Datus-agent

Generate MetricFlow metrics from natural language business descriptions

Author MetricFlow semantic model YAML from database tables with validation and Knowledge Base publishing

Optional semantic-model profiling workflow that mines historical SQL and bounded column distributions before YAML authoring

Build the project's vector-indexed knowledge base from files plus database metadata — optionally scoped to specific files / tables / datasources / domains. Scan the in-scope material, classify it into business domains, explore each domain's tables and docs in parallel with explore subagents (the validated-query SQL corpus is enumerated directly, no explore needed), then (after the user confirms a generation manifest — or directly, in the same turn, when the user has waived confirmation) route every artifact to its store via storage-classify, generating semantic_models / metrics / reference_sql (and mining any extra knowledge), and refresh AGENTS.md's KB index. The lightweight /init handles the AGENTS.md inventory plus file-based knowledge/memory; this skill owns the heavy vector-store generation.

Create new Datus skills from scratch. Use when users want to build a new skill, scaffold a skill directory, or capture a workflow as a reusable skill. Trigger phrases include "create a skill", "make a skill for", "turn this into a skill", "new skill".

Activate when the gen_job agent detects that the source and target databases differ. Covers cross-database transfer lifecycle - type mapping via adapter Mixin hints, DDL generation, data transfer via transfer_query_result, and lightweight reconciliation.

Mine shortest atomic facts from (question + gold_sql) pairs into ./knowledge/*.md; either by simulating SQL drafting (lite) or by driving the gen_sql subagent in blind iteration (deep)

Create database tables from SQL (CTAS) or natural language descriptions

Lightweight project initialization — optionally scoped to specific files / tables / datasources / domains. Infer the project goal and in-scope datasources, scan the in-scope file tree and database metadata (db/table/desc/sample), classify into business domains, then write an AGENTS.md inventory skeleton plus the cheap file-based stores (atomic facts to ./knowledge/*.md via lite extract-knowledge, durable preferences to memory). Stops short of the expensive vector-indexed stores (semantic_models / metrics / reference_sql). Single confirmation-free pass, low token cost.

Validate the column contract of a newly written table — column set, types, and nullability match expectations. Object existence and row counts are handled by the builtin layer and are out of scope. Data-content assertions belong to project-level validator skills.

Lightweight post-transfer reconciliation example — verify tool-reported row count parity and run a small target-side sanity check without re-scanning the source.

Profile data quality and distributions

Generate formatted data reports from SQL query results

Decide where a produced artifact must be persisted before writing it, then route it the prescribed way — semantic_models / metrics / reference_sql via the matching task() subagent, knowledge via extract-knowledge (lite), memory via add_memory, skills via create-skill, and AGENTS.md edited directly. Load before persisting any business fact, validated SQL, metric/model definition, session preference, project convention, or reusable workflow.

Audit and reorganize every persistent store — semantic_models, metrics, reference_sql, knowledge, memory, AGENTS.md, skills — verifying each item sits in the correct store per storage-classify, and surfacing duplicates, misclassifications, conflicts, and stale/erroneous entries. Produce a Remediation Plan, STOP for confirmation, then execute. Use ask_user only for genuine decisions during analysis. If nothing needs fixing, report it and stop.

Review the current chat session and persist its valuable takeaways — business facts/rules, validated SQL, metric/model definitions, durable preferences, project conventions, reusable workflows — by classifying each via storage-classify and routing it to the right store. Present a Summary Manifest and STOP for confirmation before any heavy generation. Use at the end of a working session or when the user asks to capture what was learned.

Execution guide for Airflow scheduled jobs — troubleshooting, updating, conn_id conventions, and cron references

Scheduler validator driven by ValidationHook — read-only static verification of scheduled jobs (schedule correctness, configuration, runtime context already collected by deterministic hook). Does not trigger test runs.

Create, view, and manage Grafana dashboards with panels and datasources

Create, view, and manage Superset dashboards with charts and datasets

BI validator driven by ValidationHook — inspects every dashboard, chart, and dataset delivered in the run, verifies config quality (chart type, metrics, dimensions, dataset wiring) and data presence via get_chart_data when supported

Optimize and improve existing Datus skills. Use when users want to edit a skill, improve its instructions, optimize its description for better triggering, or analyze skill performance based on usage sessions. Trigger phrases include "optimize skill", "improve skill", "edit skill", "fix skill", "skill not triggering".

Administrative database tools (restricted access)

Guided workflow for SQL data analysis using db_tools

SQL query optimization techniques and best practices