Interworkspace requests
Interworkspace requests let an Agent in one workspace ask another workspace owned by the same account to do bounded work. They use the same durable workspace-run machinery as scheduled automations; they do not start or inject a prompt into the target's interactive session.
Configure access
The source workspace is X and the target workspace is Y. Open More → Workspace request access in X, choose Y, select the individual permissions, and save. The current grants list lets you revoke access. Grant access only in the direction you need.
For API clients, the owner-managed API is:
GET /api/workspaces/{owner-workspace-id}/request-grants
PUT /api/workspaces/{source-workspace-id}/request-grants/{target-workspace-id}
Content-Type: application/json
{"permissions":["discover","submit","read","reply","cancel"],"enabled":true}
The grant is directional. Same ownership does not grant an Agent access to sibling workspaces. DELETE on the same grant path revokes it. Revocation blocks new discovery, submissions, reads, replies, and cancellations through the Agent broker; it does not delete already-created target data.
Y must have its model, saved instructions/context, and permitted connections configured before a request is admitted. Connections are resolved by Mapache at target-runner provisioning time; do not copy bearer tokens or connection secrets into a prompt.
Ask from workspace X
In X's normal Agent chat, ask:
Ask Customer Research to summarize onboarding complaints and return three supporting examples. Use only the permitted Customer Research workspace context and return the request ID so I can retrieve the result.
The Agent should use the image-owned workspace_request_targets tool to resolve the permitted target, then workspace_request_submit. It must retain the returned request ID. If more than one target has a similar name, resolve the target before submitting; never guess a workspace ID.
Use workspace_request_get to retrieve status and the caller-safe result/artifact projection. Use workspace_request_reply for a clarification or continuation and workspace_request_cancel for supported cancellation. Polling is bounded and should use backoff; a queued request may wait for target concurrency admission.
Paused workspaces and results
The target's interactive session may be paused: admitted work starts an isolated target worker from saved state and does not require the interactive session to remain open. X may also pause or restart after submission; retain the request ID and retrieve it after restart. Unsaved live chat state is not inherited. Saved target context is pinned at admission, source files are read-only to the background worker, and results are written to the request's separate output/artifact sink.
A request can be queued, provisioning, running, stopping, succeeded, failed, canceled, or interrupted. Queued cancellation is not a guarantee that already-running external work stopped. Sink delivery is separate from agent execution, so retrieving a result does not rerun the target agent.
Boundaries and troubleshooting
- The source receives status, the final result, and explicitly shared artifacts—not the target transcript, credentials, private history, or unrelated tickets.
- Missing tools usually mean the workspace needs a new/restarted
pi-chromerunner image. request_permission_deniedmeans the directional grant is missing, disabled, or lacks the requested operation.request_target_not_foundmeans the target is no longer visible to the source owner.- Missing model or connection errors are target configuration problems; configure Y and submit a new request.
- Delayed results normally indicate queue/concurrency admission or sink delivery; keep the request ID and check again.
Scheduled automations and interworkspace requests are two triggers over the same run machinery. See Automations for schedules, Run now, cron/timezone behavior, saved context, outputs, and recovery.
Optional HTTP reference
Owner-authenticated grant management uses the paths above. Agent calls use the short-lived runner-owned broker credential and these paths:
GET /api/agent/workspace-requests/targetsPOST /api/agent/workspace-requestswith{targetWorkspaceId, request, context?, desiredOutput?, model?, resources?}GET /api/agent/workspace-requests/{ticketId}POST /api/agent/workspace-requests/{ticketId}/replywith{message}POST /api/agent/workspace-requests/{ticketId}/cancel
These broker routes derive source workspace/session identity from the admitted runner. A user Firebase token is not a substitute, and callers must not send ownerUid or sourceWorkspaceId to authorize themselves. Use Idempotency-Key for submissions when a network failure could make the outcome ambiguous.