v1.2 — ScreenOS Phase 1 Released

AI-Native
Autonomous Agent OS

552 Immutable Contracts · 2,221+ Automated Tests · Zero I/O Kernel Determinism-first autonomous agent operating system

0
Frozen Contracts
0+
Automated Tests
0
Gate Checks
0
Core Algorithms
0
Capability Packages
0
MCP Servers

Full Governance Pipeline

10-Stage ControlPlanePipeline + 14 WorkerLoop Mixins working in concert

trace: ... · run: ... · elapsed: 0ms (simulated)
Speed:
Active
Completed
Pending
Mixin
Pipeline Stage
Infrastructure
NL Input ReceivedHTTP API
endpointPOST /intent
text"Create a TODO app HTML page"
sessiontest-agentic-final
RPC Submit → Kernel DaemonRPC
pathPrimary (RPC Daemon)
methodsubmit_run()
port9071
WorkerLoop.tick_once()Worker Loop
mixins14 active
modelocal_dev
BrainLearningCommunicationEnterpriseGovGoalSynthesisMemoryKbLlmRoutingModeDispatchResourceOrchPerceptionSafetyMonitorMetacognitionToolReasoningWebIntelHelpers
Memory Context LoadMemoryKb
eventMEMORY_READ
portSqliteMemoryPort.query()
records5 recalled
ControlPlanePipeline (10 Stages)10-Stage
typeSequential governance pipeline
1.NormalizeInput2.IntentRouting3.RiskAssess4.Governance5.ModelSelect6.CostGov7.CapSchedule8.ExecuteCap9.EvidenceCommit10.FeedbackDistill
Intent Routing (LLM-Based)Stage 2
eventRIS_ROUTING_RECOMMENDATION
routeagentic_task
layerllm (keyword fallback failed)
ms3106ms
Risk Assessment + GovernanceStage 3-4
eventPOLICY_DECISION
riskacceptable
govgovernance_required: false
Goal Synthesis + Plan GenerationGoalSynthesis
eventPLAN_GENERATED
depts2 departments
tasks2 tasks
Capability Schedule + DiscoveryStage 7
handleragentic_handler
learnedlearned_capabilities_found: 0
sourceSkill Discovery scan
Mode DispatchModeDispatch
eventMODE_DISPATCH_DECIDED
modeexecute_enqueued_intent
Safety + Metacognition GatesSafety
corrigibilityPASS
confidencePASS
perceptionmultimodal scan OK
Execute Capability (Agent Executor)Stage 8
executoragent_execute()
modelclaude-sonnet-4-6 (Anthropic)
max_tok16384
Step 1: write_fileStep 2: read_fileStep 3: done
LLM Tool Loop (3 Steps)Anthropic API
step 1write_file → /tmp/todo.html (7678 bytes)
step 2read_file → verify content
step 3done → TASK_COMPLETE
Memory Write (Execution Summary)MemoryKb
methodMemoryPort.append(MemoryPoint)
idagentic-7e33b87f413a
resultcompleted
Evidence Commit + Feedback DistillStage 9-10
events18 events committed
digeste58b8d58a604...
sinkfeedback_sink → belief update
Resource Orchestration (Plan Build)ResourceOrch
eventPLAN_BUILD_STARTED → PLAN_BUILD_FINISHED
KBRefbase.default (KnowledgePort stub)
RUN_FINISHED → HTTP ResponseOutput
statuscompleted
sourceprimary (RPC path)
total45930ms (45.9s)
file/tmp/todo.html (7678 bytes)

AIP Control Plane Protocol

AIP adds governance above MCP — policy engine, risk tiers, evidence audit — without modifying any MCP server

Agent RequestCapability ResolutionPolicy EvaluationExecution DispatchEvidence Record
Speed:
Invocation Request ReceivedAIP Core
capabilityfilesystem.read_file
actorresearch-agent (agent)
delegationuser_alice → planner → research-agent
Stage 1: Capability ResolutionRegistry
lookupfilesystem.read_file@1.0
risk_tierLOW
transportmcp (filesystem-mcp)
version_matchlifecycle_checkschema_validate
Stage 2: Policy EvaluationPolicy Engine
policies3 registered, 0 matching
risk_defaultLOW → allow (default)
decisionALLOW
target_matchcondition_evalrisk_default
Stage 3: Execution DispatchMCP
transportMCP (stdio)
methodread_file
result{ content: "[file content]" }
mcp_connecttools/callnormalize_output
Stage 4: Evidence RecordingImmutable
evidence_idev_a1b2c3d4
input_digestsha256:8f14e45f...
statusSUCCESS
Invocation ResponseComplete
statussuccess
evidenceev_a1b2c3d4

8 Core Competitive Moats

Not proof-of-concept — battle-tested engineering validated by 3,534 tests. Every feature has auditable code evidence.

Self-Evolving Capabilities

6-step pipeline auto-learns skills from GitHub / npm / PyPI / Docker with Bayesian belief tracking and auto-promotion/demotion

250+ learning tests · 3-tier safety · 7 asset types

Bayesian Trust Engine

Beta distribution probability model tracks each capability's reliability — more executions, more precision; unreliable skills auto-demote

6-level outcome grading · quality profiling · brittleness detection

ScreenOS — Screen Operating System

Unified GUI perception and actuation layer across desktop/mobile, 4-level risk assessment, evaluate-before-execute

46 frozen contracts · 4-level risk · 88 tests · 40+ ARIA roles

Self-Goal Engine

System proactively detects capability gaps, generates learning goals, schedules verification, and fills its own skill gaps

Curiosity engine · pattern mining · daily learning reports

Cryptographic Evidence Chain

Every decision generates chained hash digests; operations are byte-for-byte replayable for enterprise compliance audit

506 event kinds · chained digests · region captures · replay verification

Deterministic Zero-I/O Kernel

Pure-logic kernel with absolutely zero side effects — same input always produces same output, distributable and verifiable across machines

576 frozen contracts · 52 gates · 69,452 lines of pure domain logic

Organization Intelligence — Multi-Agent Teams

Auto-spawn agent teams with task assignment, resource arbitration, escalation, and verification aggregation for org-scale automation

Agent orchestration · subgraph scheduling · resource degrade/defer/preempt

Semantic State Plane

8 state domains × 5 fact types — understands "what happened" not just "what was output", with full provenance lineage

8 domains · 5 fact types · confidence intervals · fact lineage tracking

12 Design Constitutions, Zero Compromise

Not a GPT wrapper—an Agent OS engineered from line one for determinism, safety, and auditability

01

Determinism First

552 frozen contracts, 6 determinism gates. Same input always produces same output—predictable, replayable, auditable.

@dataclass(frozen=True) — all contracts immutable
02

Zero I/O Kernel

Absolute zero I/O in the kernel layer. Files, network, and databases all delegated via Port protocols. 1,612 kernel tests need zero external dependencies.

gate_kernel_purity() — 52 compile-time gates enforced
03

Evidence-Driven

100+ LiveEvent types record every decision. replay.jsonl fully replays any historical operation, auditing any learning conclusion's origin.

EvidenceStorePort — immutable evidence chain storage

Six-Layer Deterministic Architecture

Three-layer unidirectional dependency, 52 Gate checks enforced at compile time. 45+ pure domain modules, zero side effects.
Click any layer to view details

L6L1Unidirectional dependency, irreversible

Engineering Evidence, Not Slides

Every number comes from real codebases. Verifiable, auditable, reproducible.

552
Frozen Contracts
2,221+
Automated Tests
52
Gate Checks
12
Core Algorithms
61
Capability Packages
58
MCP Servers
46+
Role Definitions
546
API Endpoints

Covering 15 Scenario Families

46+ role definitions × 61 capability packages × 21 platform baselines form a complete scenario coverage matrix

IT Operations

Automated operations and fault repair for Linux, databases, and networks

linux-engineer database-engineer network-engineer
7 packages 35 endpoints

Software Development

Full-stack backend/frontend/QA development, CI/CD, and code quality management

programmer frontend-engineer qa-engineer
3 packages 12 endpoints

Cloud Management

AWS/Azure/GCP multi-cloud management, K8s orchestration, IaC, and cost optimization

cloud-engineer
5 packages 15 endpoints

Security

Email security, penetration testing, screen operation governance, and threat intelligence

security-analyst kali-engineer
2 packages 34 endpoints

MSP / ITSM

Alert ingestion, ticket management, SLA tracking, on-call, and runbooks

linux-engineer sre
3 packages 20 endpoints

Business Operations

CRM, finance, payments, customer support, and sales pipeline management

sales-lead finance-lead support-engineer
4 packages 25 endpoints

Knowledge Management

Document Q&A, RAG-enhanced retrieval, operations knowledge accumulation, and replay

data-engineer
3 packages 18 endpoints

Office Automation

Word/Excel/PPT processing, Teams integration, calendar, and email

personal-assistant
7 packages 30 endpoints

Cognitive Intelligence

Goal management, metacognition, curiosity engine, and multi-agent coordination

all roles
5 packages 63 endpoints

Rapid Iteration, Continuous Delivery

From AGI Enhancement to ScreenOS, all completed within weeks. Every milestone has full test coverage.

2026-03-14

ScreenOS Phase 1: DesktopOS

Screen perception and actuation layer for unified GUI interaction across browsers and desktops.

88 tests46 frozen classes29 domain modules
2026-03-13

Base Node v1.2

Production node with 9 systemd services, 58 MCP servers, and full agent loop architecture.

9 services58 MCP serversi9-14900 / 94GB / RTX 4060
2026-03

Capability Learning v1.0 - v1.3

Complete external capability acquisition pipeline with discovery, assimilation, Bayesian feedback, and quality profiling.

484 new tests6-step pipelineBayesian Beta tracking

One Backend, Five Interfaces

Unified HTTP REST backend with 546 endpoints. Full-channel coverage from CLI to native apps.

>_

CLI

Python / argparse

  • REPL Interactive Chat
  • SSE Streaming Output
  • Pipe Mode
  • Tab Completion
[ ]

TUI

Python / Textual

  • 5-Screen Dashboard
  • F1-F5 Shortcuts
  • 10s Auto-Refresh
  • Rich Text Panels
{ }

Web

React 18 / Vite 5

  • 9-Page SPA
  • Rich Output Renderer
  • Keyboard Shortcuts
  • Preference Learning
[ ]

Desktop

Tauri 2 / Rust

  • Native Window
  • System Tray
  • macOS/Win/Linux
  • Web Reuse
( )

Mobile

Expo / React Native

  • iOS + Android
  • 7 Tab Layout
  • Responsive
  • Phone/Tablet

Ready to Explore the Future of Autonomous AI?

Deterministic Kernel · Progressive Autonomy · Evidence-Driven · Learning Loop

OctopusOS
How can we help?