Cursor rule

1---
2description: "Performance: model selection, context management, build troubleshooting"
3alwaysApply: true
4---
5# Performance Optimization
6 
7## Model Selection Strategy
8 
9**Haiku 4.5** (90% of Sonnet capability, 3x cost savings):
10- Lightweight agents with frequent invocation
11- Pair programming and code generation
12- Worker agents in multi-agent systems
13 
14**Sonnet 4.6** (Best coding model):
15- Main development work
16- Orchestrating multi-agent workflows
17- Complex coding tasks
18 
19**Opus 4.5** (Deepest reasoning):
20- Complex architectural decisions
21- Maximum reasoning requirements
22- Research and analysis tasks
23 
24## Context Window Management
25 
26Avoid last 20% of context window for:
27- Large-scale refactoring
28- Feature implementation spanning multiple files
29- Debugging complex interactions
30 
31Lower context sensitivity tasks:
32- Single-file edits
33- Independent utility creation
34- Documentation updates
35- Simple bug fixes
36 
37## Extended Thinking + Plan Mode
38 
39Extended thinking is enabled by default, reserving up to 31,999 tokens for internal reasoning.
40 
41Control extended thinking via:
42- **Toggle**: Option+T (macOS) / Alt+T (Windows/Linux)
43- **Config**: Set `alwaysThinkingEnabled` in `~/.claude/settings.json`
44- **Budget cap**: `export MAX_THINKING_TOKENS=10000`
45- **Verbose mode**: Ctrl+O to see thinking output
46 
47For complex tasks requiring deep reasoning:
481. Ensure extended thinking is enabled (on by default)
492. Enable **Plan Mode** for structured approach
503. Use multiple critique rounds for thorough analysis
514. Use split role sub-agents for diverse perspectives
52 
53## Build Troubleshooting
54 
55If build fails:
561. Use **build-error-resolver** agent
572. Analyze error messages
583. Fix incrementally
594. Verify after each fix
60