| # Pinned to the versions the eval used, so apply_chat_template renders the <think>+XML-tool prompt | |
| # IDENTICALLY to training/eval (train<->serve parity). Fast tokenizer => no torch needed. | |
| transformers==5.5.0 | |
| tokenizers==0.22.2 | |
| huggingface_hub>=0.26 | |
| # Gradio 6: app-level theme/css/js moved to launch(), Chatbot buttons=, api_visibility=, MCP server. | |
| # The [mcp] extra pulls in the `mcp` package so demo.launch(mcp_server=True) can serve this Space as an MCP server. | |
| gradio[mcp]>=6.0,<7.0 | |
| # sandbox libs so the agent can produce shareable visual artifacts (charts, GIFs) - used by the agent's | |
| # `python` runs, not by app.py itself (matplotlib uses the headless Agg backend automatically). The agent | |
| # can also `pip install <pkg>` at runtime for anything else (CODEAGENT_BASH_TIMEOUT gives installs time). | |
| numpy | |
| pandas | |
| matplotlib | |
| pillow | |
| # no-key web stack (for the web self-test + future web_search/web_fetch tools): ddgs aggregates | |
| # Google/Bing/Brave/... (no API key); trafilatura fetches + extracts clean page text. | |
| ddgs | |
| trafilatura | |