v1.5.0 - Claude-Converter

Release Notes - claude-converter v1.5.0

🚀 Multi-Agent Session Support

claude-converter is no longer limited to Claude Code sessions.

Support has been added for three coding agents through a unified interface:

• Claude Code
• Codex
• Pi

All three converters normalize their native JSONL session format into the same Hugging Face messages structure, making downstream pipelines completely tool-agnostic.

✨ New Unified Converter API

A new Converter class is now the recommended entry point.

Instead of importing different functions depending on the source, every supported coding agent now exposes the same API:

from claude_converter import Converter

converter = Converter(converter="codex")   # "claude-code" (default), "codex", or "pi"

messages = converter.session_to_messages("session.jsonl")
converter.inspect_session("session.jsonl")
records = converter.load_session("session.jsonl")

This provides a consistent interface regardless of where the session originated.

🔄 Backwards Compatibility

The original Claude Code API remains fully supported.

Existing projects require no changes.

from claude_converter.claude import (
    load_session,
    session_to_messages,
    records_to_messages,
    inspect_session,
)

No functions were removed or deprecated in this release.

📚 Documentation Improvements

The documentation has been significantly expanded.

Quick Start examples

Examples are now included for:

• Claude Code
• Codex
• Pi

using publicly available datasets from Hugging Face.

Unified API documentation

The API reference now documents the new Converter interface instead of only the original Claude-specific functions.

Fine-tuning examples

Training examples have been updated to support datasets originating from multiple coding agents while preserving the same preprocessing workflow.

🔧 Unified Message Normalization

All converters now normalize their native session structures into a common representation.

Regardless of the source tool:

• reasoning blocks become <thinking>
• tool calls become <tool_use>
• tool outputs become <tool_result>
• system/developer instructions are skipped

This ensures every supported coding agent produces identical downstream messages suitable for tokenizer.apply_chat_template().

🧠 Improved Session Handling

The new converters include tool-specific handling where necessary.

Notably:

• Codex and Pi consecutive records with the same role are automatically merged into a single conversation turn.
• Session inspection now produces the same report format across all supported tools.

📦 Installation

uv pip install --upgrade claude-converter

or

uv pip install claude-converter==1.5.0

Notes

• Existing Claude Code workflows continue to work unchanged.
• The new Converter API is recommended for all new projects.
• All supported coding-agent formats now share the same downstream preprocessing and fine-tuning pipeline.