`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.
This PR adds the following field to `Config`:
```rust
pub network: Option<NetworkProxy>,
```
Though for the moment, it will always be initialized as `None` (this
will be addressed in a subsequent PR).
This PR does the work to thread `network` through to `execute_exec_env()`, `process_exec_tool_call()`, and `UnifiedExecRuntime.run()` to ensure it is available whenever we span a process.
## Summary
This PR introduces a gated Bubblewrap (bwrap) Linux sandbox path. The
curent Linux sandbox path relies on in-process restrictions (including
Landlock). Bubblewrap gives us a more uniform filesystem isolation
model, especially explicit writable roots with the option to make some
directories read-only and granular network controls.
This is behind a feature flag so we can validate behavior safely before
making it the default.
- Added temporary rollout flag:
- `features.use_linux_sandbox_bwrap`
- Preserved existing default path when the flag is off.
- In Bubblewrap mode:
- Added internal retry without /proc when /proc mount is not permitted
by the host/container.
Historically, `accept_elicitation_for_prompt_rule()` was flaky because
we were using a notification to update the sandbox followed by a `shell`
tool request that we expected to be subject to the new sandbox config,
but because [rmcp](https://crates.io/crates/rmcp) MCP servers delegate
each incoming message to a new Tokio task, messages are not guaranteed
to be processed in order, so sometimes the `shell` tool call would run
before the notification was processed.
Prior to this PR, we relied on a generous `sleep()` between the
notification and the request to reduce the change of the test flaking
out.
This PR implements a proper fix, which is to use a _request_ instead of
a notification for the sandbox update so that we can wait for the
response to the sandbox request before sending the request to the
`shell` tool call. Previously, `rmcp` did not support custom requests,
but I fixed that in
https://github.com/modelcontextprotocol/rust-sdk/pull/590, which made it
into the `0.12.0` release (see #8288).
This PR updates `shell-tool-mcp` to expect
`"codex/sandbox-state/update"` as a _request_ instead of a notification
and sends the appropriate ack. Note this behavior is tied to our custom
`codex/sandbox-state` capability, which Codex honors as an MCP client,
which is why `core/src/mcp_connection_manager.rs` had to be updated as
part of this PR, as well.
This PR also updates the docs at `shell-tool-mcp/README.md`.
helpful in the future if we want more granularity for requesting
escalated permissions:
e.g when running in readonly sandbox, model can request to escalate to a
sandbox that allows writes
This PR introduces integration tests that run
[codex-shell-tool-mcp](https://www.npmjs.com/package/@openai/codex-shell-tool-mcp)
as a user would. Note that this requires running our fork of Bash, so we
introduce a [DotSlash](https://dotslash-cli.com/) file for `bash` so
that we can run the integration tests on multiple platforms without
having to check the binaries into the repository. (As noted in the
DotSlash file, it is slightly more heavyweight than necessary, which may
be worth addressing as disk space in CI is limited:
https://github.com/openai/codex/pull/7678.)
To start, this PR adds two tests:
- `list_tools()` makes the `list_tools` request to the MCP server and
verifies we get the expected response
- `accept_elicitation_for_prompt_rule()` defines a `prefix_rule()` with
`decision="prompt"` and verifies the elicitation flow works as expected
Though the `accept_elicitation_for_prompt_rule()` test **only works on
Linux**, as this PR reveals that there are currently issues when running
the Bash fork in a read-only sandbox on Linux. This will have to be
fixed in a follow-up PR.
Incidentally, getting this test run to correctly on macOS also requires
a recent fix we made to `brew` that hasn't hit a mainline release yet,
so getting CI green in this PR required
https://github.com/openai/codex/pull/7680.
Previous to this change, large `EscalateRequest` payloads exceeded the
kernel send buffer, causing our single `sendmsg(2)` call (with attached
FDs) to be split and retried without proper control handling; this led
to `EINVAL`/broken pipe in the
`handle_escalate_session_respects_run_in_sandbox_decision()` test when
using an `env` with large contents.
**Before:** `AsyncSocket::send_with_fds()` called `send_json_message()`,
which called `send_message_bytes()`, which made one `socket.sendmsg()`
call followed by additional `socket.send()` calls, as necessary:
https://github.com/openai/codex/blob/2e4a40252157751765dff176b35c692df8a9fb4e/codex-rs/exec-server/src/posix/socket.rs#L198-L209
**After:** `AsyncSocket::send_with_fds()` now calls
`send_stream_frame()`, which calls `send_stream_chunk()` one or more
times. Each call to `send_stream_chunk()` calls `socket.sendmsg()`.
In the previous implementation, the subsequent `socket.send()` writes
had no control information associated with them, whereas in the new
`send_stream_chunk()` implementation, a fresh `MsgHdr` (using
`with_control()`, as appropriate) is created for `socket.sendmsg()` each
time.
Additionally, with this PR, stream sending attaches `SCM_RIGHTS` only on
the first chunk, and omits control data when there are no FDs, allowing
oversized payloads to deliver correctly while preserving FD limits and
error checks.
This introduces a new feature to Codex when it operates as an MCP
_client_ where if an MCP _server_ replies that it has an entry named
`"codex/sandbox-state"` in its _server capabilities_, then Codex will
send it an MCP notification with the following structure:
```json
{
"method": "codex/sandbox-state/update",
"params": {
"sandboxPolicy": {
"type": "workspace-write",
"network-access": false,
"exclude-tmpdir-env-var": false
"exclude-slash-tmp": false
},
"codexLinuxSandboxExe": null,
"sandboxCwd": "/Users/mbolin/code/codex2"
}
}
```
or with whatever values are appropriate for the initial `sandboxPolicy`.
**NOTE:** Codex _should_ continue to send the MCP server notifications
of the same format if these things change over the lifetime of the
thread, but that isn't wired up yet.
The result is that `shell-tool-mcp` can consume these values so that
when it calls `codex_core::exec::process_exec_tool_call()` in
`codex-rs/exec-server/src/posix/escalate_server.rs`, it is now sure to
call it with the correct values (whereas previously we relied on
hardcoded values).
While I would argue this is a supported use case within the MCP
protocol, the `rmcp` crate that we are using today does not support
custom notifications. As such, I had to patch it and I submitted it for
review, so hopefully it will be accepted in some form:
https://github.com/modelcontextprotocol/rust-sdk/pull/556
To test out this change from end-to-end:
- I ran `cargo build` in `~/code/codex2/codex-rs/exec-server`
- I built the fork of Bash in `~/code/bash/bash`
- I added the following to my `~/.codex/config.toml`:
```toml
# Use with `codex --disable shell_tool`.
[mcp_servers.execshell]
args = ["--bash", "/Users/mbolin/code/bash/bash"]
command = "/Users/mbolin/code/codex2/codex-rs/target/debug/codex-exec-mcp-server"
```
- From `~/code/codex2/codex-rs`, I ran `just codex --disable shell_tool`
- When the TUI started up, I verified that the sandbox mode is
`workspace-write`
- I ran `/mcp` to verify that the shell tool from the MCP is there:
<img width="1387" height="1400" alt="image"
src="https://github.com/user-attachments/assets/1a8addcc-5005-4e16-b59f-95cfd06fd4ab"
/>
- Then I asked it:
> what is the output of `gh issue list`
because this should be auto-approved with our existing dummy policy:
https://github.com/openai/codex/blob/af63e6eccc35783f1bf4dca3c61adb090efb6b8a/codex-rs/exec-server/src/posix.rs#L157-L164
And it worked:
<img width="1387" height="1400" alt="image"
src="https://github.com/user-attachments/assets/7568d2f7-80da-4d68-86d0-c265a6f5e6c1"
/>
`process_exec_tool_call()` was taking `SandboxType` as a param, but in
practice, the only place it was constructed was in
`codex_message_processor.rs` where it was derived from the other
`sandbox_policy` param, so this PR inlines the logic that decides the
`SandboxType` into `process_exec_tool_call()`.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/7122).
* #7112
* __->__ #7122
Previously, we were running into an issue where we would run the `shell`
tool call with a timeout of 10s, but it fired an elicitation asking for
user approval, the time the user took to respond to the elicitation was
counted agains the 10s timeout, so the `shell` tool call would fail with
a timeout error unless the user is very fast!
This PR addresses this issue by introducing a "stopwatch" abstraction
that is used to manage the timeout. The idea is:
- `Stopwatch::new()` is called with the _real_ timeout of the `shell`
tool call.
- `process_exec_tool_call()` is called with the `Cancellation` variant
of `ExecExpiration` because it should not manage its own timeout in this
case
- the `Stopwatch` expiration is wired up to the `cancel_rx` passed to
`process_exec_tool_call()`
- when an elicitation for the `shell` tool call is received, the
`Stopwatch` pauses
- because it is possible for multiple elicitations to arrive
concurrently, it keeps track of the number of "active pauses" and does
not resume until that counter goes down to zero
I verified that I can test the MCP server using
`@modelcontextprotocol/inspector` and specify `git status` as the
`command` with a timeout of 500ms and that the elicitation pops up and I
have all the time in the world to respond whereas previous to this PR,
that would not have been possible.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/6973).
* #7005
* __->__ #6973
* #6972
This updates `ExecParams` so that instead of taking `timeout_ms:
Option<u64>`, it now takes a more general cancellation mechanism,
`ExecExpiration`, which is an enum that includes a
`Cancellation(tokio_util::sync::CancellationToken)` variant.
If the cancellation token is fired, then `process_exec_tool_call()`
returns in the same way as if a timeout was exceeded.
This is necessary so that in #6973, we can manage the timeout logic
external to the `process_exec_tool_call()` because we want to "suspend"
the timeout when an elicitation from a human user is pending.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/6972).
* #7005
* #6973
* __->__ #6972
This PR reorganizes things slightly so that:
- Instead of a single multitool executable, `codex-exec-server`, we now
have two executables:
- `codex-exec-mcp-server` to launch the MCP server
- `codex-execve-wrapper` is the `execve(2)` wrapper to use with the
`BASH_EXEC_WRAPPER` environment variable
- `BASH_EXEC_WRAPPER` must be a single executable: it cannot be a
command string composed of an executable with args (i.e., it no longer
adds the `escalate` subcommand, as before)
- `codex-exec-mcp-server` takes `--bash` and `--execve` as options.
Though if `--execve` is not specified, the MCP server will check the
directory containing `std::env::current_exe()` and attempt to use the
file named `codex-execve-wrapper` within it. In development, this works
out since these executables are side-by-side in the `target/debug`
folder.
With respect to testing, this also fixes an important bug in
`dummy_exec_policy()`, as I was using `ends_with()` as if it applied to
a `String`, but in this case, it is used with a `&Path`, so the
semantics are slightly different.
Putting this all together, I was able to test this by running the
following:
```
~/code/codex/codex-rs$ npx @modelcontextprotocol/inspector \
./target/debug/codex-exec-mcp-server --bash ~/code/bash/bash
```
If I try to run `git status` in `/Users/mbolin/code/codex` via the
`shell` tool from the MCP server:
<img width="1589" height="1335" alt="image"
src="https://github.com/user-attachments/assets/9db6aea8-7fbc-4675-8b1f-ec446685d6c4"
/>
then I get prompted with the following elicitation, as expected:
<img width="1589" height="1335" alt="image"
src="https://github.com/user-attachments/assets/21b68fe0-494d-4562-9bad-0ddc55fc846d"
/>
Though a current limitation is that the `shell` tool defaults to a
timeout of 10s, which means I only have 10s to respond to the
elicitation. Ideally, the time spent waiting for a response from a human
should not count against the timeout for the command execution. I will
address this in a subsequent PR.
---
Note `~/code/bash/bash` was created by doing:
```
cd ~/code
git clone https://github.com/bminor/bash
cd bash
git checkout a8a1c2fac029404d3f42cd39f5a20f24b6e4fe4b
<apply the patch below>
./configure
make
```
The patch:
```
diff --git a/execute_cmd.c b/execute_cmd.c
index 070f5119..d20ad2b9 100644
--- a/execute_cmd.c
+++ b/execute_cmd.c
@@ -6129,6 +6129,19 @@ shell_execve (char *command, char **args, char **env)
char sample[HASH_BANG_BUFSIZ];
size_t larray;
+ char* exec_wrapper = getenv("BASH_EXEC_WRAPPER");
+ if (exec_wrapper && *exec_wrapper && !whitespace (*exec_wrapper))
+ {
+ char *orig_command = command;
+
+ larray = strvec_len (args);
+
+ memmove (args + 2, args, (++larray) * sizeof (char *));
+ args[0] = exec_wrapper;
+ args[1] = orig_command;
+ command = exec_wrapper;
+ }
+
```
This PR introduces an extra layer of abstraction to prepare us for the
migration to execpolicy2:
- introduces a new trait, `EscalationPolicy`, whose `determine_action()`
method is responsible for producing the `EscalateAction`
- the existing `ExecPolicy` typedef is changed to return an intermediate
`ExecPolicyOutcome` instead of `EscalateAction`
- the default implementation of `EscalationPolicy`,
`McpEscalationPolicy`, composes `ExecPolicy`
- the `ExecPolicyOutcome` includes `codex_execpolicy2::Decision`, which
has a `Prompt` variant
- when `McpEscalationPolicy` gets `Decision::Prompt` back from
`ExecPolicy`, it prompts the user via an MCP elicitation and maps the
result into an `ElicitationAction`
- now that the end user can reply to an elicitation with `Decline` or
`Cancel`, we introduce a new variant, `EscalateAction::Deny`, which the
client handles by returning exit code `1` without running anything
Note the way the elicitation is created is still not quite right, but I
will fix that once we have things running end-to-end for real in a
follow-up PR.
Jeremy Rose ·
Gabriel Peal ·
Michael Bolin ·
Matthew Zeng ·
viyatb-oai ·
iceweasel-oai ·
zhao-oai ·