0.17.0

This is the biggest EOS update to date, bringing a brand-new web UI for EOS, a powerful new default optimizer called Beacon combining traditional methods like Bayesian optimization with cutting-edge methods like LLMs, an MCP server for connecting AI agents to EOS with 50 tools, and many improvements to the EOS scheduler and EOS internals.

• New web UI for designing experiments visually, submitting tasks/experiments/campaigns, monitoring optimization, inspecting devices, browsing files, streaming logs, and managing packages.
• New Beacon optimizer that combines Bayesian optimization with LLM reasoning.
• New MCP server allowing AI assistants like Claude to connect to EOS and interact with EOS.
• New device and resource holds in the scheduler to keep allocations locked between tasks so other experiments can't claim them mid-workflow.
• New scheduling simulator (eos sim) to simulate experiment scheduling offline without hardware.
• New REST API endpoints for optimizer state, log streaming, and package management.
• Error messages are now stored for failed tasks, experiments, and campaigns.
• On-demand tasks now go through the scheduler, improving efficiency and stability.
• Optimizer sampling is now async and no longer blocks the orchestrator loop.
• CP-SAT scheduler runs in a thread executor to avoid blocking the event loop.
• Campaign optimizer parameters can be overridden at submission time
• EOS now lazy-loads optimizers, reducing startup time and memory usage.
• Fixed a bug that caused Ray to consume large amounts of disk space.
• Refactored parts of the codebase.
• Improved performance in several parts of the codebase.
• Various bug fixes and stability improvements.
• Updated documentation.
• Updated dependencies.