Part I · Overview
A nervous system for coding agents.
Single-purpose organs give an agent an accurate map of a codebase, memory that outlives the session, verification gates that refuse to go green by skipping work, and one routed path to private inference. Each organ is useful alone. The composition is worth more than the sum.
Intelligence · Verification · Knowledge
Compute behind the relay, never a hard-coded host
Fig. 2 Task lifecycle, five stages Full architecture ↗
- 2.1
Orient
Provision a seat, render the memory block and read the repository's structural maps.
- 2.2
Retrieve
Search anchors, trace call tracts and recall memory. Remote retrieval goes through the Knowledge plane.
- 2.3
Execute
Reason over the Intelligence plane, with session affinity for KV-cache reuse.
- 2.4
Closeout gate
Verify structure and concerns. The live System 1 rubric runs only after explicit opt-in.
- 2.5
Consolidate
Record outcomes, refresh maps and consolidate memory through cns curate.
Part II · Method
Rules the organs
are held to
- One organ, one faculty.
Bounded repositories and independently verifiable responsibilities.
- Contracts, not shared internals.
Explicit writers, readers and carriers for every inter-organ channel.
- Not evaluated is not passed.
A requested check that could not run must say so.
- Useful alone. Stronger together.
Local structural work does not depend on remote inference.
Part III Start here
Distribution status ↗
Public binaries are pending. Existing checkout instructions require repository and release access.
02 · ConnectWire your harness ↗
Hooks and CLI first; MCP when the client needs it.
03 · VerifyClose out a change ↗
Know what ran, what passed and what could not be evaluated.