Changelog¶

This page summarizes the major changes and additions introduced by the recent JUNIPER refactor.

Major Refactor¶

Step constructors now use explicit parameters instead of a single parameter dictionary.
Most commonly used classes and error types are re-exported from juniper.
Configurable helper objects such as Gaussian, LateralKernel, FrameGraph, and Transform still use parameter dictionaries.
Top-level architecture access is centered on get_arch() and delete_arch().
Simulation is run through Architecture.run_simulation(...); there is no public arch.tick() method in the current API.
Logging setup helpers are now exposed as init_logging(...) for console output and init_logging_to_file(...) for file output.

Added reusable nested Circuit support with context-manager construction.
Added named slot connection support through slot attributes and string paths.
Improved connection validation for duplicate connections, invalid endpoints, and max fan-in limits.
Element paths now carry nested circuit structure and are used by recordings, compile info, and errors.

Added a clearer backend split around Compiler, Engine, RuntimeState, CompileInfo, and Recording.
Compilation now gathers static, dynamic, source, sink, kernel, shape, and dtype metadata before runtime execution.
Runtime state is stored as a flat path-keyed tree with helpers for reading slots, writing sources, and recording targets.
The engine manages source pushes, PRNG key updates, JAX-jitted ticks, sink pulls, recordings, timing, and permanent buffer save/load.
Kernel state returns are normalized against the compiled state contract and raise EngineError on missing keys or shape mismatches.

Added domain-specific exception and warning hierarchies.
Compile failures now include traced upstream causes where possible, including unresolved inputs, unresolved buffers, cycles, shape inference failures, and dtype inference failures.
Added dedicated recording load/save errors and TCP errors.
Added library-wide Python logging hooks for compiler, engine, frontend, TCP, and step diagnostics.

run_simulation now returns a Recording object and timing dictionary.
Record targets can be strings, elements, slots, or buffers.
Recording supports filtering by element, filtering by time interval, slicing, appending, plotting, saving, and loading.
Saved recordings use a recording manifest plus one pickle file per time step, allowing compatible batches to append to an existing run directory.