A shipped OpenAI desktop build turns Excel ranges into queued agent assignments

The sidebar merges two thread types that do not have the same lifetime ^[21]. Native Excel and PowerPoint comments are described by the interface as saved directly into the Office file ^[2]. The ChatGPT assignments are described in the code as persistent but non-exported artifact comments ^[3]. A workbook that leaves the building therefore carries the human comments and the changed cells, and not the sentence that asked for the change.

RuntimeWire frames the feature as OpenAI turning document review into delegated agent work, where users define tasks in context, batch them, and keep control over whether the agent's edits close the assignment ^[19]. That is a fair product description. As a logging design, it puts the instruction in client state and leaves the artifact holding only the outcome.

The dispatch side is a strict contract. Hidden instructions require the model to address every assigned thread, edit the artifact when requested, and return exactly one structured artifact-comment-reply per thread [c6b], and the client intercepts those replies to route each one to its thread ^[7]. That is a state machine keyed on an exact output count. The teardown documents the requirement; it does not describe the branch that runs when a model returns two replies for one thread, or none. In my experience that branch has the thinnest test coverage in the file.

Writeback is where the document actually changes. PowerPoint comment changes are serialized back into PPTX bytes using bundled OpenXML infrastructure ^[8]. Excel changes pass through a durable workbook-update system with conflict handling and are materialized into a new XLSX file ^[9]. RuntimeWire treats that pair of paths as the evidence that native comments are meant to live in the Office files rather than exist as overlays inside ChatGPT ^[10].

For any of this to transfer to a fleet, several things have to hold. The code is controlled by feature gate 1246162292 and also requires the artifact panel expanded to full width ^[5], so two conditions sit in series and a tester who cannot see the panel has learned nothing about what is in the binary ^[13]. RuntimeWire says plainly that presence in a production client does not establish that OpenAI has enabled the feature for ordinary users ^[12]. The implementation was absent from builds checked through internal version 26.820.71523 and present by 26.825.32147 ^[11], while the same tested client also reports a Microsoft Store package number, which means the package string is not a usable key for asking which internal build you have ^[14].

Credit where it is due on method. The principal evidence is production minified JavaScript, and RuntimeWire says descriptive variable and component names counted only when executable control flow, interface copy, persistence operations or model-dispatch behavior corroborated them ^[18]; it also says the finding was only partially reproduced ^[17]. OpenAI's public documentation covers file previews, selection-based revisions, separate Excel and PowerPoint add-ins, and comments on webpages in Codex's browser ^[15], and RuntimeWire found nothing public describing this threaded review system, its private comment layer, its queue or its reply protocol ^[16]. Until such documentation exists, anyone asking what a delivered workbook was told to do has to ask the client rather than the file.