Stunt Double Studio: cloud

/stuntdouble:cloud <checklist>, one of the Studio skills (/stuntdouble:setup, /stuntdouble:run, /stuntdouble:open, /stuntdouble:cloud and /stuntdouble:doctor). It moves a local checklist to the hosted stuntdouble server's tools unchanged.

The .stuntdouble/ file format is in the setup skill's files.md (../setup/files.md from this skill's folder): read it before writing back ids or a run.

1. Disclose: the push uploads the persona, the checklist and the target URL to Stunt Double, and Free-plan cloud runs may go into the research corpus. Go on only with consent.
2. Sign in: list_workspaces(). On first use Claude Code shows the hosted connector's Connect card (an emailed code or Google). The local package never holds a Stunt Double token. Pick the workspace (ask if there are several) and write workspace_id and workspace_slug to stuntdouble.json.
3. Fix the target: the cloud cannot reach localhost. If url is local, ask for a public or preview URL, write it to url and move the local origin to local_base_url, so local runs keep using it. Or explain that devices on the Startup plan reach private targets.
4. Push the definitions, reusing cloud ids already in the files, else search(workspace_id, query) first:
   • Project: create_project(workspace_id, name, url).
   • Actor: create_actor(workspace_id, name, description), update_actor(actor_id, system_prompt), then add_actor_knowledge(actor_id, title, content, kind) per entry. With an actor_id, update_actor instead and add only the entries list_actor_knowledge lacks by title.
   • Checklist: create_checklist(workspace_id, project_id, name, actor_id, instructions, url, checks), plus each of pass_threshold, score_type, score_unit, score_bands, run_mode and record_session the file sets, sending the checks in file order without key or cloud. With a checklist_id, update_checklist(checklist_id, ...) with the same fields and each check's id from check_ids.
   • If the checklist sets browser_config, say so before the run: no hosted input or dashboard setting takes it yet, so the cloud run uses the default 1280 x 720 desktop browser, and a check about a phone layout, touch or a locale can pass there and fail locally.
5. Write ids back after every call: project_id, actor_id and checklist_id into their files, then get_checklist(checklist_id) to fill check_ids as {key: id} in order. A retry or a teammate then updates rather than duplicating.
6. Run: run_checklist(checklist_id), then poll get_checklist_run(run_id) every 30 to 60 seconds until status is completed or failed. Write runs/<YYYYMMDD-HHMMSS>-<checklist>-cloud/run.json in the local run's shape, so the viewer lists it beside local runs: format: "stuntdouble.run/1", id (the folder name), runner: {kind: "cloud", run_id, judged: true, url} with url the dashboard link https://app.stuntdouble.io/w/<workspace_slug>/checklists/<checklist_id>/runs/<run_id>, status, the checklist url, started_at (its started_at, else created_at), finished_at (its completed_at), the checklist and actor slugs, checklist_snapshot (the checklist file without cloud), results as {check: <key from check_ids>, status, evidence, at: <evaluated_at>}, score: {score, unit: <score_unit>, outcome, pass_threshold: <get_checklist's pass_threshold, or null with no weighted checks>, grade: null} (or null with no outcome), summary and cloud: null. Report as /stuntdouble:run reports a local run, with the dashboard link.
7. Relay refusals: a plan limit (the Free plan's 2 actors or 60 compute minutes) comes back as text. Relay it as written with https://app.stuntdouble.io/w/<workspace_slug>/settings/billing, never work around it, and offer the local run.