Tool catalog¶

The full routing rule (question type → tool) is the decision ladder in the workflow playbook.

Tool	Scope	Use when
`get_playbook`	`read`	Use FIRST in a fresh session: the Analyst playbook + schema reference.
`list_resources`	`read`	Use when resolving a NAME to an id, or browsing what exists.
`get_resource`	`read`	Use when you have an id and need the full detail object.
`get_workspace_brief`	`read`	Use first in any piece of work on a workspace: where it stands, what could come next.
`describe_catalog`	`read`	Use before query_sql or a recipe: the tables you can read, and their columns.
`query_sql`	`read`	Use for rows and aggregates: ONE read-only SELECT over the lakehouse.
`profile_table`	`read`	Use for one-call table triage before writing SQL.
`get_eda`	`read`	Use for dataset triage: profile, alerts, leakage — before any SQL.
`get_autodoc`	`read`	Use for big-picture orientation: the AutoDoc model card.
`get_capabilities`	`read`	Use before deep reads: which analytical surfaces an experiment supports.
`get_model_evaluation`	`read`	Use for numeric model quality: metrics, curves, confusion — not prose.
`get_insight`	`read`	Use for computed analysis artifacts beyond charts and model evaluation.
`get_rules`	`read`	Use for IF-THEN rule cards of a supervised experiment, with Train and Test numbers.
`get_chart_data`	`read`	Use for curated chart series — the numbers behind a chart, never a figure.
`get_run`	`read`	Use to observe a run: is it running, what failed and why, what config ran.
`describe_platform`	`read`	Use when planning: the config vocabulary ParleML offers, never invented.
`compare`	`read`	Use to compare like with like — and to answer "how does it compare with production?".
`probe_dataset`	`compute`	Use for cheap pre-flight probes on a dataset, before proposing a run.
`what_if`	`compute`	Use to score ONE row and see why: "what would the model say if ...".
`simulate_decision`	`compute`	Use to choose an operating point: what threshold / cost / capacity yields.
`prepare_data`	`builder`	Use to make a dataset ready to train: column roles, EDA, leakage triage.
`author_recipe`	`builder`	Use to shape a training table: a recipe is your own SELECT over the knowledge base.
`run_experiment`	`builder`	Use to train: plan an experiment, launch the reviewed plan, then explain or publish it.
`score_batch`	`builder`	Use to score a dataset with a trained experiment, and read the results.
`register_model`	`builder`	Use to register a model, record what a person approved for a release, or import predictions.
`update_workspace`	`builder`	Use to record the approved framing's goal, intended use, oversight and risk level on the workspace.

Coverage matrix¶

What the platform computes → how agents reach it. This table is the standing completeness audit.

Capability	Agent surface	Engine
The playbook: decision ladder, naming rules, schemas	`get_playbook`	reference markdown
Queryable tables, one experiment's star schema, object previews	`describe_catalog`	the workspace knowledge base
Lakehouse SQL over results ↔ inputs	`query_sql`	workspace DuckDB views
What an experiment supports, and what is ready now	`get_capabilities`	mli / forecast / unsupervised capabilities
Table triage before SQL	`profile_table`	DuckDB SUMMARIZE
Dataset EDA / quality / leakage	`get_eda` + `eda.*` charts	mli EDA
Model quality (metrics, curves, confusion)	`get_model_evaluation`	mli evaluation
Leaderboard / trust / fairness / drift / errors …	`get_insight`	mli + analysis
Per-row why	`get_insight` (kind=case) + `*__rules_fired` SQL	mli decision support
IF-THEN rule cards (+ redundancy QA)	`get_rules`	mli rule extraction
Recommendation / ranking / LTR	`get_insight` (recommendation_* / ranking / ltr)	ml4 ranking engines
Forecast / unsupervised / association digests	`get_insight` (forecast_ / unsupervised_ / association_*)	ml4 engines
Curated chart series (+PNG) and how to read them	`get_chart_data`	mli/ml4 figure builders
Model card prose	`get_autodoc`	mli AutoDoc
Run status, failures, logs, config used	`get_run`	experiment runner telemetry
Config vocabulary (presets, templates, metrics …)	`describe_platform`	ml4 describe
Experiment vs experiment / production; model vs model	`compare`	backend comparison
Dataset drift vs a reference (compute)	`compare` (kind=datasets)	dataset drift PSI
Pre-flight probes: signal, forecastability, baseline floor, temporal split (compute)	`probe_dataset`	ml4 probes
One-row what-if scoring, supervised or unsupervised (compute)	`what_if`	mli justify + pipeline
Operating points: thresholds, costs, capacity — nothing applied (compute)	`simulate_decision`	mli decision support
Navigation / ids / deep links / running jobs	`list_resources`, `get_resource`, `ui_url`	backend catalog
Where a workspace stands; resuming work	`get_workspace_brief`	the workspace brief
Column roles, EDA, leakage triage (builder)	`prepare_data`	dataset services
Training tables as recipe SQL (builder)	`author_recipe`	knowledge-base recipes
Plan, launch, explain, publish (builder)	`run_experiment`	ml4 engines via reviewed plans
Batch scoring (builder)	`score_batch`	batch predict jobs
Registration, release records; predictions import (builder)	`register_model`	the model registry; a run's release record
The approved intended use & oversight (builder)	`update_workspace`	the workspace record
Deliberately NOT exposed to agents	promoting, deleting, settings, file downloads, writing annotations (AI-written text)	a person, in the parleml UI