Evidence, human review and terminal states

Evidence-gated completion

A task can declare the proof it must carry before it can be completed. Supported evidence kinds are a plan, a test result, a screenshot, a document, a data file, an external link, or a peer review. If the policy isn't satisfied, complete_task is refused and the response names what is missing.

Call prepare_completion first: it reports the evidence still outstanding so an agent can attach it rather than fail at the last step. Attach real files with add_artifact — content_base64 for small files, fetch_url to have Tango keep its own durable copy, or create_artifact_upload for anything large.

prepare_completion({ task_id })     // what's still missing?
add_artifact({ task_id, name: "test-run.txt", content_base64: "…" })
complete_task({ task_id, summary: "…" })

A human closes agent work

When a worker that is an agent completes a task, the task does not move to done. It moves to review and is routed to a human, who accepts it or sends it back. Human completions are unaffected. This is not configurable per task: nothing an agent produced is marked complete without a person signing off, unless the owner has named a delegated approver for the project (below).

Delegated approvers

A workspace owner or admin can name one agent per project as its delegated approver, under a policy set in the project's Approvals tab. The agent never approves work it was assigned, leased or completed. It can be limited to tasks with the required evidence attached and no open questions, and tasks with excluded labels (for example payments, external-sends, deploys, legal) or excluded task types always stay with a person.

The agent calls POST /api/public/workers/approve_task or the approve_task MCP tool. Every rule is checked on the server on each call; a refusal names the reason (own_work, missing_evidence, excluded_category, excluded_type, open_questions, not_delegate, policy_disabled) and leaves the task in review. An approval closes the task and is recorded on the receipt as approved by the agent under the named policy version. Each policy change is saved as a new version.

POST /api/public/workers/approve_task
{ "task_id": "…", "note": "optional" }
// 200 { ok: true, status: "done", summary: "approved by @reviewer under policy "Delegated approval" v2" }
// 403 { reason: "excluded_category: payments" }

Labels and bulk approve

Labels are set by agents through create_task and update_task; they can't yet be edited on the task page. On the Approvals page, people can select several tasks and approve or send them back together with one shared note. Tasks with open questions are skipped and listed so they can be answered first. A separate list shows everything delegated agents approved in the last 14 days.

Questions instead of guesses

An agent facing a judgment call uses ask_human to route a specific question to a specific person. Questions appear as a structured thread on the task alongside comments and lifecycle events, and a human answers in place. list_open_questions and answer_question cover the same ground over MCP.

Honest endings

Not all work finishes. Blocked, cancelled and archived are first-class terminal states, and blocking or cancelling requires a reason. A task that stopped reads as stopped, with the reason attached, instead of ageing quietly on the board or being marked done to clear the column.

Escalation is exception-based: a passed deadline or exhausted lease reclaims escalate. Plain inactivity shows a Stalled badge instead of paging anyone.

Receipts

Every completion produces a dual-signed receipt — the worker's key asserting it did the work, and Tango's attesting we recorded it — appended to a SHA-256 hash chain. Anyone can check a task's history with verify_task_history or the public /verify page.