Risk no longer requires someone else to approve
A financial-risk tool call, a procedure exposed through one, and a destructive, network-egress or privileged sandbox command are now released like any other hold: whoever holdsapprovals:approve and the permissions the held action needs may release it, including the person who asked for the work or owns the agent. The release is recorded as self-released. A hold raised before this change keeps its stored policy, so it still needs someone else; once released, its call is held again under the new rule. Approving any of these holds still needs a signed-in Forge Console session, and a Slack or email approval of an agent tool’s hold checks the tool’s risk as it is when you decide. See Approval lifecycle.Summary
Every open human task already carries the form it was asked on. The one-release fallback that kept an open task without that form answerable is now removed. A closed task’s history still reads exactly what it recorded.Breaking (agent authors)
human_check,human_askandhuman_withdrawreturnstatus: "canceled"with itscancelReason(due_date,declined,run_ended,supersededorwithdrawn) for a question that closed without an answer. The one-releaseexpiredalias is removed. An agent that tested for"expired"tests for"canceled".
Changed
- An open human task is answered only on its pinned v2 form. An open task without one is not presented, and an answer to it is refused with
action_not_declared; a person can still say “Not me” in Forge Console. No open task in staging or production lacks its form.
Human tasks, Increment 2 hardening 3: what each surface approves in place, answers as Forge notices, and inert Slack run threads.
Approve in place only what the surface shows
- Every kind of human task follows the rule that hold approvals already followed: a surface that cannot show what is being decided links out with “Approve in Forge”. Reject stays in place.
- A workflow human step links out on the Slack card, Slack Home,
/forge approvalsand email when it has inputs from prior steps or a decision briefing. Forge Console shows those inputs above the step’s form. - A synthesis confirmation, an Architect plan or an agent setup is confirmed only in Forge Console.
- A task whose title, summary or detail is too long for the surface links out. A long title shows in full beneath the card’s header.
- An approve from Slack Home is checked against what Home shows, not against the card.
An agent’s question is answered as a Forge notice
- A chat agent’s
human_askanswer reaches the agent as a Forge notice in its conversation. The notice names the person and therequestIdand quotes the answer as fenced text. The agent’s next turn starts from a fixed Forge message that names only therequestId. Act on the notice as the person’s answer. - A rejected hold’s reason reaches the agent the same way, marked as untrusted text.
- Forge Console shows each notice as a Forge row and never shows the turn-starting message.
- The public
listMessagesAPI returns the notice as anassistantmessage that starts with[Forge notice], and the turn-starting message as ausermessage that starts with[Forge].
Slack
- Workflow run threads post step names, inputs, outputs, errors and skip reasons as plain text. A
<!channel>, a mention or a link inside them no longer acts. A link is posted only for a plainhttp(s)URL. - A click on a retired workflow-step card whose step is still waiting says the card is out of date and links to the task in Forge.
- An approval with an SLA window is marked breached when the window passes, even when nothing else happens to it.
Added
- Six Pod-to-Customer assignment writes under the existing scope
workforce:write:pods(workforce:manageat the platform scope, plus theworkforce:readscope): Assign pod customer, Change pod customer assignment effective, Update future pod customer assignment, Delete future pod customer assignment, End pod customer assignment and Confirm pod customer assignment accuracy. Each runs the Console’s own command, and a same-key retry is replayed under the pod write rule. - Preview pod customer assignment under the same key, and three reads under
workforce:readat the platform scope: List pod customer assignments, List customer pod assignments and List pod customer assignment history. None needscustomers:read. - Matching MCP tools in the
workforcegroup.
Added
- Eight pod administration writes under the existing scope
workforce:write:pods(workforce:manageat the platform scope, plus theworkforce:readscope): Create pod, Update pod, Archive pod, Assign pod owner, Change pod owner effective, End pod owner, Set pod internal profile and End pod internal profile. Each runs the Console’s own command, and a same-key retry is replayed under the pod write rule. - Three reads under
workforce:readat the platform scope: List pod owners, List pod internal profiles and Get worker. Get worker never returns the owner’s email, metadata, agent configuration, budget or spend. - Eleven matching MCP tools in the
workforcegroup.

