Stream Chain

Stream-Chain Skill

Execute multi-step workflows where each step's complete output flows into the next step as context, enabling sequential data transformation and multi-agent coordination.

Rewritten 2026-09-09: the claude-flow stream-chain CLI command this skill originally documented does not exist in the deployed CLI (verified: claude-flow stream-chain --help → "[ERROR] Unknown command: stream-chain"; confirmed absent from the full claude-flow --help command list). ruvnet-brain's upstream corpus shows this is a faithful vendor copy of a real ruflo skill, so the break is a deployment/version mismatch, not invented content — the command may return in a future image. Until then, use the MCP tools below.

Claude Code only: the mcp__claude-flow__* tools below need the claude-flow MCP server registered (already wired in this container). On Codex / GPT-6 Astra: there is no claude-flow MCP wiring unless registered in ~/.codex/config.toml — run the steps sequentially yourself in one session instead, pasting each step's output into the next prompt by hand.

When to use

• Sequential pipelines where each step builds on the previous output.
• Multi-stage data transformation (extract -> transform -> validate -> report).
• Iterative refinement or verification chains (implement -> test -> verify).

When NOT to use

• Parallel multi-agent swarm orchestration -> use swarm-advanced (this skill is sequential).
• GitHub-specific CI/CD pipelines -> use github-workflow-automation.
• Cloud workflow automation with Flow Nexus -> use flow-nexus-swarm.
• Simple single-step tasks that need no chaining -> run the task directly.

Quick start (MCP tools)

Two ways to chain, verified against the live tool list (mcp__claude-flow__workflow_create, workflow_execute, parallel_execute, task_orchestrate all present):

Ad hoc chain — feed output forward yourself

For a one-off chain, call task_orchestrate with strategy: "sequential" once per step, passing the previous step's result back in as context for the next:

const step1 = await mcp__claude-flow__task_orchestrate({
  task: "Analyze codebase structure",
  strategy: "sequential"
})

const step2 = await mcp__claude-flow__task_orchestrate({
  task: `Previous step output:\n${step1.result}\n\nNext task: Identify improvement areas`,
  strategy: "sequential"
})

const step3 = await mcp__claude-flow__task_orchestrate({
  task: `Previous step output:\n${step2.result}\n\nNext task: Generate action plan`,
  strategy: "sequential"
})

Reusable chain — a named workflow

For a chain you will reuse, define it once with workflow_create (steps run in the order listed; omit parallel: true to keep them sequential — see swarm-advanced for the parallel form) and run it with workflow_execute:

mcp__claude-flow__workflow_create({
  name: "analysis-chain",
  steps: [
    { phase: "structure",        task: "Map directory structure and identify components" },
    { phase: "issues",           task: "Find potential improvements and problems" },
    { phase: "recommendations",  task: "Generate actionable improvement report" }
  ]
})

mcp__claude-flow__workflow_execute({ workflowId: "analysis-chain" })

The CLI equivalent for named workflows, claude-flow workflow run -t <template> --task "..." and claude-flow workflow list, is real (verified: claude-flow workflow --help) but runs a workflow template file, not an inline prompt sequence — it is not the same as the retired stream-chain run/pipeline subcommands.

References

• references/pipelines.md — the four predefined pipeline shapes (analysis / refactor / test / optimize) as workflow_create step lists, and the retired CLI syntax kept as a dated historical note.
• references/cookbook.md — how context flows between steps, worked examples translated to MCP calls, best practices, and the retired CLI syntax kept as a dated historical note.
• scripts/qa-chain.sh — currently a stub: the stream-chain pipeline command it drove does not exist, and there is no single-shot CLI substitute (the MCP tools above are the working path, and they are not scriptable from bash without an MCP client). Run the four pipelines from Claude Code interactively via the workflow calls above instead.