Turning off optional Windows reports isn’t proof coding-tool checks stop
A renamed Command Prompt triggered a Codex indicator. The investigation traces local tracking to an optional…
The short version
- Xusheng Li’s Windows investigation found locally saved coding-tool indicators and, with optional diagnostics enabled, a report queued for upload.
- Renaming Command Prompt to codex.exe activated the Codex indicator, showing that this flag does not establish genuine Codex use.
- Disabling optional diagnostics controls optional reporting, but its effect on local matching and saved state remains unresolved.
Windows’ coding-agent telemetry needs a better explanation than a checkbox
Command Prompt needed only a costume change to register as Codex in Xusheng Li’s Windows test: rename a copy codex.exe, run it, and the indicator lights up. Before Windows classifies my coding tools, I’d like more than “optional diagnostic data.”

Li published his investigation on November 16. I haven’t reproduced it. He found a component that classifies coding-tool activity and saves indicators locally. With optional diagnostics enabled, his test also queued a report for upload, not proof Microsoft received it.
Whether declining optional reporting stops local observation remains unresolved. That deserves an answer without me upgrading the finding into “Microsoft is reading everyone’s code.”
How an application launch becomes a tool-use indicator
The plumbing tells me more than the report screenshot. This component classifies activity using existing Windows application-state and interactivity events. WSL execution events feed it too; running a tool inside Linux doesn’t automatically hide its identity.
Li’s reverse engineering traces DiagTrack to AggregatorHost.exe, which loads MissionControlAggregator.dll. The component extracts application identities from existing Windows and WSL events, then checks built-in executable names; some developer-tool checks also verify the publisher. Matches set slots in local state, saved on disk beyond the triggering event. A timer-driven routine converts those slots into bitsets and emits a Segments.SubCohorts diagnostic event identifying categories observed during the collection window. Li says richer incoming details, including runtime information, appear discarded.
The DLL had four command-line coding-agent slots, five desktop-agent slots and twelve broader developer-tool categories. I read those as available buckets, not counts of genuine products used or launches. Li also says timestamps describe the collection window, not agent usage time. An afternoon-long window is not an afternoon spent in Codex.
The ordinary-diagnostics argument deserves consideration. Li found no prompts or source code in this report. Sulat’s account also cites Microsoft’s existing documentation of application-launch activity within optional diagnostics. Compact categories expose substantially less than transcripts or repositories. That limits my privacy claim.
Li explains why he still objects:
Although I found no prompt text or source code in this report, collecting which agents and developer tools someone uses already has significant privacy implications.
I agree. Tool choice reveals work habits even when work stays private. Broad application-activity disclosures don’t explain why my operating system needs dedicated coding-assistant checks.
Command Prompt borrowed Codex’s name badge
My favourite experiment needed no elaborate exploit. Li copied cmd.exe, renamed it codex.exe, and ran it. The Codex slot activated. Apparently, Command Prompt passed the interview on its CV header.
I read that as a clean test of one matching rule: same program, different filename, indicator set. This flag therefore doesn’t prove genuine Codex use. Some other checks verify publishers, so the result cannot extend to every DLL category. But interpretations of this Codex flag need that filename caveat. Otherwise, it vanishes between the technical write-up and somebody’s product-adoption presentation.
Kanishka Narayan said of British AI hardware investment:
For too long, Britain has produced some of the world’s best scientific ideas, only to see too much of the technology, industry and wealth they create built elsewhere.
The queued report contained CliAgentSubCohorts: [4]. That encodes the Codex flag, not four launches or four users. Before trusting a chart, I’d ask how its analysis handles a renamed Windows executable qualifying as an AI coding agent.
The queue route matters too. After a match changes local state, the reporting routine emits an event through ETW, Windows’ event-reporting system. DiagTrack can consume it. With optional diagnostics allowed, Li found the report in DiagTrack’s upload queue, classified isRequired: false. Secondary accounts describe that observation; they don’t independently establish Microsoft’s receipt.
Nor was this an untouched recording of an ordinary day. Li shifted the saved reporting-window start backward by just over a day and restarted the service to accelerate generation. His analysis describes a roughly daily boundary. That intervention demonstrates the mechanism, not precisely when an unmodified machine would report.

I therefore reject the claim that Microsoft knows whether someone ran Codex or Cursor whenever optional diagnostics are enabled. The renamed executable undermines certainty about genuine Codex use. A separate clean-install test also found the feature disabled by image default.
A setup choice doesn’t establish a rollout
A selected checkbox makes a persuasive screenshot, not a deployment count.
In Li’s US-region Windows Home ARM64 setup test, Include Optional was preselected. That establishes that setup’s reporting choice. Yet his clean public-ISO baseline had the coding-agent feature disabled by image default.
His active environment was a Parallels Windows ARM64 virtual machine. Li reported the feature already enabled there as a service experiment, without manual activation. An active installation and a clean-image baseline are different observations. Neither provides a denominator for estimating activation’s reach.
I separate activation from reporting. The feature flag controls whether coding-agent functionality runs; optional diagnostics permits its report into the optional reporting channel. A machine can allow optional reporting while the classifier stays inactive. Where it’s active, withdrawing reporting permission raises another question: what continues locally? Neither finding the DLL nor seeing optional diagnostics selected establishes activation. A useful test would hold activation constant while changing reporting permission.
Kat Jones said of AI datacentre power requirements:
Hyperscale AI datacentres simply cannot run on renewable power alone, they absolutely need a highly reliable source of grid power.
That weakens my broadest objection. I can’t say every Windows developer is classified this way, or diagnose your laptop from somebody else’s VM. The investigation doesn’t establish deployment across editions, regions, architectures or configurations. A US setup screen cannot certify a laptop bought in Torino.
For a company assessing its fleet, I’d establish both reporting configuration and which machines ran the checks. Code present and code observed running support different conclusions.
I haven’t found a feature-specific Microsoft explanation in these accounts. Its purpose, Microsoft’s use of the indicators, retention period and access controls remain unknown. Li finding no prompts or source code in this report cannot clear other diagnostic reports; those need separate inspection.
What I’d ask before trusting the switch
I’d start with the supported control: Settings → Privacy & security → Diagnostics & feedback → Send optional diagnostic data. Turn it off if, like me, you don’t want optional reports sent. The downloadable “privacy fix” script with dozens of unexplained changes can stay with whoever enjoys spending Saturday repairing Windows.
What matters is what that buys me. Windows separates required diagnostics from optional reporting; Li found this event classified as optional. Disabling optional diagnostics removes permission to use that reporting tier. Required reporting remains separate. The setting doesn’t establish that previously collected or transmitted information disappears. Its effect on the local matcher and saved state also remains unresolved in Li’s investigation. Until Microsoft documents otherwise, I’d call it a reporting control, not proof local observation stopped.
Microsoft says Windows remains equally secure and operates normally with required-only diagnostics, as EnjoyAgentic’s guide explains. I haven’t compared security or performance myself. That assurance makes the supported setting a reasonable starting point, minus the laptop restoration project.
The same Settings area offers Diagnostic Data Viewer. EnjoyAgentic explains that it shows data captured while viewing is enabled, not a complete history. I’d search for MissionControlAggregator, keeping that capture limit in mind.
A matching event supplies evidence within that window. An empty search doesn’t rule out activity outside it or local state the viewer doesn’t expose. I wouldn’t declare a machine cleared from an empty results screen.
On an organization-managed device, I’d take an unavailable switch to IT: which diagnostic policy applies? Bypassing workplace policy adds trouble without establishing classifier activation.
My questions for Microsoft are specific. What product purpose requires coding-tool indicators? Which devices run these checks? Does opting out stop local matching, and what happens to saved state? What are the retention period and access controls? The design cannot establish advertising, competitive intelligence or model training as its purpose. I won’t substitute whichever accusation gets clicks.
A checkbox label cannot contain an implementation. But it can explain my decision, especially when software observes my work tools and saves the result before upload becomes relevant.
Give me a plain sentence beneath that switch explaining whether observation continues after I opt out, with a link to exactly what gets sent. If Microsoft needs room, I volunteer the space occupied by another invitation to try Copilot.
Frequently asked questions
Does Windows 11 send coding-agent usage data to Microsoft?
Xusheng Li’s investigation found coding-tool category indicators saved locally and a report queued for upload when optional diagnostics were enabled on an active installation. A queued report does not establish Microsoft’s receipt. The investigation also does not establish that the feature is active across all Windows 11 installations.
Does turning off optional diagnostic data stop Windows from tracking coding tools?
Turning off optional diagnostic data removes permission to use the optional reporting tier. Li’s investigation does not resolve whether that change stops the local coding-tool matcher or affects its saved state. Required diagnostics remain separate, and the setting does not establish deletion of previously collected or transmitted information.
Did the Windows coding-tool report contain prompts or source code?
Li found no prompt text or source code in the coding-tool report he inspected. The report represented observed categories through compact indicators rather than transcripts or repositories. That finding applies to this report only; it does not establish what other Windows diagnostic reports contain or how Microsoft uses the indicators.
Sources
- Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use
- Windows 11 can tell Microsoft which coding agents you run. How to check and stop it
- Windows 11 Telemetry: MissionControlAggregator Tracks Your Coding Agents
- Windows 11 Optional Telemetry Can Report Coding Agent Usage, Research Finds
- Windows 11 queues a report of which coding agents you run, Xusheng Li’s teardown finds
- It has been discovered that Windows 11 collects usage data for things like Claude Code and Codex and sends it to Microsoft.