v0.14.0

• VSCode Extension: when a test is stuck on a query, a new command allows seeing diffs with all unreachable queries of the same type, which is very useful for repairing demonstrations after a strategy change.
• Add delphyne browse command for generating browsable traces from raw traces.
• Add Run nodes that subclass Branch for extracting a single element from an opaque space without branching.
• Breaking: Overhaul example selectors for better flexibility (see select_examples argument of few_shot).
• Breaking: Overhaul Stream to make it an iterable directly and remove Stream.gen.
• Breaking: Overhaul the Experiment class to make it more ergonomic.
• Implement support for embeddings and embedding-based example retrieval (including the standard MMR algorithm).
• Remove auto_reload setting, which is fundamentally unsafe.
• Experimental: add support for feedback backpropagation.
• Add init key to delphyne.yaml to register custom initializers.
• Remove automatic reloading of Python modules, which is very error prone. Users should manually restart the server instead.
• Allow replaying failing experiment configurations in a debugger.

v0.13.0

• Breaking: Change signature of dp.compute to allow passing additional keyword arguments. You must now write dp.compute(fun)(*args, **kwargs) instead of dp.compute(fun, *args, **kwargs).
• Add override_args argument to dp.compute and dp.elim_compute. This allows policies to override timeout arguments for tool calls.
• Add elim_values and binarize_values tree transformers.
• Add new policy for handling Abduction nodes: abduct_recursively.
• Add a Data effect for loading evolving external data.
• Fix summary generation bug in experiment launcher.

v0.12.0

• Add a take test instruction to the demonstration language for manually specifying an action to take at a given node. This ensures that the demonstration language is complete (any success node can be reached via a demonstration).
• Breaking: improve navigation function for Abduction nodes to make it (conditionally) invertible.
• Add workers_setup argument to Experiment.
• Experimental: add support for external answer sources in demonstrations and in the run_strategy command, and for hindsight feedback.
• Fix path bug in experiment launcher.