https://github.com/openai/codex/pull/1086 is a work-in-progress to make
Linux sandboxing work more like Seatbelt where, for the command we want
to sandbox, we build up the command and then hand it, and some sandbox
configuration flags, to another command to set up the sandbox and then
run it.
In the case of Seatbelt, macOS provides this helper binary and provides
it at `/usr/bin/sandbox-exec`. For Linux, we have to build our own and
pass it through (which is what #1086 does), so this makes the new
`codex_linux_sandbox_exe` available on `Config` so that it will later be
available in `exec.rs` when we need it in #1086.
To date, when handling `shell` and `local_shell` tool calls, we were
spawning new processes using the environment inherited from the Codex
process itself. This means that the sensitive `OPENAI_API_KEY` that
Codex needs to talk to OpenAI models was made available to everything
run by `shell` and `local_shell`. While there are cases where that might
be useful, it does not seem like a good default.
This PR introduces a complex `shell_environment_policy` config option to
control the `env` used with these tool calls. It is inevitably a bit
complex so that it is possible to override individual components of the
policy so without having to restate the entire thing.
Details are in the updated `README.md` in this PR, but here is the
relevant bit that explains the individual fields of
`shell_environment_policy`:
| Field | Type | Default | Description |
| ------------------------- | -------------------------- | ------- |
-----------------------------------------------------------------------------------------------------------------------------------------------
|
| `inherit` | string | `core` | Starting template for the
environment:<br>`core` (`HOME`, `PATH`, `USER`, …), `all` (clone full
parent env), or `none` (start empty). |
| `ignore_default_excludes` | boolean | `false` | When `false`, Codex
removes any var whose **name** contains `KEY`, `SECRET`, or `TOKEN`
(case-insensitive) before other rules run. |
| `exclude` | array<string> | `[]` | Case-insensitive glob
patterns to drop after the default filter.<br>Examples: `"AWS_*"`,
`"AZURE_*"`. |
| `set` | table<string,string> | `{}` | Explicit key/value
overrides or additions – always win over inherited values. |
| `include_only` | array<string> | `[]` | If non-empty, a
whitelist of patterns; only variables that match _one_ pattern survive
the final step. (Generally used with `inherit = "all"`.) |
In particular, note that the default is `inherit = "core"`, so:
* if you have extra env variables that you want to inherit from the
parent process, use `inherit = "all"` and then specify `include_only`
* if you have extra env variables where you want to hardcode the values,
the default `inherit = "core"` will work fine, but then you need to
specify `set`
This configuration is not battle-tested, so we will probably still have
to play with it a bit. `core/src/exec_env.rs` has the critical business
logic as well as unit tests.
Though if nothing else, previous to this change:
```
$ cargo run --bin codex -- debug seatbelt -- printenv OPENAI_API_KEY
# ...prints OPENAI_API_KEY...
```
But after this change it does not print anything (as desired).
One final thing to call out about this PR is that the
`configure_command!` macro we use in `core/src/exec.rs` has to do some
complex logic with respect to how it builds up the `env` for the process
being spawned under Landlock/seccomp. Specifically, doing
`cmd.env_clear()` followed by `cmd.envs(&$env_map)` (which is arguably
the most intuitive way to do it) caused the Landlock unit tests to fail
because the processes spawned by the unit tests started failing in
unexpected ways! If we forgo `env_clear()` in favor of updating env vars
one at a time, the tests still pass. The comment in the code talks about
this a bit, and while I would like to investigate this more, I need to
move on for the moment, but I do plan to come back to it to fully
understand what is going on. For example, this suggests that we might
not be able to spawn a C program that calls `env_clear()`, which would
be...weird. We may still have to fiddle with our Landlock config if that
is the case.
Previously, running Codex as an MCP server required a standalone binary
in our Cargo workspace, but this PR makes it available as a subcommand
(`mcp`) of the main CLI.
Ran this with:
```
RUST_LOG=debug npx @modelcontextprotocol/inspector cargo run --bin codex -- mcp
```
and verified it worked as expected in the inspector at
`http://127.0.0.1:6274/`.
When using Codex to develop Codex itself, I noticed that sometimes it
would try to add `#[ignore]` to the following tests:
```
keeps_previous_response_id_between_tasks()
retries_on_early_close()
```
Both of these tests start a `MockServer` that launches an HTTP server on
an ephemeral port and requires network access to hit it, which the
Seatbelt policy associated with `--full-auto` correctly denies. If I
wasn't paying attention to the code that Codex was generating, one of
these `#[ignore]` annotations could have slipped into the codebase,
effectively disabling the test for everyone.
To that end, this PR enables an experimental environment variable named
`CODEX_SANDBOX_NETWORK_DISABLED` that is set to `1` if the
`SandboxPolicy` used to spawn the process does not have full network
access. I say it is "experimental" because I'm not convinced this API is
quite right, but we need to start somewhere. (It might be more
appropriate to have an env var like `CODEX_SANDBOX=full-auto`, but the
challenge is that our newer `SandboxPolicy` abstraction does not map to
a simple set of enums like in the TypeScript CLI.)
We leverage this new functionality by adding the following code to the
aforementioned tests as a way to "dynamically disable" them:
```rust
if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
println!(
"Skipping test because it cannot execute when network is disabled in a Codex sandbox."
);
return;
}
```
We can use the `debug seatbelt --full-auto` command to verify that
`cargo test` fails when run under Seatbelt prior to this change:
```
$ cargo run --bin codex -- debug seatbelt --full-auto -- cargo test
---- keeps_previous_response_id_between_tasks stdout ----
thread 'keeps_previous_response_id_between_tasks' panicked at /Users/mbolin/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/wiremock-0.6.3/src/mock_server/builder.rs:107:46:
Failed to bind an OS port for a mock server.: Os { code: 1, kind: PermissionDenied, message: "Operation not permitted" }
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
failures:
keeps_previous_response_id_between_tasks
test result: FAILED. 0 passed; 1 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s
error: test failed, to rerun pass `-p codex-core --test previous_response_id`
```
Though after this change, the above command succeeds! This means that,
going forward, when Codex operates on Codex itself, when it runs `cargo
test`, only "real failures" should cause the command to fail.
As part of this change, I decided to tighten up the codepaths for
running `exec()` for shell tool calls. In particular, we do it in `core`
for the main Codex business logic itself, but we also expose this logic
via `debug` subcommands in the CLI in the `cli` crate. The logic for the
`debug` subcommands was not quite as faithful to the true business logic
as I liked, so I:
* refactored a bit of the Linux code, splitting `linux.rs` into
`linux_exec.rs` and `landlock.rs` in the `core` crate.
* gating less code behind `#[cfg(target_os = "linux")]` because such
code does not get built by default when I develop on Mac, which means I
either have to build the code in Docker or wait for CI signal
* introduced `macro_rules! configure_command` in `exec.rs` so we can
have both sync and async versions of this code. The synchronous version
seems more appropriate for straight threads or potentially fork/exec.
Some effects of this change:
- New formatting changes across many files. No functionality changes
should occur from that.
- Calls to `set_env` are considered unsafe, since this only happens in
tests we wrap them in `unsafe` blocks
@oai-ragona and I discussed it, and we feel the REPL crate has served
its purpose, so we're going to delete the code and future archaeologists
can find it in Git history.
This introduces a standalone executable that run the equivalent of the
`codex debug landlock` subcommand and updates `rust-release.yml` to
include it in the release.
The idea is that we will include this small binary with the TypeScript
CLI to provide support for Linux sandboxing.
Previous to this PR, `SandboxPolicy` was a bit difficult to work with:
237f8a11e1/codex-rs/core/src/protocol.rs (L98-L108)
Specifically:
* It was an `enum` and therefore options were mutually exclusive as
opposed to additive.
* It defined things in terms of what the agent _could not_ do as opposed
to what they _could_ do. This made things hard to support because we
would prefer to build up a sandbox config by starting with something
extremely restrictive and only granting permissions for things the user
as explicitly allowed.
This PR changes things substantially by redefining the policy in terms
of two concepts:
* A `SandboxPermission` enum that defines permissions that can be
granted to the agent/sandbox.
* A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`,
but externally exposes a simpler API that can be used to configure
Seatbelt/Landlock.
Previous to this PR, we supported a `--sandbox` flag that effectively
mapped to an enum value in `SandboxPolicy`. Though now that
`SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single
`--sandbox` flag no longer makes sense. While I could have turned it
into a flag that the user can specify multiple times, I think the
current values to use with such a flag are long and potentially messy,
so for the moment, I have dropped support for `--sandbox` altogether and
we can bring it back once we have figured out the naming thing.
Since `--sandbox` is gone, users now have to specify `--full-auto` to
get a sandbox that allows writes in `cwd`. Admittedly, there is no clean
way to specify the equivalent of `--full-auto` in your `config.toml`
right now, so we will have to revisit that, as well.
Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy`
changed considerably, I had to overhaul how config loading works, as
well. There are now two distinct concepts, `ConfigToml` and `Config`:
* `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one
might expect, every field is `Optional` and it is `#[derive(Deserialize,
Default)]`. Consistent use of `Optional` makes it clear what the user
has specified explicitly.
* `Config` is the "normalized config" and is produced by merging
`ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw
`Option<Vec<SandboxPermission>>`, `Config` presents only the final
`SandboxPolicy`.
The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra
special attention to ensure we are faithfully mapping the
`SandboxPolicy` to the Seatbelt and Landlock configs, respectively.
Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been
renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that
`(allow file-read*)` has been removed from the `.sbpl` file as now this
is added to the policy in `core/src/exec.rs` when
`sandbox_policy.has_full_disk_read_access()` is `true`.
This PR adds a `debug landlock` subcommand to the Codex CLI for testing
how Codex would execute a command using the specified sandbox policy.
Built and ran this code in the `rust:latest` Docker container. In the
container, hitting the network with vanilla `curl` succeeds:
```
$ curl google.com
<HTML><HEAD><meta http-equiv="content-type" content="text/html;charset=utf-8">
<TITLE>301 Moved</TITLE></HEAD><BODY>
<H1>301 Moved</H1>
The document has moved
<A HREF="http://www.google.com/">here</A>.
</BODY></HTML>
```
whereas this fails, as expected:
```
$ cargo run -- debug landlock -s network-restricted -- curl google.com
curl: (6) getaddrinfo() thread failed to start
```
Originally, the `interactive` crate was going to be a placeholder for
building out a UX that was comparable to that of the existing TypeScript
CLI. Though after researching how Ratatui works, that seems difficult to
do because it is designed around the idea that it will redraw the full
screen buffer each time (and so any scrolling should be "internal" to
your Ratatui app) whereas the TypeScript CLI expects to render the full
history of the conversation every time(*) (which is why you can use your
terminal scrollbar to scroll it).
While it is possible to use Ratatui in a way that acts more like what
the TypeScript CLI is doing, it is awkward and seemingly results in
tedious code, so I think we should abandon that approach. As such, this
PR deletes the `interactive/` folder and the code that depended on it.
Further, since we added support for mousewheel scrolling in the TUI in
https://github.com/openai/codex/pull/641, it certainly feels much better
and the need for scroll support via the terminal scrollbar is greatly
diminished. This is now a more appropriate default UX for the
"multitool" CLI.
(*) Incidentally, I haven't verified this, but I think this results in
O(N^2) work in rendering, which seems potentially problematic for long
conversations.
This changes how instantiating `Config` works and also adds
`approval_policy` and `sandbox_policy` as fields. The idea is:
* All fields of `Config` have appropriate default values.
* `Config` is initially loaded from `~/.codex/config.toml`, so values in
`config.toml` will override those defaults.
* Clients must instantiate `Config` via
`Config::load_with_overrides(ConfigOverrides)` where `ConfigOverrides`
has optional overrides that are expected to be settable based on CLI
flags.
The `Config` should be defined early in the program and then passed
down. Now functions like `init_codex()` take fewer individual parameters
because they can just take a `Config`.
Also, `Config::load()` used to fail silently if `~/.codex/config.toml`
had a parse error and fell back to the default config. This seemed
really bad because it wasn't clear why the values in my `config.toml`
weren't getting picked up. I changed things so that
`load_with_overrides()` returns `Result<Config>` and verified that the
various CLIs print a reasonable error if `config.toml` is malformed.
Finally, I also updated the TUI to show which **sandbox** value is being
used, as we do for other key values like **model** and **approval**.
This was also a reminder that the various values of `--sandbox` are
honored on Linux but not macOS today, so I added some TODOs about fixing
that.
As stated in `codex-rs/README.md`:
Today, Codex CLI is written in TypeScript and requires Node.js 22+ to
run it. For a number of users, this runtime requirement inhibits
adoption: they would be better served by a standalone executable. As
maintainers, we want Codex to run efficiently in a wide range of
environments with minimal overhead. We also want to take advantage of
operating system-specific APIs to provide better sandboxing, where
possible.
To that end, we are moving forward with a Rust implementation of Codex
CLI contained in this folder, which has the following benefits:
- The CLI compiles to small, standalone, platform-specific binaries.
- Can make direct, native calls to
[seccomp](https://man7.org/linux/man-pages/man2/seccomp.2.html) and
[landlock](https://man7.org/linux/man-pages/man7/landlock.7.html) in
order to support sandboxing on Linux.
- No runtime garbage collection, resulting in lower memory consumption
and better, more predictable performance.
Currently, the Rust implementation is materially behind the TypeScript
implementation in functionality, so continue to use the TypeScript
implmentation for the time being. We will publish native executables via
GitHub Releases as soon as we feel the Rust version is usable.
