Regime detection

lazytools.connectors.regimes.RegimeTools exposes the HMM / Markov-switching regime engines of lazystats.regimes as regime_* LazyBridge tools. lazystats.regimes is designed around plain functions with Annotated[type, "description"] parameters, so each tool is a thin Tool.wrap — the math lives in lazystats, the LLM surface (and its output caps) lives here.

Unlike the stateless datahub_* tools, regime fitting is session state: loaded series, fitted results, and parameter sets live in an in-process or SQLite-backed store shared across calls within one agent run — that is the point of the regime_store_* / regime_db_* / regime_params_* tools.

Tool surface

Read (always exposed): regime_scan_state_counts, regime_get_current, regime_get_changes, regime_get_summary, regime_compare_emission_models, regime_compare_windows, regime_store_list, regime_store_load, regime_params_list, regime_params_load, and depot inspection (regime_db_list_series, regime_db_get_series_info, regime_db_list_results, regime_db_get_result_summary, regime_db_list_plots, regime_db_compare_results, regime_db_get_state_sequence).

Write (require RegimeTools(allow_write=True)): regime_load_from_datahub, regime_fit, regime_fit_categorical, regime_fit_window, regime_apply_params, regime_generate_plots, regime_store_delete, regime_params_save, and depot mutation (regime_db_export_plot, regime_db_delete_series, regime_db_delete_result).

Output is hard-capped at the bridge (state sequences ≤ 500 timesteps, change lists ≤ 200 entries, truncated flag reported) so a full posterior matrix can never flood an LLM context.

Data path

The only data loader on the agent surface is regime_load_from_datahub — symbols/dates/frequency in, a bounded summary out, market-data-hub as the sole source. The file-path loader in lazystats.regimes is a notebook-only building block and is not reachable through this connector: an agent never chooses what file to read.

Install and add to an agent

pip install "lazytoolkit @ git+https://github.com/selvaz/LazyTools.git"
pip install "lazystats[regimes] @ git+https://github.com/selvaz/LazyStats.git"
pip install "market-data-hub @ git+https://github.com/selvaz/market-data-hub.git"

from lazybridge import Agent
from lazytools.connectors.regimes import RegimeTools

# Read-only inspection of existing fits/depot:
agent = Agent("claude-opus-4-8", tools=[RegimeTools()])

# Full workflow — load from the hub, fit, summarize, plot:
agent = Agent("claude-opus-4-8", tools=[RegimeTools(allow_write=True)])

lazystats is imported lazily: the provider imports without it, and a missing install raises an ImportError with the [regimes] install hint.