v0.7.1-beta

What's Changed

• HTML parsing support by @rubyTanuki in #104
• registry seperation of concerns by @rubyTanuki in #108
• Module structure full redesign by @rubyTanuki in #109

Full Changelog: v0.6.0...v0.7.0

v0.6.0-beta

What's Changed

• file watcher incremental update tests by @rubyTanuki in #94
• Feat/ollama llm strategy by @Krshs90 in #95
• Refactor init to split config and parsing by @rubyTanuki in #97
• fix: cache carryover called during parse by @rubyTanuki in #98
• Add version control lockfile export command by @rubyTanuki in #99
• config piping llm strategy by @rubyTanuki in #100
• add-agent and remove-agent instructions commands in CLI by @rubyTanuki in #102

New Contributors

• @Krshs90 made their first contribution in #95

Full Changelog: v0.5.0...v0.6.0

v0.5.0-beta

What's Changed

• no-llm api key support by @rubyTanuki in #90
• detect language at parse by @rubyTanuki in #91
• Many misc bug fixes by @rubyTanuki

Full Changelog: 0.4.3...v0.5.0

v0.4.3-beta

Fix missing tokenizers dependency in package metadata

v0.4.2-beta

Fixed user visiblity toward huggingface download on the first usage of the embedding model.

Ensured that no partially downloaded .onnx file can be saved when stopped with a keyboard interrupt

v0.4.1-beta

Fixed catastrophic failure where embedding model weights failed to get packaged in the pip install.
The model weights are now pulled from huggingface on first use to avoid this issue.

v0.4.0-beta

What's Changed

• Add python parser by @rubyTanuki in #84
• Describer Visitor pattern to remove complexity from models.py by @rubyTanuki in #87
• Implement python dependency resolution by @rubyTanuki in #88

Full Changelog: v0.3.0...v0.4.0

v0.3.0-beta

What's Changed

The output rendering for both MCP and CLI fully overhauled to utilize rich terminal rendering. tostr.serializer refactored to build generic recursive dataclasses containing the info needed for display by either entrypoint.

tostr init refactored to display live-updating progress bars (uses rich.progress) to track completion and time-to-complete during the initialization of a project.

README updated to reflect recent changes, as well as display a demo gif and images for CLI examples. Full installation and configuration instructions were also added here.

Full Changelog: v0.2.0...v0.3.0

v0.2.0-alpha

Initial public release for closed testing.

Tostr is designed to solve LLM context bloat by mapping codebases into a combined syntax tree and dependency multigraph, allowing AI agents to surgically retrieve context rather than grepping massive files.

🚀 Core Features in v0.1.0

• Pre-compiled AST Caching: Scrapes your OOP project on initialization into a lightweight, local SQLite .db cache.
• Semantic Dependency Graph: Resolves dependencies semantically, building a graph that allows agents to traverse inbound (<), outbound (>), and heuristic (~) method calls.
• Dual Interface (CLI & MCP): * Full CLI access for human developers (tostr skeleton, tostr inspect).
  • Built-in Model Context Protocol (MCP) server for direct integration with agentic tools (like Claude Desktop).
• Automatic Incremental Diffs: The MCP daemon watches your file system and re-scrapes only the specific AST subtrees that change on save.
• Token Circuit Breakers: Hard --max-lines limits and truncation warnings protect agent context windows from accidental bloat.

⚠️ Known Limitations (Alpha)

As this is an initial 0.1.0 release, there are a few limitations we are actively working on:

• Currently functional only for Java.
• Semantic search currently only supports onnx-compiled MiniLM-L6-v2
• Description generation uses the Gemini API, requiring a google developer gemini API key saved as an environment variable GEMINI_API_KEY

📦 Installation

The recommended way to install Tostr globally without polluting your system Python environment is via pipx:

pipx install tostr