code:opencode-worker",
  model="haiku",
  prompt="<full task description with all context needed>"
)

The worker will ACK, call oc-run.sh (which runs opencode run), and SendMessage the result back.

---

Parallel workers (agent team)

TeamCreate(team_name="oc-team", description="<what the team is doing>")

Agent(
  subagent_type="swarm-code:opencode-worker",
  name="worker-review",
  team_name="oc-team",
  model="haiku",
  prompt="<task A> — report result via SendMessage to team-lead"
)

Agent(
  subagent_type="swarm-code:opencode-worker",
  name="worker-plan",
  team_name="oc-team",
  model="haiku",
  prompt="<task B> — report result via SendMessage to team-lead"
)

Workers run concurrently. Wait for both SendMessage results before synthesizing.

---

What workers can do

Workers call opencode run "<prompt>" headlessly. Prompt them with:

• Code review: "Review this diff for bugs and security issues: "
• Planning: "Create an implementation plan for: . Include files to change, tradeoffs, order."
• Q&A: "Explain how X works in this codebase. Context: "
• Analysis: "Find potential issues in: "

Always include the relevant code/context inline — workers can't read files themselves.

---

Estimated token savings

~70-80% fewer Claude tokens vs doing the analysis directly.

---

What NOT to do

• Don't spawn workers for tasks under 50 tokens — just answer directly
• Don't give workers file paths to read — paste the code inline in the prompt
• Don't use opencode-bridge.sh — that's deprecated, use oc-run.sh via the worker