OverviewProof MatrixCase StudiesMethod & AboutRequest Proof Room
Overview/Case Studies/Aria Connector
ACTIVE-LOCALCASE STUDY 01

Aria Connector: cross-CLI AI governance harness

One installable package, @aria_asi/cli, wires Claude Code, Codex, and OpenCode to a local runtime that enforces truth, quality, and safety gates before a destructive action or an unverified claim ever leaves the machine.

One command. Every coding agent under the same gate. Bring your own keys.

01 Problem & system

The problem

AI coding tools are powerful but fragmented. Each surface has its own memory, output risks, and action path. Destructive commands, unsupported completion claims, unverified deploys, and doctrine drift slip through when tools operate as isolated chats.

The system built

Aria Connector installs a local-first harness, the Coach Kernel, that wraps each coding agent with a deterministic gatekeeper, an output gate, local runtime enforcement, and action ledgers, so the same governance runs across every surface instead of each chat being on its own.

02 Architecture

Three coding agents converge on one CLI, which mounts a local runtime at port 4319. Every tool call passes through the Coach Kernel before it resolves to a bounded outcome: allow, repair once, hard block, or receipt.

DIAGRAM B · CONNECTOR HARNESSACTIVE-LOCAL
Claude Code
Codex
OpenCode
Aria Connector CLI
aria connect installer · local-first
Mounted runtime :4319
local sidecar · harness daemon · provider proxy · local memory
Coach Kernel · deterministic gatekeeper · output gate
Doctrine Trigger Map · Action Ledger · Hive Binding
Allow · Repair once · Hard block · Receipt

03 Proof cards

@aria_asi/cli

ACTIVE-LOCAL
ProvesInstallable package wires three coding agents (Claude Code, Codex, OpenCode) into one harness.
Anchoraria connect

Mounted local runtime

VERIFIED QA
ProvesRuntime mounts and responds to API/UI smokes locally.
Anchorruntime smoke

Coach Kernel decisions

ACTIVE-LOCAL
ProvesDestructive commands hard-block; missing cognition repairs once; deploys need verify evidence.
Anchorcoach/state

Doctrine trigger map sync

REPOSITORY ASSET
ProvesTyped doctrine triggers and sync targets across surfaces.
Anchordoctrine map

04 What a receipt looks like

action ledger entrylocal runtime path
{ "action": "apply_patch", "surface": "claude-code", "gate": "ALLOW", "repairs": 0, "ledger_id": "act_7c1…d92" }
Red canary

A destructive shell command is attempted from an isolated chat with no gate.

expected: blocked
Green behavior

Connector intercepts, hard-blocks the command, and writes a refusal receipt.

unsafe claims = 0

05 What it proves, and where it stops

What it proves, and where it stops
Proves
  • Installable cross-CLI harness architecture
  • Local enforcement, gates, and action ledgers
  • Doctrine-triggered repair paths
Does not prove
  • Every external client has the latest runtime active
  • …without a fresh install / readback proof

06 Business outcome & next proof

Outcome

One governance harness replaces four ungoverned chat surfaces. Destructive actions are blocked by default; every accepted action is receipted and replayable.

Next proof required

Fresh-install readback on each external client confirming the latest runtime is active in the field, not only locally.

← All case studies Next: Research Substrate