Overview
AttentionOS is the agent foundation for high-reliability systems — just as VxWorks schedules hardware tasks, AttentionOS governs how AI agents collaborate and execute. Built on this foundation, AttForge (铸智) is an L4 full-automatic AI software engineering system: humans define requirements and architecture, while AI completes programming and testing around the clock. AttForge brings the CMMI quality system (GJB5000A in defense) into agent systems — the process is enforced by the engine, the work is done by AI, checks are performed by objective tools, and acceptance remains with humans.
AttentionOS Foundation: Four Core Mechanisms
Process Engine
Every step is arranged by the engine — AI only executes. No skipped steps, no self-declared "done".
Verification Gate
Each step must pass objective checks (compile / test / coverage) — if it fails, it stops. No bluffing through.
State Ledger
The whole process is recorded automatically — progress, next steps, fully traceable across weeks-long tasks.
Independent Review
Every AI output is cross-checked by an independent AI reviewer — review-only, no edits, everything archived.
Five Engineering Disciplines, Enforced by the Engine
Designs Must Be Reviewed
The engine checks design completeness against preset rules — no pass, no go.
Code Must Be Reviewed
Static analysis checks every rule of military standards (GJB); issues are characterized and cleared level by level.
Testing Must Be Staged
Four levels — static, unit, integration, configuration item. No advancing without passing the current level.
Process Must Be Recorded
Every test case's input, result and coverage is archived automatically — fully traceable.
Delivery Must Be Accepted
Complete test documentation is generated automatically; human experts sign off final acceptance.
Division of Labor: Humans Decide, AI Works
Human
- Requirements analysis + architecture design
- Review AI module-level designs
- Review phase deliverables
- Verify & accept all CMMI expert-check items
AI
- Produce module-level design solutions
- Write code and test cases
- Run compile/test/check on real hardware 24/7
- Provides material and advice only — never decides
AI Reviewer
- Played by AI — review-only, never edits
- Cross-checks every step's output
- Verifies each fix item by item
- Phase deliverables go to human expert review
Three Core Scenarios
Document & Solution Design
- AI produces module-level designs and complete engineering documentation
- Design review gate — incomplete means no-go
- Documents generated as execution proceeds — numbers always match the ledger
- Rules & checklists keep accumulating — stricter with every project
High-Quality Code Review
- Static analysis checks every GJB8114 rule item by item
- Issues characterized first, then cleared by level — zero remaining
- Independent AI reviewers cross-check — review only, no edits
- Every change triggers automatic regression — fix issues on the spot
Code Generation & Testing
- AI completes programming and testing as one, 24/7 non-stop
- Testing level by level: static → unit → integration → configuration item
- Statement & branch coverage both 100% measured (gcov)
- Algorithm correctness verified bit-by-bit against output baselines
One Foundation, Industry Templates on Demand
CMMI Level-5 processes, GJB8114 rule sets, document and test-flow templates are packaged as standard template packs — the engine enforces whatever pack is loaded. Defense, power and automotive templates load on demand: customization without rebuilding the system. Every project's experience flows back into the library — the system grows stricter and delivery grows faster with use.
AttentionOS is like a smartphone OS; industry templates are the apps you install — one foundation, consistent quality, adapted to different customers by swapping the "compliance pack".
Air-Gapped Deployment: Code Never Leaves
All development and testing run on your own compute servers with the latest edge LLMs — inference never leaves the server, and compilation, testing and coverage run entirely on-premises. With no external service dependencies, the whole toolchain meets classified and sensitive-environment requirements.