## Summary
This fixes BUGB-17567 by preventing non-Windows command safety
classification from invoking the Windows PowerShell safelist/parser
path.
Previously, `is_known_safe_command` called the Windows PowerShell
classifier on every platform. That classifier recognizes
`pwsh`/`powershell` by basename and delegates script parsing to the
PowerShell AST parser. The parser starts the supplied executable, so on
macOS/Linux a repository-controlled `pwsh` path could execute during
safety parsing before the normal sandboxed command execution path.
The change gates the Windows PowerShell classifier and module behind
`#[cfg(windows)]`. On macOS/Linux, PowerShell-looking commands are no
longer auto-approved by the Windows classifier and instead fall through
to the normal non-Windows safe-command logic.
## Validation
- `/private/tmp/codex-tools/bin/just fmt`
- `PATH=/private/tmp/codex-tools/bin:$PATH
/private/tmp/codex-tools/bin/just test -p codex-shell-command`
The focused test run passed 135 tests with 0 skipped and completed the
crate bench-smoke step.
## Notes
This PR is scoped to the BUGB-17567 macOS/Linux path. Windows still uses
the PowerShell classifier; a separate hardening follow-up should ensure
Windows safety parsing only executes a trusted PowerShell parser binary
and does not spawn the command's `argv[0]` when that path may be
repository-controlled.
## Summary
- Treat PowerShell stop-parsing token forms as unsupported in the
AST-backed command flattener.
- Add focused regressions at the parser layer and Windows command-safety
layer.
## Why
The command-safety parser lowers PowerShell AST elements into argv-like
words. Stop-parsing syntax preserves a native-command argument shape
that this lowering does not model, so these forms should stay on the
conservative unsupported path.
## Validation
- `cargo fmt --manifest-path codex-rs/Cargo.toml --all --check`
- `cargo test --manifest-path codex-rs/Cargo.toml -p
codex-shell-command`
## Why
This is a follow-up to the Windows Git safe-command bypass fix for
BUGB-15601. Git's global `--paginate` / `-p` flags can route output
through a configured pager, so they should not be auto-approved as safe
before the subcommand. At the same time, `-p` after read-only
subcommands like `log`, `diff`, and `show` is the common patch-output
flag, so treating every `-p` as unsafe would make ordinary read-only
inspection commands prompt unnecessarily.
## What Changed
- Split Git option safety matching into explicit global-option and
subcommand-option lists.
- Treat global `git --paginate ...` and `git -p ...` as unsafe.
- Keep post-subcommand patch usage such as `git log -p`, `git diff -p`,
and `git show -p HEAD` safe.
- Keep the pagination coverage with the shared Git safe-command
implementation rather than the Windows wrapper tests.
- Remove the stale `git_global_option_requires_prompt` helper now that
safe-command Git option matching owns the prompt-required lists.
## Testing
- `cargo test -p codex-shell-command`
## Why
BUGB-15601 showed that the Windows safe-command path had drifted from
the generic Git classifier. The Windows-specific Git parser could
classify a PowerShell-wrapped `git` command as safe as soon as it found
a safelisted subcommand, without applying the generic checks for unsafe
subcommand options such as `--output`, `--ext-diff`, `--textconv`,
`--paginate`, or `cat-file --filters`.
The generic classifier already models the Git command boundary and the
read-only argument checks more carefully, so Windows should reuse that
logic instead of maintaining a smaller parallel parser.
## What Changed
- Extracted the existing generic Git classification logic into
`is_safe_git_command`.
- Updated `windows_safe_commands.rs` to call that shared helper for
parsed PowerShell `git` commands.
- Removed the Windows-only Git subcommand safelist, including the
`cat-file` allowance that was part of the reported bypass.
- Added a Windows regression test that keeps PowerShell-wrapped Git
commands with side-effecting options classified unsafe.
- Made the full-path PowerShell test discover the installed PowerShell
executable instead of depending on one hard-coded `pwsh.exe` path.
## Verification
- `cargo test -p codex-shell-command
rejects_git_subcommand_options_with_side_effects`
- `cargo test -p codex-shell-command
git_global_override_flags_are_not_safe`
- `cargo test -p codex-shell-command
windows_powershell_full_path_is_safe -- --nocapture`
Co-authored-by: Codex <codex@openai.com>
## Why
On Windows, Codex runs shell commands through a top-level
`powershell.exe -NoProfile -Command ...` wrapper. `execpolicy` was
matching that wrapper instead of the inner command, so prefix rules like
`["git", "push"]` did not fire for PowerShell-wrapped commands even
though the same normalization already happens for `bash -lc` on Unix.
This change makes the Windows shell wrapper transparent to rule matching
while preserving the existing Windows unmatched-command safelist and
dangerous-command heuristics.
## What changed
- add `parse_powershell_command_plain_commands()` in
`shell-command/src/powershell.rs` to unwrap the top-level PowerShell
`-Command` body with `extract_powershell_command()` and parse it with
the existing PowerShell AST parser
- update `core/src/exec_policy.rs` so `commands_for_exec_policy()`
treats top-level PowerShell wrappers like `bash -lc` and evaluates rules
against the parsed inner commands
- carry a small `ExecPolicyCommandOrigin` through unmatched-command
evaluation and expose `is_safe_powershell_words()` /
`is_dangerous_powershell_words()` so Windows safelist and
dangerous-command checks still work after unwrap
- add Windows-focused tests for wrapped PowerShell prompt/allow matches,
wrapper parsing, and unmatched safe/dangerous inner commands, and
re-enable the end-to-end `execpolicy_blocks_shell_invocation` test on
Windows
## Testing
- `cargo test -p codex-shell-command`
## 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
`//codex-rs/shell-command:shell-command-unit-tests` became a real
bottleneck in the Windows Bazel lane because repeated calls to
`is_safe_command_windows()` were starting a fresh PowerShell parser
process for every `powershell.exe -Command ...` assertion.
PR #16056 was motivated by that same bottleneck, but its test-only
shortcut was the wrong layer to optimize because it weakened the
end-to-end guarantee that our runtime path really asks PowerShell to
parse the command the way we expect.
This PR attacks the actual cost center instead: it keeps the real
PowerShell parser in the loop, but turns that parser into a long-lived
helper process so both tests and the runtime safe-command path can reuse
it across many requests.
## What Changed
- add `shell-command/src/command_safety/powershell_parser.rs`, which
keeps one mutex-protected parser process per PowerShell executable path
and speaks a simple JSON-over-stdio request/response protocol
- turn `shell-command/src/command_safety/powershell_parser.ps1` into a
long-running parser server with comments explaining the protocol, the
AST-shape restrictions, and why unsupported constructs are rejected
conservatively
- keep request ids and a one-time respawn path so a dead or
desynchronized cached child fails closed instead of silently returning
mixed parser output
- preserve separate parser processes for `powershell.exe` and
`pwsh.exe`, since they do not accept the same language surface
- avoid a direct `PipelineChainAst` type reference in the PowerShell
script so the parser service still runs under Windows PowerShell 5.1 as
well as newer `pwsh`
- make `shell-command/src/command_safety/windows_safe_commands.rs`
delegate to the new parser utility instead of spawning a fresh
PowerShell process for every parse
- add a Windows-only unit test that exercises multiple sequential
requests against the same parser process
## Testing
- adds a Windows-only parser-reuse unit test in `powershell_parser.rs`
- the main end-to-end verification for this change is the Windows CI
lane, because the new service depends on real `powershell.exe` /
`pwsh.exe` behavior
## Summary
- block git global options that can redirect config, repository, or
helper lookup from being auto-approved as safe
- share the unsafe global-option predicate across the Unix and Windows
git safety checks
- add regression coverage for inline and split forms, including `bash
-lc` and PowerShell wrappers
## Root cause
The Unix safe-command gate only rejected `-c` and `--config-env`, even
though the shared git parser already knew how to skip additional
pre-subcommand globals such as `--git-dir`, `--work-tree`,
`--exec-path`, `--namespace`, and `--super-prefix`. That let those
arguments slip through safe-command classification on otherwise
read-only git invocations and bypass approval. The Windows-specific
safe-command path had the same trust-boundary gap for git global
options.
## Why
[#12964](https://github.com/openai/codex/pull/12964) added
`host_executable()` support to `codex-execpolicy`, and
[#13046](https://github.com/openai/codex/pull/13046) adopted it in the
zsh-fork interception path.
The remaining gap was the preflight execpolicy check in
`core/src/exec_policy.rs`. That path derives approval requirements
before execution for `shell`, `shell_command`, and `unified_exec`, but
it was still using the default exact-token matcher.
As a result, a command that already included an absolute executable
path, such as `/usr/bin/git status`, could still miss a basename rule
like `prefix_rule(pattern = ["git"], ...)` during preflight even when
the policy also defined a matching `host_executable(name = "git", ...)`
entry.
This PR brings the same opt-in `host_executable()` resolution to the
preflight approval path when an absolute program path is already present
in the parsed command.
## What Changed
- updated
`ExecPolicyManager::create_exec_approval_requirement_for_command()` in
`core/src/exec_policy.rs` to use `check_multiple_with_options(...)` with
`MatchOptions { resolve_host_executables: true }`
- kept the existing shell parsing flow for approval derivation, but now
allow basename rules to match absolute executable paths during preflight
when `host_executable()` permits it
- updated requested-prefix amendment evaluation to use the same
host-executable-aware matching mode, so suggested `prefix_rule()`
amendments are checked consistently for absolute-path commands
- added preflight coverage for:
- absolute-path commands that should match basename rules through
`host_executable()`
- absolute-path commands whose paths are not in the allowed
`host_executable()` mapping
- requested prefix-rule amendments for absolute-path commands
## Verification
- `just fix -p codex-core`
- `cargo test -p codex-core --lib exec_policy::tests::`
### Motivation
- Git subcommand matching was being classified as "dangerous" and caused
benign developer workflows (for example `git push --force-with-lease`)
to be blocked by the preflight policy.
- The change aligns behavior with the intent to reserve the dangerous
checklist for truly destructive shell ops (e.g. `rm -rf`) and avoid
surprising developer-facing blocks.
### Description
- Remove git-specific subcommand checks from
`is_dangerous_to_call_with_exec` in
`codex-rs/shell-command/src/command_safety/is_dangerous_command.rs`,
leaving only explicit `rm` and `sudo` passthrough checks.
- Deleted the git-specific helper logic that classified `reset`,
`branch`-delete, `push` (force/delete/refspec) and `clean --force` as
dangerous.
- Updated unit tests in the same file to assert that various `git
reset`/`git branch`/`git push`/`git clean` variants are no longer
classified as dangerous.
- Kept `find_git_subcommand` (used by safe-command classification)
intact so safe/unsafe parsing elsewhere remains functional.
### Testing
- Ran formatter with `just fmt` successfully.
- Ran unit tests with `cargo test -p codex-shell-command` and all tests
passed (`144 passed; 0 failed`).
------
[Codex
Task](https://chatgpt.com/codex/tasks/task_i_698d19dedb4883299c3ceb5bbc6a0dcf)
Adrian ·
Chris Bookholt ·
iceweasel-oai ·
Owen Lin ·
pakrym-oai ·
Michael Bolin ·
Josh McKinney ·