[liam-machine]

Why

The Physical stage only compares against dbt today. To decide which other sources to support next (a DBML file, an information_schema snapshot, SQL projects, …), we need to know how ERD Studio users actually build their tables and on which platform.

What

Two optional heartbeat fields, each a subset of a fixed allowlist (contract stays additive, HEARTBEAT_VERSION stays 1):

Field	Values
physicalTools	dbt, sqlmesh, dataform, databricksBundle, fabric, powerBi, sqlProject, flyway, liquibase, alembic, prisma, django, rails, notebooks, ddlScripts, dbml
platforms	databricks, snowflake, bigquery, postgres, redshift, fabric, synapse, sqlserver, duckdb, spark, trino, athena, mysql, oracle, clickhouse, other

Absent = not looked that day (stored NULL); [] = looked, found nothing.

• src/services/physicalTooling.ts: pure, vscode-free, every file access through an injected probe. Detection uses marker file names plus a fixed pattern in the head of a few small files: CREATE TABLE, Databricks notebook header, SQLMesh config keys, a SQL project's <DSP>, dbt-<adapter> packages, and .dbml files that aren't our own export. The dbt warehouse is read from the first 8 KB of manifest.json (metadata.adapter_type). profiles.yml is never read.
• extension.ts: recordPhysicalTooling(context) runs 15 s after both activation outcomes, once per workspace per UTC day, and not at all while telemetry is off. Not carried over by nextDay, because it describes workspaces, not the machine.
• Worker: allowlists, validation, upsert, D1 migration 0003_physical_tooling.sql (applied by deploy.yml before the Worker deploy), tests.
• telemetry.json, telemetry/README.md (contract + two ready queries) and CHANGELOG updated.

Privacy

Only allowlisted values leave the machine: never a path, name, count or content. test/unit/physicalTooling.test.ts runs detection over an in-memory workspace through the real include/exclude globs, including one full of secret-looking names, and asserts nothing else comes back. The payload privacy test feeds junk through recordPhysicalTooling too.

Testing

• npm run compile, npm test (192 files, 4162 tests), npm run build: green
• telemetry/: npm test (160) + npm run check: green
• Ran the bundled detector over real local repos (Databricks bundle, Power BI PBIP, Alembic app, dbt fixture): correct flags; about 1.2 s of background ripgrep I/O on a 10.5k-file repo, 6–70 ms on normal ones
• Migrations 0001–0003 plus the README queries run against a throwaway SQLite DB

Note

ERD Studio activates only on workspaceContains:**/dbt_project.yml, so non-dbt repos appear only when someone opens ERD Studio there. The data will skew towards dbt.

🤖 Generated with Claude Code