A Renamed cmd.exe Counted as Codex

On 16 November Xusheng Li published a reverse-engineering write-up of a Windows 11 system file called MissionControlAggregator.dll. Inside it are the names of coding agents: copilot.exe, claude.exe, codex.exe and agy.exe in four slots for command-line agents, and github.exe, claude.exe, codex.exe, cursor.exe and antigravity.exe in five slots for desktop agents. Alongside them are developer tools such as python.exe, node.exe, gcc.exe, devenv.exe and code.exe, sorted into twelve broader categories. The feature that guards the agent checks is named CLICodingAgentUsage.

The post made the Hacker News front page, and the obvious question is whether Microsoft should be counting coding agents at all. I’m more interested in two narrower ones. What exactly does the count record? And how many machines does it end up coming from? The write-up answers the first precisely. The second turns on a checkbox, and there is a large literature on what checkboxes do.

What the bit is

Li followed the code instead of guessing from strings. The component doesn’t scan processes. It subscribes to event summaries Windows already produces (KernelProcess.AppStateChangeSummary and Win32kTraceLogging.AppInteractivitySummary for Windows programs, and the WSL exec events for Linux programs), passes the application names to a set of matchers, and sets a slot when a name matches. The slots are saved to disk under C:\ProgramData\Microsoft\Diagnosis\AggregatorStorage\. Roughly once a day they are packed into bitsets and emitted as an event called Segments.SubCohorts.

How a program name becomes an uploaded bit Flow diagram of the path described by Xusheng Li on 16 November 2026. Existing Windows event summaries carry application names to a name matcher inside MissionControlAggregator.dll. A match sets a slot that is saved to disk. About once a day the slots are packed into a bitmap event called Segments.SubCohorts. That event reaches DiagTrack's upload queue only when optional diagnostic data is allowed. A second gate sits before the matcher: the CLICodingAgentUsage feature, which was on in Li's own virtual machine and off by default in the public ISO image. MISSIONCONTROLAGGREGATOR · DATA PATH · PER LI, 16 NOV 2026 Gate 1: CLICodingAgentUsage feature (ID 62580080) on in Li's VM as a service experiment; off by default in the public ISO image Event summaries app names that Windows already logs Name matcher compares the executable name Saved slot on disk under ProgramData Bitmap report SubCohorts event, about daily Upload queue DiagTrack's EventStore.db Gate 2: only if optional diagnostic data is on In Li's reading, the report keeps no prompts, code or run time: only which slots matched.
Two gates decide whether any of this leaves a machine: a feature flag Microsoft controls and a diagnostics setting the user controls. The matcher in the middle decides what the bit means.

The report is small. The source events include run time, time in focus, interaction counts and resource use, but in Li’s reading the component “appears to discard those details and record only a bitmap of which agents were observed.” He found no prompt text or source code in it. He still thinks it matters: “collecting which agents and developer tools someone uses already has significant privacy implications.” I agree, though the reason I’d give is different from his. A bitmap of tool names is the kind of record nobody thinks of as sensitive until it’s joined to something else.

What the bit measures

To test the matcher, Li copied cmd.exe to a file named codex.exe and ran it. “The name alone was enough for the component to recognize it as Codex.” The saved state showed CLI slot 2 set, and the forced report came out as CliAgentSubCohorts: 4, which is 1 << 2. He did the same with a copy of cmd.exe named gcc.exe. It set developer-tool slot 3 and produced SubCohorts: [8], and that event was sitting in DiagTrack’s upload queue.

Two different programs, one Codex bit Diagram of the four command-line agent slots in the Segments.SubCohorts report, as tested by Xusheng Li. Both the real Codex CLI and a copy of cmd.exe renamed to codex.exe map to bit 2, the Codex slot, and the report value becomes 4, which is 1 shifted left by 2. The other three slots, which Li lists as copilot, claude and agy, stay empty. In a separate test, cmd.exe renamed to gcc.exe set developer-tool bit 3, value 8, a compiler and debugger category shared by GCC, Clang and related tools. CLIAGENTSUBCOHORTS · 4 SLOTS · LI'S RENAMED-BINARY TEST The Codex CLI file name codex.exe cmd.exe, copied and renamed file name codex.exe bit 3 bit 2 bit 1 bit 0 Codex slot Bit 2 set: CLI slot 2, the Codex slot. Li, 16 Nov 2026 0 1 0 0 value 4 1 << 2 Same test, developer-tool path: cmd.exe renamed gcc.exe set slot 3, value 8 (1 << 3), found in the upload queue. That slot is a shared compiler/debugger category (GCC, Clang and related tools), not one bit per program. Li lists the other CLI slots as copilot, claude and agy; his post confirms only the Codex position.
The report can't tell the two programs on the left apart. What it records is "a file with this name ran in the interactive session," which is a narrower fact than "this person used Codex."

Some slots do more than compare names. According to Li, the Node.js slot expects the publisher “Node.js Foundation,” Python expects “Python Software Foundation,” and the .NET and Visual Studio slots expect Microsoft Corporation. The compiler and debugger group has no publisher check on that path, and the renamed cmd.exe test shows the Codex slot accepted a copy of the Windows command shell as long as it carried the right name. So the Codex slot, which is the part getting the attention, is at least as loose as the compiler group.

The looseness works in two directions. Anyone who wants to pollute the count can rename a binary. Anyone who runs an agent under a wrapper, a different file name or a shim may not be counted. Neither case breaks the feature. It’s a cohort flag, built to sort machines into rough buckets, and a rough bucket only has to be right in aggregate. But it does mean the honest description of the record is “an executable with one of nine names was seen.” People who don’t want that record can’t get out of it by being careful about which tools they use, only by turning off the setting that sends it.

Who sends it

Two gates control whether the bit leaves a machine. Li is careful about the first one. The CLICodingAgentUsage feature was already on in his ARM64 virtual machine “as a service experiment,” and he hadn’t enabled it. He also writes that “the public ISO’s baseline VM had the coding-agent feature disabled by image default, so the evidence does not support saying that it is active on every Windows installation.” So the feature runs on some set of machines that Microsoft chooses, and the post can’t say how big that set is.

The second gate belongs to the user. In DiagTrack’s queue the agent event is marked "isRequired" : false, so it travels only with optional diagnostic data. Microsoft’s documentation for organizations describes optional data as including “data about the websites you browse, device activity, and enhanced error reporting,” and says of the lower tier, required diagnostic data, that “this is the default setting for Windows 10, version 1903 and later.” Li’s clean install of Windows 11 Home ARM64 from the current official ISO, with the region set to the United States, showed something different on the consumer setup screen. The diagnostic data toggle was already on and labeled “Include Optional,” and after he accepted the defaults the AllowTelemetry value was 3, which Microsoft’s policy table maps to “Optional (Full).”

I don’t think those two statements contradict each other. One describes what a managed device gets when no policy is set. The other describes what a person clicking through setup on a new laptop is shown. But only the second one matters for how many machines send this bit, and on that screen the choice is made in advance.

What a preset does

Jachimowicz, Duncan, Weber and Johnson pooled 58 default studies with 73,675 participants in a 2019 meta-analysis. Making an option the default, as opposed to requiring people to opt in, raised uptake by d = 0.68 (95% CI 0.53 to 0.83). For the 38 observations with yes-or-no outcomes, they translate that into “an absolute average of 27.24%.” The best-known example is older. In Johnson and Goldstein’s 2003 table of European organ-donor registries, effective consent ran from 4.25% to 28.79% in the opt-in countries they list and from 91.73% to 99.997% in the opt-out ones.

Effective consent under opt-in and opt-out defaults Horizontal bar chart of effective organ-donor consent rates from Johnson and Goldstein's 2003 supplementary table. Opt-in countries: Denmark 4.25 percent, Germany 12 percent, United Kingdom 16.9 percent, Netherlands 28.79 percent. Opt-out countries: Sweden 91.73 percent, Belgium 98 percent, Austria 99.99 percent. A footnote gives the 2019 meta-analysis by Jachimowicz and colleagues: across 58 studies the average default effect was d equals 0.68, or an absolute 27.24 percentage points for binary outcomes. EFFECTIVE CONSENT RATE · ORGAN-DONOR REGISTRIES · JOHNSON & GOLDSTEIN 2003, TABLE S1 OPT-IN (must choose to register) OPT-OUT (registered unless you object) Denmark Germany United Kingdom Netherlands Sweden Belgium Austria 4.25% — Denmark, opt-in, Johnson & Goldstein 2003 12% — Germany, opt-in, Johnson & Goldstein 2003 16.9% — United Kingdom, opt-in, Johnson & Goldstein 2003 28.79% — Netherlands, opt-in, Johnson & Goldstein 2003 91.73% — Sweden, opt-out, Johnson & Goldstein 2003 98% — Belgium, opt-out, Johnson & Goldstein 2003 99.99% — Austria, opt-out, Johnson & Goldstein 2003 4.25% 12% 16.9% 28.79% 91.73% 98% 99.99% Jachimowicz et al. 2019, 58 studies: average default effect d = 0.68, about 27 points on yes/no choices.
Organ donation is the extreme case, and most defaults move people far less. The meta-analysis average is the fairer prior for a setup-screen toggle, and it is still large.

I wouldn’t carry the organ-donor gap over to a Windows toggle. The meta-analysis itself warns against that. Its heterogeneity is extreme (I² of 98.21%): 46 observations were positive and significant, 10 were null and 2 went the other way. What it does offer is a rule for when defaults work best. Of the three mechanisms it tested, two moderated the effect significantly: endorsement, where people read the default as what the choice architect thinks they should do, and endowment, where the preset feels like the status quo. A setup screen written by the operating system’s own vendor, with the box already checked before the user arrives, looks designed to trigger both. The 2019 paper also found larger effects for consumer decisions.

The research on privacy settings specifically points the same way. Bahirat and colleagues, re-analyzing smart-home privacy decisions at CHI 2021, found that defaults and framing change the outcome and also the process: they weaken the link between what people think of a scenario and what they decide, so the decision gets more heuristic. Steffel, Williams and Pogacar ran seven experiments on whether telling people about a default protects them, and found it doesn’t. Disclosure changed how fair people judged the default to be, but did not make it any less influential. What did reduce the effect was making people state a preference, as in a forced choice. A Settings page that explains how to turn optional diagnostics off later is a disclosure, and in their experiments disclosure alone left the default’s pull intact.

What I’d want to know

None of this tells us how many US Windows 11 machines send the coding-agent bit, and I won’t guess. What I can say is which numbers would settle it, and none of them are public. We’d need the share of consumer installs that finish setup with optional diagnostics still on, the share of machines enrolled in the CLICodingAgentUsage experiment, and the sampling rate. Microsoft’s table says sampling applies to optional data where required data is collected at 100%.

The literature gives a strong prior for the first number. Any count of coding-agent use that Microsoft builds from this bit will lean toward people who left the setup preset alone, and it will count file names. Both caveats belong next to any chart built from it. For a developer reading Li’s post, the practical step is the one he gives: Settings, then Privacy & security, then Diagnostics & feedback, and switch off optional diagnostic data.

References

  1. Li, X. (2026). Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use. xusheng.dev, 16 November 2026.
  2. Microsoft. Configure Windows diagnostic data in your organization. Microsoft Learn, Windows Privacy documentation, page dated 11 September 2025.
  3. Jachimowicz, J. M., Duncan, S., Weber, E. U. and Johnson, E. J. (2019). When and why defaults influence decisions: a meta-analysis of default effects. Behavioural Public Policy, 3(2), 159-186, published online 24 January 2019.
  4. Johnson, E. J. and Goldstein, D. (2003). Do Defaults Save Lives?. Science, 302(5649), 1338-1339, 21 November 2003.
  5. Bahirat, P., Willemsen, M., He, Y., Sun, Q. and Knijnenburg, B. (2021). Overlooking Context: How do Defaults and Framing Reduce Deliberation in Smart Home Privacy Decision-Making?. Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, 1-18, 6 May 2021.
  6. Steffel, M., Williams, E. F. and Pogacar, R. (2016). Ethically Deployed Defaults: Transparency and Consumer Protection through Disclosure and Preference Articulation. Journal of Marketing Research, 53(5), 865-880, October 2016.