trellum
Docs / Recommended workflow / Build an analysis locally

Build a local analysis with an agent

Applies to: latest

Use a report as evidence, then write the finding into a Git-backed article or presentation. The local analysis workspace keeps the source report, the destination analysis, and your feedback in one review session.

Open the analysis workspace

Name the source report in your request when you know which one to use. If you do not know its name, the agent can show the available report titles and ask which one to open. Articles are the default; ask for a presentation when you want slides. Include the reporting period and any filters that matter to the finding.

From the project root, the agent creates the destination analysis and opens a workspace with the chosen source report selected:

python -m trellum analysis new cart-funnel-finding --format article
python -m trellum analysis session reports/cart-funnel-finding --source cart-funnel

Use --format presentation for slides. The destination and source can have different slugs. The first --source sets the initial report; the source picker lists all visible reports in the project. Omit --source to choose one in the workspace. The workspace builds the analysis and serves it with the source reports. If a source report is missing or out of date, the workspace builds it in the background and opens it when ready.

The printed Analysis workspace URL opens the combined review. Its tabs show the analysis, source reports, and feedback queue. You can return to the same workspace while the session remains active. There is one review queue and one agent listener for this destination analysis.

Select and route feedback

Switch between the full-width Analysis and Reports pages. Choose a source from the Reports page picker, then use Select element in the workspace toolbar. On the Reports page, select a whole chart or section and explain what it should contribute to the analysis. On the Analysis page, select an individual heading, text block, or other element and describe the edit. Queue adds the request to the shared queue; Queue & send adds it and sends the queue to the agent. Open Queue to review or edit queued notes, read replies, or send a message. It overlays the current page without resizing its charts; Escape or Close returns to the same report. Filters, selection drafts and scroll positions are retained when switching pages or source reports.

Use Report only or Analysis only for a view without workspace controls. On a keyboard, press Escape to return. Back to workspace is also available at the top right; it fades on desktop until hovered or focused, and remains visible on touch devices. This fills the browser viewport; it does not change the browser’s fullscreen setting.

Both pages target the active analysis, with the selected source attached to each request. For example, a source selection can say:

Add this chart to the analysis: explain why conversion fell over the selected dates, using the current filters.

The note and captured chart context travel together. Source selections add evidence to the analysis by default; analysis selections edit the authored result. To edit the source report itself, open its ordinary report review. That independent review keeps its familiar feedback and selection behavior.

The queued item preserves the page, filters, build context, and chart snapshot from the time you selected it. Later filter changes do not alter that queued evidence. If the source report is rebuilding or the view changes before its snapshot is ready, wait for the background build to finish and select the element again. For a presentation, a selection may refer to an enclosing section rather than a specific heading; name the intended slide or section in your note. If the wording is unclear, provide the exact text you want.

You can queue several selections and notes, then send them together. Each item keeps its source and purpose. Keep the workspace session active while the agent works through the queue; the agent replies in the workspace and continues listening until you end the session. The agent uses one poll scoped to the destination analysis for the whole workspace, processes each batch, replies, and polls again until the session ends.

For chart comparisons, include the dates and filters you used. Ask the agent to check the underlying data before stating that a value increased or fell, and distinguish a percent change from a percentage-point change. Ratios across groups should use total numerators divided by total denominators, not the average of each group's ratio.

Use the workspace from a phone

The printed URL uses localhost, which points to the device opening it. On a phone, localhost means the phone itself, not the computer running Trellum. The review API is loopback-only, so changing the link to a bare LAN address does not make remote review work. Phone access requires a protected, authenticated proxy that routes the review requests back to the local server. Do not expose the development server or review endpoints directly to the LAN.

Evidence and interrupted sessions

A captured chart is a frozen image with report context, active filters, and capture/build times when known; it contains no raw dataset. The agent imports the capture into the analysis and edits its source in Git. Rebuilding the source report later does not change evidence already in the analysis. Local review does not publish the analysis or change its sharing permissions.

If a session is interrupted, the agent can list pending capture IDs and resume importing them:

python -m trellum analysis inbox reports/cart-funnel-finding
python -m trellum analysis import reports/cart-funnel-finding --inbox <capture-id>

Without an active analysis session, Capture for analysis downloads a JSON file that can be imported by file path. For a source-report-only preview, see Refine reports in your browser.