ypollak2/llm-router

What's new — Adaptive Universal Router v5.0

The router now dynamically selects models based on real-time budget pressure, benchmark quality scores, measured latency, and user feedback — replacing the static routing table.

Five new pillars

| Pillar | Module | What it does |
|--------|--------|-------------|
| Budget Oracle | budget.py | Pressure [0.0–1.0] across all provider types. Claude quota uses highest_pressure (max of session/weekly/sonnet dims) |
| Universal Discovery | discover.py | Scans Ollama, HuggingFace, API keys, Codex CLI. 16-family alias registry for local models |
| Live Benchmark Registry | benchmarks.py | get_quality_score() for API + Ollama models. Weekly background refresh from session-start |
| Unified Scorer | scorer.py | Weighted formula: simple tasks = 45% budget + 20% quality; complex = 55% quality + 20% budget |
| Dynamic Chain Builder | chain_builder.py | Assembles best-first chain from discovered models. Free-first invariant preserved |

New MCP tool: llm_budget

Shows real-time Budget Oracle pressure bars for all providers.

Feature flag

LLM_ROUTER_DYNAMIC=true activates dynamic routing. Default false — full v4.x backward compatibility.

Hook improvements

• enforce-route.py: session-type tracking — once Claude edits a file, enforcement downgrades to soft for the session
• auto-route.py: _is_build_task() fast-path skips routing for implementation prompts
• Pending route TTL: 300s → 60s

Upgrade

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