Design a configurable agent execution contract

[ineedjet]

Summary

Design a declarative contract for launching an agent with process-specific configuration.

Required dimensions

• model and inference settings;
• regulations or system instructions;
• role and playbook selection;
• skills enabled for a playbook;
• MCP servers and local command-line tools;
• additional packages or Homebrew software required by those tools;
• input artifact and expected output or transition;
• credentials passed by the consumer.

The contract should make process configuration reviewable in Git and usable by more than one runtime. Skills and tools should be selected explicitly by the playbook rather than loaded globally.

Acceptance criteria

• A configuration schema and validation strategy are proposed.
• Responsibilities are separated between StarCast, the consumer repository, and the runtime.
• Secrets are references or caller inputs, never values committed to configuration.
• Tool installation and MCP configuration are reproducible.
• The design supports both ephemeral Actions jobs and a future persistent runtime where practical.
• One end-to-end example demonstrates model, regulations, skills, MCP, and tools.

Origin

Consolidates former dupmachine/ground-control backlog items #449, #450, #451, and #454.