Skip to content

v0.2.2

Released: 2026-10-09

Headline: project-local environments.

Two problems shaped this release. Adding a source could succeed while its dependencies never got installed, so the first run failed on a missing import with nothing having warned you. And a filesystem load that found nothing to read could still empty a table that already held rows, then report "load complete".

Every project now gets its own uv-managed environment, with source code and dependencies tracked inside the project. A load that finds nothing leaves your table alone, and a run that loaded zero rows is flagged instead of reading as a healthy sync. Around those two: MotherDuck warehouse targets work end to end, --max-records caps every source type, and saving tycoon.yml keeps your comments.

Landed

Each project gets its own environment

  • tycoon init builds a project-local .venv. It seeds a pyproject.toml and runs uv sync right after scaffolding, so there is no separate tycoon setup step to forget. tycoon setup and tycoon doctor --fix build the same way, and tycoon init --no-venv opts out.
  • Source code and dependencies live in the project. A catalog source downloads into <project>/.tycoon/sources/ rather than one ~/.tycoon/sources/ shared by every project on the machine. Its dependencies go into the project's pyproject.toml with uv add, so deleting .venv and running uv sync rebuilds the same environment, source dependencies included. When an install fails, you see uv's own error, so a version conflict and a timeout no longer look the same.
  • Every source's requirements get installed, not just the first one's. dlt init writes a sources directory's shared requirements.txt only once, so a second catalog source's dependencies were never recorded or installed, and its first run failed on an import. The file now holds every source's requirements.
  • tycoon data sources migrate <type> copies a source installed in the old shared location into the project, along with its requirements, and installs them. It never overwrites a directory that already holds your files. tycoon data sources run points at it when it finds a source sitting in the old directory, instead of reporting "not installed".
  • You're told where an install is going. In a project with a hand-made .venv but no pyproject.toml, sources add now warns that a dlt extra will go into the shared environment and names that destination in the prompt, instead of installing there silently.
  • A failed download says why. sources add printed only "Failed to download" whether you were offline or had mistyped the source name. It now shows dlt's own error line, and the retry hint names a command that exists.
  • tycoon doctor tells you which model a project is on. It reports whether the project has its own .venv and whether tycoon is actually running from it. The same warning fires before a run or an install falls back to the old shared behaviour. Nothing forces a migration.
  • --no-prompt installs a source's dependencies. tycoon data sources add <type> --no-prompt used to register the source without downloading or installing anything, because the install step was skipped along with the prompt. It now installs, and a failed install rolls the tycoon.yml change back.

A load that finds nothing leaves your data alone

  • A glob that matches no local files no longer empties the table. Filesystem sources load with replace, so a rerun whose glob matched nothing (a typo, or a folder that emptied out) used to truncate the table and report success. That resource is now left out of the run with a warning, and the other resources in the source still load. If nothing matched at all, the run exits 0 with "nothing to load".
  • A path that names a single file loads that file. Answering data/sales.csv at the path prompt used to produce a source that loaded zero rows, every time. A file path, plain or file://, now reads as its directory plus that file name. A file_glob set alongside a file path is an error, since it could never match.
  • Zero-row runs are flagged. The run_completed ledger event carries zero_rows and a warnings list. A run is flagged when a replace resource loaded zero rows, even if others loaded some, or when a glob matched nothing. An append, merge or incremental run with no new records is a normal sync and is not flagged. tycoon data status measures freshness from the last run that loaded something, and tycoon data history marks a flagged run with a yellow !.
  • --fail-on-empty on data sources run, data sources run-all and data run-all turns those cases into a failure: the command exits 1 with an error naming the table, and the ledger records run_failed. Use it on a schedule, where an exit code is the only thing anyone sees.

Fixes

  • MotherDuck warehouses work with data query, data schema and data clean. With database.warehouse: md:<name> in tycoon.yml, tycoon used to look for a local file of that name, and data clean --all could delete a local file called md:<name>_raw. An md: target now goes to DuckDB unchanged. Commands that still need a local raw database, such as data sources run, stop with a clear error instead.
  • --max-records caps filesystem, rest_api and sql_database sources. These printed the cap and then loaded every row.
  • Saving tycoon.yml keeps your comments, blank lines and indentation. Only the values tycoon changes are rewritten, and a ${ENV} reference stays a reference. tycoon register now saves the same way.
  • Rill source files use relative paths. Generated Rill sources used to carry the absolute path to the Parquet export, which broke on every other clone and in CI.
  • tycoon city builds faster from cold. A cold city.json on an 87-object catalog dropped from 43.6s to 10.3s, with identical lineage.
  • A quoted motherduck_token is masked in printed errors. motherduck_token='...' used to pass through the redaction with the token intact.
  • The data run-all docs list the options the command has. The page described flags that didn't exist.

Upgrading from 0.2.1

  • An existing project keeps working as it did. Until it has its own .venv, sources stay in the shared ~/.tycoon/sources/ and installs go where they went before. tycoon doctor warns about this. To move a project over, run tycoon setup, then tycoon data sources migrate <type> for each catalog source it uses.
  • --no-prompt now installs dependencies. A CI job that ran tycoon data sources add ... --no-prompt and then installed packages itself can drop its own install step. If you relied on --no-prompt not touching the environment, that has changed.
  • tycoon data status may show an older Last Sync. A zero-row run no longer counts as a sync, so a source whose recent runs loaded nothing shows the last run that did. Ledger events written before 0.2.2 read back as not zero-row.
  • A filesystem run that matches nothing exits 0 with "nothing to load", not "load complete", and leaves the table as it was. A script that parsed the old message needs updating. Add --fail-on-empty if you want it to fail instead.
  • A negative --max-records is now a usage error.
  • A MotherDuck query is not opened read-only. DuckDB supports read_only only for local files, so what data query can change on MotherDuck depends on your token's permissions.

Known limitations

  • The legacy-table warning does not reach multi-resource sources. Migrating a flat filesystem source to the resources: shape leaves the old generic table behind without a warning. Tracked in #244.
  • The city is a plan, not yet a simulation. Vehicles and guests are simulated but not drawn, and coverage milestones are not surfaced. The fixes (#250, #251) and the renderer's CI gate (#256, #252) move to 0.2.3.
  • Remote object storage is still unsupported. s3://, gs:// and az:// are recognised only well enough to be skipped by the local-glob check.

For contributors

  • The build is reproducible. hatchling is pinned in [build-system] requires, since an unpinned upgrade changed the wheel's metadata version between 0.1.11 and 0.2.0 and the publisher rejected it. The sdist now holds src/ and the package metadata only, down from 2.9 MB to 523 KB.
  • The release docs match the publish workflow. docs/publishing-to-pypi.md describes the four jobs that actually run and the fix-forward-and-re-tag procedure for a failed publish.
  • A root CLAUDE.md points coding agents at CONTRIBUTING.md, the PR template and the writing-style skill. It holds no rules of its own.
  • Merging into a release branch is documented. CONTRIBUTING now covers merge commits, pinning a merge to the reviewed head, merging stacked PRs and checking a CHANGELOG after a rebase.
  • Upgrade safety is part of the contributor guide. CONTRIBUTING covers testing that a change doesn't break an existing project, and the PR template asks for a short mental model of each change: input, change, outcome.
  • Project tracking moved to GitHub. Issues, per-release milestones and the public Tycoon CLI project board are the record of what's planned. The Jira board is no longer updated. The decision is in Discussion #353.

What's next

0.2.3 picks up the city stack (#248 through #256): the renderer's tests in the default pytest run, the vehicle, guest and achievements fixes, and a CI gate for the web front end. It also brings #260, which creates a new dbt project inside the tycoon project instead of next to it.