Architecture

The .developer/ tree is the single source of truth. Everything else — runtime adapters, CI, and docs — points back to it.

Layered overview

Each runtime invokes skills through a thin, generated adapter. Adapters carry no logic; they resolve to a canonical SKILL.md. Skills then drive external systems through a CLI-first, MCP-fallback tooling layer.

Layered architecture diagram. — Runtimes → adapters → canonical library → tooling.

Closed-loop lifecycle (EASE-MAS)

A supervisor plans the work as a DAG and delegates each stage to the owning skill or lead. Failures feed root-cause analysis and a bounded repair loop; every repair is re-validated against a regression gate before a human merges.

Closed-loop lifecycle from requirements through repair back to re-validation. — Forward flow (orange) builds; the return flow (green) diagnoses, repairs, and re-validates. Merge is human-only.

Traceability graph

The traceability-keeper maintains a graph so every artifact links back to a requirement and forward to its tests and status. It is written by .developer/scripts/traceability.js and is queryable for coverage and change-impact.

Code-review critic pipeline

The code-review skill dispatches a diff to the relevant specialist critics by file type and risk. Critics are independent and read-only; their findings are weighed by deliberate into a single verdict.

Code review pipeline dispatching to specialist critics and a deliberate step. — Dispatch → independent critics → deliberate → human-owned merge.

Design principles

Ask, don’t hallucinate — skills load config/profile first and ask when intent is unclear.
Human-owned merge — autonomy ladder L0–L4; autonomous CI is disabled by default, draft PRs only.
Open-source only — no vendor-internal content; generic equivalents throughout.
Minimal by default — smallest correct change; no gold-plating.

Getting started Skills