Portable agent definitions compiled to four runtimes

Problem

Every coding-agent runtime uses its own configuration format. Maintaining the same agent in four places means four files that drift apart, and finding out at runtime that a target does not support what you asked for.

Solution

One declarative definition as the single source of truth, pushed through a five-stage pipeline: parse, validate whether it is structurally correct, lint whether it is also a good idea, analyse what each target runtime really supports, and compile to its native configuration.

Separating validation, lint and capability analysis is deliberate and recorded as an architecture decision: they answer three different questions, and merging them produces confusing errors.

Architecture

Technical decisions

• The core does not execute agents, manage worktrees or ship a UI: scope closed on purpose.
• Documented CLI/JSON contract so other tooling can build on top.
• Permission model and skill format captured as ADRs, including the choice to use real SKILL.md instead of inventing a format.
• Hexagonal architecture: isolated domain, one adapter per runtime.

Result

Working core with architecture, contract and capability documentation, plus example projects: a repository maintainer and a SAP Business One pair of developer and reviewer agents.