# ResearchStudio: start and recover

Select the connected ResearchStudio app and invoke idea_spark(topic=...).
For a server-backed run, read adapter/SERVER.md and create_research_run with
a fresh request_id. Keep its run_id, run_token and revision. Follow the original
navigator, submit actual phase artifacts, and execute its deterministic steps.
Ask for followup_scoop=true to hand the final candidate to the original Scoop
Check workflow. This does not automatically pass novelty checks.

Skills: idea-spark, paper-search, scoop-check, paper2assets, paper2blog,
paper2poster, paper2video, paper2reel. Only standalone IdeaQuality is excluded.
Use get_research_index to obtain their exact original source IDs. Prompts must
be read whole with their hashes; an index is not a substitute for a prompt.

If backend access, runtime download or DNS fails:

- Stop automatic retries. Do not guess a minimal runtime or rebuild it by hand.
- Use get_research_index and read only the current phase's original materials.
- Prepare evidence/inputs with actually available Web Search and file tools.
- Save this TODO: topic; workflow; run_id; last acknowledged revision;
  pending files; request_id of an uncertain write; missing checks; next action.
- Keep run_token private. After service recovery read status; recover a lost
  acknowledgement using the same request_id and exact request, not a new run.
- Label unexecuted checks and unsupported capabilities. A draft is not a pass.

If the entire MCP endpoint is unreachable, use your saved artifacts and this
offline checklist. Do not fabricate source reads or repeat failed calls.