The Workforce HTTP API and existing MCP workforce group now provide read-only Engineers, Pods and Customers planning. Employee pool summaries, capacity boundaries, classification slices and dated source percentages share the Console’s domain calculations.
listEngineers/forge_list_engineersreturns a whole visible Engineering pool summary and one detail page. Search, modes and pagination preserve pool totals.listPodPlanning/forge_list_pod_planningandlistCustomerPlanning/forge_list_customer_planningreturn visible delivery staffing and typed type/category/stage context.- All three require existing platform
workforce:read. Hidden Pod commitments cause an Engineer’s omission before totals, never false bench. Customers come only from minimal workforce summaries linked to visible Pods. - Strict inputs and outputs reject unknown fields. Responses contain no primary email, audit identities, unrestricted CRM, contracts, signatures, SOW or revenue evidence. Bounded-source limits refuse with
422rather than returning partial totals.
Summary
Every human task now carries one form, pinned when it is asked: the fields the person fills in and the actions they can take. Forge Console, Slack and email render a task from that form, and the server checks every answer against it for every kind of task. Where a surface cannot render a form honestly, it sends a link to the task in Forge instead. An agent that asks a person in chat no longer holds its turn open: the answer arrives later as a new message that resumes the conversation.Breaking (agent authors)
- A chat agent’s
human_askreturnsstatus: "pending"at once. When the person answers, the answer arrives as a system message that starts a new turn on the agent’s own thread, run as the chat author whose turn asked (the task’s accountable subject), never as the person who answered. - A workflow or factory agent’s
human_askwaits up to 120 seconds in its turn. If nobody answers, the question is withdrawn and the tool returnsstatus: "withdrawn".waitSecondsis ignored. human_asktakesfields(at most 10) to ask for a small form, and the newhuman_withdrawtool withdraws a question the agent asked. An answered result addsactionandoutput;answeris kept.- When a workflow node or an agent dispatch ends (detached agents included), its open questions are withdrawn. When a run ends, its open questions are canceled with
cancelReason: "run_ended".
Deprecated
responseContracton the approval outputs (GET /api/v2/governance/pendingApprovals, theforge_get_approvalandforge_list_approvalstools) is frozen and deprecated. It is removed at the next forge-mcp major release. A task’s form reaches the public API with the human-task resource of a later increment, as new tools and routes. This release changes no REST or forge-mcp schema.- Seven Console mutations are deprecated aliases of
workersMyWork.completeMyHumanTaskand are removed next release:workersMyWork.resolveMyApprovalandcompleteMyHumanNode,runsActions.resolveApprovalandcompleteHumanNode,governanceInternal.resolveApproval, andhumanInteraction.resolveForSourceandmarkAgentSetupDone. Each keeps its arguments, returns and checks until then.
Changed
- Forge Console answers every human task on its own form, through one door,
workersMyWork.completeMyHumanTask. The server sets the door and builds the person; no argument names either.workersMyWork.getHumanTaskViewreturns a task’s form, what was decided, and what the caller may do. humanInteraction.getActiveForSourcereturns the open task’s prompt aspromptV2and no longer returns the recorded v1prompt.- Take over: an approver with
approvals:approvewhere a hold or a workflow step lives can take it over from the Console task view with a reason. It is a reassignment to themselves, audited as a take-over. - Slack and email render a task from its form. A field or action they cannot render (a file, a JSON value, a long list, or an approve that needs a signed-in Console session) becomes a link to the task in Forge.
- The server checks every answer against the task’s form for every kind: bounds, formats (
monthincluded), patterns, options, required fields, a JSON value against its declared schema, and a file the person uploaded. - Approving a held chat prompt (
dispatch:chat_response) re-sends it automatically. Retry stays as the fallback and does nothing once the approval is spent. - A chat question’s task now names the asking agent by its agent id (
sourceEntityId). No question was open in production or staging, so none is stranded.
Known limit
A personal assistant’s workflow-launch tools refuse to launch from a turn that an answer resumed, because that turn was triggered by a system message rather than a person’s message. Ask the assistant again to launch it.Deploy note
Thehuman-task-forms-v2 backfill adds the form and prompt to tasks that are open at deploy; a workflow step’s form is compiled again from its node’s declared outputs. Closed tasks keep exactly what they recorded.Engineer rows now include an authoritative capacity timeline with explicit historical uncertainty,
exclusive interval ends and eligibility-aware availability. Current group labels follow the evaluation
date; invalid Engineering group selectors reject instead of showing a misleading zero.Pod and Customer reports qualify staffing affiliation with
proven, incomplete or
historical_unproven coverage. Machine reads require a proven visible Pod home before computing
rows or totals, and merged identities count once against the canonical People ceiling.These contracts support Forge becoming the ongoing tracker. Production setup, one-time tracker
reconciliation and business acceptance remain required before retirement of the spreadsheet.Workforce Planning now leads with available FTE and engineers with
capacity. Compact rows separate current commitments, temporary opening duration and proven
sustained capacity. Engineer, Pod, Customer and whole-pool timelines share one UTC window and an
as-of marker, with exact textual intervals and explicit uncertainty.Customer ownership points to serving Pods, delivery labels are readable, and affiliation coverage
follows server evidence. Keyboard perspective navigation, responsive scans and removable filter
chips keep the shared planning context usable. Forge owns ongoing planning after one-time tracker
reconciliation and acceptance; the code change does not import or retire a spreadsheet.

