ypollak2/llm-router

What's New

Smart Enforcement Mode (new default)

Balances cost savings with developer productivity:

• query / research / generate / analyze tasks → hard block (forces cheap model routing)
• code tasks → soft (file tools are never blocked for actual editing)
• Target: >80% of question-answering goes through the router to free/cheap models

# Switch enforcement mode instantly
llm-router set-enforce smart   # (default for new users) balanced
llm-router set-enforce hard    # maximum savings, all tasks enforced
llm-router set-enforce soft    # suggestions only, never blocks

User-Acceptance Feedback Loop

• llm_rate feedback now influences model ordering — models users rate poorly are pushed down the chain
• Acceptance rate thresholds: ≥70% → no penalty, ≥50% → 20% penalty, <50% → 40% penalty
• Model Performance section added to llm_usage dashboard (P50/P95 latency + acceptance rate)

Reliability Fixes

• Stale pending state cleared on session start — crashed/killed sessions no longer block tools in the next session
• Orphaned pending_route_*.json files are now cleaned up automatically at SessionStart

How to upgrade

pip install --upgrade claude-code-llm-router
llm-router install --force