Claude Pipe

claude-pipe

[](OPENCLAW.md)

**Operational pipelines for Claude Code**

Readable agents. Inspectable state. Predictable costs.

**Building a pipeline with Claude Code?** Start with [CLAUDE.md](CLAUDE.md).

Claude Code is great for tasks. But when you need multi-step pipelines — research → analyze → generate, batch processing overnight, workflows that run unattended — you hit gaps:

• No state persistence between steps
• No resume after failure
• No coordination for parallel workers
• No cost control

claude-pipe is a set of conventions that fills these gaps. File-based state you can `cat`. Flat orchestration you can debug. Agents you can `git blame`.

---

How It Works

Claude Code has architectural constraints. claude-pipe works with them:

Constraint	Why It Exists	Convention
Subagents can't spawn subagents	Forces flat hierarchy → easier debugging	Manager = Skill (ROOT reads instructions)
Context is finite	Prevents runaway costs	Workers write to files, not context
Model is "blind" until Read()	Explicit I/O = traceable	File-based state: cat state.yaml
Skills are steering, not code	Flexible, not brittle	Conventions over enforcement

**Result**: Flat orchestration, file-based state, predictable costs.

---

Before / After

Problem	Without claude-pipe	With claude-pipe
Pipeline fails at step 4	Re-run from scratch	Resume from step 4
Agent loops forever	$50 bill	Circuit breaker stops it
"What happened?"	Dig through logs	cat artifacts/state.yaml
Share with team	Copy-paste prompts	git push

---

Core Pattern

**ROOT orchestrates, Workers execute, Files persist.**

┌─────────────────────────────────────────────────────┐
│                     USER                            │
│                  (slash co