Give an unfinished task a record that can resume

A new session receives a helpful summary of yesterday’s work. It says the draft is nearly finished, the tests mostly passed, and one upload remains. The assistant still cannot safely resume until it knows which draft, which tests, and whether the upload was merely planned or already attempted.

A summary is useful reading material. It compresses decisions and gives a person or model a quick account of the task. Durable task state has a different job. It preserves the specific information needed to decide which operation can execute next, even when the surrounding conversation is unavailable.

The specification for Boris, a native agent interface, separates executable task contracts from conversational transcripts. It calls for versioned objectives, dependencies, acceptance criteria, resource scope, approvals, artifacts, and evidence. The source establishes a design for durable state; it does not prove that the complete runtime recovered real tasks successfully.

Reconstruct an illustrative interrupted job

Suppose a task creates a reference guide from six supplied documents. The assistant reads all six, drafts the guide, runs structural checks, and prepares an upload. The session ends unexpectedly. A conversational summary says the guide is done and awaiting delivery.

A durable record should identify the six source versions, the current guide artifact, and the criteria already verified against it, while preserving the delivery action’s state so that recovery can distinguish a request still being prepared from one already dispatched to the destination. If the upload was dispatched before the interruption, recovery must inspect the destination. If it was only planned, the next permitted operation may be different.

The difference is practical. “Guide complete” can omit an unresolved citation, while a criterion record can identify the exact source-backed claim still requiring review. “Tests passed” can omit a later edit, while a receipt tied to the earlier artifact exposes the mismatch. “Upload pending” can conceal whether a request already reached the service.

In this illustrative exercise, resume with the durable record and current resource observations while withholding the full conversation. Ask the assistant to identify the next authorized action and the evidence supporting that choice. The test fails if it needs to invent a missing artifact version or infer permission from the general tone of the summary.

Keep the original user objective and amendments so that later work remains attached to the requested outcome. Record acceptance criteria separately from the prose plan. A plan can change as the assistant learns; the criteria establish what must be demonstrated before the task is complete.

Preserve dependencies between phases. A rendered-document review depends on the current document, and a publication step depends on both approved content and the required authorization. When a dependency changes, the downstream phase must reconsider its readiness against the new artifact or approval, retaining the earlier result as history while preventing an obsolete summary of progress from authorizing a step whose required inputs are no longer current.

Artifact references should resolve to retained versions or identifiable current files. Evidence references should lead to the command output, source span, screenshot, or review record that supports a criterion. Store large evidence outside the compact task state if necessary, but detect missing evidence when loading the record.

The pending action deserves particular care. Save its intended effect before dispatch, then preserve its observations afterward. This allows a resumed session to reconcile an uncertain outcome. Writing only a final completion flag loses the interval in which an external effect can occur without a recorded response.

Keep the summary derived and revisable

A human-readable summary can be generated from the durable record, with unresolved items and current artifact identities included, but a model’s later paraphrase should remain a separate account whose omissions cannot overwrite the criteria and evidence that determine what the workflow may do next. If the summary disagrees with the evidence, investigate the discrepancy and correct the summary.

Test this deliberately by introducing a stale summary. Mark one paragraph as complete in the narrative while leaving its acceptance criterion unresolved. The resumed workflow should notice the unresolved criterion. It should also survive a missing log or unreadable artifact by naming the missing dependency rather than accepting the narrative claim.

Durability requires maintenance. Version the record format, validate required fields, and preserve recoverable previous states. A checkpoint that exists but cannot be interpreted after an application update does not support resumption. Include a load-and-resume test when changing the workflow’s stored structure.

For the reference-guide exercise, the resumed task should open the current artifact, verify any stale criteria, and resolve the delivery action from its recorded state and destination evidence. The completion report can then summarize those results, with direct references to the exact guide and receipts that supported them.

Source notes

Based on the September 2026 Boris specification’s separation of contract state from conversational history. The reference-guide recovery exercise is illustrative.