Speculative Exec

Speculative execution engine — predict and pre-load context the agent is likely to need next. Distilled from Claude Code speculation module.

Distilled from Claude Code speculation module. Predicts the next likely action and pre-loads the needed context in parallel with LLM response generation. Reduces perceived latency by 30-60% for common workflows.

When to use

• User asks to "analyze" something → pre-load data sources
• User asks to "read" a file → pre-load the file content
• User asks to "search" → pre-load search index
• User asks to "compare" → pre-load both items
• User asks to "write" a report → pre-load template
• Any repetitive workflow where the next step is predictable

How it works

1. Pattern Recognition — Match user intent against known workflow patterns
2. Context Prediction — Predict what files/data/tools will be needed
3. Pre-load Queue — Start loading predicted resources in background
4. Cache Delivery — When the agent actually needs the resource, serve from cache

Usage

bash

# Start speculative pre-load based on user message
python3 {baseDir}/speculator.py --message "帮我分析这个股票"

# Predict next actions from conversation history
python3 {baseDir}/speculator.py --history conversation.json

# Train on custom workflow patterns
python3 {baseDir}/speculator.py --train workflows.json

# Check cache status
python3 {baseDir}/speculator.py --status

Output

json

{
  "prediction": "read_stock_data",
  "confidence": 0.92,
  "preloaded": ["stock_price", "financial_report", "news"],
  "estimated_save_ms": 1200
}

Built-in workflow patterns

Intent Pattern	Predicted Next Action	Pre-loads
分析/analyze	Read data source	Data files, reports
对比/compare	Read both items	Item A, Item B
搜索/search	Search index	Index files
写/write	Template + context	Templates, examples
修改/edit	Read target file	Target file
安装/install	Read docs	Documentation
测试/test	Read test files	Test files, fixtures
部署/deploy	Read config	Config files

Algorithm reference

Based on Claude Code speculation module:

• Intent classification via keyword + regex patterns
• Weighted prediction scoring
• Parallel pre-load with timeout
• Cache invalidation on context change
• Confidence threshold gating (default: 0.7)