Agent Routing & Delegation

Agent Pipeline (ordered workflow)

1. requirements-interviewer (Sonnet) — Gather and clarify requirements from the user
2. explorer (Sonnet) — Research libraries, APIs, prior art, technical approaches
3. architect (Sonnet) — Design module boundaries, data flow, type definitions
4. planner (Opus) — Write step-by-step implementation plan (ONLY agent that writes local plan files)
5. red-teamer (Opus) — Critique the plan, find bugs/edge cases/risks before implementation
6. junior-coder (Haiku) — Scaffolding, boilerplate, mechanical refactors from fully-specified plans
7. coder (Sonnet) — Standard implementation with TDD
8. senior-coder (Opus) — Complex/cross-cutting/performance-critical implementation
9. tech-lead (Sonnet) — Cross-cutting review AFTER each implementation step, BEFORE commit (see .claude/agents/tech-lead.md)
10. reviewer (Sonnet) — Code review after implementation (reads code, checks quality)
11. qa (Sonnet) — QA testing after implementation (runs things, verifies behavior, tests hooks/workflows)
12. documentation (Haiku) — Update README, doc comments, guides
13. explainer (Haiku) — Explain code at different expertise levels (junior → staff/architect)
14. optimizer (Haiku) — Meta-workflow audit (run after every major task completion)

When to Use junior-coder vs coder vs senior-coder

• junior-coder (Haiku): Scaffolding new files from spec, struct/enum definitions, adding imports/mod declarations, moving functions between modules, writing boilerplate (constructors, getters, Display impls). Task MUST be fully specified with zero ambiguity.
• coder (Sonnet): Single-module changes, straightforward features, bug fixes with clear cause, test writing, anything requiring logic or design decisions
• senior-coder (Opus): Cross-module refactors, performance-critical paths, subtle/intermittent bugs, architecture-sensitive changes, tasks a coder failed at

Orchestrator Rules

• The orchestrator (main session) MUST NOT write implementation code directly
• The orchestrator coordinates: spawns agents, assigns tasks, reviews results
• ALL code changes go through junior-coder, coder, or senior-coder agents
• The orchestrator MAY edit non-code files: CLAUDE.md, agent specs, hook scripts, plans
• Use the decision framework: subagent for independent tasks, team for peer communication
• After each implementation step: spawn tech-lead (subagent or team member) to review before committing
• The orchestrator commits ONLY after tech-lead approves (or for trivial/no-code changes)

Model Routing

• Agent specs (.claude/agents/*.md) define the authoritative model for each agent type
• Claude-Flow routing provides suggestions based on task complexity and past performance
• In case of conflict: Agent spec always wins (manual configuration > automated suggestion)
• Model tiers: Opus (3: planner, red-teamer, senior-coder), Sonnet (7: requirements-interviewer, explorer, architect, coder, tech-lead, reviewer, qa), Haiku (4: junior-coder, documentation, explainer, optimizer)
• Example: coder uses Sonnet (per spec), even if Claude-Flow suggests Opus for a task

Portability Notes

• SKILL.md standard works across Claude Code, ChatGPT, Cursor, Copilot
• Agent specs (.claude/agents/) are Claude Code-specific
• Hooks (.claude/hooks/) are Claude Code-specific
• Keep skills portable; keep enforcement tool-specific