Skip to content

0.9.0

Choose a tag to compare

@chauduyphanvu chauduyphanvu released this 18 Nov 21:31
· 26 commits to main since this release

The language for models and tools. The plugin works as in 0.8.0; what is new is everything around it that lets an AI assistant write, check and run strategies.

  • A guide any model can read. llms.txt and llms-full.txt in the repository give a model the whole language: what a block is, every setting, rule and value, what not to do, worked examples and how to check a block. They carry the same text the drafting command sends, made from the plugin's code, so they list exactly what the parser takes.
  • Worked examples and mistakes. docs/corpus/ holds requests in plain English with the blocks that answer them, using every word of the language, and blocks with a common mistake, the parser's exact messages and the fix. Every block is checked by the tests.
  • A word list and grammars. docs/language.json has every function, setting, value and word with its form, meaning and example. docs/grammar/ has the text of a block as a Lark grammar (OpenAI's grammar-constrained tools, vLLM, llguidance) and a GBNF grammar (llama.cpp), for models that generate under a grammar.
  • A command line and MCP server. backtest-blocks.mjs, attached to this release, runs with Node.js 20 or later. check checks the strategy blocks of a note and prints each problem with its line, run backtests a block on live prices, explain says what a word means, and mcp serves the same as tools for Claude Code, Claude Desktop, Codex or any MCP client, with the guide, examples and grammars as resources and a write_strategy prompt.
  • Tried from a terminal and over MCP with live prices; its figures match the plugin's for the same block.