The Closed Loop
How a conversation becomes deliberate work, survives execution, and returns evidence or a human decision.
Oblive separates conversation, work intent, execution attempts, human blockers, and external effects. That separation makes recovery possible and keeps each important state inspectable.
Message
Explore, clarify, or request work.
Task
Record deliberate ownership and expected outcome.
Attempt
Execute in an isolated, recoverable run.
Return
Deliver evidence, ask a person, or stage an action.
Conversation
A human turn is stored before live delivery wakes the chat runtime. If the live signal is lost, reconciliation can recover the pending response.
A conversation does not automatically become a task. Quick questions can remain chat, while work that needs ownership, dependencies, retries, or artifacts becomes deliberate durable work.
Tasks And Runs
A task describes the work and its current business state. Queue and retry intent are separate from the immutable history of execution attempts.
Workers re-read durable state before acting. Duplicate or stale wakeups do not become duplicate attempts merely because transport delivered more than one notification.
Human Checkpoints
When execution needs judgment, the task opens a typed question and waits. The answer advances the task version and queues fresh execution instead of attempting to continue an expired process.
External Actions
Consequential effects are staged as actions. Policy determines whether they may execute or require approval. Confirmed effects retain receipts; uncertain effects wait for reconciliation and are not retried blindly.
Durable Output
A run is not successful until validated output has been synchronized. Operational summaries, evidence, blockers, checkpoints, and next actions may persist; hidden model reasoning does not.
For the contributor-level lifecycle, see Architecture and Tasks, Schedules, and Actions.