Human Inbox
Handle questions, approvals, and uncertain external outcomes without losing task context.
The Human Inbox is where Oblive returns work that cannot safely continue without a person.
Informational agent updates live in the separate Notifications tab. They do not block tasks and remain open until acknowledged. Inbox counts cover questions, approvals, and reconciliation only; Notifications has its own count and status filter.
Agent updates should lead with the result, its practical effect, and any next step. Questions should ask for the actual decision or setup needed. Exact errors and other implementation details belong under Technical details or in the linked engineering handoff. This writing contract also applies to task summaries, failure explanations, progress updates and resumed chat; a successful test is not reported as a deployment, and unfinished work is never described as complete.
Questions
Questions request missing information or judgment. The item should explain what is blocked, what decision is needed, and enough context to answer without reconstructing the task manually.
After an answer, Oblive closes the exact blocker, advances the task version, and creates fresh execution intent when no other blocker remains.
Items labeled Learning input ask for an essential business definition, goal, constraint, or source interpretation that the agents could not establish from available evidence. They do not block ongoing work. An answer is saved and used in a fresh learning review; the original completed task stays complete. Each team can have at most three open learning inputs.
Structured Context revisions show their verified unified diff directly above the response form. Added lines are green and removed lines are red. Approval applies to the complete displayed revision; request a revision when any path or line is not acceptable.
Approvals
Approvals are proposed consequential actions that policy requires a person to review before execution. Review:
- the provider and account;
- the plain-language effect;
- the owning task;
- the risk and reversibility;
- the intended recipient or target; and
- whether the underlying context is still current.
Reject an action when the target, authorization, or task intent has changed. Approval is not a general grant; it applies to the staged action under the recorded policy and task version.
Reconciliation
Reconciliation appears when an external effect may or may not have happened—for example, a network failure after a provider received a request. Oblive does not retry such an action blindly.
Verify the provider directly, then mark the action according to the observed result. This prevents duplicate emails, payments, invitations, or other effects.
Cleanup
Select answered or cancelled questions and choose Move resolved questions to Trash. Select all includes eligible questions across the current filters; open questions stay available. Restoring a question preserves its answer and never resumes or cancels its source work. Notifications have separate bulk acknowledge and Trash actions. Uncertain outcomes are reconciled individually.
Good Inbox Hygiene
- Answer or reject items promptly so tasks do not remain blocked indefinitely.
- Verify uncertain effects in the provider before reconciling.
- Do not approve a changed target because the summary looks similar.
- Use the task and action context instead of relying on memory.
- Acknowledge informational Notifications separately; do not treat them as blocking questions.
Agent questions explain the business goal or original request, the work being done, and why your input is needed. A task link supports that explanation. Internal execution details stay in the separate technical view; missing evidence, real costs and unfinished work remain clear. Questions also explain what happens today and what has already been checked or completed, so you can understand the proposed change before answering. An explanation of the current setup is separate from permission to change it. Each question asks for one decision. When there is a choice, it offers clear options and still lets you write your own answer. Separate decisions appear as separate questions; a follow-up that depends on your choice is asked only when it is needed.