delegate

name	delegate
description	Orchestrate multiple CLI agents in parallel — delegate tasks to claude/gemini/codex/opencode/rust-code via .tasks/ system. Use when "delegate to", "spawn agents", "parallel analysis", "send task to claude", "run in background", "orchestrate agents". Do NOT use for in-process sub-agents (use spawn_agent) or single bash commands (use bash_bg).
allowed-tools	delegate_task, delegate_status, delegate_result, task, bash, read_file, finish
argument-hint	<task description or .tasks/ path>

/delegate — Multi-Agent Orchestration

You are the orchestrator. You run on a cheap/fast model. You delegate complex work to powerful CLI agents and collect results.

Available Agents

Agent	Best for	Speed	Cost
claude	Complex code changes, refactoring, multi-file edits	Slow	$$$
gemini	Analysis, architecture review, large context	Fast	$
codex	Focused code tasks, single-file fixes	Medium	$$
opencode	Multi-model tasks (uses its own model config)	Medium	$$
rust-code	Autonomous loops (BigHead mode, --loop 5)	Slow	$

Reference: references/agent-guide.md

Two Modes

1. Free-text task

delegate_task {agent: "claude", task: "fix the auth bug in src/auth.rs"}

2. Task file (preferred for tracking)

task {operation: "create", title: "fix auth bug", priority: "high"}
delegate_task {agent: "claude", task_path: ".tasks/005-fix-auth-bug.md"}

The agent reads the task file, executes, updates status: done, writes ## Results.

Workflow

1. Create tasks in .tasks/ for each unit of work
2. Delegate to the right agent for each task
3. Monitor with delegate_status (check every 30-60s)
4. Collect results with delegate_result when done
5. Review — read .tasks/ files to see what each agent did

Gotchas

1. Agents inherit CLAUDE.md automatically — don't paste project conventions into the task. Just point to the task file, the agent reads CLAUDE.md on its own.
2. Check availability first — delegate_task checks if the CLI is installed, but auth issues (expired tokens, missing API keys) only surface at runtime. If a delegate fails immediately, check tmux logs.
3. task_path > free-text — with task_path, the agent updates the file with results. With free-text, results are only in the tmux buffer (ephemeral). Always prefer task files for anything important.
4. Don't poll too fast — delegates are multi-minute processes. Poll every 30-60s, not every 5s. Use bash {command: "sleep 30"} between checks.
5. Gemini CLI adds noise — gemini-cli prints warnings about API keys to stderr. The actual analysis is in the output, ignore the warnings.

Parallel Analysis Pattern

For code review / refactoring analysis, spawn 3 agents with different angles:

task {operation: "create", title: "code quality analysis", priority: "high"}
task {operation: "create", title: "architecture analysis", priority: "high"}
task {operation: "create", title: "performance analysis", priority: "high"}

delegate_task {agent: "claude", task_path: ".tasks/001-code-quality-analysis.md"}
delegate_task {agent: "gemini", task_path: ".tasks/002-architecture-analysis.md"}
delegate_task {agent: "opencode", task_path: ".tasks/003-performance-analysis.md"}

Then wait, collect results, synthesize, and delegate fixes.