Skip to content

Data analysis

The web workspace's Pages list shows saved results pages under Analyses. Ask Attaché to analyse a dataset and it can use the optional bundled analysis skill to keep calculations in a persistent Hailer notebook. Attaché remains the chat and uses your existing model configuration; no second Hailer model or API key is needed.

After upgrading an existing installation, run attache init to add the bundled analysis skill, using the same --home if you use a custom home. Existing files are preserved.

For example: “Analyse sales.csv in a saved notebook and show monthly revenue.” Put the dataset in the chat's files folder first and tell Attaché its filename. The skill accepts CSV, TSV, Parquet, JSON, JSONL and Excel files. It copies each selected file into a dedicated analysis workspace and refuses to replace an existing copy.

Saved results and live notebooks

Saved results are HTML pages named analysis-<name>, listed under Analyses. These are snapshots of verified calculations; ask in chat to refresh them after data changes. They use the same sandbox as every other Attaché page.

The live editable notebook opens in a separate browser window on the machine running Attaché. It is not embedded in the workspace. Keep the notebook open while working; opening a notebook establishes the session its calculations use. Tables, charts and Python cells remain saved in the workspace, so Attaché can reopen them later.

This first bridge uses Hailer's versioned runtime and notebook interfaces without exposing the notebook server through Attaché's artifact sandbox. A future inline editor needs a separately authenticated notebook transport and WebSocket proxy, not a relaxation of the rules for generated pages.

Requirements and lifecycle

  • Docker must be installed and running. The adapter uses Docker only, with network access off, copied datasets mounted read-only, and only the notebooks folder writable.
  • The skill declares hailer==0.2.8, which pins marimo==0.24.2, and Python 3.12 or newer. uv run installs these separately on first use. Attaché's core dependencies are unchanged. The first notebook start may also download Hailer's versioned Docker kernel image. If that registry image is unavailable, ask Attaché to build the analysis kernel, or run the skill with args: ["sales", "build"] after initialising the sales workspace. This uses Hailer's bundled image builder with host package configuration disabled, then start uses the locally built matching image. Allow several minutes and about 1 GB of disk.
  • Notebook execution does not receive host model credentials. Summaries returned to Attaché are part of its normal model conversation, so your configured model receives those summaries.
  • The skill is owner-only and cannot be invoked by members or from the owner's group turns.
  • A started Docker kernel persists across short script calls and Attaché restarts. Ask “Stop the sales analysis” when done; saved notebooks and datasets remain. A later start reuses a matching running kernel or starts one, with no silent fallback to host execution.

The workspace layout is files/<chat>/analysis/<name>/: data/ contains imported copies, notebooks/ contains ordinary marimo Python notebooks, and .hailer/ holds runtime and active-notebook state outside the container mounts. Browser authentication tokens stay in that private runtime state and the browser launcher, never in Attaché results pages.

Adapter contract and verification

The bundled script builds its own fixed Docker configuration rather than loading Hailer environment overrides or model settings. It rejects mismatched versions, runtime settings and mounts before executing. Shutdown selects only Docker resources labelled for that workspace, including resources left by an interrupted start; a failed cleanup retains its runtime record for retry. Notebook creation and active selection use Hailer's native APIs.

The adapter targets the published PyPI Hailer 0.2.8 wheel, not an arbitrary checkout reporting the same version. That release exposes HailerTools(config, server=...), persistent kernel state and DockerRuntime; newer main-branch code uses a different sandbox contract. The wheel's SHA-256 is 5031b1911d9cea7d7b7237d8cad04d7594d20d445d234852ac92c559f289eeac. Upgrading this bridge requires checking the new API and lifecycle together. A signature guard rejects the incompatible same-version checkout rather than attempting a fallback.

The implementation is covered by offline tests with fake Hailer modules and fake runtime objects, including ownership, containment, resume, shutdown failures and token redaction. Those tests do not install Hailer, start Docker, call a model, or analyse real user data. A separate contract smoke check imported the exact published wheel, bound the adapter's runtime and tool signatures, and created/listed a notebook in a disposable directory using the actual library, without starting a kernel or importing Hailer's model agent. A separate Windows Docker/headless-Chrome smoke check built that exact release's image from its public wheel assets, imported a synthetic three-row CSV and computed its total of 42 in a saved notebook cell. It verified runtime reuse, stop/restart and notebook reopen, read-only data, blocked outbound networking and withholding of a synthetic host secret. The test removed its containers, network and temporary workspace; it made no model calls.

At verification time the public 0.2.8 registry image was unavailable without authentication. The local build fallback succeeded and produced a roughly 914 MB image. Registry access and available disk space are prerequisites to check on each machine; the local builder needs access to its public base image and Python package downloads.