Systems catalog · live index

Advanced features,
tools & systems.

One command surface for everything ClearGlass runs autonomously — governed operators, named defenders, the PERCIVAL agent stack, and the append-only audit spine underneath them. Read-only by default, human-approved by design.

6Named agent systems
0–100Risk score / action
100%Material changes logged
01The catalog

Systems that run themselves — safely

Each is an independent subtree with its own CI gate. All of them fail closed, and all of them log.

Governed
clearglass-commerce/

Autonomous Commerce Operator

A governed commerce engine — FastAPI control plane + Next.js storefront/admin. Pricing, payments, refunds and fulfillment stay blocked until a human approves.

  • Risk-scored action router (0–100)
  • Approvals queue for money-moving paths
  • Safe mock mode with no live key
Autonomous
sentinel/ · PERCIVAL

PERCIVAL Agent Stack

The keyless governor / identity / capability / mission-memory core — stdlib-only, fail-closed Python agents that reason under strict capability scoping.

  • Governor + identity + capability layers
  • Mission-memory continuity
  • No external keys, deterministic
Defender
sentinel/ · SENTINEL · AEGIS

SENTINEL & AEGIS Defenders

Named-agent defenders in the keyless sentinel index — continuous posture checks that surface risk without ever taking a destructive action on their own.

  • Continuous posture & health signals
  • Read-first, escalate-second
  • Feeds the defender handoff reports
Mesh
sentinel/ · Agent Mesh

Agent Mesh

The coordination fabric that lets named agents pass missions, share memory, and hand off work while every step stays inside the same governance envelope.

  • Mission routing between agents
  • Shared, auditable context
  • Target-state v9 architecture docs
Queue
control-plane/app/audit.py

Audit Ledger & Approvals

An append-only events ledger records every material change with its risk score; the approvals table is the only door high-risk actions can pass through.

  • Immutable event trail
  • Per-event risk annotation
  • Human-in-the-loop gating
Scheduled
bots/ · .github/workflows/

Scheduled Bot Loops

Standalone automation bots (like the store smoke bot) run on scheduled workflows — daily governance self-checks and executive reports that keep the CI gates green.

  • Storefront smoke + health checks
  • Daily governance self-check
  • Generated executive report
02Architecture

The governed data model

A small, typed schema — but the load-bearing edge isn't a foreign key. It's the one that routes every proposed action through governance before it can ever touch money.

Users
PKuser_idemailrolecreated_at
Orders
PKorder_idFKuser_idstatustotal
Products
PKproduct_idnamepricestock_level
Payments
PKpayment_idFKorder_idamountmethod
Approvals
PKapproval_idFKevent_idstateapprover
Events (ledger)
PKevent_idactionrisk_scorets
Every material action travels this path
step 01observe
step 02draft
governancescore 0–100
low · auto-execute medium · queue high / critical · approval
▤ audit ledger
03The safety model

Every action earns its route

0–30

Low. Generate copy, read metrics, reconcile.

auto + log
31–70

Medium. Content publish, non-price catalog edits.

queue approval
71–100

High / critical. Pricing, payments, refunds, fulfillment, reorders, mass outbound.

blocked · approve

No high or critical action executes until an approvals row reaches approved. The self-check fails, by design, if that door is ever left open.

  • Never fabricate inventory, reviews, sales, or urgency
  • Never move live pricing, tax, payment, refund or fulfillment without approval
  • Log every material action to the append-only ledger
04Tools

The operator toolchain

ToolSurfaceWhat it doesStatus
daily_loopcontrol-planeGovernance self-check + executive report (stdlib only)live
governancecontrol-planeScores every action 0–100 and routes it by risk tierlive
auditcontrol-planeWrites every material change to the append-only ledgerlive
paymentscontrol-planeGoverned money movement — mock mode until a key is presentgated
store_smoke_botbots/Storefront smoke test run on scheduled workflowslive
PERCIVALsentinel/Keyless governor/identity/capability/mission-memory stacklive
Agent Meshsentinel/Mission routing & shared memory between named agentstarget-state
Explore

Governed autonomy, on the record.

Advanced doesn't mean unchecked. Every ClearGlass system reads first, drafts second, asks a human before it moves anything that matters — and logs all of it.