Changelog¶

All notable changes to MCParser. The format follows Keep a Changelog.

[0.2.0a1]¶

First alpha release on PyPI. Install with pip install --pre mcparser.

mcparser --version.
Evaluation now follows the official CoNLL 2018 scorer: UFeats counts universal features only, LAS ignores subtypes, and CLAS, MLAS and BLEX are reported. --strict adds the full-feature and full-relation variants and --official PATH runs the official script on the same files.
--max-cost / max_cost: stop starting agent tasks at an estimated spend; run again to continue.
A conventions hint that tells the FEATS agent where the treebank writes combined feature values.
Cost estimates for non-Anthropic models (through genai-prices), and extras for other providers (openai, google, bedrock, groq, mistral, xai, openrouter, all-providers).
Documentation: models and providers, results, troubleshooting. A project logo.
LICENSE, CONTRIBUTING.md, package metadata and a py.typed marker.

When an agent task needed a retry, answers for sentences that had already passed were dropped, which left sentences without any HEAD. They are now kept.
Tests no longer depend on terminal colour (they failed under GitHub Actions).

Renamed from llmparser to MCParser. Corpus store, MCP server, agent team, pipeline, rules, baseline, evaluation and documentation.