task-router

Role

Route tasks to the appropriate model tier and execution mode. Most work should happen inline in the main conversation. Spawn subagents only when there's a clear benefit.

When to Spawn a Subagent

Spawn a subagent via the Task tool ONLY when:

1. Context isolation — verbose output that would pollute main context (large codegen, deep search)
2. Different model tier — task needs opus reasoning or haiku speed
3. Parallel work — multiple independent tasks can run simultaneously
4. Tool restrictions — read-only exploration that shouldn't accidentally edit files

For everything else, work inline. Inline is faster (no startup overhead) and cheaper (~4x fewer tokens than a subagent).

Routing Table

Task Class	Model	Agent Type	Key Skills to Reference
Quick triage, simple Q&A	haiku	Explore	--
Codebase exploration	haiku	Explore	--
Standard planning, task breakdown	sonnet	general-purpose	planning, arch-core
Architecture design, deep research	opus	general-purpose	architecture-design, go-arch, go-api
Code implementation	sonnet	general-purpose	go-code
Write/fix tests	sonnet	general-purpose	go-test
Standard debugging	sonnet	general-purpose	debug-core, go-test
Complex debugging (concurrency, perf)	opus	general-purpose	debug-core, go-perf, go-concurrency
Code review	opus	Explore	go-review, security-core
Acceptance validation	opus	Explore	acceptance
API/endpoint testing	sonnet	general-purpose	qa-api
Browser E2E testing	sonnet	general-purpose	qa-browser
Flutter/mobile testing	sonnet	general-purpose	qa-flutter
Performance/load testing	sonnet	general-purpose	qa-performance
Visual regression, a11y	sonnet	general-purpose	qa-visual

Slim Agents (OpenCode / Named Agents)

For platforms like OpenCode where agent name determines model+tools, use named slim agents instead of ad-hoc routing:

Agent	Model Tier	Role	Use For
scout	fast	research	File search, quick lookup, simple Q&A
coder	main	execution	Writing, editing, refactoring code
planner	main	planning	Task breakdown, implementation plans
researcher	heavy	research	Architecture design, deep codebase analysis
reviewer	heavy	review	Code review, security audit, quality check

In OpenCode, reference agents by name: the agent definition in pkg/agents/meta/ binds the correct model and skills automatically.

Task Tool Invocation Patterns

Quick exploration (read-only, fast)

Task(model: "haiku", subagent_type: "Explore", prompt: "Find all usages of UserRepository interface in the codebase")

Implementation (needs editing tools)

Task(model: "sonnet", subagent_type: "general-purpose", prompt: "Implement the CreateUser use case following Clean Architecture. Domain entity in internal/domain/entity/user.go, repository contract in internal/domain/contract/user.go, use case in internal/usecase/user/create.go")

Deep analysis (needs best reasoning)

Task(model: "opus", subagent_type: "general-purpose", prompt: "Design the authentication system architecture. Create design.md with component diagram, layer decisions, data flow, and database schema")

Review (read-only, best reasoning)

Task(model: "opus", subagent_type: "Explore", prompt: "Review the changes in internal/usecase/order/ for bugs, security issues, and adherence to Clean Architecture. Only report findings with confidence >= 80%")

Complexity Assessment Signals

Use haiku (fast) when:

• Simple lookup, file search, or pattern matching
• Task has a clear, single-step answer
• No reasoning or judgment needed

Use sonnet (standard) when:

• Standard implementation following established patterns
• Multi-step but well-defined tasks
• Test writing, standard debugging, planning

Use opus (heavy) when:

• Architecture decisions with trade-offs
• Complex debugging (concurrency, performance, multi-component)
• Code review requiring deep reasoning
• Ambiguous requirements needing interpretation

Decision Flow

Is this a quick lookup or search?
  YES → inline (or haiku Explore if large)
  NO ↓

Does this need editing tools?
  YES → sonnet general-purpose (or opus for complex)
  NO → Explore agent type

Is this complex/ambiguous?
  YES → opus
  NO → sonnet (or haiku for trivial)

Can multiple tasks run in parallel?
  YES → spawn separate Task calls
  NO → work sequentially