Support for rustup category home mode

[Cloud0310]

Related to #247.

Summary

Rustup historically stores configuration, state, data, and caches under a single RUSTUP_HOME (defaulting to $HOME/.rustup).
This prevents rustup from adhering to platform-standard directories (XDG on Unix, Known Folders on Windows).

This PR introduces an opt-in category-home layout guarded by RUSTUP_USE_CATEGORY_HOME, while preserving the legacy single-directory layout by default.

When category mode is enabled (RUSTUP_USE_CATEGORY_HOME=1 or any non-empty value other than "0"):

Category	Contents	Platform Default
Cache	downloads/, tmp/, update-hashes/	$XDG_CACHE_HOME/rustup or ~/.cache/rustup
Config	settings.toml	$XDG_CONFIG_HOME/rustup or ~/.config/rustup
Data	toolchains/, fallback/	$XDG_DATA_HOME/rustup or ~/.local/share/rustup
State	state.toml	$XDG_STATE_HOME/rustup or ~/.local/state/rustup
Bin	Proxies (rustc, cargo, etc.)	Unix: ~/.local/bin Windows: %USERPROFILE%/.local/bin (tentative)

Important

Why gated?
Gating both category overrides and platform defaults avoids split-brain installations where external tools (e.g. older rust-analyzer) still look exclusively in RUSTUP_HOME for toolchains.

---

Resolution Precedence

Note

This order still needs discussion, upon whether we should consider RUSTUP_HOME and CARGO_HOME. See open questions.

Design Guidance

We consider this as a breaking change, and RUSTUP_HOME as a purely legacy env
var, so in new mode, we don't use it anymore, so, on category mode available,
the fallback paths are ignored.

Categories (Cache / Config / Data / State)

1. RUSTUP_<CATEGORY>_HOME (if category mode is enabled)
2. RUSTUP_HOME (legacy shared override)
3. Platform category default (if category mode is enabled)
4. $HOME/.rustup (legacy fallback)

Note

On Unix, explicit absolute XDG_variables take precedence over $HOME-derived paths. Empty or relative XDG values are ignored.
Windows does not consult XDG variables.

Bin Directory

1. RUSTUP_BIN_HOME (if category mode is enabled)
2. $CARGO_HOME/bin
3. Platform default bin directory (if category mode is enabled: ~/.local/bin)
4. $HOME/.cargo/bin

---

Important Commits

feat(home): resolve category homes

Implements the core path resolution logic structured for future home crate compatibility across three layers:

1. Underlying platform resolution: Handles Unix XDG environment variables and Windows native directories.
2. home crate facade (mod home): Re-exports standard home APIs, implements RUSTUP_<CATEGORY>_HOME resolution and RUSTUP_HOME fallback without rollout logic.
3. Rollout strategy: Implements the gating logic driven by RUSTUP_USE_CATEGORY_HOME.

feat(uninstall): remove legacy and category rustup homes

Removes rustup home directories (config, cache, data, state) for both legacy and category layouts. (Excludes bin directory).

feat(uninstall): clean legacy and category cargo bin homes

Cleans up rustup-managed proxy binaries and symlinks across both legacy and category bin directories, leaving unrelated user binaries untouched.

feat(installer): migrate bin and env paths to split homes

Completes the installer-side migration from Cargo-owned paths to Rustup's resolved homes. In category mode, rustup binaries, proxy links, self-update artifacts, executable checks, and child-process PATH setup now use rustup_bin_home, while generated shell environment scripts use the config home through rustup_env_home. Legacy mode preserves the existing $CARGO_HOME layout.

The commit also updates shell profile integration, Windows PATH handling, installer messages, and end-to-end coverage for the split layout.

Open Questions

1. Windows Bin Location:
   • Currently set to %USERPROFILE%/.local/bin. Needs community consensus on whether this is the appropriate platform default for Windows.
   • Requires validating PATH precedence and migration UX from %USERPROFILE%\.cargo\bin.
2. Env Script Localtion
   • source env script should be in data home, config home or somewhere else?

Following tasks

• Create a cli interface for migration, rustup migration category-mode, for
  creating symlinks in legacy dir, pointing to coorresponding category dirs.
• Add warning message for asking user for migrating to new category mode.
• Update user-guide for new mode

[Cloud0310]

Fixed a minor CI/CD problem on Windows.

[Cloud0310]

About Windows' bin directory location problem, here's more context:

Directory	Pros	Cons	Precedent
%USERPROFILE%\.local\bin	Consistent across platforms; shared PATH entry	Not a Windows Known Folder; shared filenames complicate collision handling and cleanup	uv and uv tool install use this directory
%LOCALAPPDATA%\rustup\bin	Clear ownership; relatively short path	Separate PATH entry; application-defined layout	pnpm 11+ uses %LOCALAPPDATA%\pnpm\bin
%LOCALAPPDATA%\Programs\Rustup\bin	Standard per-user programs parent directory; clear ownership	Longer path; separate PATH entry	VS Code user setup installs under %LOCALAPPDATA%\Programs\Microsoft VS Code

@ChrisDenton @rami3l , Current PR's defaulting to use %USERPROFILE%\.local\bin, but I think we need to discuss upon this.

[ChrisDenton]

Hm, since I feel one of the points of this feature is to use more platform native paths, I don't think %USERPROFILE%\.local\bin is the best choice. That seems quite alien to Windows. I suppose a more Windows-like equivalent would be %LOCALAPPDATA%\Programs\bin but I don't think there's much (if any) precedent for that.

I think either of the other two options are justifiable. Putting it under "Programs" is the most technically correct I guess but if anything there's a weaker precedent for that, albeit applications that used to support XP wouldn't have used it because "Programs" wasn't a thing back then.

[djc]

What are we doing for macOS here? The XDG stuff definitely doesn't feel native for it.

[ChrisDenton]

There appears to be a fairly strong use of XDG for macOS cli utilities. I cannot find any officially documented conventions that aren't intended for app bundles but unofficially almost everything in the shell seems to treat it as a Unix (which it is).

[Cloud0310]

There appears to be a fairly strong use of XDG for macOS cli utilities. I cannot find any officially documented conventions that aren't intended for app bundles but unofficially almost everything in the shell seems to treat it as a Unix (which it is).

That's what @rami3l and I wants as well.
I personally agree with this idea to consider macOS as Unix, which also agrees with current repo's state: we only maintain two major platforms: Unix/Windows.
Currently introducing another macOS specific code would not fit in.

[rami3l]

In terms of precedence of different configuration options, as I have discussed with @Cloud0310, the only remaining concern is what to do when RUSTUP_USE_CATEGORY_HOME=1 meets explicit RUSTUP_HOME or CARGO_HOME overrides.

As per the previous summary in https://blog.rust-lang.org/inside-rust/2025/10/01/this-development-cycle-in-cargo-1.90/#all-hands-xdg-paths:

The most immediate problem is Rustup setting CARGO_HOME. To not break compatibility, Cargo should always respect it if CARGO_HOME is set.

This means CARGO_HOME must have higher precedence than RUSTUP_USE_CATEGORY_HOME=1. To ensure consistency across both rustup and cargo, I propose that we treat RUSTUP_HOME similarly, hence the following fallback order:

1. RUSTUP_XXX_HOME. Setting this means that the user has explicitly deviated from the old paths. RUSTUP_USE_CATEGORY_HOME=1 only.
2. The old override path, whether that is CARGO_HOME/bin or RUSTUP_HOME/xxx.
3. The platformdirs-derived path that is derived from XDG on Unix, or the registry on Windows. RUSTUP_USE_CATEGORY_HOME=1 only.
4. The old default path, whether that is ~/.cargo/bin or ~/.rustup/xxx.

[rami3l]

What are we doing for macOS here? The XDG stuff definitely doesn't feel native for it.

AFAIK the standard library currently uses the same logic for macOS and Linux, so I think that is a very good default to start with: https://doc.rust-lang.org/stable/std/os/unix/xdg/fn.cache_home_dir.html

Also, this is the convention followed by other existing tools such as neovim and uv.

I am aware of the frequently-quoted macOS documentation page which to some may suggest paths like ~/Library/ApplicationSupport/rustup/toolchains etc.

I would again point out (as @ChrisDenton has mentioned above) that these guidelines are specific to app bundles (that page is full of app bundle-specific paths, with Library being the only out-of-bundle directory), for which using this convention may facilitate migration to iOS apps, sandboxed bundles or iCloud-synced data. However, none of the above would be interesting to rustup usage on macOS (also note that uv has explicitly migrated from macOS bundle-specific paths to general XDG paths).

If the user really prefers to place stuff in ApplicationSupport, we have explicit overrides for that on both rustup and XDG levels. Oftentimes they would only need to set an $XDG_CONFIG_HOME override and that will work for all their CLI apps, not just rustup.

[Cloud0310]

In terms of precedence of different configuration options, as I have discussed with @Cloud0310, the only remaining concern is what to do when RUSTUP_USE_CATEGORY_HOME=1 meets explicit RUSTUP_HOME or CARGO_HOME overrides.

As per the previous summary in https://blog.rust-lang.org/inside-rust/2025/10/01/this-development-cycle-in-cargo-1.90/#all-hands-xdg-paths:

The most immediate problem is Rustup setting CARGO_HOME. To not break compatibility, Cargo should always respect it if CARGO_HOME is set.

This means CARGO_HOME must have higher precedence than RUSTUP_USE_CATEGORY_HOME=1. To ensure consistency across both rustup and cargo, I propose that we treat RUSTUP_HOME similarly, hence the following fallback order:

1. RUSTUP_XXX_HOME. Setting this means that the user has explicitly deviated from the old paths. RUSTUP_USE_CATEGORY_HOME=1 only.
2. The old override path, whether that is CARGO_HOME/bin or RUSTUP_HOME/xxx.
3. The platformdirs-derived path that is derived from XDG on Unix, or the registry on Windows. RUSTUP_USE_CATEGORY_HOME=1 only.
4. The old default path, whether that is ~/.cargo/bin or ~/.rustup/xxx.

This also has an effect on the env var recursive forwarding system, I guess once we're in new mode, I need to stop forwarding cargo_home and rustup_home in case of causing underlaying program having confusing about if this var is from manually setting or from rustup resolution.

[rami3l]

This also has an effect on the env var recursive forwarding system, I guess once we're in new mode, I need to stop forwarding cargo_home and rustup_home in case of causing underlaying program having confusing about if this var is from manually setting or from rustup resolution.

@Cloud0310 Exactly, and I'd like to see more regression tests on this point to make sure that we have done the right thing.

[djc]

For macOS, I'm particularly concerned about the cache directories. If we don't put them in the location that the OS understands, we basically don't get to benefit from the OS managing caches, which seems bad.

For binaries, won't it cause issues if we share a bin dir with other tooling as opposed to the rustup/Cargo-specific bin we have today?

[rami3l]

For binaries, won't it cause issues if we share a bin dir with other tooling as opposed to the rustup/Cargo-specific bin we have today?

@djc What potential issues do you have in mind to be precise?

• In terms of directory nuking the semantics has already been changed in Stop nuking $CARGO_HOME (~/.cargo) on rustup self uninstall #285, as many would prefer to keep the cargo-installed binaries after rustup self uninstall in the first place.
• In terms of defaults I think it'd be okay to default to a common directory as everyone else (such as UV and cabal) but we keep the possibility of having a specific one.

I'm particularly concerned about the cache directories. If we don't put them in the location that the OS understands, we basically don't get to benefit from the OS managing caches, which seems bad.

I am aware of special macOS mechanisms that may force the shrinking of ~/Library/Caches from time to time, but the point here, at least for me, is that we are actively cleaning up RUSTUP_CACHE_HOME so the size of that dir should be below 1MiB almost all the time (after a transaction has successfully finished), so this is not a concern here. However I'm curious what the std/uv people think about this.

[Cloud0310]

About Windows' bin directory location problem, here's more context:

Directory Pros Cons Precedent
%USERPROFILE%\.local\bin Consistent across platforms; shared PATH entry Not a Windows Known Folder; shared filenames complicate collision handling and cleanup uv and uv tool install use this directory
%LOCALAPPDATA%\rustup\bin Clear ownership; relatively short path Separate PATH entry; application-defined layout pnpm 11+ uses %LOCALAPPDATA%\pnpm\bin
%LOCALAPPDATA%\Programs\Rustup\bin Standard per-user programs parent directory; clear ownership Longer path; separate PATH entry VS Code user setup installs under %LOCALAPPDATA%\Programs\Microsoft VS Code
@ChrisDenton @rami3l , Current PR's defaulting to use %USERPROFILE%\.local\bin, but I think we need to discuss upon this.

@epage what's cargo's decision here? Maybe we want to sync this "favor" choice with cargo team.

[rami3l]

@epage Note that we are not pressuring cargo to do the final decision because in the opt-in period we can still change the default value, but it'd be interesting to see if we can land on something in common first.

@Cloud0310 For Windows’ bin dir my personal preference would be the option 2 (PNPM) above.

[djc]

I think the commit history here could stand to be improved. Would suggest:

• Move all commits that are not directly related to this feature into another PR (chore(process.rs): re-organize imports, potentially some of refactor(installer): generalize path formatting helpers)
• Then have one commit per each of the 5 categories here, from least to most intrusive

[Cloud0310]

@Cloud0310 For Windows’ bin dir my personal preference would be the option 2 (PNPM) above.
Why option 2 here? I want your reason.

Untill we finally out of opt-in period and settle on this choice, I would just keep using %USERPROFILE%\.local\bin with a TODO comment for this in current PR.

[Cloud0310]

Self reviewed and checked again, also changed the resolution order as comment. And as for the forwarding system, I added extra conditions for making sure backward compatibility.
And as for split PRs, I guess maybe today or tomorrow it would be done.

[rustbot]

☔ The latest upstream changes (possibly #5096) made this pull request unmergeable. Please resolve the merge conflicts.