Crate llm_coding_tools_core 

llm-coding-tools-core

Lightweight, high-performance core types and utilities for coding tools - framework agnostic.

Overview

This crate provides the foundational building blocks for coding tool implementations:

• ToolError - Unified error type for all tool operations
• ToolResult<T> - Result type alias using ToolError
• ToolOutput - Wrapper for tool responses with truncation metadata
• Utility functions for text processing and formatting
• context module - LLM guidance strings for tool usage

Features

• tokio (default): Async mode with tokio runtime. Enables async function signatures.
• blocking: Sync/blocking mode. Mutually exclusive with tokio/async.
• async: Base async signatures (internal). Requires a runtime; use tokio instead.

The async and blocking features are mutually exclusive - enabling both causes a compile error.

Future runtimes (smol, async-std) can be added following the same pattern as tokio.

Usage

use llm_coding_tools_core::{ToolError, ToolResult, ToolOutput};
use llm_coding_tools_core::util::{truncate_text, format_numbered_line};

Context Module

The context module provides embedded strings containing usage guidance for LLM agents. These can be appended to tool descriptions or system prompts.

Path-based tools have two variants:

• *_ABSOLUTE: For unrestricted filesystem access (absolute paths required)
• *_ALLOWED: For sandboxed access (paths relative to allowed directories)

use llm_coding_tools_core::context::{BASH, READ_ABSOLUTE, READ_ALLOWED};

// Non-path tools have a single variant
println!("{}", BASH);

// Path-based tools have absolute and allowed variants
println!("{}", READ_ABSOLUTE);
println!("{}", READ_ALLOWED);

Available context strings:

• BASH, TASK, TODO_READ, TODO_WRITE, WEBFETCH - standalone tools
• READ_ABSOLUTE, READ_ALLOWED - file reading
• WRITE_ABSOLUTE, WRITE_ALLOWED - file writing
• EDIT_ABSOLUTE, EDIT_ALLOWED - file editing
• GLOB_ABSOLUTE, GLOB_ALLOWED - pattern matching
• GREP_ABSOLUTE, GREP_ALLOWED - content search

Design Principles

• No framework-specific dependencies, plug and play into any LLM framework/library
  • See llm-coding-tools-rig for an integration example with rig
• Minimal dependency footprint
• Performance-oriented (optimized) with zero-cost abstractions

Re-exports

pub use context::ToolContext;

pub use error::ToolError;

pub use error::ToolResult;

pub use output::ToolOutput;

pub use path::AbsolutePathResolver;

pub use path::AllowedPathResolver;

pub use path::PathResolver;

pub use system_prompt::Substitute;

pub use system_prompt::SystemPromptBuilder;

pub use operations::edit_file;

pub use operations::execute_command;

pub use operations::glob_files;

pub use operations::grep_search;

pub use operations::read_file;

pub use operations::read_todos;

pub use operations::write_file;

pub use operations::write_todos;

pub use operations::BashOutput;

pub use operations::EditError;

pub use operations::GlobOutput;

pub use operations::GrepFileMatches;

pub use operations::GrepLineMatch;

pub use operations::GrepOutput;

pub use operations::Todo;

pub use operations::TodoPriority;

pub use operations::TodoState;

pub use operations::TodoStatus;

pub use operations::fetch_url;

pub use operations::format_json;

pub use operations::html_to_markdown;

pub use operations::WebFetchOutput;

Modules

context

Tool context strings for LLM agents.

error

Common error types for coding tools.

fs

Filesystem abstraction layer.

operations

Core operations for file systems and utilities.

output

Common output types for tool responses.

path

Path resolution strategies for tool security.

system_prompt

System prompt generation for LLM agents.

tool_names

Canonical tool name constants for consistent naming across crates.

util

Shared utilities for tool implementations.