OpenAI built a hidden task system for ChatGPT inside Excel and PowerPoint files
The unreleased desktop feature turns comments on cells, charts, slides and shapes into threaded assignments that ChatGPT can complete in batches.
By Ryan Merket · Published
RUNTIMEWIRE INVESTIGATION — Scoop
Original reporting by RuntimeWire, based on reverse engineering.
Why it matters
OpenAI is turning document review into delegated agent work: users define tasks in context, batch them, and retain control over whether ChatGPT's edits close the assignment.
Reporting record
Finding
OpenAI has built an unannounced, gated review system in the ChatGPT/Codex desktop client that lets users attach native Office comments or private ChatGPT assignments to specific Excel ranges and PowerPoint elements, batch the assignments, and receive threaded replies after the agent works on the underlying file.
How we verified
Methods: reverse engineering.
RuntimeWire analyzed the production app.asar distributed with OpenAI’s Windows desktop package OpenAI.Codex_26.825.4187.0. The client contains a feature called “Artifact comments” with two parallel comment systems: Native Excel and PowerPoint comments, described by the interface as being saved directly into the Office file. Private ChatGPT comment threads, described in the code as persistent but non-exported artifact comments assigned to ChatGPT. Users can select an Excel range, chart, PowerPoint slide, shape, or presentation element; leave a file comment; or choose “Ask ChatGPT for a change.” Multiple ChatGPT requests can be placed into an “Add to review” queue and sent together. The client submits these requests using a dedicated artifact_comment agent-turn trigger. Hidden instructions require ChatGPT to address every assigned thread, edit the artifact when requested, and return exactly one structured artifact-comment-reply for each thread. The client intercepts those replies and places them into the corresponding sidebar thread with pending, unread, resolved, and reopened states. PowerPoint comment changes are serialized back into PPTX bytes using bundled OpenXML infrastructure. Excel comment changes pass through a durable workbook-update system and are materialized into a new XLSX file. This establishes that native comments are intended to exist in the Office files, rather than only as visual overlays inside ChatGPT. The system is controlled by feature gate 1246162292 and requires the artifact panel to be expanded to full width. Its presence in a production client does not establish that OpenAI has enabled it for ordinary users. OpenAI publicly documents file previews, selection-based revisions, separate Excel and PowerPoint add-ins, and comments on webpages in Codex’s browser. RuntimeWire found no public OpenAI documentation describing this threaded Office review system, its private ChatGPT comment layer, its review queue, or its model-reply protocol.
Preserved the production app.asar from the installed Microsoft Store package and calculated its SHA-256 hash. Extracted the Electron archive without modifying the installed application. Located the Artifact comments interface by tracing user-facing strings including “Ask ChatGPT for a change,” “Add to review,” “PowerPoint comment,” “Excel comment,” and “New ChatGPT reply.” Followed comment creation from the artifact renderer into the desktop application’s state and message-handling code. Traced ChatGPT assignments into the dedicated artifact_comment turn trigger and inspected the hidden response contract used to route agent replies back to individual threads. Traced native PowerPoint comments through presentation export, OpenXML serialization, and workspace-file writeback. Traced native Excel comments through workbook-worker operations, durable document updates, conflict handling, XLSX export, and file materialization. Confirmed that native and ChatGPT-only threads are merged into one sidebar while using different persistence and export behavior. Compared archived extracted clients. The implementation was absent from builds checked through internal version 26.820.71523 and present by 26.825.32147. Reviewed OpenAI’s current product documentation and searched the public record for the feature’s exact interface and protocol terminology. The principal evidence comes from production minified JavaScript and bundled runtime files. Descriptive variable and component names were treated as supporting evidence only when corroborated by executable control flow, interface copy, persistence operations, or model-dispatch behavior.
Tested versions: OpenAI Codex Windows package 26.825.4187.0 — internal application version 26.825.32147.
Reproduction
RuntimeWire partially reproduced the finding.
Reviewed the unmodified production Windows client and traced the feature across its interface, agent workflow, and Office-file handling code.
File hashes
sha256:22DC7276D4A5616DAA504CE7519955AE420FFA4FB10A234B4DA38F74A4AC403F app.asar — OpenAI.Codex_26.825.4187.0
Company response
RuntimeWire requested comment; the company had not responded by publication time.
OpenAI has built a review system inside its shipping desktop client that turns selected parts of Excel workbooks and PowerPoint presentations into threaded assignments for ChatGPT, according to RuntimeWire's examination of the application package.
The interface, labeled "Artifact comments," combines native Microsoft Office comments with a second class of private threads assigned to ChatGPT. Users can queue requests against specific cells, charts, slides, shapes, or other selected elements, send several requests in a batch, and have ChatGPT edit the underlying file before replying inside each corresponding thread.
The implementation is present in the Windows package OpenAI.Codex_26.825.4187.0, with internal application version 26.825.32147. It is controlled by a rollout gate and does not establish that OpenAI has made the feature generally available. RuntimeWire found the implementation in that build after it was absent from archived desktop bundles through internal version 26.820.71523.
Two comment systems in one sidebar
Artifact comments merges two technically separate review systems.
The first writes conventional comments into .xlsx and .pptx files. These threads carry authors, timestamps, replies, resolution status, editing and deletion controls, and references to the selected Office object. PowerPoint targets can identify a slide, shape, selected element, frame, and anchor point. Excel targets can point to a range such as Sheet1!A1:B2 or to a floating object such as a chart.
The interface labels the composer according to the file type, using "PowerPoint comment" or "Excel comment," and tells users that the comment will be saved to the presentation or spreadsheet. The desktop client then reconstructs the Office package with the comment included rather than leaving a visual marker inside ChatGPT's preview.
The second system is private to ChatGPT. Its composer reads "Ask ChatGPT for a change," and the implementation describes those threads as persistent artifact comments assigned to ChatGPT that are not exported into the Office file. Each thread is tied to the conversation, local host, artifact path, a unique thread ID, the selected object or cell range, and an available text or image preview.
That distinction creates two review layers over the same file. A recipient can receive the native Office comments embedded in an exported workbook or presentation without receiving the separate instructions and replies exchanged with ChatGPT.
Comments become an agent queue
Users can submit an assignment immediately or choose "Add to review." The latter accumulates multiple requests and presents a single action to send the waiting comments to ChatGPT. The resulting instruction is framed as "Address these comments."
The client submits the batch through a dedicated artifact_comment turn type rather than as an ordinary chat message. Each attachment contains structured information including the file path, thread ID, comment text, target metadata, range label, and any available preview. Authorized workspace context is also available to the agent, allowing it to modify the file instead of returning suggested text alone.
The system accounts for additional assignments arriving while ChatGPT is already working. Its instructions treat later requests as additive and require the agent to complete every assigned thread before ending the turn. The design resembles an inbox for delegated document work, with selections serving as task scopes.
For each completed assignment, ChatGPT emits a structured artifact-comment-reply directive containing the exact artifact path, thread ID, and response body. The client validates the identifiers, consumes the directive, and inserts the answer under a ChatGPT identity in the appropriate sidebar thread. Users see dedicated states for pending work, "Thinking," new replies, unread messages, and active or resolved threads.
ChatGPT can edit the file and answer the assignment, but the implementation reserves resolution for the user. The agent cannot close its own review thread.
The edits persist in Office files
PowerPoint comment operations export the presentation's internal representation, serialize it back into the OpenXML format, and write the resulting bytes to the workspace path. The package includes OpenXML WebAssembly components and an internal serialization runtime called Walnut. Comment writes are serialized, and the interface includes a specific error for a failed save.
Excel uses a more elaborate durability path. Comment operations pass through the workbook worker and a collaborative document record that supports adding threads, replying, resolving, reopening, editing, deleting, and reacting. The client applies checkpoints and queued updates, recalculates the workbook, exports new XLSX bytes, and replaces the materialized file.
The Excel implementation also contains handling for stale updates, write conflicts, external file changes, oversized workbooks, and failed saves. That machinery is designed for edits that may outlast a single preview interaction or encounter interruptions while ChatGPT works.
OpenAI's public products stop short of the queue
OpenAI's documentation for ChatGPT Work says the product can create and edit documents, spreadsheets, presentations, reports, and analyses. It also describes selecting part of a file in the desktop app and requesting a change. The same page says PowerPoint is excluded from the Work desktop flow at launch.
OpenAI separately distributes ChatGPT for Excel and ChatGPT for PowerPoint, which operate as sidebars within Microsoft's applications. The Excel add-in can inspect and update formulas, values, formatting, and charts. The PowerPoint add-in can create and revise editable slides.
In its April 16th Codex update, OpenAI disclosed file previews and a separate system for placing comments on pages in Codex's in-app browser. Those descriptions establish selection-based editing and comment-driven browser work. They do not describe native Office threads, private @ChatGPT assignments, batched review requests, or the protocol that routes agent replies back to individual comments.
A rollout gate leaves policy questions open
Artifact comments activates only when a numeric feature gate is enabled and other conditions are met, including an authenticated account, an available conversation, an eligible artifact, a full-width artifact panel, and permission to use annotations. The implementation shows native comment persistence for PowerPoint and Excel. RuntimeWire found no equivalent persistence path for Word, PDF, CSV, TSV, HTML, or notebook files in the examined build.
The code also creates native-comment authors from the signed-in user's identity, including an email address. Because native threads are written into Office packages, the implementation raises a direct metadata question: which identity fields become visible to people who later receive the workbook or presentation.
Private ChatGPT assignments can include the local artifact path, thread and comment identifiers, target metadata, selected text or an image preview, and workspace context available to the agent. The client code does not establish the retention period for those threads, whether administrators can disable them separately, or how they appear in enterprise compliance records.
OpenAI's public app documentation says workspace administrators can control connected apps and that eligible Enterprise and Edu administrators may receive some conversation and app activity through its compliance systems. Artifact comments adds a separate supervisory layer inside local Office work. Its eventual rollout will require OpenAI to explain where that layer is stored and which parts travel with the file.
RuntimeWire contacted OpenAI for comment and is awaiting a reply.