Fix rate limiting for Anthropic Claude API with robust multi-agent support

[SFRDevelopment]

Problem

Strix encounters RateLimitError when using Anthropic Claude models, especially when multiple agents run in parallel. The default rate limiting configuration (1.0s delay, 6 concurrent requests) is too aggressive for Anthropic's tier 1 limits (50 requests/minute), causing failures when multiple agents spawn simultaneously.

Solution

Implemented robust rate limiting specifically for Anthropic/Claude models:

• Increased delay: 1.0s → 3.0s (~20 req/min, 60% under 50 req/min limit)
• Reduced concurrency: 6 → 1 (prevents burst requests from parallel agents)
• Added 10% safety buffer: Each request waits 10% longer than calculated (3.3s effective = ~18 req/min)
• Enhanced error handling: 60-second backoff on rate limit errors before retry
• Thread-safe implementation: Lock-based synchronization ensures global queue works correctly across all agents

Changes

• Modified strix/llm/request_queue.py:
  • Auto-detect Anthropic models via STRIX_LLM environment variable
  • Apply conservative rate limits (3.0s delay, 1 concurrent) for Anthropic
  • Add 10% safety buffer to delay calculation for Anthropic models
  • Implement 60s backoff on rate limit errors for Anthropic
  • Support manual override via STRIX_RATE_LIMIT_DELAY and STRIX_RATE_LIMIT_CONCURRENT environment variables
  • Log rate limiting configuration on initialization

Testing

• Tested with anthropic/claude-sonnet-4-20250514 and anthropic/claude-3-5-haiku-20241022
• Verified with multiple parallel agents (5+ agents running simultaneously)
• No rate limit errors observed during extended scans
• Successfully handles burst scenarios when multiple agents spawn at once

Rate Limit Math

• Anthropic Limit: 50 requests/minute = ~1.2s per request
• Configured: 3.0s delay + 10% buffer = 3.3s effective = ~18 req/min
• Safety margin: 64% under limit (very conservative for multi-agent scenarios)

Backward Compatibility

• Fully backward compatible
• Only affects Anthropic/Claude models automatically
• Other providers (OpenAI, etc.) use default settings unless manually configured
• Manual configuration via environment variables works for all providers

Related Issues

This addresses rate limiting issues reported by users when using Anthropic Claude models, especially in multi-agent scenarios.

[0xallam]

LGTM. Thanks! @SellMeFish