Built-in research tools
Built-in operations connect the conversation to project files, Notebook execution, literature records and saved results. They are exposed through the active agent framework; the interface may group several operations under an Agent SDK or Notebook activity card.
Operations by research task
| Family | Supply | Operation and visible result | Check before relying on it |
|---|---|---|---|
| File discovery and reading | Project/session and exact input | Lists accessible files or reads supported content | Full name, source and current version; a catalog row is not file-content inspection |
| Notebook execution | Python/R code, runtime and valid input references | Code cell, output, timing and execution status | Actual language/runtime, errors and complete-data checks |
| Artifact publication | A supported output path or inline content | Versioned result card in the conversation/Files | Open the output, compare dimensions/content and retain its version |
| Literature search/import | Query, DOI/PMID or reference files | Retrieved records, Inbox candidates and library entries | Source identifiers, duplicates, review decision and full-text access |
| Plan operations | Research stages and completion criteria | Session Plan with step states | Status matches the outputs actually produced |
| Memory | A fact or instruction intended for reuse | Saved memory in the configured scope | Content and scope, separate from project file storage |
| Skill loading | An available package ID | Method instructions and supporting resources | The intended package was read; loading is not computation |
| Specialist delegation | Role, task and allowed inputs | Child activity and Subagents transcript | Child actually started, ran the requested checks and reported limitations |
| Review | Eligible response and review configuration | Reviewer checks, findings and correction state | Which response/version was reviewed and what evidence was available |
Follow an operation through to its result
- Send a request with the input and desired output. Follow First project for a small table-to-report task.
- Expand the activity card and check the requested operation, input and permission scope.
- After execution, read the tool result. Resolve an error before relying on a summary of it.
- Open the generated file card and check its contents. For a calculation, open Notebook to inspect the producing code and runtime.
The data-quality workflow supplies a complete scientific example. Environment controls are explained in Notebook and Runtimes.
Read tool activity
| UI element | How to use it |
|---|---|
| Activity group | Expands the operations grouped in that run. |
| Tool / Agent SDK / Notebook run | Opens operation details where available. Read inputs, response and error rather than only the group label. |
| Copy code / code disclosure | Copies or reveals the submitted code; it does not execute a second run. |
| Allow / authorization scope / Deny | Controls the displayed operation and chosen scope. Inspect which agent is requesting it. |
| Generated file card | Opens the actual saved result. |
| Open notebook | Opens session execution records and variables. |
| Child task chip | Opens a real delegated task; inspect its own transcript. |
Avoid common input/output confusion
An uploaded file, a Notebook working file and an artifact version are related but distinct. A file visible in one list may not yet be mounted in a child kernel. Use the current application-provided reference; do not repair an unavailable input by guessing a path or replacing an artifact version ID with its filename.
A Connector-created file may have no captured Python producer block. No producer block, No review for this version, partial environment and bounded evidence are meaningful states, not blanks to fill with generated prose. Use Troubleshooting for exact errors and recovery steps.
Implementation reference: notebook.ts, artifacts.ts, WorkspaceMessageItem.tsx.
For supported long-running work, follow Background tasks and result delivery. Inspect the actual run and saved outputs after delivery. Environment & Packages, Compute Environment Setup and Remote Compute (SSH) stay enabled, but their runtime, network and host requirements still apply.