stream-chainlisted

# Stream-Chain Pattern Sequential multi-agent execution where step N receives the complete output of step N-1 as its primary context. ## When to use - ✅ Ordered pipelines with data handoff (research → synthesize → write) - ✅ Progressive refinement (draft → critique → revise) - ✅ Phase transitions needing context propagation (analyze repo → design change → implement → verify) - ✅ Tasks too big for one agent context but naturally sequential ## When NOT to use - ❌ Independent tasks → use **parallel** Agent calls in one message (see `refs/orchestration.md`) - ❌ 2-step "do X then verify X" → just do it inline - ❌ Tasks where each step is <1 minute → pipeline overhead exceeds value - ❌ When steps need to iterate or backtrack — use a single persistent agent instead ## The pattern ``` Step 1 Agent: raw input → structured output ↓ (full output passed as context) Step 2 Agent: receives step 1 output → transforms → structured output ↓ Step 3 Agent: receives steps 1+2 concatenated → final output ``` Each step is a fresh subagent with **only** the previous output(s) in its context — no accumulated conversation noise. This is the key advantage over monolithic execution. ## How to execute in Claude Code Use the `Agent` tool sequentially, passing each agent's result into the next prompt: ``` 1. Call Agent(subagent_type=Explore, prompt="Analyze /opt/p2p auth layer...") 2. Read result R1 3. Call Agent(subagent_type=Plan, prompt="Given this analysis: <R1>. Design refactor to...