Work with projects and runs

Outcome

You can identify the project that owns the work, create one research run, and distinguish durable run state from a browser tab or process lifetime.

Prerequisites

Complete the first local run and keep the API reachable.

Steps

1. Open the development project shown by the workbench.
2. Review its name, version, theme, and supported panel settings.
3. Compose one objective. Type @ to insert a structured reference to saved work when it is relevant; the submitted objective renders that reference as readable @Label text. The idempotency key binds retries of that same command.
4. Generate the plan. The planning attempt can discover frozen local context with read, list, and search, inspect canonical research and web sources, read task state, and propose one complete revision through approval-gated tasks.revise_plan. It cannot write files, run shell/Python, delegate, or put ordinary tool/environment policy into a task. Review that exact plan approval, then review and approve the immutable draft DAG before research work starts. Read the proposed claim at the top of the persisted story, open its exact evidence, and accept or reject it through the review queue. Expand technical source, output, or record details only when you need them.
5. Open Research activity and confirm its durable updates use the same run identity. Follow the nested project, run, task, attempt, claim, and artifact links, then refresh the page and verify the selected object's heading remains in the contextual inspector. Expand Technical identifier to compare the underlying ID. An invalid object ID returns to the project overview with an explicit recovery notice.

Project settings use compare-and-swap. A stale expected version is rejected instead of overwriting a newer update.

Verify

pnpm check exercises the composer by keyboard, verifies that a structured reference submits as readable text, repeats the exact command with the same idempotency key, and requires the existing planner-first run. Changing the payload under that key must produce an explicit conflict rather than a second run. The proof also requires explicit plan and claim decisions; neither is inferred from an agent's successful attempt.

Recover

On a stale project version, fetch the current projection, reapply only the intended setting, and submit with the observed version. On an uncertain run command, retry the original key before inventing a new one.

Next task

Read plans, tasks, and attempts.