When an agent says done: what a task receipt can prove

The question we tested

On 29 October 2026 we released Tvijo DSHREM 0.10.25. The question was practical: when an agent writes ‘done’, what evidence can a reader actually inspect? A task receipt can record what a tool observed. It cannot by itself establish that the user's wider objective is correct or complete.

What we changed

The agent distinguishes completed, blocked and partial outcomes. A local receipt can list observed commands and exit codes, edits made through file tools, service task identifiers, and the cost reported for model calls. Missing cost stays unknown and stops subsequent model steps until reconciled; an explicitly measured zero remains zero. Account CLI review in Full needs a separate project approval.

The distinction matters in a simple example. A command returning exit code zero shows that this particular command ended successfully. It does not show that every required workflow was tested, that a deployment happened, or that an external dependency was available. The receipt describes the evidence it has, and the release leaves independent acceptance separate.

What we observed

For the release candidate, the extension test suite passed 983 tests on Node 22 and 983 on Node 24. We installed the exact VSIX in a separate non-root Linux VS Code 1.143.1 profile, compared 75 packaged files, activated the extension, opened its learning guide, and exercised a file-tool write and a command. The installer added its own package metadata, which was checked separately. We also downloaded the public archive without authentication and compared its size and SHA-256 with the tested archive.

These checks have boundaries. The file and command exercise used an isolated test project, not a customer project. A live paid model coding task was not run while an existing provider accounting freeze remained unresolved. Passing the VSIX checks does not accept a Windows, macOS or Linux Tvijo Code installer; each is a separate artifact and needs its own native test.

Access is a separate decision

Commands use the operating-system rights of the editor host. The non-root Extra mode is not an operating-system sandbox; on a root extension host Extra is refused. An older Full grant does not automatically enable account CLI reviews. The existing Copilot integration sends work to Tvijo; an outbound Copilot reviewer is not implemented. These are boundaries of the shipped behavior, not a promise that an agent can never make a mistake.

How to use the record

Read the outcome and remaining steps first. Then inspect which commands ran, their exit codes, which files were edited, and whether usage is measured or unknown. Add the task-specific checks that matter for your own project. A receipt is useful evidence for review, not a substitute for it.

The public release and download show the exact archive and verification limits. This article is an automated project record based on observed release evidence; it is not a security certification or a claim of human independent review.