Run the first research story

Outcome

The local workbench turns one objective into an agent-proposed typed plan, an approved research task, an immutable attempt, an executed notebook block, an artifact, an evidence-backed claim awaiting review, and a writeup. Reloading the API or browser does not erase the story.

Prerequisites

Complete local setup. Use the seeded development workspace only; this path is not a production authentication profile.

Steps

Start the API and workbench:

pnpm dev

Open http://127.0.0.1:5173, enter a research objective of at least 12 characters, and choose Generate research plan. The first isolated agent may only inspect the project task snapshot and propose one complete typed DAG. Review the pending tasks.revise_plan approval, allow it only when its scope and budget are correct, then inspect and approve the resulting draft plan revision. No research task starts before both decisions.

Allow the exact notebook edit when the research attempt requests it. Open the proposed claim's evidence link, then open the artifact link. Confirm that the claim, notebook output, and artifact agree and that the interface shows the durable event sequence. Accept or reject the claim through the review queue; agent publication alone never accepts a claim.

Verify

Keep the page open, stop and restart the API, and confirm live updates reconnect while the visible story remains. After a full browser reload, the durable event history returns and the saved run route restores its selected object. pnpm check independently kills the API process and proves the exact saved story endpoint and filesystem bytes survive. A resumed event request must continue after its opaque cursor without duplicating earlier UI state.

Recover

If the form reports a command rejection, keep the message visible and correct the stated input. If the API is offline, restart pnpm dev and retry; do not create a second run merely to hide a failed first command. If an evidence link fails, run pnpm check and preserve the failure for diagnosis.

Next task

Understand projects and runs.