This PR adds support for a model-based summary and risk assessment for
commands that violate the sandbox policy and require user approval. This
aids the user in evaluating whether the command should be approved.
The feature works by taking a failed command and passing it back to the
model and asking it to summarize the command, give it a risk level (low,
medium, high) and a risk category (e.g. "data deletion" or "data
exfiltration"). It uses a new conversation thread so the context in the
existing thread doesn't influence the answer. If the call to the model
fails or takes longer than 5 seconds, it falls back to the current
behavior.
For now, this is an experimental feature and is gated by a config key
`experimental_sandbox_command_assessment`.
Here is a screen shot of the approval prompt showing the risk assessment
and summary.
<img width="723" height="282" alt="image"
src="https://github.com/user-attachments/assets/4597dd7c-d5a0-4e9f-9d13-414bd082fd6b"
/>
## Summary
- make the plan tool available by default by removing the feature flag
and always registering the handler
- drop plan-tool CLI and API toggles across the exec, TUI, MCP server,
and app server code paths
- update tests and configs to reflect the always-on plan tool and guard
workspace restriction tests against env leakage
## Testing
Manually tested the extension.
------
https://chatgpt.com/codex/tasks/task_i_68f67a3ff2d083209562a773f814c1f9
Add a `--add-dir` CLI flag so sessions can use extra writable roots in
addition to the ones specified in the config file. These are ephemerally
added during the session only.
Fixes#3303Fixes#2797
## Summary
- Factor `load_config_as_toml` into `core::config_loader` so config
loading is reusable across callers.
- Layer `~/.codex/config.toml`, optional `~/.codex/managed_config.toml`,
and macOS managed preferences (base64) with recursive table merging and
scoped threads per source.
## Config Flow
```
Managed prefs (macOS profile: com.openai.codex/config_toml_base64)
▲
│
~/.codex/managed_config.toml │ (optional file-based override)
▲
│
~/.codex/config.toml (user-defined settings)
```
- The loader searches under the resolved `CODEX_HOME` directory
(defaults to `~/.codex`).
- Managed configs let administrators ship fleet-wide overrides via
device profiles which is useful for enforcing certain settings like
sandbox or approval defaults.
- For nested hash tables: overlays merge recursively. Child tables are
merged key-by-key, while scalar or array values replace the prior layer
entirely. This lets admins add or tweak individual fields without
clobbering unrelated user settings.
This is a very large PR with some non-backwards-compatible changes.
Historically, `codex mcp` (or `codex mcp serve`) started a JSON-RPC-ish
server that had two overlapping responsibilities:
- Running an MCP server, providing some basic tool calls.
- Running the app server used to power experiences such as the VS Code
extension.
This PR aims to separate these into distinct concepts:
- `codex mcp-server` for the MCP server
- `codex app-server` for the "application server"
Note `codex mcp` still exists because it already has its own subcommands
for MCP management (`list`, `add`, etc.)
The MCP logic continues to live in `codex-rs/mcp-server` whereas the
refactored app server logic is in the new `codex-rs/app-server` folder.
Note that most of the existing integration tests in
`codex-rs/mcp-server/tests/suite` were actually for the app server, so
all the tests have been moved with the exception of
`codex-rs/mcp-server/tests/suite/mod.rs`.
Because this is already a large diff, I tried not to change more than I
had to, so `codex-rs/app-server/tests/common/mcp_process.rs` still uses
the name `McpProcess` for now, but I will do some mechanical renamings
to things like `AppServer` in subsequent PRs.
While `mcp-server` and `app-server` share some overlapping functionality
(like reading streams of JSONL and dispatching based on message types)
and some differences (completely different message types), I ended up
doing a bit of copypasta between the two crates, as both have somewhat
similar `message_processor.rs` and `outgoing_message.rs` files for now,
though I expect them to diverge more in the near future.
One material change is that of the initialize handshake for `codex
app-server`, as we no longer use the MCP types for that handshake.
Instead, we update `codex-rs/protocol/src/mcp_protocol.rs` to add an
`Initialize` variant to `ClientRequest`, which takes the `ClientInfo`
object we need to update the `USER_AGENT_SUFFIX` in
`codex-rs/app-server/src/message_processor.rs`.
One other material change is in
`codex-rs/app-server/src/codex_message_processor.rs` where I eliminated
a use of the `send_event_as_notification()` method I am generally trying
to deprecate (because it blindly maps an `EventMsg` into a
`JSONNotification`) in favor of `send_server_notification()`, which
takes a `ServerNotification`, as that is intended to be a custom enum of
all notification types supported by the app server. So to make this
update, I had to introduce a new variant of `ServerNotification`,
`SessionConfigured`, which is a non-backwards compatible change with the
old `codex mcp`, and clients will have to be updated after the next
release that contains this PR. Note that
`codex-rs/app-server/tests/suite/list_resume.rs` also had to be update
to reflect this change.
I introduced `codex-rs/utils/json-to-toml/src/lib.rs` as a small utility
crate to avoid some of the copying between `mcp-server` and
`app-server`.
## 📝 Review Mode -- Core
This PR introduces the Core implementation for Review mode:
- New op `Op::Review { prompt: String }:` spawns a child review task
with isolated context, a review‑specific system prompt, and a
`Config.review_model`.
- `EnteredReviewMode`: emitted when the child review session starts.
Every event from this point onwards reflects the review session.
- `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review
finishes or is interrupted, with optional structured findings:
```json
{
"findings": [
{
"title": "<≤ 80 chars, imperative>",
"body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>",
"confidence_score": <float 0.0-1.0>,
"priority": <int 0-3>,
"code_location": {
"absolute_file_path": "<file path>",
"line_range": {"start": <int>, "end": <int>}
}
}
],
"overall_correctness": "patch is correct" | "patch is incorrect",
"overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>",
"overall_confidence_score": <float 0.0-1.0>
}
```
## Questions
### Why separate out its own message history?
We want the review thread to match the training of our review models as
much as possible -- that means using a custom prompt, removing user
instructions, and starting a clean chat history.
We also want to make sure the review thread doesn't leak into the parent
thread.
### Why do this as a mode, vs. sub-agents?
1. We want review to be a synchronous task, so it's fine for now to do a
bespoke implementation.
2. We're still unclear about the final structure for sub-agents. We'd
prefer to land this quickly and then refactor into sub-agents without
rushing that implementation.
We're trying to migrate from `session_id: Uuid` to `conversation_id:
ConversationId`. Not only does this give us more type safety but it
unifies our terminology across Codex and with the implementation of
session resuming, a conversation (which can span multiple sessions) is
more appropriate.
I started this impl on https://github.com/openai/codex/pull/3219 as part
of getting resume working in the extension but it's big enough that it
should be broken out.
When item ids are sent to Responses API it will load them from the
database ignoring the provided values. This adds extra latency.
Not having the mode to store requests also allows us to simplify the
code.
## Breaking change
The `disable_response_storage` configuration option is removed.
Adds web_search tool, enabling the model to use Responses API web_search
tool.
- Disabled by default, enabled by --search flag
- When --search is passed, exposes web_search_request function tool to
the model, which triggers user approval. When approved, the model can
use the web_search tool for the remainder of the turn
<img width="1033" height="294" alt="image"
src="https://github.com/user-attachments/assets/62ac6563-b946-465c-ba5d-9325af28b28f"
/>
---------
Co-authored-by: easong-openai <easong@openai.com>
The existing `wire_format.rs` should share more types with the
`codex-protocol` crate (like `AskForApproval` instead of maintaining a
parallel `CodexToolCallApprovalPolicy` enum), so this PR moves
`wire_format.rs` into `codex-protocol`, renaming it as
`mcp-protocol.rs`. We also de-dupe types, where appropriate.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/2423).
* #2424
* __->__ #2423
## Summary
We've been seeing a number of issues and reports with our synthetic
`apply_patch` tool, e.g. #802. Let's make this a real tool - in my
anecdotal testing, it's critical for GPT-OSS models, but I'd like to
make it the standard across GPT-5 and codex models as well.
## Testing
- [x] Tested locally
- [x] Integration test
This PR:
* Added the clippy.toml to configure allowable expect / unwrap usage in
tests
* Removed as many expect/allow lines as possible from tests
* moved a bunch of allows to expects where possible
Note: in integration tests, non `#[test]` helper functions are not
covered by this so we had to leave a few lingering `expect(expect_used`
checks around
This introduces a new set of request types that our `codex mcp`
supports. Note that these do not conform to MCP tool calls so that
instead of having to send something like this:
```json
{
"jsonrpc": "2.0",
"method": "tools/call",
"id": 42,
"params": {
"name": "newConversation",
"arguments": {
"model": "gpt-5",
"approvalPolicy": "on-request"
}
}
}
```
we can send something like this:
```json
{
"jsonrpc": "2.0",
"method": "newConversation",
"id": 42,
"params": {
"model": "gpt-5",
"approvalPolicy": "on-request"
}
}
```
Admittedly, this new format is not a valid MCP tool call, but we are OK
with that right now. (That is, not everything we might want to request
of `codex mcp` is something that is appropriate for an autonomous agent
to do.)
To start, this introduces four request types:
- `newConversation`
- `sendUserMessage`
- `addConversationListener`
- `removeConversationListener`
The new `mcp-server/tests/codex_message_processor_flow.rs` shows how
these can be used.
The types are defined on the `CodexRequest` enum, so we introduce a new
`CodexMessageProcessor` that is responsible for dealing with requests
from this enum. The top-level `MessageProcessor` has been updated so
that when `process_request()` is called, it first checks whether the
request conforms to `CodexRequest` and dispatches it to
`CodexMessageProcessor` if so.
Note that I also decided to use `camelCase` for the on-the-wire format,
as that seems to be the convention for MCP.
For the moment, the new protocol is defined in `wire_format.rs` within
the `mcp-server` crate, but in a subsequent PR, I will probably move it
to its own crate to ensure the protocol has minimal dependencies and
that we can codegen a schema from it.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/2264).
* #2278
* __->__ #2264
## Summary
#1865 added `AskForApproval::OnRequest`, but missed adding it to our
custom struct in `mcp-server`. This adds the missing configuration
## Testing
- [x] confirmed locally
This adds support for easily running Codex backed by a local Ollama
instance running our new open source models. See
https://github.com/openai/gpt-oss for details.
If you pass in `--oss` you'll be prompted to install/launch ollama, and
it will automatically download the 20b model and attempt to use it.
We'll likely want to expand this with some options later to make the
experience smoother for users who can't run the 20b or want to run the
120b.
Co-authored-by: Michael Bolin <mbolin@openai.com>
This adds a tool the model can call to update a plan. The tool doesn't
actually _do_ anything but it gives clients a chance to read and render
the structured plan. We will likely iterate on the prompt and tools
exposed for planning over time.
## Summary
Adds a new mcp tool call, `codex-reply`, so we can continue existing
sessions. This is a first draft and does not yet support sessions from
previous processes.
## Testing
- [x] tested with mcp client
This PR introduces a single integration test for `cargo mcp`, though it
also introduces a number of reusable components so that it should be
easier to introduce more integration tests going forward.
The new test is introduced in `codex-rs/mcp-server/tests/elicitation.rs`
and the reusable pieces are in `codex-rs/mcp-server/tests/common`.
The test itself verifies new functionality around elicitations
introduced in https://github.com/openai/codex/pull/1623 (and the fix
introduced in https://github.com/openai/codex/pull/1629) by doing the
following:
- starts a mock model provider with canned responses for
`/v1/chat/completions`
- starts the MCP server with a `config.toml` to use that model provider
(and `approval_policy = "untrusted"`)
- sends the `codex` tool call which causes the mock model provider to
request a shell call for `git init`
- the MCP server sends an elicitation to the client to approve the
request
- the client replies to the elicitation with `"approved"`
- the MCP server runs the command and re-samples the model, getting a
`"finish_reason": "stop"`
- in turn, the MCP server sends the final response to the original
`codex` tool call
- verifies that `git init` ran as expected
To test:
```
cargo test shell_command_approval_triggers_elicitation
```
In writing this test, I discovered that `ExecApprovalResponse` does not
conform to `ElicitResult`, so I added a TODO to fix that, since I think
that should be updated in a separate PR. As it stands, this PR does not
update any business logic, though it does make a number of members of
the `mcp-server` crate `pub` so they can be used in the test.
One additional learning from this PR is that
`std::process::Command::cargo_bin()` from the `assert_cmd` trait is only
available for `std::process::Command`, but we really want to use
`tokio::process::Command` so that everything is async and we can
leverage utilities like `tokio::time::timeout()`. The trick I came up
with was to use `cargo_bin()` to locate the program, and then to use
`std::process::Command::get_program()` when constructing the
`tokio::process::Command`.
This updates the schema in `generate_mcp_types.py` from `2025-03-26` to
`2025-06-18`, regenerates `mcp-types/src/lib.rs`, and then updates all
the code that uses `mcp-types` to honor the changes.
Ran
```
npx @modelcontextprotocol/inspector just codex mcp
```
and verified that I was able to invoke the `codex` tool, as expected.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/1621).
* #1623
* #1622
* __->__ #1621
On a high-level, we try to design `config.toml` so that you don't have
to "comment out a lot of stuff" when testing different options.
Previously, defining a sandbox policy was somewhat at odds with this
principle because you would define the policy as attributes of
`[sandbox]` like so:
```toml
[sandbox]
mode = "workspace-write"
writable_roots = [ "/tmp" ]
```
but if you wanted to temporarily change to a read-only sandbox, you
might feel compelled to modify your file to be:
```toml
[sandbox]
mode = "read-only"
# mode = "workspace-write"
# writable_roots = [ "/tmp" ]
```
Technically, commenting out `writable_roots` would not be strictly
necessary, as `mode = "read-only"` would ignore `writable_roots`, but
it's still a reasonable thing to do to keep things tidy.
Currently, the various values for `mode` do not support that many
attributes, so this is not that hard to maintain, but one could imagine
this becoming more complex in the future.
In this PR, we change Codex CLI so that it no longer recognizes
`[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`,
and `[sandbox_workspace_write]` is used to further configure the sandbox
when when `sandbox_mode = "workspace-write"` is used:
```toml
sandbox_mode = "workspace-write"
[sandbox_workspace_write]
writable_roots = [ "/tmp" ]
```
This feels a bit more future-proof in that it is less tedious to
configure different sandboxes:
```toml
sandbox_mode = "workspace-write"
[sandbox_read_only]
# read-only options here...
[sandbox_workspace_write]
writable_roots = [ "/tmp" ]
[sandbox_danger_full_access]
# danger-full-access options here...
```
In this scheme, you never need to comment out the configuration for an
individual sandbox type: you only need to redefine `sandbox_mode`.
Relatedly, previous to this change, a user had to do `-c
sandbox.mode=read-only` to change the mode on the command line. With
this change, things are arguably a bit cleaner because the equivalent
option is `-c sandbox_mode=read-only` (and now `-c
sandbox_workspace_write=...` can be set separately).
Though more importantly, we introduce the `-s/--sandbox` option to the
CLI, which maps directly to `sandbox_mode` in `config.toml`, making
config override behavior easier to reason about. Moreover, as you can
see in the updates to the various Markdown files, it is much easier to
explain how to configure sandboxing when things like `--sandbox
read-only` can be used as an example.
Relatedly, this cleanup also made it straightforward to add support for
a `sandbox` option for Codex when used as an MCP server (see the changes
to `mcp-server/src/codex_tool_config.rs`).
Fixes https://github.com/openai/codex/issues/1248.
For the `approval_policy` config option, renames `unless-allow-listed`
to `untrusted`. In general, when it comes to exec'ing commands, I think
"trusted" is a more accurate term than "safe."
Also drops the `AskForApproval::AutoEdit` variant, as we were not really
making use of it, anyway.
Fixes https://github.com/openai/codex/issues/1250.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/1378).
* #1379
* __->__ #1378
This is a major redesign of how sandbox configuration works and aims to
fix https://github.com/openai/codex/issues/1248. Specifically, it
replaces `sandbox_permissions` in `config.toml` (and the
`-s`/`--sandbox-permission` CLI flags) with a "table" with effectively
three variants:
```toml
# Safest option: full disk is read-only, but writes and network access are disallowed.
[sandbox]
mode = "read-only"
# The cwd of the Codex task is writable, as well as $TMPDIR on macOS.
# writable_roots can be used to specify additional writable folders.
[sandbox]
mode = "workspace-write"
writable_roots = [] # Optional, defaults to the empty list.
network_access = false # Optional, defaults to false.
# Disable sandboxing: use at your own risk!!!
[sandbox]
mode = "danger-full-access"
```
This should make sandboxing easier to reason about. While we have
dropped support for `-s`, the way it works now is:
- no flags => `read-only`
- `--full-auto` => `workspace-write`
- currently, there is no way to specify `danger-full-access` via a CLI
flag, but we will revisit that as part of
https://github.com/openai/codex/issues/1254
Outstanding issue:
- As noted in the `TODO` on `SandboxPolicy::is_unrestricted()`, we are
still conflating sandbox preferences with approval preferences in that
case, which needs to be cleaned up.
This PR introduces support for `-c`/`--config` so users can override
individual config values on the command line using `--config
name=value`. Example:
```
codex --config model=o4-mini
```
Making it possible to set arbitrary config values on the command line
results in a more flexible configuration scheme and makes it easier to
provide single-line examples that can be copy-pasted from documentation.
Effectively, it means there are four levels of configuration for some
values:
- Default value (e.g., `model` currently defaults to `o4-mini`)
- Value in `config.toml` (e.g., user could override the default to be
`model = "o3"` in their `config.toml`)
- Specifying `-c` or `--config` to override `model` (e.g., user can
include `-c model=o3` in their list of args to Codex)
- If available, a config-specific flag can be used, which takes
precedence over `-c` (e.g., user can specify `--model o3` in their list
of args to Codex)
Now that it is possible to specify anything that could be configured in
`config.toml` on the command line using `-c`, we do not need to have a
custom flag for every possible config option (which can clutter the
output of `--help`). To that end, as part of this PR, we drop support
for the `--disable-response-storage` flag, as users can now specify `-c
disable_response_storage=true` to get the equivalent functionality.
Under the hood, this works by loading the `config.toml` into a
`toml::Value`. Then for each `key=value`, we create a small synthetic
TOML file with `value` so that we can run the TOML parser to get the
equivalent `toml::Value`. We then parse `key` to determine the point in
the original `toml::Value` to do the insert/replace. Once all of the
overrides from `-c` args have been applied, the `toml::Value` is
deserialized into a `ConfigToml` and then the `ConfigOverrides` are
applied, as before.
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.
This introduces a much-needed "profile" concept where users can specify
a collection of options under one name and then pass that via
`--profile` to the CLI.
This PR introduces the `ConfigProfile` struct and makes it a field of
`CargoToml`. It further updates
`Config::load_from_base_config_with_overrides()` to respect
`ConfigProfile`, overriding default values where appropriate. A detailed
unit test is added at the end of `config.rs` to verify this behavior.
Details on how to use this feature have also been added to
`codex-rs/README.md`.
Adds `expect()` as a denied lint. Same deal applies with `unwrap()`
where we now need to put `#[expect(...` on ones that we legit want. Took
care to enable `expect()` in test contexts.
# Tests
```
cargo fmt
cargo clippy --all-features --all-targets --no-deps -- -D warnings
cargo test
```
This is the first step in supporting other model providers in the Rust
CLI. Specifically, this PR adds support for the new entries in `Config`
and `ConfigOverrides` to specify a `ModelProviderInfo`, which is the
basic config needed for an LLM provider. This PR does not get us all the
way there yet because `client.rs` still categorically appends
`/responses` to the URL and expects the endpoint to support the OpenAI
Responses API. Will fix that next!
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
This PR replaces the placeholder `"echo"` tool call in the MCP server
with a `"codex"` tool that calls Codex. Events such as
`ExecApprovalRequest` and `ApplyPatchApprovalRequest` are not handled
properly yet, but I have `approval_policy = "never"` set in my
`~/.codex/config.toml` such that those codepaths are not exercised.
The schema for this MPC tool is defined by a new `CodexToolCallParam`
struct introduced in this PR. It is fairly similar to `ConfigOverrides`,
as the param is used to help create the `Config` used to start the Codex
session, though it also includes the `prompt` used to kick off the
session.
This PR also introduces the use of the third-party `schemars` crate to
generate the JSON schema, which is verified in the
`verify_codex_tool_json_schema()` unit test.
Events that are dispatched during the Codex session are sent back to the
MCP client as MCP notifications. This gives the client a way to monitor
progress as the tool call itself may take minutes to complete depending
on the complexity of the task requested by the user.
In the video below, I launched the server via:
```shell
mcp-server$ RUST_LOG=debug npx @modelcontextprotocol/inspector cargo run --
```
In the video, you can see the flow of:
* requesting the list of tools
* choosing the **codex** tool
* entering a value for **prompt** and then making the tool call
Note that I left the other fields blank because when unspecified, the
values in my `~/.codex/config.toml` were used:
https://github.com/user-attachments/assets/1975058c-b004-43ef-8c8d-800a953b8192
Note that while using the inspector, I did run into
https://github.com/modelcontextprotocol/inspector/issues/293, though the
tip about ensuring I had only one instance of the **MCP Inspector** tab
open in my browser seemed to fix things.
