Project AumOS - Aja
Mission Control
Ready to start implementation and now bootstrapped as a no-Docker Phase 0 vertical slice; not ready for pilot or production claims.
Readiness
phase-0-implementation
Agent runs
0
Persisted AI workbench runs combine policy, retrieval, routing, invocation, and evidence checkpoints.
Capabilities
10
Capability grants are filtered before tools become visible to an actor.
Models
6
Registry entries keep GLM, Claude, GPT, and local routes replaceable.
Config stores
16
Tenant, provider, and connector records are persisted in the local no-Docker runtime.
Agent Workbench
Governed AI run planner
Policy Kernel
Capability visibility simulator
Run Ledger
Append-only execution timeline
Plan sovereign code-harness patch without exposing secrets
blocked-for-approvalcode-edit · restricted · 4 events
Generate implementation evidence pack from governed run
completedregulated-evidence · confidential · 6 events
Goal accepted: Plan sovereign code-harness patch without exposing secrets
goal.created · 11:06:45 AM
Policy blocked run
policy.evaluated · 11:06:45 AM
Model route selected
model.route_planned · 11:06:45 AM
Approval gate recorded for reversible or destructive action
approval.requested · 11:06:45 AM
Credential Broker
Scoped secret handles
The broker validates policy, narrows scopes, and returns an auditable handle. Raw secrets do not enter model context, UI state, or run evidence.
Context Engine
ACL-aware retrieval
Tenant Boundary
Configured tenants
Provider Registry
Adapter readiness
Connector Ingestion
Governed context write
Model Fabric
Registry and route plan
GLM-5.2 Sovereign Route
Z.ai compatible / self-hostable candidate
Best initial pressure-test for open-weight, long-context, sovereign deployment workflows.
Claude Opus 4.8 Enterprise Route
Anthropic
High-integrity enterprise coding, review, and analysis route when policy permits hosted execution.
GPT-5.5 Knowledge Work Route
OpenAI
Broad document, data, and knowledge-work route. Registry keeps it replaceable as newer releases arrive.
Gemini 3 Pro Frontier Route
Fourth provider family (Google) reachable via the Vercel AI Gateway; keeps hosted routing diversified across Anthropic/OpenAI/Google.
Local Small Control Model
Local runtime
Cheap local route for low-risk drafts, classifiers, and guard tasks.
MiniMax M2 Generalist Route
MiniMax
Fifth provider family (MiniMax) served by provider.minimax-http, rank 4 in the non-sovereign fallback chain. Deliberately graded B so it never outranks the frontier A routes.
- Claude Opus 4.8 Enterprise Route selected for regulated-evidence.
- Task class regulated-evidence requires a passing safety eval grade.
Model Adapter
Structured offline invocation
Sandbox
No-Docker command policy
Evidence
Proof projection
Evidence pack for Plan sovereign code-harness patch without exposing secrets
4 events · 1 policy decisions · 1 model routes
Control mappings
- authority-kernel
- model-route
- event-ledger
- approval-gate
Residual risks
- Postgres is the production run ledger when DATABASE_URL or AUMOS_AJA_LEDGER_DATABASE_URL is configured; local JSON remains a dev fallback.
- Model calls are route projections only; no provider invocation in this slice.
- Sandbox T1 isolation is fail-closed behind a container and egress stub until the deploy profile lands.
Build Route
Implementation gates
Phase 0
in progressSkeleton vertical thread
Contracts, durable local ledger, policy, credential broker, context ACL, model adapter, sandbox T0/T1, evidence, UI shell
Phase 1
nextSovereign Codex
Code harness, worktree isolation, sandbox tier, tests, route conformance
Phase 2
queuedMonday Morning Ops
Workflow harnesses, connector scopes, credential broker, approval inbox
Phase 3
queuedRegulator Evidence Pack
Compliance mapping, SIEM export, model trials, external audit package