Skip to content

Overview

A PIPELINE.md file describes a multi-step workflow that orchestrates skills, models, and APIs to produce a final output. One file, one pipeline, any marketplace

v0.4.0overview

What is this?

A PIPELINE.md file describes a multi-step workflow that orchestrates skills, models, and APIs to produce a final output. One file, one pipeline, any marketplace.

The stages array is a sequential DAG (MVP). Each stage is one of:

  • llm stage — direct LLM API call (provider: <name> + model: <name> + prompt:); v0.3 decouples provider routing from model name
  • uses stage — invokes a SKILL by id, optionally @version-pinned; skill becomes system prompt for the LLM call
  • uses: pipeline:NAME@VERSION (v0.3+) — invokes another published pipeline by id; stages inlined with id-prefixing (see Composition)
  • (v0.4+) agent_runner — sandboxed agent loop with MCP tool use (Phase 3, deferred from v0.3)

Why this version (v0.3)?

Real AI workflows need more than a linear script:

  • Conditional control flow — skip a stage when upstream signals make it unnecessary (when DSL)
  • Parallelism — fan out independent analyses (parallel:, parallel DAG via depends_on)
  • Crash recovery — long pipelines must survive the 30s Cloudflare Workers wall-clock (checkpoints)
  • Resilience — retry transient provider errors (429, 503) without aborting the run
  • Composability — reuse published pipelines as stages instead of copy-pasting (uses: pipeline:NAME@VERSION)
  • Secret injection — ${VAR} / ${VAR:-default} lets CI / hosted runner inject API keys without committing them
  • Decoupled routing — provider: openai + model: gpt-4o separates which gateway from which model

v0.3 ships all seven. Agents-as-stages (agent_runner) are deferred to v0.4+ because they require a long-running subprocess, which Cloudflare Workers cannot host; cheap LLM-API + skills + (eventually) MCP cover ~90% of use cases.