Skip to content

Product Architecture

Sinaris is a local-first control layer for AI coding agents.

Its architecture is built around one boundary: product code owns facts, state, tools, and invariants; agents use SKILLs to decide how to ask, design, implement, verify, continue, or stop.

Control Layer

Persistent Surfaces

SurfaceOwns
Workspace markdownHuman-readable and agent-readable project facts: goals, proposals, milestones, phase designs, Apply work packets.
SQLiteRuntime state: tasks, ownership, dispatches, sessions, AgentEvent, evaluator state, indexes.
CLI protocolStructured tools for agents, hooks, Hub, and automation.
Generated assetsHost-agent commands, SKILLs, hooks, and workflow boundaries.
HubHuman projection over state and explicit operations.

Design Constraints

  • Guide and Proposal are separate entry paths.
  • Every implementation path converges at Design before Apply.
  • Apply works one task at a time and records evidence.
  • AgentEvent is the structured handoff unit for assignment, continuation, decision resume, auto-claim, and user messages.
  • Hub is a projection, not a parallel implementation of business rules.
  • Workspace asset upgrades follow the public Sinaris tool version.
  • Host agent Plan or Build Plan modes must not bypass Sinaris workflow gates.

Deep Dive

Read the full architecture document for implementation details:

Full Product Architecture

Released under the MIT License.