llx 0.1.3

Intelligent LLM model router driven by real code metrics — successor to preLLM

llx

Intelligent LLM model router driven by real code metrics.

Successor to preLLM — rebuilt with modular architecture, no god modules, and metric-driven routing.

llx analyzes your codebase with code2llm, redup, and vallm, then selects the optimal LLM model based on actual project metrics — file count, complexity, coupling, duplication — not abstract scores.

Principle: larger + more coupled + more complex → stronger (and more expensive) model.

Why llx? (Lessons from preLLM)

preLLM proved the concept but had architectural issues that llx resolves:

Problem in preLLM	llx Solution
cli.py: 999 lines, CC=30 (main), CC=27 (query)	CLI split into app.py + formatters.py, max CC ≤ 8
core.py: 893 lines god module	Config, analysis, routing in separate modules (≤250L each)
trace.py: 509 lines, CC=28 (to_stdout)	Output formatting as dedicated functions
Hardcoded model selection	Metric-driven thresholds from code2llm .toon data
No duplication/validation awareness	Integrates redup + vallm for richer metrics

Architecture

┌─────────────────────────────────────────────────────────────┐
│                    IDE / Agent Layer                        │
│  Roo Code │ Cline │ Continue.dev │ Aider │ Claude Code      │
│  (point at localhost:4000 as OpenAI-compatible API)         │
└─────────────────┬───────────────────────────────────────────┘
                  │
┌─────────────────▼───────────────────────────────────────────┐
│              LiteLLM Proxy (localhost:4000)                 │
│  ┌──────────┐  ┌──────────────┐  ┌────────────────────┐     │
│  │ Router   │  │ Semantic     │  │ Cost Tracking      │     │
│  │ (metrics)│  │ Cache (Redis)│  │ + Budget Limits    │     │
│  └────┬─────┘  └──────────────┘  └────────────────────┘     │
└───────┼─────────────────────────────────────────────────────┘
        │
   ┌────┼────────────────────────────────────────┐
   │    │           Model Tiers                   │
   │    ├── premium:  Claude Opus 4               │
   │    ├── balanced: Claude Sonnet 4 / GPT-5     │
   │    ├── cheap:    Claude Haiku 4.5            │
   │    ├── free:     Gemini 2.5 Pro              │
   │    ├── openrouter: 300+ models (fallback)    │
   │    └── local:    Ollama (Qwen2.5-Coder)      │
   └──────────────────────────────────────────────┘
        │
┌───────▼─────────────────────────────────────────────────────┐
│            Code Analysis Pipeline                           │
│  code2llm → redup → vallm → llx                             │
│  (metrics → duplication → validation → model selection)     │
└─────────────────────────────────────────────────────────────┘

Installation

pip install llx

# With integrations
pip install llx[all]        # Everything
pip install llx[litellm]    # LiteLLM proxy
pip install llx[code2llm]   # Code analysis
pip install llx[redup]      # Duplication detection
pip install llx[vallm]      # Code validation

Quick Start

# Analyze project and get model recommendation
llx analyze ./my-project

# Quick model selection
llx select .

# With task hint
llx select . --task refactor

# Point to pre-existing .toon files
llx analyze . --toon-dir ./analysis/

# JSON output for CI/CD
llx analyze . --json

# Chat with auto-selected model
llx chat . --prompt "Refactor the god modules"

# Force local model
llx select . --local

Model Selection Logic

Metric	Premium (≥)	Balanced (≥)	Cheap (≥)	Free
Files	50	10	3	<3
Lines	20,000	5,000	500	<500
Avg CC	6.0	4.0	2.0	<2.0
Max fan-out	30	10	—	—
Max CC	25	15	—	—
Dup groups	15	5	—	—
Dep cycles	any	—	—	—

Real-World Selection Examples

Project	Files	Lines	CC̄	Max CC	Fan-out	Tier
Single script	1	80	2.0	4	0	free
Small CLI	5	600	3.0	8	3	cheap
preLLM	31	8,900	5.0	28	30	premium
vallm	56	8,604	3.5	42	—	balanced
code2llm	113	21,128	4.6	65	45	premium
Monorepo	500+	100K+	5.0+	30+	50+	premium

LiteLLM Proxy

llx proxy config     # Generate litellm_config.yaml
llx proxy start      # Start proxy on :4000
llx proxy status     # Check if running

Configure IDE tools to point at http://localhost:4000:

Tool	Config
Roo Code / Cline	"apiBase": "http://localhost:4000/v1"
Continue.dev	"apiBase": "http://localhost:4000/v1"
Aider	OPENAI_API_BASE=http://localhost:4000
Claude Code	ANTHROPIC_BASE_URL=http://localhost:4000
Cursor / Windsurf	OpenAI-compatible endpoint

Configuration

llx init  # Creates llx.toml with defaults

Environment variables: LLX_LITELLM_URL, LLX_DEFAULT_TIER, LLX_PROXY_PORT, LLX_VERBOSE.

Python API

from llx import analyze_project, select_model, LlxConfig

metrics = analyze_project("./my-project")
result = select_model(metrics)
print(result.model_id)   # "claude-opus-4-20250514"
print(result.explain())   # Human-readable reasoning

Integration with wronai Toolchain

Tool	Role	llx Uses
code2llm	Static analysis	CC, fan-out, cycles, hotspots
redup	Duplication detection	Groups, recoverable lines
vallm	Code validation	Pass rate, issue count
llx	Model routing	Consumes all above

Package Structure

llx/
├── __init__.py              # Public API (30L)
├── config.py                # Config loader (160L)
├── analysis/
│   ├── collector.py         # Metrics from .toon, filesystem (280L)
│   └── runner.py            # Tool invocation (80L)
├── routing/
│   ├── selector.py          # Metric → tier mapping (200L)
│   └── client.py            # LiteLLM client wrapper (150L)
├── integrations/
│   ├── context_builder.py   # .toon → LLM context (130L)
│   └── proxy.py             # LiteLLM proxy management (100L)
└── cli/
    ├── app.py               # Commands (230L, max CC ≤ 8)
    └── formatters.py        # Output formatting (120L)

Total: ~1,280 lines across 10 modules. No file exceeds 300L. Max CC ≤ 8.

Compare: preLLM had 8,900 lines with 3 god modules (cli.py: 999L, core.py: 893L, trace.py: 509L).

License

Apache License 2.0 - see LICENSE for details.

Author

Created by Tom Sapletta - [e-mail]