v0.44.0

0.44.0

Breaking Changes

• RobotoContextTooLongException no longer accepts estimated_tokens and max_tokens constructor arguments, no longer exposes estimated_tokens / max_tokens properties, and no longer extends the to_dict() payload with those fields. The exception now serializes as the standard {error_code, message} shape like every other RobotoDomainException. The internal heuristic estimate and the model's context limit were never billing-grade values and were not consumed by any in-tree caller; they remain visible in CloudWatch failure logs (see BedrockLLMBackbone). Callers that need usage telemetry will get it through a separate, dedicated channel rather than by introspecting the error.

Features Added

• Sessions are queryable (experimental): Sessions join datasets, files, topics, and events as a queryable resource. You can sort and paginate listings server-side by start time or duration. Higher-level query helpers in the SDK will arrive in a follow-up release. Sessions remain experimental, and Roboto enables access per organization; contact us to opt your org in.

Bugs Fixed

• Concurrent Topic.get_data_as_df calls no longer race on cache-directory creation (which previously raised FileExistsError from mkdir) or on Parquet downloads. Within a process, downloads of the same representation are deduped via a per-path lock; across processes, an atomic temp-file-plus-rename write guarantees readers never observe a partial file.