llm-code-sdk

Rust SDK for LLM APIs with agentic tool use. Inspired by Anthropic's Python SDK, built for Palace.

Purpose

Provides a unified interface to LLM APIs with first-class support for tool use, streaming, and code intelligence features.

Features

• Messages API: Chat completions with system/user/assistant messages
• Streaming: Real-time token streaming with async iterators
• Tool Use: Define tools, execute them, feed results back
• Extended Thinking: Support for reasoning traces
• Token Counting: Estimate usage before sending
• LCS Module and Bin: Drive DeepWiki-style model-driven repository understanding, sharpened by SmartRead and Hugo-backed live bases from inside the crate

Optional Features

• smart: Five-layer code analysis (AST, Call Graph, CFG, DFG, PDG)
• e2e: Playwright integration for browser testing
• tree-sitter-*: Language-specific syntax parsing

LCS Module and Bin

The crate now includes a first-cut lcs binary for DeepWiki-style model-driven repository understanding from the same semantic substrate that powers smart_read, smart_write, ask_docs, ask_code, and mr_search.

The intended shape is simple:

• model-driven structure generation
• model-driven page generation
• SmartRead-enhanced repository decomposition while those steps happen
• live Hugo bases kept hot by an lcs daemon
• the same substrate exposed back out through Palace Call

cargo run -p llm-code-sdk --bin lcs -- site /mnt/riffcastle/castle/garage/lcs
cargo run -p llm-code-sdk --bin lcs -- jetpack /home/wings/projects/jetpack /mnt/riffcastle/garage/jetpack-docs
cargo run -p llm-code-sdk --bin lcs -- daemon jetpack /home/wings/projects/jetpack /mnt/riffcastle/garage/jetpack-docs --poll-seconds 2

This is the current dogfood path for Living Code inside Palace. DeepWiki-open is the product-shape reference. The differentiator is that LCS gives the generator a Swiss army knife instead of only a chunked document pile. The deep semantic graph work comes next; the current cut proves the live documentation spine end to end while the imperative tool surface and the repo-scoped lcs daemon evolve around it.

llm-code-sdk already supports Anthropic-style integrations, so local models, Anthropic-compatible endpoints, OpenRouter, and Minimax-style Anthropic-compatible providers can all sit behind the same higher-level LCS flow through runtime configuration.

Smart Tools (5-Layer Analysis)

When smart feature is enabled:

Layer 1: AST        - Syntax structure (functions, types)
Layer 2: Call Graph - Who calls whom
Layer 3: CFG        - Control flow, cyclomatic complexity
Layer 4: DFG        - Data flow, variable def-use chains
Layer 5: PDG        - Program dependence, slicing

Key Components

• SmartRead: Token-efficient code reading (60-95% savings)
• SmartWrite: Structural edits with complexity analysis
• ChangeTransaction: Staged commits with validation gates
• CodeQuery: Granular, composable code queries

Usage

use llm_code_sdk::{Client, Message};

let client = Client::new("http://localhost:1234/v1");

let response = client
    .messages()
    .model("glm-4.7-flash")
    .system("You are a helpful assistant.")
    .user("Hello!")
    .send()
    .await?;

License

AGPL-3.0