Added
POST /api/v2/workforce/createPerson,PATCH updatePerson,POST mergePersons,POST offboardPersonandPOST reactivatePersonunderworkforce:write:pods(workforce:manageat the platform scope, plus theworkforce:readscope), andPOST previewPersonMerge, which writes nothing and needs the same two scopes because it names both people. A merge cannot be undone and an offboard is only partly reversible; preview first and send anIdempotency-Key. See People.GET /api/v2/workforce/getPersonandGET /api/v2/workforce/listWorkersunderworkforce:readat the platform scope. A person never carries its avatar URL; the worker list never returns email, metadata, agent configuration, budget or spend.POST /api/v2/workforce/updateWorkerSeatunderworkforce:write:orgsets a human seat’s monthly budget.- MCP:
forge_create_person,forge_update_person,forge_preview_person_merge,forge_merge_persons,forge_offboard_person,forge_reactivate_person,forge_get_person,forge_list_workersandforge_update_worker_seatin theworkforcegroup.
Changed
- A person with any seat in a customer organization or under one is not found by the people routes and tools, exactly as an unknown person.
- Changing a worker seat’s name, status or budget in Forge Console is now audited, on the person’s history for an employee or contractor seat.
Added
POST /api/v2/workflows/createFiscalCalendar,PATCH updateFiscalCalendarandPOST deleteFiscalCalendar, andPOST createStep,PATCH updateStepandPOST deleteStep, underworkflows:manageat the platform scope (the elevatedworkflows:managescope, plusworkflows:read). See Fiscal calendars and Step library.GET /api/v2/workflows/listSteps(workflows:readat the platform scope) andGET /api/v2/workflows/listIngressInvocations(an ingress trigger’s recent requests, under the trigger read rule).- MCP:
forge_create_fiscal_calendar,forge_update_fiscal_calendar,forge_delete_fiscal_calendar,forge_list_workflow_steps,forge_create_workflow_step,forge_update_workflow_step,forge_delete_workflow_stepandforge_list_ingress_invocationsin theworkflowsgroup.
Changed
- A step library entry’s author is always the person or credential that created it; a
createdByvalue sent by the caller is ignored. Creating, changing and deleting a step is now audited, in Forge Console and the API.
Added
GET /api/v2/workforce/listPersonIdentities,POST linkPersonIdentityandPOST unlinkPersonIdentity;GET listPersonEmailClaims,POST upsertPersonEmailClaimandPOST revokePersonEmailClaim;GET getPersonCommunicationProfile,POST savePersonCommunicationProfileandGET listPersonCommunicationProfileHistory, underworkforce:write:pods(workforce:manageat the platform scope; changes also need theworkforce:readscope). WorkOS sign-in bindings stay in the Console and Freshservice links are not accepted. See Identities, email claims and profiles.GET /api/v2/workforce/listPersonAuditEventsundercompliance:readandworkforce:read(audit:readandworkforce:readat the platform scope).- MCP:
forge_list_person_identities,forge_link_person_identity,forge_unlink_person_identity,forge_list_person_email_claims,forge_upsert_person_email_claim,forge_revoke_person_email_claim,forge_get_person_communication_profile,forge_save_person_communication_profile,forge_list_person_communication_profile_historyandforge_list_person_audit_eventsin theworkforcegroup.
Changed
- Reading the audit events of one resource now requires
audit:readfor every resource type except a project (which keepsprojects:readand its visibility); before, any signed-in member could read them for most types. - Adding or revoking a person’s email claim, in Forge Console or the API, is now recorded on the person’s history, and revoking a claim that is already revoked keeps its original revocation time and reason.
Added
POST /api/v2/workforce/createPositionAssignment,PATCH /api/v2/workforce/updatePositionAssignment,POST /api/v2/workforce/endPositionAssignmentandPOST /api/v2/workforce/reassignPrimaryPositionrun the Console people page’s assignment commands. They needorg:manage(scopeworkforce:write:org) andworkforce:readat the platform scope, take an auditedreason, and answer assignment history rows. Workers, positions and assignments in or under a customer organization are404. See Position assignment commands.GET /api/v2/workforce/listWorkerRoleslists every worker role (Job) without its agent configuration, underworkforce:readat the platform scope.- MCP tools
forge_create_position_assignment,forge_update_position_assignment,forge_end_position_assignment,forge_reassign_primary_positionandforge_list_worker_rolesin theworkforcegroup.
Fixed
POST /api/inference/v1/files/confirm used to accept any storage_id whose bytes existed, so an app that knew another upload’s storage id could add that file to its own knowledge. The upload_url that POST /api/inference/v1/files returns now points at Forge. Forge stores the bytes you upload there and records their storage id against the pending file. Confirm then accepts only that storage id. Any other answers 400 storage_id_mismatch with a sentence ending Reference: <id>. and a decision_id, and the refusal is recorded.Changed (integrators)
- Keep using the flow as before: upload the raw bytes with
POSTtoupload_url, readstorageIdfrom its answer, and pass it asstorage_idto confirm. The request and response shapes are unchanged. upload_urltakes one upload of at most 20,971,520 bytes (20 MiB), within an hour of thePOST /filescall, and needs noAuthorizationheader. A second upload to the same URL answers409upload_already_received, and a late one403upload_url_expired.- A file started before this change and not yet confirmed cannot be confirmed, even if its bytes were uploaded. Start it again with
POST /api/inference/v1/files.
Changed
- Adding a file to knowledge from chat or a factory dispatch, an External App’s confirmed upload, and the Rippling org-chart import now write one
file_readaudit event before the file is read. The event (categoryfile_access, eventfile_read) names the person or app that asked, withmode: "ingest". - When that event cannot be written, the ingest refuses and reads nothing:
- in chat, the conversation shows
"<file name>" was not added to knowledge, because Forge could not record reading it. Attach the file again to retry.; - an External App’s file ends in status
error(GET /api/inference/v1/files/get); - the org-chart import answers
FILE_READ_NOT_RECORDEDwith “The file read could not be recorded, so it was refused. Try again.”
- in chat, the conversation shows
Changed
- Forge now writes an event to the event log before it sends any webhook for it, trying the write up to three times. Every webhook’s
idis the event’s row inGET /api/v2/events, and one event has the sameidon every subscription. - If the write still fails, no webhook is sent for that event and operators see a
notification.event_log_write_failedaudit event naming it. In-app, Slack and email notifications still go out. - An operator-only (
admin) event is never in the event log, by design. Its webhookidis now the same for every subscription.
Fixed
Forge Console functions that take an uploaded file’s storage id used to accept any storage id whose bytes existed. That covers chat and knowledge attachments, workflow run inputs, files returned on a human task, workflow documents, avatars, trace snapshots and the Rippling org-chart import. A person who learned another file’s storage id could adopt that file as their own. Now the Console hashes the file before asking for an upload URL, and Forge issues that person a one-hour upload ticket for exactly those bytes. A storage id is accepted only with a ticket of the same person, issued before the file existed. Any other storage id is refused with “This file was not uploaded through Forge by you, so it cannot be used here. Upload it again.”A storage id belongs to the first person who uses it. A workflow launch accepts a raw storage id as a file input only from the person who uploaded it; Slack attachments now reach a run as platform files. Uploads still go straight to storage, so the upload size limit is unchanged. A Console tab opened before this release needs a reload before its next upload.Changed
orgWorkers.listWorkers,orgWorkers.listWorkersPageandorgWorkers.workerFacetsnow requireworkforce:readat the platform scope; before, any signed-in member could read them.- My Work → Reassign now searches people by name, and needs
approvals:approve. - The planning assignee, analytics worker, and environment and vault member pickers list people only for a person with
workforce:read, and otherwise say so.
Changed
workers.listnow requiresworkforce:readinstead ofworkflows:read, andchatDM.listWorkersForDMrequiresworkforce:readinstead ofchat:read. Both return the whole worker directory.- The workflow builder canvas names a node’s specific person by reading that one worker, so building a workflow no longer needs the worker directory.
Changed
- A workflow agent node’s
human_asknow returnsstatus: "pending"at once, as a chat agent’s does. When the agent ends its turn, its node pauses: it holds no agent slot and no timer reaps it. When every question it asked has closed (answered, canceled at its due date, or declined), the node resumes on the same conversation, and each answer arrives as a Forge notice that names the person and therequestIdand quotes the answer as fenced text. - A factory worker’s
human_askpauses its dispatch the same way, and the worker resumes on its own thread. A dispatch that needed an approval resumes on that approval: the same approval is never spent twice. - For these agents,
human_checkandhuman_withdrawreport a question’s status without its answer. The answer arrives only as the Forge notice. Chat agents and agent nodes placed detached keep today’s results. - An agent node placed detached (in a sandbox) still waits up to 2 minutes for an answer in its turn, then withdraws the question.
- Rerunning a workflow from a step cancels the questions its earlier attempt left open, as
superseded. - A question answered through the API with a Personal key resumes the agent exactly as an answer in Forge Console does.
- A workflow node paused on a person shows a card in the run view with a link to each question in My Work.

