ypollak2/llm-router

Key Fixes

1. Claude Opus Cost Misclassification

• Problem: Expensive Claude Opus ($15/1M) was incorrectly promoted to the front of routing chains when subscription quota was available
• Root Cause: _CLAUDE_MODELS frozenset included both cheap models (Haiku/Sonnet at $3/1M) and expensive Opus
• Fix: Renamed to _CLAUDE_CHEAP_MODELS and excluded Opus entirely. Hard cap logic (≥99% pressure) now blocks ALL Claude models, not just cheap ones

2. Test Environment Isolation (no_providers_env Fixture)

• Problem: Tests expecting zero configured providers were picking up Ollama from .env files
• Root Cause: monkeypatch.delenv wasn't enough because Pydantic Settings reads from both env vars AND .env files
• Fix: Create a minimal mock config object that bypasses file-based configuration entirely

Changes

• src/llm_router/profiles.py: Renamed _CLAUDE_MODELS → _CLAUDE_CHEAP_MODELS, removed Opus, updated hard cap logic
• src/llm_router/router.py: Enhanced QUOTA_BALANCED logging with provider priority indicators
• tests/conftest.py: Improved no_providers_env fixture with manual config mocking

Test Status

✅ Core router tests: 100% passing (25+ tests)
✅ Codex routing tests: 100% passing in isolation (3 tests)
✅ test_no_providers_configured: Now passing
⚠️ Full suite: Environment isolation issues with pre-existing shared state (affects Codex routing tests only when run with full suite)

Availability

📦 PyPI
🐙 GitHub Release