vtcode-llm

June 17, 2026 · View on GitHub

LLM provider abstraction, client implementations, and streaming for VT Code.

Overview

Provides a unified interface for multiple LLM providers with streaming, tool calling, and provider-specific adaptations. This is a partial extraction from vtcode-core; integration-point files remain there.

Key Modules

Module	Purpose
`providers/`	Per-provider implementations (gemini, openai, anthropic, deepseek, ollama, etc.)
`provider/`	Core trait (`LLMProvider`), message types, request/response
`factory_types.rs`	`ProviderConfig` struct, `infer_provider_from_model`
`system_prompt.rs`	System prompt injection via `OnceLock` callbacks
`http_client.rs`	`HttpClientFactory` with timeout configuration
`lightweight_routing.rs`	Model routing for lightweight/fast inference
`model_resolver.rs`	Model name resolution and provider detection
`tool_bridge.rs`	Tool execution correlation and intent tracking
`rig_adapter.rs`	Rig-core adapter for structured output
`capabilities.rs`	Provider capability detection

Supported Providers

Provider	Module	Key Models
Google Gemini	`providers/gemini/`	Gemini 3.1 Pro, Gemini 3.5 Flash
OpenAI	`providers/openai/`	GPT-5.4, GPT-5.5, GPT-5.3 Codex
Anthropic	`providers/anthropic/`	Claude Opus 4.8, Claude Sonnet 4.6
DeepSeek	`providers/deepseek.rs`	DeepSeek V4 Pro, V4 Flash
Z.AI	`providers/zai.rs`	GLM-5.2, GLM-5.1, GLM-4.7
Moonshot	`providers/moonshot.rs`	Kimi K2.7 Code, K2.6, K2.5
StepFun	`providers/stepfun.rs`	Step-3.7-Flash
MiniMax	`providers/minimax.rs`	MiniMax-M3, M2.7, M2.5
Ollama	`providers/ollama/`	Local and cloud models
OpenRouter	`providers/openrouter/`	Marketplace models
Evolink	`providers/evolink.rs`	Multi-model gateway
HuggingFace	`providers/huggingface.rs`	Router-based models
Mistral	`providers/mistral.rs`	Mistral models
Qwen	`providers/qwen.rs`	Qwen models
MiMo	`providers/mimo.rs`	MiMo V2.5, V2.5 Pro

Architecture Notes

ProviderConfig is defined here AND in vtcode-core (identical fields) — will merge when CGP integration is decoupled
system_prompt.rs provides stub getters with OnceLock setters; vtcode-core overrides at init
Provider trait uses feature gates: #[cfg(feature = "copilot")], #[cfg(feature = "anthropic-api")]

Dependencies

vtcode-commons — HTTP, CGP, error types, model families, CompactStr
vtcode-config — provider config, timeouts, auth
vtcode-utility-tool-specs — apply_patch schemas

Coding Conventions

Provider implementations go in providers/<name>/mod.rs
Use anyhow::Result for fallible operations
Use tracing for logging, not println!
Provider-specific types stay in their provider module
Shared types go in types.rs or provider/