v0.0.11

Added

Changed

• README usage documents the health check and how to stop the app

Fixed

v0.0.10

Added

• math-ai-agent --version prints the installed version and exits

Changed

• README simplified to PyPI installation, configuration, and usage; Git
  clone run instructions moved to docs/DEVELOPMENT_SETUP.md
• Build requires poetry-core>=2.5

Fixed

v0.0.9

Added

Changed

• README documents running the PyPI package against the remote
  calculator-mcp on Prefect Horizon with GitHub OAuth
• App logs when the MCP connection opens and the agent is created
• GET / serves index.html as a file; the unused /static mount is
  removed

Fixed

• Tool errors, such as division by zero, and invalid tool arguments are
  returned to the LLM as the tool output instead of failing the prompt
  with HTTP 500

v0.0.8

Added

• docs/NVIDIA-Nemotron-Open-Model-License-12-12-25.pdf

Changed

Fixed

v0.0.7

Added

Changed

• Web page shows 6 rows in the Response box and no gap below the title
• Web page notes that Display Reasoning requires the Responses API
• Web page shows errors, including network and non-JSON responses, in
  red text below the form instead of in the Response box
• llm.reasoning_summary setting (auto, concise, or detailed)
  requests reasoning summaries; reasoning shows the summary when a
  reasoning item has no content
• docs/CLASS_DIAGRAM.md shows 0..1 multiplicity for
  CalcMCPConnection._client
• Token limits, content filters, and failed Responses API requests raise
  TokenLimitError, ContentFilterError, and LLMRequestFailedError;
  POST /prompt/ returns 502, 422, and 502 for them, and logs any other
  error and returns 500
• Responses loop raises ValueError for queued and in_progress
  statuses instead of resending the request

Fixed

v0.0.6

Added

Changed

• LLM request and response logs replace the tool definitions with a
  placeholder message

Fixed

v0.0.5

Added

• llm.stateful setting (default false); when true, the Responses
  loop stores responses and sends only new items with
  previous_response_id
• llm.temperature setting (0 to 2, set to 0.2); when unset, the provider
  default applies
• "Display Reasoning" Yes/No choice on the web page; POST /prompt/
  takes display_reasoning (default true) and, when false, returns
  only the final response

Changed

• prompt.py renamed to payload.py and Prompt to Payload
• to_openai_tools() renamed to to_chat_completions_tools() in
  CalcMCPClient and CalcMCPConnection

Fixed

v0.0.4

Added

• fastmcp.client logger in config_local.yaml and config_remote.yaml
• Debug logs of the full context (instructions, history, tools) sent to
  the LLM on each call

Changed

• With llm.api_style: responses, answers show reasoning: and
  final response: sections, collecting the reasoning from every turn
• docs/PERFORMANCE.md renamed to docs/PYTHON_COROUTINE.md

Fixed

• uvicorn startup, shutdown, and access logs use the config.yaml log
  format instead of uvicorn's default

v0.0.3

Added

• llm.timeout_seconds (default 120) and llm.max_concurrent_prompts
  (default 10) settings; prompts beyond the limit get HTTP 503
• Automatic reconnect and one retry when the calculator MCP connection
  drops during a tool call
• flake8 dev dependency and .flake8 config (80-character lines)
• docs/PERFORMANCE.md explaining the async event loop and shared clients

Changed

• The app opens one calculator MCP connection and one LLM client at startup
  and reuses them for every prompt, instead of reconnecting per prompt and
  per tool call

Fixed

• Web page shows the error message instead of undefined when a prompt
  request fails with an error response

v0.0.2

Added

• GET /health endpoint returning 200 with plain text OK
• math-ai-agent command that starts the web app on the web.host and
  web.port set in config.yaml (default 127.0.0.1:9090)

Changed

Fixed