Skip to content

changelog.md

Latest commit

 

History

History
127 lines (81 loc) · 5.15 KB

changelog.md

File metadata and controls

127 lines (81 loc) · 5.15 KB

Changelog

All notable changes to machine-core.

v0.4.1 (2025)

Mixed-mode tool filtering

  • Added filter_mcp_toolsets() standalone function — wraps MCP toolsets with ToolFilterWrapper to hide irrelevant tools based on a set of relevant names
  • Added toolsets= parameter to AgentCore.rebuild_agent() — allows swapping MCP toolsets alongside dynamic tools in a single call
  • Exported filter_mcp_toolsets from the top-level package

This enables a "mixed mode" where both OpenAPI tools and MCP toolsets can be RAG-filtered per request.

Commit: 57a2eaa

v0.4.0 (2025)

Shared infrastructure migration from ai-accounting-agent

Migrated reusable components from ai-accounting-agent into machine-core as shared infrastructure:

  • FileProcessor (core/file_processor.py) — static utility for PDF text extraction (pdfplumber + PyPDF2 fallback), image OCR (pytesseract), VLM data URL preparation, and batch file upload processing
  • ProcessedFile — dataclass for file processing results (text, data_url, mime_type, pages, error)
  • generate_tools_from_openapi (core/openapi_tools.py) — generates pydantic-ai Tool objects from OpenAPI 3.x specs with automatic $ref resolution, schema depth limiting (3 levels for Gemini), and auth header injection
  • fetch_openapi_spec — async function to fetch OpenAPI specs from API endpoints
  • VectorStore (core/vector_store.py) — LanceDB-backed vector database with table management, single-table and cross-table similarity search
  • Embedder — wrapper around embedding providers (supports ResolvedEmbedding and raw providers)
  • SearchResult — dataclass for vector search results
  • ToolFilterManager (core/tool_filter.py) — RAG-based tool filtering that indexes OpenAPI and MCP tool descriptions, then filters by task prompt relevance
  • ToolFilterResult — dataclass for filter results with tool names grouped by source
  • DocumentStore (core/document_store.py) — high-level facade for document storage with automatic embedding and natural language search

Additional changes:

  • BaseAgent._process_image() now delegates to FileProcessor.prepare_for_vlm()
  • Updated __init__.py to export all 16 public symbols
  • New dependencies: lancedb>=0.20.0, pdfplumber, PyPDF2, Pillow, pytesseract, httpx

Commit: 6d70e4a

v0.3.0 (2025)

Dynamic tools support

  • Added rebuild_agent(tools=, system_prompt=, retries=) method to AgentCore — allows recreating the agent with new tools per request without re-initializing MCP connections or model providers
  • Extracted _build_agent() from __init__() to make agent creation reusable
  • Added run_query_iter() to BaseAgent — async generator yielding (node, step_num) tuples for step-by-step execution control
  • Added _validate_agent_tools() for schema validation on dynamically-created tools

This enables the "dynamic tools mode" where tools can be swapped per request, used by ai-accounting-agent for per-prompt OpenAPI tool generation.

Commit: fc8721a

v0.2.0 (2025)

Provider-agnostic model objects

  • Simplified AgentCore.__init__() to use ResolvedProvider.model directly — AgentCore no longer imports OpenAIChatModel, GoogleModel, or AnthropicModel
  • All model construction logic now lives entirely in model-providers
  • AgentCore stores self.provider_type for downstream logic that needs to know (e.g., schema depth limiting for Gemini)

Commit: 5a1f772

v0.1.8 (2025)

  • Updated version in pyproject.toml and uv.lock
  • Updated model-providers dependency to v0.0.8

Commits: d286ea9, 12f4fca

v0.1.7 (2025)

  • Bumped dependencies
  • ToolFilterWrapper extended to inherit from AbstractToolset and implement all abstract methods (get_tools, id, call_tool)
  • Fixed ToolFilterWrapper.id property to handle missing wrapped toolset ID gracefully
  • Fixed BaseAgent.run_query() to let pydantic-ai handle tool errors properly instead of interfering with retry logic

Commits: d4e4685, a43fbb0, df35141, 6661e7e, a55cbcc

v0.1.3 (2025)

  • Updated Dockerfile for Python 3.13
  • Enhanced deployment documentation
  • Refactored code structure for readability

Commits: ce758dd, 1700581, 835ce34

v0.1.1 (2025)

  • Added support for Google model in AgentCore
  • Updated model-providers Git source URL
  • Various improvements for Google/Vertex provider support

Commits: 5480123, 6b236ab, e5c0357, c80e0f6

v0.1.0 (2024)

Initial release

  • AgentCore — core infrastructure for model/embedding resolution and MCP toolset loading
  • BaseAgent — abstract base class with run_query() and run_query_stream()
  • MCP toolset integration with ToolFilterWrapper for schema validation
  • Configuration via AgentConfig and environment variables
  • 6 built-in agents: ChatAgent, CLIAgent, ReceiptProcessorAgent, TwitterBotAgent, RAGChatAgent, MemoryMasterAgent
  • FastAPI service with Prometheus metrics (/metrics), health check, and static frontend
  • Docker and Docker Compose deployment support
  • Integration with model-providers for 7 LLM + 3 embedding backends

Commits: 2c4749d through 027aeed