cohortBuilder

• Custom filter types can now be registered via register_filter_type(), enabling extensions without modifying core package code.
• New describe() helper to attach metadata (descriptions) to datasets and filters. describe() also accepts a label argument; when set on a variable, autofilter() reuses it as the generated filter's name.
• shape() filter entries now include a name field, and the description field combines the filter- and variable-level descriptions.
• New shape() generic extracts structured filter/dataset metadata from a source, including statistics (min/max for range, choices for discrete).
• New autofilter() generic auto-generates filters based on column types (character/factor → discrete, numeric → range, Date → date_range, POSIXct → datetime_range). Supports attach_as = "step" (add as filtering step) or attach_as = "meta" (store as available filters).
• New .class parameter in tblist() to prepend custom S3 classes for method dispatch customization.
• Steps now track pending status — only pending steps trigger statistics recalculation, improving performance for multi-step workflows.
• update_filter() now supports pre/post hooks via hook_args.
• New Cohort$new() propagate_domains argument controls how filter domains are narrowed between steps: "none" (default), "filter" (from previous step filter values), "stats" (from stored statistics; requires compute_stats = TRUE), or "data" (scan filtered data; the stats-free equivalent). Backed by the .propagate_domains() source method and the S7 generics cb_intersect_domain(), cb_intersect_domain_values(), cb_domain_from_stats(), and cb_domain_from_data(). Filter values are now intersected against their domain (with trimming messages) before filtering.
• set_source() gains a compute_meta_stats argument (default from the cb.source_filters_meta_stats option) controlling whether metadata statistics for available_filters are pre-computed; when FALSE, filter domains fall back to live computation. Sources also accept an available_filters definition directly.
• Filter ids are now deterministic by default (derived from dataset and variable names), enabling cross-step filter matching for domain propagation. Override with explicit id =.

AI/LLM integration

• New cb_tool() system for defining LLM-compatible tool specifications (requires ellmer).
• Built-in tools: cb_tool_filters_meta(), cb_tool_add_filters(), cb_tool_set_filter_values(), cb_tool_apply_filters() (combined add + set values).
• cb_register_tool() and cb_register_tools() register tools with an ellmer chat object.
• AI tool invocations can be traced by setting options(cb_tool_verbose = TRUE), which logs the invoked tool and its arguments via message(). Logging is off by default (replaces earlier unconditional print() debug output).

Improvements

• Use collapse for binding operations (joins), with verbose option for diagnostics.
• Reorganized breaks argument for date_range filter plots.
• Extensive test coverage improvements including vdiffr snapshot tests for all filter plot types.
• cohort() now exposes the compute_stats and propagate_domains arguments, and add_filter()/rm_filter() now expose and forward the hook argument (previously these were silently ignored).
• Comprehensive roxygen documentation added across exported and internal functions, plus a new source-intelligence vignette covering describe(), autofilter(), shape(), and the AI/LLM tools.