# Context Caching Optimization Specification

## 1. Prompt Structure (Strict)
Prompts MUST follow this sequence:

### [A. STATIC REGION] - Top (Always Cached)
1. System Persona & Rules (e.g., .geminirules)
2. Background Knowledge (API specs, architecture, DB schema)
3. Output Format Definitions

### [B. DYNAMIC REGION] - Bottom (Cache Boundary)
1. Variable Parameters (Current Time, Session Metadata)
2. User Request / Query

## 2. Token Compression
- Data Flattening: Use Markdown tables, CSV, or JSON instead of prose
- Metadata Sanitization: Remove code comments, redundant empty lines, duplicate spacing, and non-essential tracking properties

## 3. Session Controls
- History Truncation: Do not concatenate raw dialog turns (invalidates cache)
- Checkpoint Aggregation: Summarize history at checkpoints; cache summary, flush raw turns
- Request Batching: Align multi-agent tasks within same TTL cycles for Cache Hit ratio

# Agentic Coding Workflow - Model Orchestration

Role-based model dispatching strategy for rate limit optimization.

## 1. Agent Architecture
4-layer workflow routing by context size, reasoning difficulty, speed, and remaining quota:

- Layer 1. Architect Agent (Design / Bug Tracking)
  Models: DeepSeek V4 Pro, Qwen3.6 Plus (Sub: GLM-5.1)
  Role: Requirements analysis and structural design specification

- Layer 2. Repository Context Agent (Large-scale Source Analysis)
  Models: MiMo-V2.5-Pro, Kimi K2.6
  Role: Long context codebase analysis and task splitting

- Layer 3. Worker Agent (Code Generation / Boilerplate)
  Models: Qwen3.5 Plus, MiniMax M2.7
  Role: Business logic and function-level implementation

- Layer 4. Inspector Agent (Compile/Lint Error Self-Correction)
  Models: DeepSeek V4 Flash
  Role: Build monitoring, syntax error fix, output format validation

## 2. Layer Routing & Limits

| Layer | Models | Hourly | Weekly | Primary Tasks |
|---|---|---|---|---|
| Architect | DeepSeek V4 Pro, Qwen3.6 Plus, GLM-5.1 | 3450, 3300, 880 | 8550, 8200, 2150 | Domain modeling, DB schema, complex debugging. GLM-5.1: approval only |
| Context Manager | MiMo-V2.5-Pro, Kimi K2.6 | 1290, 1150 | 3225, 2880 | Long context analysis, legacy refactoring scope, API doc ingestion |
| Worker | Qwen3.5 Plus, MiniMax M2.7 | 10200, 3400 | 25200, 8500 | Unit function impl, API binding, DTO/Entity boilerplate, unit tests |
| Inspector | DeepSeek V4 Flash | 31650 | 79050 | Compile/lint error fix, output validation, self-correction loop |

## 3. Implementation Guide

```python
architect_llm = CustomLLM("deepseek-v4-pro")
context_llm = CustomLLM("mimo-v2.5-pro")
worker_llm = CustomLLM("qwen-3.5-plus")
inspector_llm = CustomLLM("deepseek-v4-flash")

architect = Agent(role="Software Architect", llm=architect_llm)
code_worker = Agent(role="Senior Backend Developer", llm=worker_llm)
linter_inspector = Agent(role="QA & Code Inspector", llm=inspector_llm)
```

## 4. Operational Rules

1. DeepSeek V4 Flash Firewall
   - Typos, semicolons, import errors: NEVER escalate to higher models
   - Minimum 3 self-correction loops must resolve within Flash (31650/hr)

2. GLM-5.1 Escalation Threshold
   - 880/hr limit is tight. Exclude from default pipeline
   - Use only as last-resort fallback after 3+ failures from other models

3. Context Split for Token Savings
   - After MiMo/Kimi scans full repo, extract only needed snippets for Qwen3.5 Plus

4. MiniMax Surge Switching
   - During boilerplate/test code surges, temporarily route Worker from Qwen3.5 Plus to MiniMax M2.5 (6300/hr) or M2.7 to prevent weekly/monthly lock

## 5. Development Docs
1. Base docs: ./docs/
2. Real-time dev logs: ./work_logs/
3. New session: check ./work_logs/work_log_*.md to resume
4. Milestone complete: use /graphify to map milestone relationships
5. Design elements: ./premium_enterprise/DESIGN.md