Skip to content

Multi-harness control plane

Status: open/planned · Kind: system · Audience: user

  • SDLC authors an agentic harness ONCE as a canonical, host-neutral model and emits it to multiple control planes — Claude Code, Codex, Cursor, Gemini — so one definition drives every host and no target is privileged.
  • The model is rich: capabilities (skills + commands), lifecycle hooks, MCP servers, tool permissions, and shared context documents — not just prompts.
  • Realizes D-VSLI-distributed-work-runner-architecture’s “the Claude plugin ends as one surface among several”: the harness becomes portable across agent hosts.

An agentic harness is the surface a coding agent runs inside: its invocable capabilities, the context it carries, the tools it may use, the servers it can reach, and the lifecycle hooks that fire around it. Each host (Claude Code, Codex, Cursor, Gemini) expresses that surface in its own file layout and vocabulary. This capability holds a single canonical model of the surface and a set of exporters that project it onto each host, so a team maintains one harness definition instead of one per tool — and adding a host is a new exporter module, not a re-authoring of every capability.

  • A canonical HarnessModel (metadata, capabilities, hooks, MCP servers, permissions, documents), validated with markdown-contract.
  • An importer that bootstraps the model from an existing Claude surface.
  • Exporters that emit the native surface for Claude Code, Codex, Cursor, and Gemini, plus a portable harness.json interchange form.
  • sdlc harness model (derive/emit the model) and sdlc harness export <target> (materialize a host’s surface) on the deterministic op substrate, so the same effect runs from a shell, CI, or a skill.

← Back to Capabilities