Runtime Capabilities Registry

The runtime capability registry is a table that encodes in code "what each runtime can observe, and what lifecycle policy it should have when the session ends". The actual source of truth is packages/core/src/runtime/capabilities.defaults.ts and packages/core/src/runtime/capabilities.helpers.ts.

Core Files

• packages/core/src/runtime/capabilities.defaults.ts
• packages/core/src/runtime/capabilities.types.ts
• packages/core/src/runtime/capabilities.helpers.ts
• docs/guide/runtime-capabilities.md

Currently Registered Adapters

Adapter	Raw prompt	Tool calls	Subagents	Native skill discovery	Event stream	Session close policy
claude-plugin	Yes	Yes	Yes	.claude/skills	No	never

Note:

• The runtimeSource in the server HTTP schema is open as a string for extensibility.
• claude-hook is an alias for backward compatibility with historical data, and the canonical value for documentation and new events is claude-plugin.

Why Capability is Needed

• Whether raw user prompts can be captured mechanically
• Whether tool / terminal / MCP activity can be observed automatically
• Whether subagent/background lineage can be tracked automatically
• Whether to close a task or leave it in waiting state when the session closes

By making this difference explicit in the registry, server lifecycle policy, observability evidence, and guide documentation can share the same expectations.

Operational Points

• The capability table should document "what actual code guarantees", not "documentation recommendations".
• When adding a new runtime, update the registry before README to reduce drift.
• Manual HTTP/MCP clients can use the server API as-is even if not registered as a built-in adapter in the registry.

Related Documentation

• Runtime Adapters & Integration
• Claude Code Plugin Adapter