| # AbteeX SovereignCode Architecture |
|
|
| ## North Star |
|
|
| SovereignCode should feel like a capable local coding agent, but every action must be accountable to data sovereignty and AI sovereignty controls. The product should never silently send sensitive code or governed data to a remote model, execute an external command, or publish a change without a visible decision trail. |
|
|
| ## Control Plane |
|
|
| ```text |
| User intent |
| -> Workspace indexer |
| -> Data Capsule resolver |
| -> Sovereignty policy decision point |
| -> LumynaX MaramaRoute model selection |
| -> Tool broker |
| -> Human review gate |
| -> Audit ledger |
| ``` |
|
|
| ## Core Concepts |
|
|
| ### Data Capsule |
|
|
| A Data Capsule is the policy envelope attached to a workspace, dataset, tenant, case, source file set, or prompt context. It carries: |
|
|
| - `allowed_purposes` |
| - `denied_purposes` |
| - `resident_regions` |
| - `retention_days` |
| - `training_allowed` |
| - `export_allowed` |
| - `data_classes` |
| - `schema_context` |
| - `consent_record` |
|
|
| ### Policy Decision Point |
|
|
| The policy decision point answers one question before every sensitive action: can this actor, for this purpose, in this region, using this model/tool, touch this capsule? |
|
|
| The first implementation lives at `src/tinyluminax/products/sovereigncode/policy.py`. |
|
|
| ### Tool Broker |
|
|
| The broker is the enforcement layer for: |
|
|
| - Shell commands |
| - File writes |
| - Git commits |
| - Network calls |
| - Package installs |
| - Model calls |
| - Retrieval queries |
| - Training or distillation jobs |
|
|
| Each tool call receives a decision: allow, deny, or allow with obligations. |
|
|
| ### Audit Ledger |
|
|
| Every decision creates a record containing: |
|
|
| - Capsule id |
| - Actor |
| - Purpose |
| - Action |
| - Model id |
| - Decision |
| - Reasons |
| - Obligations |
| - Request hash |
| - Timestamp |
|
|
| The first implementation lives at `src/tinyluminax/products/sovereigncode/audit.py`. |
|
|
| ## Launch Milestones |
|
|
| | Milestone | Outcome | |
| | --- | --- | |
| | P0 scaffold | Policy engine, audit records, CLI, examples, docs. | |
| | P1 terminal loop | Local terminal agent with plan/edit/test workflow. | |
| | P2 tool broker | Policy wrappers for shell, git, file writes, package installs, and HTTP. | |
| | P3 MaramaRoute integration | Sovereign model routing for every model call. | |
| | P4 workspace UI | Browser console showing plan, policy, diffs, tests, and approvals. | |
| | P5 enterprise controls | Tenant policies, SSO hooks, signed audit exports, policy packs. | |
|
|
| ## Aesthetic Direction |
|
|
| The product should follow the AbteeX/LumynaX visual system: |
|
|
| - White or warm paper background. |
| - Obsidian text. |
| - Warm amber accent. |
| - Thin rule-based layouts. |
| - Editorial headings. |
| - Mono labels for governance, provenance, and runtime details. |
| - No generic purple AI gradients. |
|
|
