Skill

opencode-orchestrate

Delegates code tasks like review, planning, Q&A, analysis to OpenCode Haiku subagents via bash. Single worker for focused tasks; parallel teams for complex ones. Token-efficient, no tmux needed.

Bash

automation

developer-tools

npx claudepluginhub apoapps/swarm-code-plugin --plugin swarm-code

Tool Access

This skill uses the workspace's default tool permissions.

Preview

SKILL.md

Similar Skills

codeagent

2.6k

Executes codeagent-wrapper CLI for AI code tasks using Codex, Claude, Gemini, OpenCode backends. Supports agent presets, skill injection, git worktree isolation, parallel execution, structured output.

cexll-myclaude-1

claude-code-agent-teams

Coordinates multiple Claude Code instances as agent teams with shared tasks, inter-agent messaging, and management for parallel research, feature development, debugging, and cross-layer coordination.

2 files2 tools

toolkit

orchestrate

Orchestrates runtime teams of Claude Code agents for parallel code review, debate, planning, building, research via /orchestrate commands or team requests.

1 file12 tools

cadre

Stats

Parent Repo Stars5

Parent Repo Forks0

Last CommitApr 5, 2026

Actions

View Source View Plugin View on GitHub View README

Help us improve

Share bugs, ideas, or general feedback.

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