Architecture diagrams
Visual flows for the mechanisms that build trust: harness, approval, memory, channels and skills.
Action-or-Terminate harness
A user or channel supplies a task and exit criteria.
The model can only emit a valid tool-call or terminal signal.
The harness validates schema, approval, sandbox and observable effect.
task_complete passes only when gates and exit criteria pass.
Approval and sandbox
danger classifies safe, sensitive or dangerous.
manual/smart/yolo decides go/no-go.
The tool runs on the allowed backend and emits audit events.
Writes pass through checkpoints and can be reverted.
Memory
Explicit facts and conversation signals become candidates.
FTS5, embeddings or Honcho organize retrieval.
The context builder adds task-relevant memory.
Consolidation and rollback prevent context drift.
Channels
Terminal, Telegram, Paperclip or API normalizes input.
Origin selects the right agent and session.
Queue, nonce, approval and resume stay centralized.
Text, voice or status returns through the same channel.
Skills
The prompt receives lightweight name, description and triggers.
use_skill loads full SKILL.md when relevant.
Some skills become mandatory gates, such as UI.
Real file, scan and individual page document the capability.