Sphinx 0.8 — Plan Mode, Engines, and More!

Sphinx 0.8 is now available! This update contains a myriad of features that make Sphinx more powerful than ever. Here’s the list:

• Plan Mode: Sphinx can be ran in Plan Mode to generate a thorough strategy for how to approach complex or underspecified tasks. Note that in Plan Mode, Sphinx may still write and execute several cells before generating the plan in order to better understand the attributes of the available data first. Sphinx can also ask the user questions, and the plan it generates is fully user-editable
  
• Sphinx Engines: Users can now select from one of three different Engines. An Engine defines a whole configuration scaffold for Sphinx, consisting of a set of models, settings, and tuning for the agent. With 0.8, Sphinx supports three agents. Thebes, the original Sphinx Engine backed by GPT 4.1; Siwa, our fastest Engine, backed by Claude Haiku 4.5; and Cairo, our most powerful Engine, backed by Claude Sonnet 4.5.
• Jupyter-aware File Search: A long-requested feature, the Sphinx agent is now able to perform file system read operations. These operations are performed in the file system of the connected Jupyter kernel. This matters in situations where a remote kernel has been selected, such as when using Google Colab.
  
• Structured Output in CLI: The Sphinx CLI now supports structured output via the --output-schema flag. This can be used to force the Sphinx CLI to output a JSON object that satisfies a given schema in cases where you are trying to extract certain fields as the output of an analysis. This is particularly useful when the Sphinx CLI is used as part of automated data processing pipelines!
• Interactive Mode for Sphinx CLI: The Sphinx CLI can be ran with no arguments to enable interactive mode. This mode will feel familiar to users of other CLI-based copilot tools like Claude Code, and is useful for users who are more comfortable using copilots from the confines of a terminal.
• Runtime Interrupt Tuning: Sphinx has long had the ability to interrupt cells that are taking too long to execute. This is important to stop Sphinx from getting blocked by cells that have logical errors that lead to hangs. However, we have recently found that this is overtuned, and Sphinx has been erroneously interrupting cells that are expected to take a long time to execute. We’ve dialed back Sphinx’s aggressiveness here, and this feature can also be disabled via a VSCode setting.