nano-orchestration

name

nano-orchestration

description

Use ONLY when the user wants to produce a large-scale manuscript (30–500 pages) using LiteRT-LM models (Gemma4, Qwen3) via nanogranular decomposition and 3-pass coherence fusion. Trigger keywords: nano, orchestration, manuscript, 500 pages, thesis, dissertation, book, large-scale writing, nano-orchestration, nano_orchestration, SPEC-935-R53. Do NOT activate for single-query writing, small documents (< 30 pages), or general Q&A.

Nano-Orchestration Skill

Produces large-scale academic manuscripts (up to 500 pages / 5,000+ nanoblocks) using LiteRT-LM on-device models (Qwen3 0.6B, Gemma4 2B E2B, Gemma4 4B E4B) with a 20K-token context window, overcoming the context limitation via nanogranular decomposition + 3-pass coherence fusion.

Architecture (7-Phase Pipeline)

[Input: title + sections + pages]
        │
        ▼
┌──────────────────────────────────────┐
│ 1. NanoPlanner                       │  specs/SPEC-935-R53.md
│    pages → nanoblocks (10/page)      │  nano_orchestration/planner.py
│    dependency graph, token estimates  │
└──────────────────────────────────────┘
        │
        ▼
┌──────────────────────────────────────┐
│ 2. NanoSDD Engine                    │  nano_orchestration/nano_sdd.py
│    3–7 criteria per nanoblock        │
│    prompts with type/tone/length     │
└──────────────────────────────────────┘
        │
        ▼
┌──────────────────────────────────────┐
│ 3. ContextWindowManager              │  nano_orchestration/context_window.py
│    minimal context (~300 tok/block)  │
│    neighbors + citations             │
└──────────────────────────────────────┘
        │
        ▼
┌──────────────────────────────────────┐
│ 4. NanoWriter Pool (parallel)        │  nano_orchestration/writer.py
│   ┌─────────┬──────────┬──────────┐  │
│   │Qwen3 0.6│Gemma4 2B │Gemma4 4B │  │
│   │ descr.  │ argument.│ analít.  │  │
│   │ ~3s     │ ~8s      │ ~20s     │  │
│   └─────────┴──────────┴──────────┘  │
└──────────────────────────────────────┘
        │
        ▼
┌──────────────────────────────────────┐
│ 5. Quality Checker                   │  nano_orchestration/quality_checker.py
│    validates against mini-SDD        │
│    fail → rewrite with model scale   │
│    max 3 attempts                    │
└──────────────────────────────────────┘
        │
        ▼
┌──────────────────────────────────────┐
│ 6. CoherenceEngine (3 passes)        │  nano_orchestration/coherence.py
│    Pass 1: local (bi-neighbor)       │
│    Pass 2: global (per section)      │
│    Pass 3: fluency (full text)       │
└──────────────────────────────────────┘
        │
        ▼
┌──────────────────────────────────────┐
│ 7. CrossValidator                    │  nano_orchestration/cross_validator.py
│    transitions, terminology,         │
│    contradictions, cohesion score    │
└──────────────────────────────────────┘
        │
        ▼
[Coherent manuscript up to 500 pages]

SDD — Specification-Driven Development

Every feature in this skill is gated by SPEC-935-R53 (47 acceptance criteria). Never modify nano_orchestration/ code without updating the spec first.

Validation command:

python3 scripts/validate_spec_r53.py --scale 500 --output /tmp/validacao.json

TDD — Test-Driven Development

Test suite: tests/test_nano_orchestration.py (76 tests, coverage > 95%)

RED → GREEN → REFACTOR cycle:

# Before implementing any change:
python3 -m pytest tests/test_nano_orchestration.py -q

# After implementing:
# 1. Verify all existing tests still pass
# 2. Add new test(s) for the new behavior
# 3. Run full suite
python3 -m pytest tests/test_nano_orchestration.py -v --tb=short

Model Routing

Block Type	Model	Avg Time	Timeout
Descritivo	Qwen3 0.6B	~3s	30s
Transição	Qwen3 0.6B	~2s	20s
Citação	Qwen3 0.6B	~3s	25s
Argumentativo	Gemma4 2B	~8s	60s
Metodologia	Gemma4 2B	~10s	60s
Resultado	Gemma4 2B	~6s	45s
Analítico	Gemma4 4B	~20s	120s
Discussão	Gemma4 4B	~15s	120s
Conclusão	Gemma4 2B	~8s	90s

Quality Gates

Gate	Metric	Target
SDD Gate	Criteria coverage	100%
Quality Gate	Checker score	≥ 9.5/10
Coherence Gate	Composite score	≥ 9.5/10
Cross-Validation Gate	Cohesion score	≥ 9.5/10
Success Gate	Blocks written	> 98%

Failover Strategy

Attempt 1: Ideal model (routed by block type)
    ↓ fail
Attempt 2: Fallback to next-lower model
    ↓ fail
Attempt 3: Fallback to Qwen3 0.6B (last resort)
    ↓ fail
Block marked for manual review (status="failed")

References

• Spec: specs/SPEC-935-R53-nano-orchestration.md
• Source: nano_orchestration/ (9 modules)
• Tests: tests/test_nano_orchestration.py (76 tests)
• Validator: scripts/validate_spec_r53.py (34 CAs)
• Manual: docs/MANUAL_NANO_ORCHESTRATION.md
• Server: scripts/litert-lm-serve.sh
• Provider spec: specs/SPEC-935-R210-litertlm-plugin-provider.md