Add Codex config for static trace span attributes and structured W3C
tracestate field upserts. The config flows through OtelSettings so
callers can attach trace metadata without touching every span call site.
Apply span attributes with an SDK span processor so every exported
trace span carries the configured metadata. Model tracestate as nested
member fields so configured keys can be upserted while unrelated
propagated state in the same member is preserved.
Validate configured tracestate before installing provider-global state,
including header-unsafe values the SDK does not reject by itself. This
keeps Codex from propagating malformed trace context from config.
Update the config schema, public docs, and OTLP loopback coverage for
config parsing, span export, propagation, and invalid-header rejection.
## Why
Codex desktop copies bundled Windows binaries out of `WindowsApps` into
a LocalAppData runtime cache before launching `codex.exe`. Sandboxed
commands can then need to execute helpers from that cache, but the
sandbox user group may not have read/execute access to the runtime bin
directory.
This makes the Windows sandbox refresh path repair that access directly
so the packaged desktop runtime remains usable from sandboxed sessions.
## What changed
- Added `setup_runtime_bin` to locate `%LOCALAPPDATA%\OpenAI\Codex\bin`,
matching the desktop bundled-binaries destination path, with the same
`USERPROFILE\AppData\Local` fallback shape.
- During refresh setup, check whether `CodexSandboxUsers` already has
read/execute access to the runtime bin directory.
- If access is missing, grant `CodexSandboxUsers` `OI/CI/RX` inheritance
on that directory.
- If the runtime bin directory does not exist, no-op cleanly.
## Verification
- `cargo build -p codex-windows-sandbox --bin
codex-windows-sandbox-setup`
- `cargo test -p codex-windows-sandbox --bin
codex-windows-sandbox-setup`
- Manual Windows ACL exercise against the installed packaged runtime
bin:
- existing inherited `CodexSandboxUsers:(I)(OI)(CI)(RX)` no-ops without
changing SDDL
- after disabling inheritance and removing the group ACE, setup adds
`CodexSandboxUsers:(OI)(CI)(RX)`
- with `LOCALAPPDATA` pointed at a fake location without
`OpenAI\Codex\bin`, setup exits successfully and does not create the
directory
- restored the real runtime bin with inherited ACLs and confirmed the
final SDDL matched the baseline exactly
## Why
Windows sandboxed commands run as a sandbox user, while workspace
repositories are usually owned by the real user. The sandbox compensates
by injecting a temporary Git `safe.directory` entry into the child
environment.
That injection was still broken for linked worktrees because the helper
followed the `.git` file's `gitdir:` pointer and injected the internal
`.git/worktrees/...` location. Git's dubious-ownership check expects the
worktree root instead, so sandboxed Git commands still failed in
worktree-based Codex checkouts.
## What changed
- Treat any `.git` marker, directory or file, as the worktree root for
`safe.directory` injection.
- Keep the safe-directory logic in
`windows-sandbox-rs/src/sandbox_utils.rs` and have the one-shot elevated
path reuse it.
- Add regression coverage for both normal `.git` directories and
gitfile-based worktrees.
## Validation
- `cargo test -p codex-windows-sandbox sandbox_utils::tests`
- `cargo test -p codex-windows-sandbox` built and ran; the new
`sandbox_utils` tests passed, while two pre-existing legacy sandbox
tests failed locally with `Access is denied`:
`session::tests::legacy_non_tty_cmd_emits_output` and
`spawn_prep::tests::legacy_spawn_env_applies_offline_network_rewrite`.
## Why
On Windows, background terminals could stay visible after their shell
process had already exited. The elevated runner waits for the PTY output
reader to reach EOF before it sends the final exit message, but the
ConPTY helper was reducing ownership down to raw handles too early. That
left the pseudoconsole's borrowed pipe handles alive past teardown, so
EOF never propagated and the session stayed `running`.
## What changed
- change `utils/pty/src/win/conpty.rs` to hand off owned ConPTY
resources instead of leaking only raw handles
- make `windows-sandbox-rs/src/conpty/mod.rs` keep the pseudoconsole
owner and the backing pipe handles together until teardown
- update the elevated runner and the legacy unified-exec backend to keep
that `ConptyInstance` alive, take only the specific pipe handles they
need, and drop the owner at teardown instead of trying to close a
detached pseudoconsole handle later
## Testing
- desktop app in `Auto-review`: 11 x `cmd /c "ping -n 3 google.com"` all
exited cleanly and did not accumulate in the UI
- desktop app in `Auto-review`: 5 x `cmd /c "ping -n 30 google.com"`
appeared in the UI and drained back out on their own
## Status
This is the Bazel PR-CI cross-compilation follow-up to #20485. It is
intentionally split from the Cargo/cargo-xwin release-build PoC so
#20485 can stay as the historical release-build exploration. The
unrelated async-utils test cleanup has been moved to #20686, so this PR
is focused on the Windows Bazel CI path.
The intended tradeoff is now explicit in `.github/workflows/bazel.yml`:
pull requests get the fast Windows cross-compiled Bazel test leg, while
post-merge pushes to `main` run both that fast cross leg and a fully
native Windows Bazel test leg. The native main-only job keeps full
V8/code-mode coverage and gets a 40-minute timeout because it is less
latency-sensitive than PR CI. All other Bazel jobs remain at 30 minutes.
## Why
Windows Bazel PR CI currently does the expensive part of the build on
Windows. A native Windows Bazel test job on `main` completed in about
28m12s, leaving very little headroom under the 30-minute job timeout and
making Windows the slowest PR signal.
#20485 showed that Windows cross-compilation can be materially faster
for Cargo release builds, but PR CI needs Bazel because Bazel owns our
test sharding, flaky-test retries, and integration-test layout. This PR
applies the same high-level shape we already use for macOS Bazel CI:
compile with remote Linux execution, then run platform-specific tests on
the platform runner.
The compromise is deliberately signal-aware: code-mode/V8 changes are
rare enough that PR CI can accept losing the direct V8/code-mode
smoke-test signal temporarily, while `main` still runs the native
Windows job post-merge to catch that class of regression. A follow-up PR
should investigate making the cross-built Windows gnullvm V8 archive
pass the direct V8/code-mode tests so this tradeoff can eventually go
away.
## What Changed
- Adds a `ci-windows-cross` Bazel config that targets
`x86_64-pc-windows-gnullvm`, uses Linux RBE for build actions, and keeps
`TestRunner` actions local on the Windows runner.
- Adds explicit Windows platform definitions for
`windows_x86_64_gnullvm`, `windows_x86_64_msvc`, and a bridge toolchain
that lets gnullvm test targets execute under the Windows MSVC host
platform.
- Updates the Windows Bazel PR test leg to opt into the cross-compile
path via `--windows-cross-compile` and `--remote-download-toplevel`.
- Adds a `test-windows-native-main` job that runs only for `push` events
on `refs/heads/main`, uses the native Windows Bazel path, includes
V8/code-mode smoke tests, and has `timeout-minutes: 40`.
- Keeps fork/community PRs without `BUILDBUDDY_API_KEY` on the previous
local Windows MSVC-host fallback, including
`--host_platform=//:local_windows_msvc` and `--jobs=8`.
- Preserves the existing integration-test shape on non-gnullvm
platforms, while generating Windows-cross wrapper targets only for
`windows_gnullvm`.
- Resolves `CARGO_BIN_EXE_*` values from runfiles at test runtime,
avoiding hard-coded Cargo paths and duplicate test runfiles.
- Extends the V8 Bazel patches enough for the
`x86_64-pc-windows-gnullvm` target and Linux remote execution path.
- Makes the Windows sandbox test cwd derive from `INSTA_WORKSPACE_ROOT`
at runtime when Bazel provides it, because cross-compiled binaries may
contain Linux compile-time paths.
- Keeps the direct V8/code-mode unit smoke tests out of the Windows
cross PR path for now while native Windows CI continues to cover them
post-merge.
## Command Shape
The fast Windows PR test leg invokes the normal Bazel CI wrapper like
this:
```shell
./.github/scripts/run-bazel-ci.sh \
--print-failed-action-summary \
--print-failed-test-logs \
--windows-cross-compile \
--remote-download-toplevel \
-- \
test \
--test_tag_filters=-argument-comment-lint \
--test_verbose_timeout_warnings \
--build_metadata=COMMIT_SHA=${GITHUB_SHA} \
-- \
//... \
-//third_party/v8:all \
-//codex-rs/code-mode:code-mode-unit-tests \
-//codex-rs/v8-poc:v8-poc-unit-tests
```
With the BuildBuddy secret available on Windows, the wrapper selects
`--config=ci-windows-cross` and appends the important Windows-cross
overrides after rc expansion:
```shell
--host_platform=//:rbe
--shell_executable=/bin/bash
--action_env=PATH=/usr/bin:/bin
--host_action_env=PATH=/usr/bin:/bin
--test_env=PATH=${CODEX_BAZEL_WINDOWS_PATH}
```
The native post-merge Windows job intentionally omits
`--windows-cross-compile` and does not exclude the V8/code-mode unit
targets:
```shell
./.github/scripts/run-bazel-ci.sh \
--print-failed-action-summary \
--print-failed-test-logs \
-- \
test \
--test_tag_filters=-argument-comment-lint \
--test_verbose_timeout_warnings \
--build_metadata=COMMIT_SHA=${GITHUB_SHA} \
--build_metadata=TAG_windows_native_main=true \
-- \
//... \
-//third_party/v8:all
```
## Research Notes
The existing macOS Bazel CI config already uses the model we want here:
build actions run remotely with `--strategy=remote`, but `TestRunner`
actions execute on the macOS runner. This PR mirrors that pattern for
Windows with `--strategy=TestRunner=local`.
The important Bazel detail is that `rules_rs` is already targeting
`x86_64-pc-windows-gnullvm` for Windows Bazel PR tests. This PR changes
where the build actions execute; it does not switch the Bazel PR test
target to Cargo, `cargo-nextest`, or the MSVC release target.
Cargo release builds differ from this Bazel path for V8: the normal
Windows Cargo release target is MSVC, and `rusty_v8` publishes prebuilt
Windows MSVC `.lib.gz` archives. The Bazel PR path targets
`windows-gnullvm`; `rusty_v8` does not publish a prebuilt Windows
GNU/gnullvm archive, so this PR builds that archive in-tree. That
Linux-RBE-built gnullvm archive currently crashes in direct V8/code-mode
smoke tests, which is why the workflow keeps native Windows coverage on
`main`.
The less obvious Bazel detail is test wrapper selection. Bazel chooses
the Windows test wrapper (`tw.exe`) from the test action execution
platform, not merely from the Rust target triple. The outer
`workspace_root_test` therefore declares the default test toolchain and
uses the bridge toolchain above so the test action executes on Windows
while its inner Rust binary is built for gnullvm.
The V8 investigation exposed a Windows-client gotcha: even when an
action execution platform is Linux RBE, Bazel can still derive the
genrule shell path from the Windows client. That produced remote
commands trying to run `C:\Program Files\Git\usr\bin\bash.exe` on Linux
workers. The wrapper now passes `--shell_executable=/bin/bash` with
`--host_platform=//:rbe` for the Windows cross path.
The same Windows-client/Linux-RBE boundary also affected
`third_party/v8:binding_cc`: a multiline genrule command can carry CRLF
line endings into Linux remote bash, which failed as `$'\r'`. That
genrule now keeps the `sed` command on one physical shell line while
using an explicit Starlark join so the shell arguments stay readable.
## Verification
Local checks included:
```shell
bash -n .github/scripts/run-bazel-ci.sh
bash -n workspace_root_test_launcher.sh.tpl
ruby -e "require %q{yaml}; YAML.load_file(%q{.github/workflows/bazel.yml}); puts %q{ok}"
RUNNER_OS=Linux ./scripts/list-bazel-clippy-targets.sh
RUNNER_OS=Windows ./scripts/list-bazel-clippy-targets.sh
RUNNER_OS=Linux ./tools/argument-comment-lint/list-bazel-targets.sh
RUNNER_OS=Windows ./tools/argument-comment-lint/list-bazel-targets.sh
```
The Linux clippy and argument-comment target lists contain zero
`*-windows-cross-bin` labels, while the Windows lists still include 47
Windows-cross internal test binaries.
CI evidence:
- Baseline native Windows Bazel test on `main`: success in about 28m12s,
https://github.com/openai/codex/actions/runs/25206257208/job/73907325959
- Green Windows-cross Bazel run on the split PR before adding the
main-only native leg: Windows test 9m16s, Windows release verify 5m10s,
Windows clippy 4m43s,
https://github.com/openai/codex/actions/runs/25231890068
- The latest SHA adds the explicit PR-vs-main tradeoff in `bazel.yml`;
CI is rerunning on that focused diff.
## Follow-Up
A subsequent PR should investigate making a cross-built Windows binary
work with V8/code-mode enabled. Likely options are either making the
Linux-RBE-built `windows-gnullvm` V8 archive correct at runtime, or
evaluating whether a Bazel MSVC target/toolchain can reuse the same
prebuilt MSVC `rusty_v8` archive shape that Cargo release builds already
use.
## Summary
This PR installs a first wave of WFP (Windows Filtering Platform)
filters that reduce the surface area of network egress vulnerabilities
for the Windows Sandbox.
- Add persistent Windows Filtering Platform provider, sublayer, and
filters for the Windows sandbox offline account.
- Install WFP filters during elevated full setup, log failures
non-fatally, and emit setup metrics when analytics are enabled.
- Bump the Windows sandbox setup version so existing users rerun full
setup and receive the new filters.
## What WFP is
Windows Filtering Platform (WFP) is the low-level Windows networking
policy engine underneath things like Windows Firewall. It lets
privileged code install persistent filtering rules at specific network
stack layers, with conditions like "only traffic from this Windows
account" or "only this remote port," and an action like block.
In this change, we create a Codex-owned persistent WFP provider and
sublayer, then install block filters scoped to the Windows sandbox's
offline user account via `ALE_USER_ID`. That means the filters are
targeted at sandboxed processes running as that account, rather than
globally affecting the host.
## Initial filter set
We are starting with 12 concrete WFP filters across a few high-value
bypass surfaces. The table below describes the filter families rather
than one filter per row:
| Area | Concrete filters | Purpose |
| --- | --- | --- |
| ICMP | 4 filters: ICMP v4/v6 on `ALE_AUTH_CONNECT` and
`ALE_RESOURCE_ASSIGNMENT` | Block direct ping-style network reachability
checks from the offline account. |
| DNS | 2 filters: remote port `53` on `ALE_AUTH_CONNECT_V4/V6` | Block
direct DNS queries that bypass our intended proxy/offline path. |
| DNS-over-TLS | 2 filters: remote port `853` on
`ALE_AUTH_CONNECT_V4/V6` | Block encrypted DNS attempts that could
bypass ordinary DNS interception. |
| SMB / NetBIOS | 4 filters: remote ports `445` and `139` on
`ALE_AUTH_CONNECT_V4/V6` | Block Windows file-sharing/network share
traffic from sandboxed processes. |
For IPv4/IPv6 coverage, the port-based filters are installed on both
`ALE_AUTH_CONNECT_V4` and `ALE_AUTH_CONNECT_V6`. ICMP also gets both
connect-layer and resource-assignment-layer coverage because ICMP
traffic is shaped differently from ordinary TCP/UDP port traffic.
## Validation
- `cargo fmt -p codex-windows-sandbox` (completed with existing
stable-rustfmt warnings about `imports_granularity = Item`)
- `cargo test -p codex-windows-sandbox wfp::tests`
- `cargo test -p codex-windows-sandbox` (fails in existing legacy
PowerShell sandbox tests because `Microsoft.PowerShell.Utility` could
not be loaded; WFP tests passed before that failure)
## Summary
- add elevated-only token constructors that include the current token
user SID in the restricted SID list
- switch the elevated Windows command runner to use those constructors
- leave the unelevated restricted-token path unchanged
## Why
Windows named pipes created by tools like Ninja use the platform's
default named-pipe ACL when no explicit security descriptor is provided.
In the elevated sandbox, the pipe owner has access, but the
write-restricted token can still fail its restricted-SID access check
because the sandbox user SID was not in the restricting SID set. That
causes child processes to exit successfully while Ninja never receives
the expected pipe completion/close behavior and hangs.
Including the elevated sandbox user's SID in the restricting SID list
lets the restricted check succeed for these owner-scoped pipe objects
without broadening the unelevated sandbox to the real signed-in user.
## Impact
- fixes the minimal Ninja hang repro in the elevated Windows sandbox
- preserves the existing unelevated sandbox behavior and write
protections
- keeps the change scoped to the elevated runner rather than changing
shared token semantics
- this does not affect file-writes for the sandbox because the sandbox
users themselves do not receive any additional permissions over what the
capability SIDs already have. In fact we don't even explicitly grant the
sandbox user ACLs anywhere.
## Validation
- `cargo build -p codex-windows-sandbox --quiet`
- verified the stock `ninja.exe` minimal repro exits normally on host
and in the elevated sandbox
- verified the same repro still hangs in the unelevated sandbox, which
is the intended scope of this change
## Summary
Some improvements to Windows process-management issues from
https://github.com/openai/codex/pull/15578
- bound the elevated runner pipe-connect handshake instead of waiting
forever on blocking pipe connects
- terminate the spawned runner if that handshake fails, so timeout/error
paths do not leave a stray `codex-command-runner.exe`
- loop on partial `WriteFile` results when forwarding stdin in the
elevated runner, so input is not silently truncated
- fix the concrete HANDLE/SID cleanup paths in the runner setup code
- keep draining driver-backed stdout/stderr after exit until the backend
closes, instead of dropping the tail after a fixed 200ms grace period
- reuse `LocalSid` for SID ownership and add more explanatory comments
around the ownership/concurrency-sensitive code paths
## Why
The original PR fixed a lot of Windows session plumbing, but there were
still a few sharp process-lifecycle edges:
- some elevated runner handshakes could block forever
- the new timeout path could still orphan the spawned runner process
- stdin forwarding still assumed a single `WriteFile` consumed the whole
buffer
- a few raw HANDLE/SID error paths still leaked
- driver-backed output could still lose the last chunk of stdout/stderr
on slower backends
## Validation
- `cargo fmt -p codex-windows-sandbox -p codex-utils-pty`
- `cargo test -p codex-utils-pty`
- `cargo test -p codex-windows-sandbox finish_driver_spawn`
- `cargo test -p codex-windows-sandbox runner_`
Ran a local test matrix of unified-exec and shell_tool tests, all
passing
## Summary
Fix the Windows sandbox PTY spawn path to pass the pseudoconsole handle
value directly into `UpdateProcThreadAttribute`.
## Why
Sandboxed `unified_exec` PTY sessions on Windows were failing during
child process startup with `0xc0000142` (`STATUS_DLL_INIT_FAILED`). In
practice this showed up as PowerShell DLL init popups when the sandboxed
background-terminal path tried to launch an interactive shell.
The root cause was that we were passing a pointer to a local `isize`
variable instead of the pseudoconsole handle value in the form Windows
expects for `PROC_THREAD_ATTRIBUTE_PSEUDOCONSOLE`.
## Validation
- `cargo build -p codex-windows-sandbox --bins`
- Reproduced the real sandboxed `codex exec` flow with
`windows.sandbox_private_desktop=true`
- Verified a `tty=true` interactive session launched through the normal
PowerShell wrapper, printed `READY`, accepted follow-up stdin, and
exited cleanly
- Confirmed no new `0xc0000142` / `Application Popup` events appeared
after the successful repro
## Why
`ReadOnlyAccess` was a transitional legacy shape on `SandboxPolicy`:
`FullAccess` meant the historical read-only/workspace-write modes could
read the full filesystem, while `Restricted` tried to carry partial
readable roots. The partial-read model now belongs in
`FileSystemSandboxPolicy` and `PermissionProfile`, so keeping it on
`SandboxPolicy` makes every legacy projection reintroduce lossy
read-root bookkeeping and creates unnecessary noise in the rest of the
permissions migration.
This PR makes the legacy policy model narrower and explicit:
`SandboxPolicy::ReadOnly` and `SandboxPolicy::WorkspaceWrite` represent
the old full-read sandbox modes only. Split readable roots, deny-read
globs, and platform-default/minimal read behavior stay in the runtime
permissions model.
## What changed
- Removes `ReadOnlyAccess` from
`codex_protocol::protocol::SandboxPolicy`, including the generated
`access` and `readOnlyAccess` API fields.
- Updates legacy policy/profile conversions so restricted filesystem
reads are represented only by `FileSystemSandboxPolicy` /
`PermissionProfile` entries.
- Keeps app-server v2 compatible with legacy `fullAccess` read-access
payloads by accepting and ignoring that no-op shape, while rejecting
legacy `restricted` read-access payloads instead of silently widening
them to full-read legacy policies.
- Carries Windows sandbox platform-default read behavior with an
explicit override flag instead of depending on
`ReadOnlyAccess::Restricted`.
- Refreshes generated app-server schema/types and updates tests/docs for
the simplified legacy policy shape.
## Verification
- `cargo check -p codex-app-server-protocol --tests`
- `cargo check -p codex-windows-sandbox --tests`
- `cargo test -p codex-app-server-protocol sandbox_policy_`
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/19449).
* #19395
* #19394
* #19393
* #19392
* #19391
* __->__ #19449
## Why
The elevated Windows command runner currently trusts the first process
that connects to its parent-created named pipes. Tightening the pipe ACL
already narrows who can reach that boundary, but verifying the connected
client PID gives the parent one more fail-closed check: it only accepts
the exact runner process it just spawned.
## What changed
- validate `GetNamedPipeClientProcessId` after `ConnectNamedPipe` and
reject clients whose PID does not match the spawned runner
- also did some code de-duplication to route the one-shot elevated
capture flow in `windows-sandbox-rs/src/elevated_impl.rs` through
`spawn_runner_transport()` so both elevated codepaths use the same pipe
bootstrap and PID validation
Using the transport unification here also reduces duplication in the
elevated Windows IPC bootstrap, so future hardening to the runner
handshake only needs to land in one place.
## Validation
- `cargo test -p codex-windows-sandbox`
- manual testing: one-shot elevated path via `target/debug/codex.exe
exec` running a randomized shell command and confirming captured output
- manual testing: elevated session path via `target/debug/codex.exe -c
'windows.sandbox="elevated"' sandbox windows -- python -u -c ...` with
stdin/stdout round-trips (`READY`, then `GOT:...` for two input lines)
---------
Co-authored-by: viyatb-oai <viyatb@openai.com>
We used to attempt a read-ACL on the same dir as `codex.exe` to grant
the sandbox user the ability to invoke `codex-command-runner.exe`. That
worked for the CLI case but it always fails for the installed desktop
app.
We have another solution already in place that copies
`codex-command-runner.exe` to `CODEX_HOME/.sandbox-bin` so we don't even
need this anymore. It causes a scary looking error in the logs that is a
non-issue and is therefore confusing
we copy `codex-command-runner.exe` into `CODEX_HOME/.sandbox-bin/` so
that it can be executed by the sandbox user.
We also detect if that version is stale and copy a new one in if so.
This can fail when you are running multiple versions of the app - the
file in `.sandbox-bin` can look stale because it comes from another app
build.
This change allows us to have multiple versions in there for different
CLI versions, and it fallsback to a `size+mtime` hash in the filename
for dev builds that don't report a real CLI version.
`codex sandbox windows` previously did a one-shot spawn for all
commands.
This change uses the `unified_exec` session to spawn long-lived
processes instead, and implements a simple bridge to forward stdin to
the spawned session and stdout/stderr from the spawned session back to
the caller.
It also fixes a bug with the new shared spawn context code where the
"no-network env" was being applied to both elevated and unelevated
sandbox spawns. It should only be applied for the unelevated sandbox
because the elevated one uses firewall rules instead of an env-based
network suppression strategy.
## Summary
This is the runtime/foundation half of the Windows sandbox unified-exec
work.
- add Windows sandbox `unified_exec` session support in
`windows-sandbox-rs` for both:
- the legacy restricted-token backend
- the elevated runner backend
- extend the PTY/process runtime so driver-backed sessions can support:
- stdin streaming
- stdout/stderr separation
- exit propagation
- PTY resize hooks
- add Windows sandbox runtime coverage in `codex-windows-sandbox` /
`codex-utils-pty`
This PR does **not** enable Windows sandbox `UnifiedExec` for product
callers yet because hooking this up to app-server comes in the next PR.
Windows sandbox advertising is intentionally kept aligned with `main`,
so sandboxed Windows callers still fall back to `ShellCommand`.
This PR isolates the runtime/session layer so it can be reviewed
independently from product-surface enablement.
---------
Co-authored-by: jif-oai <jif@openai.com>
Co-authored-by: Codex <noreply@openai.com>
## Stack
1. Base PR: #18443 stops granting ACLs on `USERPROFILE`.
2. This PR: filters additional SSH-owned profile roots discovered from
SSH config.
## Bug
The base PR removes the broadest bad grant: `USERPROFILE` itself.
That still leaves one important case. A user profile child can be
SSH-owned even when its name is not one of our fixed exclusions.
For example:
```sshconfig
Host devbox
IdentityFile ~/.keys/devbox
CertificateFile ~/.certs/devbox-cert.pub
UserKnownHostsFile ~/.known_hosts_custom
Include ~/.ssh/conf.d/*.conf
```
After profile expansion, the sandbox might see these as normal profile
children:
```text
C:\Users\me\.keys
C:\Users\me\.certs
C:\Users\me\.known_hosts_custom
C:\Users\me\.ssh
```
Those paths have another owner: OpenSSH and the tools that manage SSH
identity and host-key state. Codex should not add sandbox ACLs to them.
OpenSSH describes this dependency tree in
[`ssh_config(5)`](https://man.openbsd.org/ssh_config.5), and the client
parser follows the same shape in `readconf.c`:
- `Include` recursively reads more config files and expands globs
- `IdentityFile` and `CertificateFile` name authentication files
- `UserKnownHostsFile`, `GlobalKnownHostsFile`, and `RevokedHostKeys`
name host-key files
- `ControlPath` and `IdentityAgent` can name profile-owned sockets or
control files
- these path directives can use forms such as `~`, `%d`, and `${HOME}`
## Change
This PR adds a small SSH config dependency scanner.
It starts at:
```text
~/.ssh/config
```
Then it returns concrete paths named by `Include` and by path-valued SSH
config directives:
```text
IdentityFile
CertificateFile
UserKnownHostsFile
GlobalKnownHostsFile
RevokedHostKeys
ControlPath
IdentityAgent
```
For example:
```sshconfig
IdentityFile ~/.keys/devbox
CertificateFile ~/.certs/devbox-cert.pub
Include ~/.ssh/conf.d/*.conf
```
returns paths like:
```text
C:\Users\me\.keys\devbox
C:\Users\me\.certs\devbox-cert.pub
C:\Users\me\.ssh\conf.d\devbox.conf
```
The setup code then maps those paths back to their top-level
`USERPROFILE` child and filters matching sandbox roots out of both the
writable and readable root lists.
## Why this shape
The parser reports what SSH config references. The sandbox setup code
decides which `USERPROFILE` roots are unsafe to grant.
That keeps the policy simple:
1. expand broad profile grants
2. remove the profile root
3. remove fixed sensitive profile folders
4. remove profile folders referenced by SSH config dependencies
If a path has two possible owners, the sandbox steps back. SSH keeps
control of SSH config, keys, certificates, known-hosts files, sockets,
and included config files.
## Tests
- `cargo test -p codex-windows-sandbox --lib`
- `just bazel-lock-check`
- `just fix -p codex-windows-sandbox`
- `git diff --check`
## Stack
1. This PR: expand and filter `USERPROFILE` roots.
2. Follow-up: #18493 filters SSH config dependency roots on top of this
base.
## Bug
On Windows, Codex can grant the sandbox ACL access to the whole user
profile directory.
That means the sandbox ACL can be applied under paths like:
```text
C:\Users\me\.ssh
C:\Users\me\.tsh
```
This breaks SSH. Windows OpenSSH checks permissions on SSH config and
key material. If Codex adds a sandbox group ACL to those files, OpenSSH
can reject the config or keys.
The bad interaction is:
1. Codex asks the Windows sandbox to grant access to `USERPROFILE`.
2. The sandbox applies ACLs under that root.
3. SSH-owned files get an extra ACL entry.
4. OpenSSH rejects those files because their permissions are no longer
strict enough.
## Why this happens more now
Codex now has more flows that naturally start in the user profile:
- a new chat can start in the user directory
- a project can be rooted in the user directory
- a user can start the Codex CLI from the user directory
Those are valid user actions. The bug is that `USERPROFILE` is too broad
a sandbox root.
## Change
This PR keeps the useful behavior of starting from the user profile
without granting the profile root itself.
The new flow is:
1. collect the normal read and write roots
2. if a root is exactly `USERPROFILE`, replace it with the direct
children of `USERPROFILE`
3. remove `USERPROFILE` itself from the final root list
4. apply the existing user-profile read exclusions to both read and
write roots
5. add `.tsh` and `.brev` to that exclusion list
So this input:
```text
C:\Users\me
```
becomes roots like:
```text
C:\Users\me\Desktop
C:\Users\me\Documents
C:\Users\me\Downloads
```
and does not include:
```text
C:\Users\me
C:\Users\me\.ssh
C:\Users\me\.tsh
C:\Users\me\.brev
```
If `USERPROFILE` cannot be listed, expansion falls back to the profile
root and the later filter removes it. That keeps the failure mode closed
for this bug.
## Why this shape
The sandbox still gets access to ordinary profile folders when the user
starts from home.
The sandbox no longer grants access to the profile root itself.
All filtering happens after expansion, for both read and write roots.
That gives us one simple rule: expand broad profile grants first, then
remove roots the sandbox must not own.
## Tests
- `just fmt`
- `cargo test -p codex-windows-sandbox`
- `just fix -p codex-windows-sandbox`
- `git diff --check`
## Summary
This PR significantly improves the standalone installer experience.
The main changes are:
1. We now install the codex binary and other dependencies in a
subdirectory under CODEX_HOME.
(`CODEX_HOME/packages/standalone/releases/...`)
2. We replace the `codex.js` launcher that npm/bun rely on with logic in
the Rust binary that automatically resolves its dependencies (like
ripgrep)
## Motivation
A few design constraints pushed this work.
1. Currently, the entrypoint to codex is through `codex.js`, which
forces a node dependency to kick off our rust app. We want to move away
from this so that the entrypoint to codex does not rely on node or
external package managers.
2. Right now, the native script adds codex and its dependencies directly
to user PATH. Given that codex is likely to add more binary dependencies
than ripgrep, we want a solution which does not add arbitrary binaries
to user PATH -- the only one we want to add is the `codex` command
itself.
3. We want upgrades to be atomic. We do not want scenarios where
interrupting an upgrade command can move codex into undefined state (for
example, having a new codex binary but an old ripgrep binary). This was
~possible with the old script.
4. Currently, the Rust binary uses heuristics to determine which
installer created it. These heuristics are flaky and are tied to the
`codex.js` launcher. We need a more stable/deterministic way to
determine how the binary was installed for standalone.
5. We do not want conflicting codex installations on PATH. For example,
the user installing via npm, then installing via brew, then installing
via standalone would make it unclear which version of codex is being
launched and make it tough for us to determine the right upgrade
command.
## Design
### Standalone package layout
Standalone installs now live under `CODEX_HOME/packages/standalone`:
```text
$CODEX_HOME/
packages/
standalone/
current -> releases/0.111.0-x86_64-unknown-linux-musl
releases/
0.111.0-x86_64-unknown-linux-musl/
codex
codex-resources/
rg
```
where `standalone/current` is a symlink to a release directory.
On Windows, the release directory has the same shape, with `.exe` names
and Windows helpers in `codex-resources`:
```text
%CODEX_HOME%\
packages\
standalone\
current -> releases\0.111.0-x86_64-pc-windows-msvc
releases\
0.111.0-x86_64-pc-windows-msvc\
codex.exe
codex-resources\
rg.exe
codex-command-runner.exe
codex-windows-sandbox-setup.exe
```
This gives us:
- atomic upgrades because we can fully stage a release before switching
`standalone/current`
- a stable way for the binary to recognize a standalone install from its
canonical `current_exe()` path under CODEX_HOME
- a clean place for binary dependencies like `rg`, Windows sandbox
helpers, and, in the future, our custom `zsh` etc
### Command location
On Unix, we add a symlink at `~/.local/bin/codex` which points directly
to the `$CODEX_HOME/packages/standalone/current/codex` binary. This
becomes the main entrypoint for the CLI.
On Windows, we store the link at
`%LOCALAPPDATA%\Programs\OpenAI\Codex\bin`.
### PATH persistence
This is a tricky part of the PR, as there's no ~super reliable way to
ensure that we end up on PATH without significant tradeoffs.
Most Unix variants will have `~/.local/bin` on PATH already, which means
we *should* be fine simply registering the command there in most cases.
However, there are cases where this is not the case. In these cases, we
directly edit the profile depending on the shell we're in.
- macOS zsh: `~/.zprofile`
- macOS bash: `~/.bash_profile`
- Linux zsh: `~/.zshrc`
- Linux bash: `~/.bashrc`
- fallback: `~/.profile`
On Windows, we update the User `Path` environment variable directly and
we don't need to worry about shell profiles.
### Standalone runtime detection
This PR adds a new shared crate, `codex-install-context`, which computes
install ownership once per process and caches it in a `OnceLock`.
That context includes:
- install manager (`Standalone`, `Npm`, `Bun`, `Brew`, `Other`)
- the managed standalone release directory, when applicable
- the managed standalone `codex-resources` directory, when present
- the resolved `rg_command`
The standalone path is detected by canonicalizing `current_exe()`,
canonicalizing CODEX_HOME via `find_codex_home()`, and checking whether
the binary is running from under
`$CODEX_HOME/packages/standalone/releases`.
We intentionally do not use a release metadata file. The binary path is
the source of truth.
### Dependency resolution
For standalone installs, `grep_files` now resolves bundled `rg` from
`codex-resources` next to the Codex binary.
For npm/bun/brew/other installs, `grep_files` falls back to resolving
`rg` from PATH.
For Windows standalone installs, Windows sandbox helpers are still found
as direct siblings when present. If they are not direct siblings, the
lookup also checks the sibling `codex-resources` directory.
### TUI update path
The TUI now has `UpdateAction::StandaloneUnix` and
`UpdateAction::StandaloneWindows`, which rerun the standalone install
commands.
Unix update command:
```sh
sh -c "curl -fsSL https://chatgpt.com/codex/install.sh | sh"
```
Windows update command:
```powershell
powershell -c "irm https://chatgpt.com/codex/install.ps1|iex"
```
The Windows updater runs PowerShell directly. We do this because `cmd
/C` would parse the `|iex` as a cmd pipeline instead of passing it to
PowerShell.
## Additional installer behavior
- standalone installs now warn about conflicting npm/bun/brew-managed
`codex` installs and offer to uninstall them
- same-version reruns do not redownload the release if it is already
staged locally
## Testing
Installer smoke tests run:
- macOS: fresh install into isolated `HOME` and `CODEX_HOME` with
`scripts/install/install.sh --release latest`
- macOS: reran the installer against the same isolated install to verify
the same-version/update path and PATH block idempotence
- macOS: verified the installed `codex --version` and bundled
`codex-resources/rg --version`
- Windows: parsed `scripts/install/install.ps1` with PowerShell via
`[scriptblock]::Create(...)`
- Windows: verified the standalone update action builds a direct
PowerShell command and does not route the `irm ...|iex` command through
`cmd /C`
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
This updates the Windows elevated sandbox setup/refresh path to include
the legacy `compute_allow_paths(...).deny` protected children in the
same deny-write payload pipe added for split filesystem carveouts.
Concretely, elevated setup and elevated refresh now both build
deny-write payload paths from:
- explicit split-policy deny-write paths, preserving missing paths so
setup can materialize them before applying ACLs
- legacy `compute_allow_paths(...).deny`, which includes existing
`.git`, `.codex`, and `.agents` children under writable roots
This lets the elevated backend protect `.git` consistently with the
unelevated/restricted-token path, and removes the old janky hard-coded
`.codex` / `.agents` elevated setup helpers in favor of the shared
payload path.
## Root Cause
The landed split-carveout PR threaded a `deny_write_paths` pipe through
elevated setup/refresh, but the legacy workspace-write deny set from
`compute_allow_paths(...).deny` was not included in that payload. As a
result, elevated workspace-write did not apply the intended deny-write
ACLs for existing protected children like `<cwd>/.git`.
## Notes
The legacy protected children still only enter the deny set if they
already exist, because `compute_allow_paths` filters `.git`, `.codex`,
and `.agents` with `exists()`. Missing explicit split-policy deny paths
are preserved separately because setup intentionally materializes those
before applying ACLs.
## Validation
- `cargo fmt --check -p codex-windows-sandbox`
- `cargo test -p codex-windows-sandbox`
- `cargo build -p codex-cli -p codex-windows-sandbox --bins`
- Elevated `codex exec` smoke with `windows.sandbox='elevated'`: fresh
git repo, attempted append to `.git/config`, observed `Access is
denied`, marker not written, Deny ACE present on `.git`
- Unelevated `codex exec` smoke with `windows.sandbox='unelevated'`:
fresh git repo, attempted append to `.git/config`, observed `Access is
denied`, marker not written, Deny ACE present on `.git`
## Summary
- preserve legacy Windows elevated sandbox behavior for existing
policies
- add elevated-only support for split filesystem policies that can be
represented as readable-root overrides, writable-root overrides, and
extra deny-write carveouts
- resolve those elevated filesystem overrides during sandbox transform
and thread them through setup and policy refresh
- keep failing closed for explicit unreadable (`none`) carveouts and
reopened writable descendants under read-only carveouts
- for explicit read-only-under-writable-root carveouts, materialize
missing carveout directories during elevated setup before applying the
deny-write ACL
- document the elevated vs restricted-token support split in the core
README
## Example
Given a split filesystem policy like:
```toml
":root" = "read"
":cwd" = "write"
"./docs" = "read"
"C:/scratch" = "write"
```
the elevated backend now provisions the readable-root overrides,
writable-root overrides, and extra deny-write carveouts during setup and
refresh instead of collapsing back to the legacy workspace-only shape.
If a read-only carveout under a writable root is missing at setup time,
elevated setup creates that carveout as an empty directory before
applying its deny-write ACE; otherwise the sandboxed command could
create it later and bypass the carveout. This is only for explicit
policy carveouts. Best-effort workspace protections like `.codex/` and
`.agents/` still skip missing directories.
A policy like:
```toml
"/workspace" = "write"
"/workspace/docs" = "read"
"/workspace/docs/tmp" = "write"
```
still fails closed, because the elevated backend does not reopen
writable descendants under read-only carveouts yet.
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
- reduce public module visibility across Rust crates, preferring private
or crate-private modules with explicit crate-root public exports
- update external call sites and tests to use the intended public crate
APIs instead of reaching through module trees
- add the module visibility guideline to AGENTS.md
## Validation
- `cargo check --workspace --all-targets --message-format=short` passed
before the final fix/format pass
- `just fix` completed successfully
- `just fmt` completed successfully
- `git diff --check` passed
## Why
Follow-up to #16345, the Bazel clippy rollout in #15955, and the cleanup
pass in #16353.
`cargo clippy` was enforcing the workspace deny-list from
`codex-rs/Cargo.toml` because the member crates opt into `[lints]
workspace = true`, but Bazel clippy was only using `rules_rust` plus
`clippy.toml`. That left the Bazel lane vulnerable to drift:
`clippy.toml` can tune lint behavior, but it cannot set
allow/warn/deny/forbid levels.
This PR now closes both sides of the follow-up. It keeps `.bazelrc` in
sync with `[workspace.lints.clippy]`, and it fixes the real clippy
violations that the newly-synced Windows Bazel lane surfaced once that
deny-list started matching Cargo.
## What Changed
- added `.github/scripts/verify_bazel_clippy_lints.py`, a Python check
that parses `codex-rs/Cargo.toml` with `tomllib`, reads the Bazel
`build:clippy` `clippy_flag` entries from `.bazelrc`, and reports
missing, extra, or mismatched lint levels
- ran that verifier from the lightweight `ci.yml` workflow so the sync
check does not depend on a Rust toolchain being installed first
- expanded the `.bazelrc` comment to explain the Cargo `workspace =
true` linkage and why Bazel needs the deny-list duplicated explicitly
- fixed the Windows-only `codex-windows-sandbox` violations that Bazel
clippy reported after the sync, using the same style as #16353: inline
`format!` args, method references instead of trivial closures, removed
redundant clones, and replaced SID conversion `unwrap` and `expect`
calls with proper errors
- cleaned up the remaining cross-platform violations the Bazel lane
exposed in `codex-backend-client` and `core_test_support`
## Testing
Key new test introduced by this PR:
`python3 .github/scripts/verify_bazel_clippy_lints.py`
## Why
The initial `argument-comment-lint` rollout left Windows on
default-target coverage because there were still Windows-only callsites
failing under `--all-targets`. This follow-up cleans up those remaining
Windows-specific violations so the Windows CI lane can enforce the same
stricter coverage, leaving Linux as the remaining platform-specific
follow-up.
## What changed
- switched the Windows `rust-ci` argument-comment-lint step back to the
default wrapper invocation so it runs full-target coverage again
- added the required `/*param_name*/` annotations at Windows-gated
literal callsites in:
- `codex-rs/windows-sandbox-rs/src/lib.rs`
- `codex-rs/windows-sandbox-rs/src/elevated_impl.rs`
- `codex-rs/tui_app_server/src/multi_agents.rs`
- `codex-rs/network-proxy/src/proxy.rs`
## Validation
- Windows `argument comment lint` CI on this PR
## Why
`codex-utils-pty` and `codex-windows-sandbox` were the remaining crates
in `codex-rs` that still overrode the workspace's Rust 2024 edition.
Moving them forward in a separate PR keeps the baseline edition update
isolated from the follow-on Bazel clippy workflow in #15955, while
making linting and formatting behavior consistent with the rest of the
workspace.
This PR also needs Cargo and Bazel to agree on the edition for
`codex-windows-sandbox`. Without the Bazel-side sync, the experimental
Bazel app-server builds fail once they compile `windows-sandbox-rs`.
## What changed
- switch `codex-rs/utils/pty` and `codex-rs/windows-sandbox-rs` to
`edition = "2024"`
- update `codex-utils-pty` callsites and tests to use the collapsed `if
let` form that Clippy expects under the new edition
- fix the Rust 2024 fallout in `windows-sandbox-rs`, including the
reserved `gen` identifier, `unsafe extern` requirements, and new Clippy
findings that surfaced under the edition bump
- keep the edition bump separate from a larger unsafe cleanup by
temporarily allowing `unsafe_op_in_unsafe_fn` in the Windows entrypoint
modules that now report it under Rust 2024
- update `codex-rs/windows-sandbox-rs/BUILD.bazel` to `crate_edition =
"2024"` so Bazel compiles the crate with the same edition as Cargo
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/15954).
* #15976
* #15955
* __->__ #15954
## Summary
This PR makes Windows sandbox proxying enforceable by routing proxy-only
runs through the existing `offline` sandbox user and reserving direct
network access for the existing `online` sandbox user.
In brief:
- if a Windows sandbox run should be proxy-enforced, we run it as the
`offline` user
- the `offline` user gets firewall rules that block direct outbound
traffic and only permit the configured localhost proxy path
- if a Windows sandbox run should have true direct network access, we
run it as the `online` user
- no new sandbox identity is introduced
This brings Windows in line with the intended model: proxy use is not
just env-based, it is backed by OS-level egress controls. Windows
already has two sandbox identities:
- `offline`: intended to have no direct network egress
- `online`: intended to have full network access
This PR makes proxy-enforced runs use that model directly.
### Proxy-enforced runs
When proxy enforcement is active:
- the run is assigned to the `offline` identity
- setup extracts the loopback proxy ports from the sandbox env
- Windows setup programs firewall rules for the `offline` user that:
- block all non-loopback outbound traffic
- block loopback UDP
- block loopback TCP except for the configured proxy ports
- optionally allow broader localhost access when `allow_local_binding=1`
So the sandboxed process can only talk to the local proxy. It cannot
open direct outbound sockets or do local UDP-based DNS on its own.The
proxy then performs the real outbound network access outside that
restricted sandbox identity.
### Direct-network runs
When proxy enforcement is not active and full network access is allowed:
- the run is assigned to the `online` identity
- no proxy-only firewall restrictions are applied
- the process gets normal direct network access
### Unelevated vs elevated
The restricted-token / unelevated path cannot enforce per-identity
firewall policy by itself.
So for Windows proxy-enforced runs, we transparently use the logon-user
sandbox path under the hood, even if the caller started from the
unelevated mode. That keeps enforcement real instead of best-effort.
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
- keep legacy Windows restricted-token sandboxing as the supported
baseline
- support the split-policy subset that restricted-token can enforce
directly today
- support full-disk read, the same writable root set as legacy
`WorkspaceWrite`, and extra read-only carveouts under those writable
roots via additional deny-write ACLs
- continue to fail closed for unsupported split-only shapes, including
explicit unreadable (`none`) carveouts, reopened writable descendants
under read-only carveouts, and writable root sets that do not match the
legacy workspace roots
## Example
Given a filesystem policy like:
```toml
":root" = "read"
":cwd" = "write"
"./docs" = "read"
```
the restricted-token backend can keep the workspace writable while
denying writes under `docs` by layering an extra deny-write carveout on
top of the legacy workspace-write roots.
A policy like:
```toml
"/workspace" = "write"
"/workspace/docs" = "read"
"/workspace/docs/tmp" = "write"
```
still fails closed, because the unelevated backend cannot reopen the
nested writable descendant safely.
## Stack
-> fix: support split carveouts in windows restricted-token sandbox
#14172
fix: support split carveouts in windows elevated sandbox #14568
## Why
The argument-comment lint now has a packaged DotSlash artifact from
[#15198](https://github.com/openai/codex/pull/15198), so the normal repo
lint path should use that released payload instead of rebuilding the
lint from source every time.
That keeps `just clippy` and CI aligned with the shipped artifact while
preserving a separate source-build path for people actively hacking on
the lint crate.
The current alpha package also exposed two integration wrinkles that the
repo-side prebuilt wrapper needs to smooth over:
- the bundled Dylint library filename includes the host triple, for
example `@nightly-2025-09-18-aarch64-apple-darwin`, and Dylint derives
`RUSTUP_TOOLCHAIN` from that filename
- on Windows, Dylint's driver path also expects `RUSTUP_HOME` to be
present in the environment
Without those adjustments, the prebuilt CI jobs fail during `cargo
metadata` or driver setup. This change makes the checked-in prebuilt
wrapper normalize the packaged library name to the plain
`nightly-2025-09-18` channel before invoking `cargo-dylint`, and it
teaches both the wrapper and the packaged runner source to infer
`RUSTUP_HOME` from `rustup show home` when the environment does not
already provide it.
After the prebuilt Windows lint job started running successfully, it
also surfaced a handful of existing anonymous literal callsites in
`windows-sandbox-rs`. This PR now annotates those callsites so the new
cross-platform lint job is green on the current tree.
## What Changed
- checked in the current
`tools/argument-comment-lint/argument-comment-lint` DotSlash manifest
- kept `tools/argument-comment-lint/run.sh` as the source-build wrapper
for lint development
- added `tools/argument-comment-lint/run-prebuilt-linter.sh` as the
normal enforcement path, using the checked-in DotSlash package and
bundled `cargo-dylint`
- updated `just clippy` and `just argument-comment-lint` to use the
prebuilt wrapper
- split `.github/workflows/rust-ci.yml` so source-package checks live in
a dedicated `argument_comment_lint_package` job, while the released lint
runs in an `argument_comment_lint_prebuilt` matrix on Linux, macOS, and
Windows
- kept the pinned `nightly-2025-09-18` toolchain install in the prebuilt
CI matrix, since the prebuilt package still relies on rustup-provided
toolchain components
- updated `tools/argument-comment-lint/run-prebuilt-linter.sh` to
normalize host-qualified nightly library filenames, keep the `rustup`
shim directory ahead of direct toolchain `cargo` binaries, and export
`RUSTUP_HOME` when needed for Windows Dylint driver setup
- updated `tools/argument-comment-lint/src/bin/argument-comment-lint.rs`
so future published DotSlash artifacts apply the same nightly-filename
normalization and `RUSTUP_HOME` inference internally
- fixed the remaining Windows lint violations in
`codex-rs/windows-sandbox-rs` by adding the required `/*param*/`
comments at the reported callsites
- documented the checked-in DotSlash file, wrapper split, archive
layout, nightly prerequisite, and Windows `RUSTUP_HOME` requirement in
`tools/argument-comment-lint/README.md`
## Summary
- support legacy `ReadOnlyAccess::Restricted` on Windows in the elevated
setup/runner backend
- keep the unelevated restricted-token backend on the legacy full-read
model only, and fail closed for restricted read-only policies there
- keep the legacy full-read Windows path unchanged while deriving
narrower read roots only for elevated restricted-read policies
- honor `include_platform_defaults` by adding backend-managed Windows
system roots only when requested, while always keeping helper roots and
the command `cwd` readable
- preserve `workspace-write` semantics by keeping writable roots
readable when restricted read access is in use in the elevated backend
- document the current Windows boundary: legacy `SandboxPolicy` is
supported on both backends, while richer split-only carveouts still fail
closed instead of running with weaker enforcement
## Testing
- `cargo test -p codex-windows-sandbox`
- `cargo check -p codex-windows-sandbox --tests --target
x86_64-pc-windows-msvc`
- `cargo clippy -p codex-windows-sandbox --tests --target
x86_64-pc-windows-msvc -- -D warnings`
- `cargo test -p codex-core windows_restricted_token_`
## Notes
- local `cargo test -p codex-windows-sandbox` on macOS only exercises
the non-Windows stubs; the Windows-targeted compile and clippy runs
provide the local signal, and GitHub Windows CI exercises the runtime
path
## Summary
This is PR 2 of the Windows sandbox runner split.
PR 1 introduced the framed IPC runner foundation and related Windows
sandbox infrastructure without changing the active elevated one-shot
execution path. This PR switches that elevated one-shot path over to the
new runner IPC transport and removes the old request-file bootstrap that
PR 1 intentionally left in place.
After this change, ordinary elevated Windows sandbox commands still
behave as one-shot executions, but they now run as the simple case of
the same helper/IPC transport that later unified_exec work will build
on.
## Why this is needed for unified_exec
Windows elevated sandboxed execution crosses a user boundary: the CLI
launches a helper as the sandbox user and has to manage command
execution from outside that security context. For one-shot commands, the
old request-file/bootstrap flow was sufficient. For unified_exec, it is
not.
Unified_exec needs a long-lived bidirectional channel so the parent can:
- send a spawn request
- receive structured spawn success/failure
- stream stdout and stderr incrementally
- eventually support stdin writes, termination, and other session
lifecycle events
This PR does not add long-lived sessions yet. It converts the existing
elevated one-shot path to use the same framed IPC transport so that PR 3
can add unified_exec session semantics on top of a transport that is
already exercised by normal elevated command execution.
## Scope
This PR:
- updates `windows-sandbox-rs/src/elevated_impl.rs` to launch the runner
with named pipes, send a framed `SpawnRequest`, wait for `SpawnReady`,
and collect framed `Output`/`Exit` messages
- removes the old `--request-file=...` execution path from
`windows-sandbox-rs/src/elevated/command_runner_win.rs`
- keeps the public behavior one-shot: no session reuse or interactive
unified_exec behavior is introduced here
This PR does not:
- add Windows unified_exec session support
- add background terminal reuse
- add PTY session lifecycle management
## Why Windows needs this and Linux/macOS do not
On Linux and macOS, the existing sandbox/process model composes much
more directly with long-lived process control. The parent can generally
spawn and own the child process (or PTY) directly inside the sandbox
model we already use.
Windows elevated sandboxing is different. The parent is not directly
managing the sandboxed process in the same way; it launches across a
different user/security context. That means long-lived control requires
an explicit helper process plus IPC for spawn, output, exit, and later
stdin/session control.
So the extra machinery here is not because unified_exec is conceptually
different on Windows. It is because the elevated Windows sandbox
boundary requires a helper-mediated transport to support it cleanly.
## Validation
- `cargo test -p codex-windows-sandbox`
# Summary
This PR introduces the Windows sandbox runner IPC foundation that later
unified_exec work will build on.
The key point is that this is intentionally infrastructure-only. The new
IPC transport, runner plumbing, and ConPTY helpers are added here, but
the active elevated Windows sandbox path still uses the existing
request-file bootstrap. In other words, this change prepares the
transport and module layout we need for unified_exec without switching
production behavior over yet.
Part of this PR is also a source-layout cleanup: some Windows sandbox
files are moved into more explicit `elevated/`, `conpty/`, and shared
locations so it is clearer which code is for the elevated sandbox flow,
which code is legacy/direct-spawn behavior, and which helpers are shared
between them. That reorganization is intentional in this first PR so
later behavioral changes do not also have to carry a large amount of
file-move churn.
# Why This Is Needed For unified_exec
Windows elevated sandboxed unified_exec needs a long-lived,
bidirectional control channel between the CLI and a helper process
running under the sandbox user. That channel has to support:
- starting a process and reporting structured spawn success/failure
- streaming stdout/stderr back incrementally
- forwarding stdin over time
- terminating or polling a long-lived process
- supporting both pipe-backed and PTY-backed sessions
The existing elevated one-shot path is built around a request-file
bootstrap and does not provide those primitives cleanly. Before we can
turn on Windows sandbox unified_exec, we need the underlying runner
protocol and transport layer that can carry those lifecycle events and
streams.
# Why Windows Needs More Machinery Than Linux Or macOS
Linux and macOS can generally build unified_exec on top of the existing
sandbox/process model: the parent can spawn the child directly, retain
normal ownership of stdio or PTY handles, and manage the lifetime of the
sandboxed process without introducing a second control process.
Windows elevated sandboxing is different. To run inside the sandbox
boundary, we cross into a different user/security context and then need
to manage a long-lived process from outside that boundary. That means we
need an explicit helper process plus an IPC transport to carry spawn,
stdin, output, and exit events back and forth. The extra code here is
mostly that missing Windows sandbox infrastructure, not a conceptual
difference in unified_exec itself.
# What This PR Adds
- the framed IPC message types and transport helpers for parent <->
runner communication
- the renamed Windows command runner with both the existing request-file
bootstrap and the dormant IPC bootstrap
- named-pipe helpers for the elevated runner path
- ConPTY helpers and process-thread attribute plumbing needed for
PTY-backed sessions
- shared sandbox/process helpers that later PRs will reuse when
switching live execution paths over
- early file/module moves so later PRs can focus on behavior rather than
layout churn
# What This PR Does Not Yet Do
- it does not switch the active elevated one-shot path over to IPC yet
- it does not enable Windows sandbox unified_exec yet
- it does not remove the existing request-file bootstrap yet
So while this code compiles and the new path has basic validation, it is
not yet the exercised production path. That is intentional for this
first PR: the goal here is to land the transport and runner foundation
cleanly before later PRs start routing real command execution through
it.
# Follow-Ups
Planned follow-up PRs will:
1. switch elevated one-shot Windows sandbox execution to the new runner
IPC path
2. layer Windows sandbox unified_exec sessions on top of the same
transport
3. remove the legacy request-file path once the IPC-based path is live
# Validation
- `cargo build -p codex-windows-sandbox`
## Summary
- launch Windows sandboxed children on a private desktop instead of
`Winsta0\Default`
- make private desktop the default while keeping
`windows.sandbox_private_desktop=false` as the escape hatch
- centralize process launch through the shared
`create_process_as_user(...)` path
- scope the private desktop ACL to the launching logon SID
## Why
Today sandboxed Windows commands run on the visible shared desktop. That
leaves an avoidable same-desktop attack surface for window interaction,
spoofing, and related UI/input issues. This change moves sandboxed
commands onto a dedicated per-launch desktop by default so the sandbox
no longer shares `Winsta0\Default` with the user session.
The implementation stays conservative on security with no silent
fallback back to `Winsta0\Default`
If private-desktop setup fails on a machine, users can still opt out
explicitly with `windows.sandbox_private_desktop=false`.
## Validation
- `cargo build -p codex-cli`
- elevated-path `codex exec` desktop-name probe returned
`CodexSandboxDesktop-*`
- elevated-path `codex exec` smoke sweep for shell commands, nested
`pwsh`, jobs, and hidden `notepad` launch
- unelevated-path full private-desktop compatibility sweep via `codex
exec` with `-c windows.sandbox=unelevated`
We do this for codex-command-runner.exe as well for the same reason.
Windows sandbox users cannot execute binaries in the WindowsApp/
installed directory for the Codex App. This causes apply-patch to fail
because it tries to execute codex.exe as the sandbox user.
• Keep Windows sandbox runner launches working from packaged installs by
running the helper from a user-owned runtime location.
On some Windows installs, the packaged helper location is difficult to
use reliably for sandboxed runner launches even though the binaries are
present. This change works around that by copying codex-
command-runner.exe into CODEX_HOME/.sandbox-bin/, reusing that copy
across launches, and falling back to the existing packaged-path lookup
if anything goes wrong.
The runtime copy lives in a dedicated directory with tighter ACLs than
.sandbox: sandbox users can read and execute the runner there, but they
cannot modify it. This keeps the workaround focused on the
command runner, leaves the setup helper on its trusted packaged path,
and adds logging so it is clear which runner path was selected at
launch.
The Mac and Linux implementations of the sandbox recently added write
protections for `.codex` and `.agents` subdirectories in all writable
roots. When adding documentation for this, I noticed that this change
was never made for the Windows sandbox.
Summary
- make compute_allow_paths treat .codex/.agents as protected alongside
.git, and cover their behavior in new tests
- wire protect_workspace_agents_dir through the sandbox lib and setup
path to apply deny ACEs when `.agents` exists
- factor shared ACL logic for workspace subdirectories
There is an edge case where a directory is not readable by the sandbox.
In practice, we've seen very little of it, but it can happen so this
slash command unlocks users when it does.
Future idea is to make this a tool that the agent knows about so it can
be more integrated.
`SandboxPolicy::ReadOnly` previously implied broad read access and could
not express a narrower read surface.
This change introduces an explicit read-access model so we can support
user-configurable read restrictions in follow-up work, while preserving
current behavior today.
It also ensures unsupported backends fail closed for restricted-read
policies instead of silently granting broader access than intended.
## What
- Added `ReadOnlyAccess` in protocol with:
- `Restricted { include_platform_defaults, readable_roots }`
- `FullAccess`
- Updated `SandboxPolicy` to carry read-access configuration:
- `ReadOnly { access: ReadOnlyAccess }`
- `WorkspaceWrite { ..., read_only_access: ReadOnlyAccess }`
- Preserved existing behavior by defaulting current construction paths
to `ReadOnlyAccess::FullAccess`.
- Threaded the new fields through sandbox policy consumers and call
sites across `core`, `tui`, `linux-sandbox`, `windows-sandbox`, and
related tests.
- Updated Seatbelt policy generation to honor restricted read roots by
emitting scoped read rules when full read access is not granted.
- Added fail-closed behavior on Linux and Windows backends when
restricted read access is requested but not yet implemented there
(`UnsupportedOperation`).
- Regenerated app-server protocol schema and TypeScript artifacts,
including `ReadOnlyAccess`.
## Compatibility / rollout
- Runtime behavior remains unchanged by default (`FullAccess`).
- API/schema changes are in place so future config wiring can enable
restricted read access without another policy-shape migration.
1. Move Windows Sandbox NUX to right after trust directory screen
2. Don't offer read-only as an option in Sandbox NUX.
Elevated/Legacy/Quit
3. Don't allow new untrusted directories. It's trust or quit
4. move experimental sandbox features to `[windows]
sandbox="elevated|unelevatd"`
5. Copy tweaks = elevated -> default, non-elevated -> non-admin
calculated a hashed user ID from either auth user id or API key
Also correctly populates OS.
These will make our metrics more useful and powerful for analysis.
Today, there is a single capability SID that allows the sandbox to write
to
* workspace (cwd)
* tmp directories if enabled
* additional writable roots
This change splits those up, so that each workspace has its own
capability SID, while tmp and additional roots, which are
installation-wide, are still governed by the "generic" capability SID
This isolates workspaces from each other in terms of sandbox write
access.
Also allows us to protect <cwd>/.codex when codex runs in a specific
<cwd>
Currently `apply_patch` will fail on Windows if the file contents happen
to have a multi-byte character at the point where the `preview` function
truncates.
I've used the existing `take_bytes_at_char_boundary` helper and added a
regression test (that fails without the fix).
This is related to #4013 but doesn't fix it.
## Summary
- Remove elevated runner request files after read (best-effort cleanup
on errors)
- Add a unit test to cover request file lifecycle
## Testing
- `cargo test -p codex-windows-sandbox` (Windows)
Fixes#9315
This fixes a bug where the elevated sandbox setup encrypts sandbox user
passwords as an admin user, but normal command execution attempts to
decrypt them as a different user.
Machine scope allows all users to encyrpt/decrypt
this PR also moves the encrypted file to a different location
.codex/.sandbox-secrets which the sandbox users cannot read.
