Log

2026-05-06: Name LLM Harness As Agent-Side Prompt Boundary

Decision:

• Use LLM Harness for the prompt-building layer inside an LLM-backed Agent.
• Keep Connectors responsible for preserving platform facts and signed identity evidence.
• Keep Transport Modules responsible for verifying and normalizing wire input into NormalizedMessage.
• Keep the LLM Harness responsible for rendering NormalizedMessage into model input with explicit participants, identities, history, and current request boundaries.

Reason:

Multi-party threads can include several humans and agents. Passing history as a flat user/assistant chat transcript makes models confuse old turns with pasted content or misidentify who is speaking. Naming the harness as a separate layer keeps this as an agent-side concern without moving prompt policy into Connectors or Transport Modules.

2026-05-06: Split Transport and Connector

Decision:

• Use Transport for protocol surfaces: ActivityPub, A2A, Email, REST.
• Use Transport Module for packages that implement those surfaces.
• Use Connector for deployable platform/client integrations such as the Slack Connector.
• Use Connector Card for public Connector Instance metadata and signing keys.
• Treat Adapter as deprecated umbrella terminology.

Reason:

Adapter was overloaded. It referred both to protocol implementations and to platform products that relay users into Mentionable. Identity trust needs the distinction: a receiving agent may trust a Transport Module in its own runtime boundary, but it must make an explicit issuer-policy decision for a remote Connector Instance.