Deepnote Data Execution

Deepnote Data And Execution

Available Context

Use the connected Deepnote app tools for the execution and context they currently expose:

• get_notebook for notebook blocks, input variables, last-run metadata, and integration references visible in blocks.
• list_integrations for workspace integration names and types.
• get_integration for integration details and cached table structure, optionally filtered by databaseName, schemaName, or exact tableName.
• list_integration_project_usages, list_integration_notebook_usages, and list_integration_block_usages for direct integration usage mapping.
• create_run to start full-notebook execution, optionally with input values.
• list_notebook_runs for historical notebook runs, newest first, with pageSize and pageToken pagination.
• get_run for run status, errors, completion time, and run snapshots. When snapshotDelivery is omitted, it returns a short-lived snapshotDownloadUrl when a snapshot is available; this is equivalent to snapshotDelivery: "downloadUrl". Request snapshotDelivery: "inline" only when you need snapshotContent in the tool response.
• get_me for the authenticated workspace, connected user, and caller access level when execution permissions or workspace identity matter.

Use get_integration for cached table and column structure when the user asks about integration schemas, tables, columns, or whether a table exists. Do not claim live database introspection, table previews, query previews, file metadata, or environment configuration unless a current Deepnote app tool explicitly exposes that data.

Execution Workflow

1. Read the notebook context first with get_notebook.
2. Identify whether the user needs a fresh run, a specific run status, or notebook run history.
3. If the notebook has inputs and the user supplied values, map values to the exact input name fields returned by get_notebook.
4. Check whether execution may mutate data, call external services, trigger schedules, or consume significant compute.
5. Use create_run only for full-notebook execution by notebookId; the Deepnote app does not currently expose single-block execution.
6. If create_run returns a tool error, report that error and stop; do not call get_run unless a run ID was returned.
7. Use get_run to inspect status, errors, completion time, and snapshot availability before reporting results. For routine status checks, omit snapshotDelivery so the default download URL delivery is used; request snapshotDelivery: "inline" when you need to inspect notebook outputs, snapshot errors, or result details.

Before starting a run, inspect the notebook for cells that print environment variables, secrets, credentials, or entire configuration objects. If found, warn the user and get explicit confirmation before running. Also warn before running notebooks that start servers, send bulk requests, call external services, or mutate data.

Run History

Use list_notebook_runs when the user asks about recent runs, failed runs, run history, or anything older than the latest run metadata returned by get_notebook. The tool returns runs newest first with runId, notebookId, status, createdAt, completedAt, and pagination.

For short history requests, call list_notebook_runs with the default pageSize of 20. For broader audits, use pageSize: 100 and follow pagination.nextPageToken while pagination.hasMore is true, stopping once you have enough evidence for the user's question.

Use get_run only after selecting a specific run that needs detail, snapshot availability, output inspection, or failure debugging. If the user asks "why did the last run fail?" and no run ID is provided, list recent runs first, pick the newest failed run, then call get_run for that run.

Cached Integration Structure

Use list_integrations to resolve an integration name or type to an ID, then call get_integration for cached table structure. The response includes integration details plus tables, where each table has name, schema, optional database, and cached columns with names and database-native types.

Use databaseName, schemaName, and tableName filters when the user asks about a specific database, schema, or exact table. These filters apply to cached structure rows; an empty table list means no matching cached structure is visible through the app tools, not proof that the live database has no such table.

When reporting cached structure, say it is cached. Do not present it as a fresh live database scan, and do not claim access to row previews or query results unless you obtained them from a notebook run snapshot or another exposed app tool.

Notebook Run Inputs

create_run accepts an optional inputs object. Keys must be notebook input names from get_notebook, not labels or block IDs. Values must match the input block type:

• Text, textarea, file, date, slider, and single-select inputs use strings.
• Checkbox inputs use booleans.
• Multi-select inputs use arrays of strings.
• Date-range inputs use a string or an array of exactly two strings.
• Slider values must be numeric strings.

If the user provides a label instead of a name, inspect get_notebook inputs and map it to the closest input name only when the match is unambiguous. Otherwise ask for clarification. Run input values apply only to the new run and do not update the notebook's saved defaults.

Run Snapshots

When snapshotDelivery is omitted, get_run returns snapshotDownloadUrl for available .snapshot.deepnote files and snapshotContent: null; this is equivalent to snapshotDelivery: "downloadUrl". The URL is short-lived and grants access to the run snapshot, so do not paste it into the final answer unless the user asks for a download link or file handoff.

Use snapshotDelivery: "inline" when the user asks you to inspect outputs, summarize results, diagnose a failed run from snapshot details, map visible references from the snapshot, or otherwise reason over the snapshot content. Inline snapshots can be large and sensitive, so summarize the relevant blocks, outputs, failures, or data shape instead of dumping raw content.

If the current Deepnote app tool schema does not expose snapshotDelivery, use the fields returned by get_run as-is and do not invent snapshotContent or snapshotDownloadUrl.

Sensitive Outputs

When snapshot content, snapshot download URLs, or errors include sensitive, proprietary, personal, or production-like data, minimize exposure in the response. Summarize the result, shape, quality issues, aggregates, or failure mode instead of dumping raw records, presigned URLs, or long logs.

Environment Changes

The Deepnote app tools currently do not expose environment mutation tools. Do not claim to change package versions, environment images, hardware, integrations, credentials, secrets, scheduled runs, or shared app settings unless a current tool explicitly supports that action.

Reporting Results

For successful runs, include the executed notebook name or ID, run ID, status, any input overrides that are safe to mention, and the important result from inline snapshot content when you requested it. If you only have snapshotDownloadUrl, mention that a snapshot is available without exposing the URL by default. For failures, include concise error detail and the next fix to try. If a run fails before it starts, such as a workspace or parallel run limit, report the user-facing API error directly. Avoid pasting long logs unless the user asks for them.

Keep run reports brief and information-dense unless the user asks for detail. Prefer one compact run table plus the most important result or first actionable error. Do not paste long logs, raw snapshots, or full notebook outputs by default.

Prefer this run summary shape:

For failed or stuck runs, use a debugging report:

When inspecting a large run snapshot, request inline delivery only when necessary, then summarize block counts, failed blocks, final outputs, and the first actionable error instead of pasting the snapshot.