STEP — Structured Task Execution Protocol

STEP is a small, filesystem-native protocol for durable work across agents, humans, tools, and execution sessions.

A STEP WorkNode keeps three things separate:

• PLAN.md — intended work and stable step identity
• STATE.md — current execution state
• logs/ — per-step evidence and outcomes

The active conversation, model, runner, branch, or process may disappear. The WorkNode remains available for inspection, resumption, and handoff.

PROTOCOL.md is authoritative. examples/ contains an unstarted WorkNode and an illustrative cross-session handoff. applications/ maps STEP to long-running agents, deep research, and human-agent handoff without extending the protocol.

STEP is pre-1.0. A private production-scale development corpus has exercised it across more than 2,200 current step records, including recovery after context compaction, runner and test failures, blocked work, branch divergence, and worktree reconciliation. The corpus has no current users, is not public, and is not independent evidence of universal conformance.

RFC-0001.md turns the ambiguities found in that history into a compact proposal for hardening STEP Core before 1.0. The RFC is non-normative.

There is no canonical implementation. Implementations may use any language, agent, toolchain, or isolation strategy that preserves STEP's semantics. STEP does not define models, prompts, tool transport, distributed scheduling, process checkpoints, or automatic routing between WorkNodes.

The first STEP interpreter was generated using STEP. Originally exposed as worktree, it selected actionable work and managed WorkNode state. Its historical WorkNode remains under provenance/; the implementation is evidence, not protocol authority.