Overview

Stelow brings product methodology to AI coding agents. Instead of open-ended feature lists, you shape proposals with clear scope boundaries, validate them through adversarial critique, and generate typed technical scopes ready for autonomous execution.

Pre-1.0 status: Stelow is under active product and market validation. Its public release line is 0.x; APIs, workflow contracts, and skills may change incompatibly before a stable 1.0.0.

Two ways to run it

Both paths run the identical workflow. The plugin is a visual layer; it never replaces the workflow state (stelow.json, .stelow/, state.md remain the source of truth). What the plugin adds is listed in plugin/what-plugin-adds.md.

How it works, in five lines

1. Shape the idea into a proposal with IN/OUT scope boundaries and an

run knobs (quality, supervision, exploration).

1. Critique it adversarially and pass the review gates you asked for

(product spec, interface, tech plan, code diff).

1. Scope it into typed delivery units (feature, spike, optimize, test-*)

with dependencies sequenced.

1. Execute each scope against an acceptance contract, then verify and

audit the result. Escalated gaps become new scopes until none remain.

1. Every approval is a portable receipt (.stelow/approvals/); every finish

leaves an audit-trail.md receipt hashed against the exact repository tree that was verified.

The full pipeline is 18 stages (triage → … → audit); lighter review modes skip gates without changing the shape. Details: architecture.md and the workflow pages.

What Stelow does not do

• No scheduler of its own — each host drives invocations from its native

event surface (the bb plugin is the reference implementation).

• No TUI in this repo — workflow status is /sw-status; visual overlays are

host-owned.

• No extension code, no compiled plugin, no per-host adapter here — just

skills plus the scripts/stelow shell helper (bash + python3).