Refine reports in your browser

Once a report is open locally, you can request a focused change from the report itself. Trellum relays that feedback to the coding agent working in your repository; the agent edits the source, rebuilds, and the preview reloads.

The live review loop

1. Build and open a report locally. Follow Try Trellum locally for the sample project, or open your own report's local preview.
2. Ask your coding agent to start review and keep listening. The review session is a relay between your browser and an active external agent.
3. Select a chart or table, or type a general request, and send it. Both an element selection and an element-free message are supported.
4. Let the agent edit and rebuild. It changes the Python or SQL source, runs the build, and replies in the panel. The browser refreshes when the rebuild is ready.
5. Repeat, then inspect the source diff. Keep the resulting source in your normal Git review and publishing process.

Use this prompt with your coding agent:

Start a live review of reports/store-health. Keep listening for my browser feedback, update and rebuild the report, and reply in the review panel.

Commands for the coding agent

The agent can run this loop from the repository root:

python -m trellum review start reports/store-health
python -m trellum review poll --report store-health
# Apply requested edits to reports/store-health source files.
python -m trellum.run reports/store-health --no-serve
python -m trellum review poll --report store-health --reply "Updated the report and rebuilt the preview."

poll waits for feedback, so the agent must keep that process alive (or restart it if its timeout expires). Running start alone serves the report and enables review; it does not create a model, editor, or coding agent. The --report option isolates feedback when several reports share a project. When the user presses End, poll exits with code 3; Send & end delivers the final batch before that exit. Do not restart a session after it has ended. When the review is finished from the agent side, use python -m trellum review end when the user asks you to wrap up. The installed framework's python -m trellum guide review is the canonical reference for current flags and lifecycle details.

The agent edits files under reports/ and rebuilds the report. Do not hand-edit the generated output/ HTML or expose the loopback review endpoints. The feedback relay itself needs no provider key; the coding agent you choose has its own account, setup, and runtime requirements.

Local review and the portal assistant

Local review is for changing report source in your repository with a coding agent. The optional portal assistant answers questions about portal reports and can propose supported portal actions with approval; it never edits your Git repository. Public or exported reports do not retain the local editing chat.