## Summary
- extend rollout format to store all session data in JSON
- add resume/write helpers for rollouts
- track session state after each conversation
- support `LoadSession` op to resume a previous rollout
- allow starting Codex with an existing session via
`experimental_resume` config variable
We need a way later for exploring the available sessions in a user
friendly way.
## Testing
- `cargo test --no-run` *(fails: `cargo: command not found`)*
------
https://chatgpt.com/codex/tasks/task_i_68792a29dd5c832190bf6930d3466fba
This video is outdated. you should use `-c experimental_resume:<full
path>` instead of `--resume <full path>`
https://github.com/user-attachments/assets/7a9975c7-aa04-4f4e-899a-9e87defd947a
## Summary
- add OpenAI retry and timeout fields to Config
- inject these settings in tests instead of mutating env vars
- plumb Config values through client and chat completions logic
- document new configuration options
## Testing
- `cargo test -p codex-core --no-run`
------
https://chatgpt.com/codex/tasks/task_i_68792c5b04cc832195c03050c8b6ea94
---------
Co-authored-by: Michael Bolin <mbolin@openai.com>
This PR implements server name validation for MCP (Model Context
Protocol) servers to ensure they conform to the required pattern
^[a-zA-Z0-9_-]+$. This addresses the TODO comment in
mcp_connection_manager.rs:82.
+ Added validation before spawning MCP client tasks
+ Invalid server names are added to errors map with descriptive messages
I have read the CLA Document and I hereby sign the CLA
---------
Co-authored-by: Michael Bolin <bolinfest@gmail.com>
- Added support for message and reasoning deltas
- Skipped adding the support in the cli and tui for later
- Commented a failing test (wrong merge) that needs fix in a separate
PR.
Side note: I think we need to disable merge when the CI don't pass.
In order to to this, I created a new `chatgpt` crate where we can put
any code that interacts directly with ChatGPT as opposed to the OpenAI
API. I added a disclaimer to the README for it that it should primarily
be modified by OpenAI employees.
https://github.com/user-attachments/assets/bb978e33-d2c9-4d8e-af28-c8c25b1988e8
https://github.com/openai/codex/pull/1524 introduced the new `config`
field on `ModelClient`, so this does the post-PR cleanup to remove the
now-unnecessary `model` field.
As noted in the updated docs, this makes it so that you can set:
```toml
model_supports_reasoning_summaries = true
```
as a way of overriding the existing heuristic for when to set the
`reasoning` field on a sampling request:
341c091c5b/codex-rs/core/src/client_common.rs (L152-L166)
Some users have proxies or other setups where they are ultimately
hitting OpenAI endpoints, but need a custom `base_url` rather than the
default value of `"https://api.openai.com/v1"`. This PR makes it
possible to override the `base_url` for the `openai` provider via the
`OPENAI_BASE_URL` environment variable.
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.
This adds support for two new model provider config options:
- `http_headers` for hardcoded (key, value) pairs
- `env_http_headers` for headers whose values should be read from
environment variables
This also updates the built-in `openai` provider to use this feature to
set the following headers:
- `originator` => `codex_cli_rs`
- `version` => [CLI version]
- `OpenAI-Organization` => `OPENAI_ORGANIZATION` env var
- `OpenAI-Project` => `OPENAI_PROJECT` env var
for consistency with the TypeScript implementation:
bd5a9e8ba9/codex-cli/src/utils/agent/agent-loop.ts (L321-L329)
While here, this also consolidates some logic that was duplicated across
`client.rs` and `chat_completions.rs` by introducing
`ModelProviderInfo.create_request_builder()`.
Resolves https://github.com/openai/codex/discussions/1152
As discovered in https://github.com/openai/codex/issues/1365, the Azure
provider needs to be able to specify `api-version` as a query param, so
this PR introduces a generic `query_params` option to the
`model_providers` config so that an Azure provider can be defined as
follows:
```toml
[model_providers.azure]
name = "Azure"
base_url = "https://YOUR_PROJECT_NAME.openai.azure.com/openai"
env_key = "AZURE_OPENAI_API_KEY"
query_params = { api-version = "2025-04-01-preview" }
```
This PR also updates the docs with this example.
While here, we also update `wire_api` to default to `"chat"`, as that is
likely the common case for someone defining an external provider.
Fixes https://github.com/openai/codex/issues/1365.
As we are [close to releasing the Rust CLI
beta](https://github.com/openai/codex/discussions/1405), for the moment,
let's take a more neutral stance on what it takes to be a "built-in"
provider.
* For example, there seems to be a discrepancy around what the "right"
configuration for Gemini is: https://github.com/openai/codex/pull/881
* And while the current list of "built-in" providers are all arguably
"well-known" names, this raises a question of what to do about
potentially less familiar providers, such as
https://github.com/openai/codex/pull/1142. Do we just accept every pull
request like this, or is there some criteria a provider has to meet to
"qualify" to be bundled with Codex CLI?
I think that if we can establish clear ground rules for being a built-in
provider, then we can bring this back. But until then, I would rather
take a minimalist approach because if we decided to reverse our position
later, it would break folks who were depending on the presence of the
built-in providers.
When using the OpenAI Responses API, we now record the `usage` field for
a `"response.completed"` event, which includes metrics about the number
of tokens consumed. We also introduce `openai_model_info.rs`, which
includes current data about the most common OpenAI models available via
the API (specifically `context_window` and `max_output_tokens`). If
Codex does not recognize the model, you can set `model_context_window`
and `model_max_output_tokens` explicitly in `config.toml`.
When then introduce a new event type to `protocol.rs`, `TokenCount`,
which includes the `TokenUsage` for the most recent turn.
Finally, we update the TUI to record the running sum of tokens used so
the percentage of available context window remaining can be reported via
the placeholder text for the composer:
We could certainly get much fancier with this (such as reporting the
estimated cost of the conversation), but for now, we are just trying to
achieve feature parity with the TypeScript CLI.
Though arguably this improves upon the TypeScript CLI, as the TypeScript
CLI uses heuristics to estimate the number of tokens used rather than
using the `usage` information directly:
296996d74e/codex-cli/src/utils/approximate-tokens-used.ts (L3-L16)
Fixes https://github.com/openai/codex/issues/1242
This PR reworks `assess_command_safety()` so that the combination of
`AskForApproval::Never` and `SandboxPolicy::DangerFullAccess` ensures
that commands are run without _any_ sandbox and the user should never be
prompted. In turn, it adds support for a new
`--dangerously-bypass-approvals-and-sandbox` flag (that cannot be used
with `--approval-policy` or `--full-auto`) that sets both of those
options.
Fixes https://github.com/openai/codex/issues/1254
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 does not implement the full Login with ChatGPT experience, but it
should unblock people.
**What works**
* The `codex` multitool now has a `login` subcommand, so you can run
`codex login`, which should write `CODEX_HOME/auth.json` if you complete
the flow successfully. The TUI will now read the `OPENAI_API_KEY` from
`auth.json`.
* The TUI should refresh the token if it has expired and the necessary
information is in `auth.json`.
* There is a `LoginScreen` in the TUI that tells you to run `codex
login` if both (1) your model provider expects to use `OPENAI_API_KEY`
as its env var, and (2) `OPENAI_API_KEY` is not set.
**What does not work**
* The `LoginScreen` does not support the login flow from within the TUI.
Instead, it tells you to quit, run `codex login`, and then run `codex`
again.
* `codex exec` does read from `auth.json` yet, nor does it direct the
user to go through the login flow if `OPENAI_API_KEY` is not be found.
* The `maybeRedeemCredits()` function from `get-api-key.tsx` has not
been ported from TypeScript to `login_with_chatgpt.py` yet:
a67a67f325/codex-cli/src/utils/get-api-key.tsx (L84-L89)
**Implementation**
Currently, the OAuth flow requires running a local webserver on
`127.0.0.1:1455`. It seemed wasteful to incur the additional binary cost
of a webserver dependency in the Rust CLI just to support login, so
instead we implement this logic in Python, as Python has a `http.server`
module as part of its standard library. Specifically, we bundle the
contents of a single Python file as a string in the Rust CLI and then
use it to spawn a subprocess as `python3 -c
{{SOURCE_FOR_PYTHON_SERVER}}`.
As such, the most significant files in this PR are:
```
codex-rs/login/src/login_with_chatgpt.py
codex-rs/login/src/lib.rs
```
Now that the CLI may load `OPENAI_API_KEY` from the environment _or_
`CODEX_HOME/auth.json`, we need a new abstraction for reading/writing
this variable, so we introduce:
```
codex-rs/core/src/openai_api_key.rs
```
Note that `std::env::set_var()` is [rightfully] `unsafe` in Rust 2024,
so we use a LazyLock<RwLock<Option<String>>> to store `OPENAI_API_KEY`
so it is read in a thread-safe manner.
Ultimately, it should be possible to go through the entire login flow
from the TUI. This PR introduces a placeholder `LoginScreen` UI for that
right now, though the new `codex login` subcommand introduced in this PR
should be a viable workaround until the UI is ready.
**Testing**
Because the login flow is currently implemented in a standalone Python
file, you can test it without building any Rust code as follows:
```
rm -rf /tmp/codex_home && mkdir /tmp/codex_home
CODEX_HOME=/tmp/codex_home python3 codex-rs/login/src/login_with_chatgpt.py
```
For reference:
* the original TypeScript implementation was introduced in
https://github.com/openai/codex/pull/963
* support for redeeming credits was later added in
https://github.com/openai/codex/pull/974
This fixes a longstanding error in the Rust CLI where `codex.rs`
contained an errant `is_first_turn` check that would exclude the user
instructions for subsequent "turns" of a conversation when using the
responses API (i.e., when `previous_response_id` existed).
While here, renames `Prompt.instructions` to `Prompt.user_instructions`
since we now have quite a few levels of instructions floating around.
Also removed an unnecessary use of `clone()` in
`Prompt.get_full_instructions()`.
As explained in detail in the doc comment for `ParseMode::Lenient`, we
have observed that GPT-4.1 does not always generate a valid invocation
of `apply_patch`. Fortunately, the error is predictable, so we introduce
some new logic to the `codex-apply-patch` crate to recover from this
error.
Because we would like to avoid this becoming a de facto standard (as it
would be incompatible if `apply_patch` were provided as an actual
executable, unless we also introduced the lenient behavior in the
executable, as well), we require passing `ParseMode::Lenient` to
`parse_patch_text()` to make it clear that the caller is opting into
supporting this special case.
Note the analogous change to the TypeScript CLI was
https://github.com/openai/codex/pull/930. In addition to changing the
accepted input to `apply_patch`, it also introduced additional
instructions for the model, which we include in this PR.
Note that `apply-patch` does not depend on either `regex` or
`regex-lite`, so some of the checks are slightly more verbose to avoid
introducing this dependency.
That said, this PR does not leverage the existing
`extract_heredoc_body_from_apply_patch_command()`, which depends on
`tree-sitter` and `tree-sitter-bash`:
5a5aa89914/codex-rs/apply-patch/src/lib.rs (L191-L246)
though perhaps it should.
Previous to this PR, we always set `reasoning` when making a request
using the Responses API:
d7245cbbc9/codex-rs/core/src/client.rs (L108-L111)
Though if you tried to use the Rust CLI with `--model gpt-4.1`, this
would fail with:
```shell
"Unsupported parameter: 'reasoning.effort' is not supported with this model."
```
We take a cue from the TypeScript CLI, which does a check on the model
name:
d7245cbbc9/codex-cli/src/utils/agent/agent-loop.ts (L786-L789)
This PR does a similar check, though also adds support for the following
config options:
```
model_reasoning_effort = "low" | "medium" | "high" | "none"
model_reasoning_summary = "auto" | "concise" | "detailed" | "none"
```
This way, if you have a model whose name happens to start with `"o"` (or
`"codex"`?), you can set these to `"none"` to explicitly disable
reasoning, if necessary. (That said, it seems unlikely anyone would use
the Responses API with non-OpenAI models, but we provide an escape
hatch, anyway.)
This PR also updates both the TUI and `codex exec` to show `reasoning
effort` and `reasoning summaries` in the header.
Prior to this PR, there were two big misses in `chat_completions.rs`:
1. The loop in `stream_chat_completions()` was only including items of
type `ResponseItem::Message` when building up the `"messages"` JSON for
the `POST` request to the `chat/completions` endpoint. This fixes things
by ensuring other variants (`FunctionCall`, `LocalShellCall`, and
`FunctionCallOutput`) are included, as well.
2. In `process_chat_sse()`, we were not recording tool calls and were
only emitting items of type
`ResponseEvent::OutputItemDone(ResponseItem::Message)` to the stream.
Now we introduce `FunctionCallState`, which is used to accumulate the
`delta`s of type `tool_calls`, so we can ultimately emit a
`ResponseItem::FunctionCall`, when appropriate.
While function calling now appears to work for chat completions with my
local testing, I believe that there are still edge cases that are not
covered and that this codepath would benefit from a battery of
integration tests. (As part of that further cleanup, we should also work
to support streaming responses in the UI.)
The other important part of this PR is some cleanup in
`core/src/codex.rs`. In particular, it was hard to reason about how
`run_task()` was building up the list of messages to include in a
request across the various cases:
- Responses API
- Chat Completions API
- Responses API used in concert with ZDR
I like to think things are a bit cleaner now where:
- `zdr_transcript` (if present) contains all messages in the history of
the conversation, which includes function call outputs that have not
been sent back to the model yet
- `pending_input` includes any messages the user has submitted while the
turn is in flight that need to be injected as part of the next `POST` to
the model
- `input_for_next_turn` includes the tool call outputs that have not
been sent back to the model yet
This PR introduces a `hide_agent_reasoning` config option (that defaults
to `false`) that users can enable to make the output less verbose by
suppressing reasoning output.
To test, verified that this includes agent reasoning in the output:
```
echo hello | just exec
```
whereas this does not:
```
echo hello | just exec --config hide_agent_reasoning=false
```
The main motivator behind this PR is that `stream_chat_completions()`
was not adding the `"tools"` entry to the payload posted to the
`/chat/completions` endpoint. This (1) refactors the existing logic to
build up the `"tools"` JSON from `client.rs` into `openai_tools.rs`, and
(2) updates the use of responses API (`client.rs`) and chat completions
API (`chat_completions.rs`) to both use it.
Note this PR alone is not sufficient to get tool calling from chat
completions working: that is done in
https://github.com/openai/codex/pull/1167.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/1177).
* #1167
* __->__ #1177
The output of an MCP server tool call can be one of several types, but
to date, we treated all outputs as text by showing the serialized JSON
as the "tool output" in Codex:
25a9949c49/codex-rs/mcp-types/src/lib.rs (L96-L101)
This PR adds support for the `ImageContent` variant so we can now
display an image output from an MCP tool call.
In making this change, we introduce a new
`ResponseInputItem::McpToolCallOutput` variant so that we can work with
the `mcp_types::CallToolResult` directly when the function call is made
to an MCP server.
Though arguably the more significant change is the introduction of
`HistoryCell::CompletedMcpToolCallWithImageOutput`, which is a cell that
uses `ratatui_image` to render an image into the terminal. To support
this, we introduce `ImageRenderCache`, cache a
`ratatui_image::picker::Picker`, and `ensure_image_cache()` to cache the
appropriate scaled image data and dimensions based on the current
terminal size.
To test, I created a minimal `package.json`:
```json
{
"name": "kitty-mcp",
"version": "1.0.0",
"type": "module",
"description": "MCP that returns image of kitty",
"main": "index.js",
"dependencies": {
"@modelcontextprotocol/sdk": "^1.12.0"
}
}
```
with the following `index.js` to define the MCP server:
```js
#!/usr/bin/env node
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { readFile } from "node:fs/promises";
import { join } from "node:path";
const IMAGE_URI = "image://Ada.png";
const server = new McpServer({
name: "Demo",
version: "1.0.0",
});
server.tool(
"get-cat-image",
"If you need a cat image, this tool will provide one.",
async () => ({
content: [
{ type: "image", data: await getAdaPngBase64(), mimeType: "image/png" },
],
})
);
server.resource("Ada the Cat", IMAGE_URI, async (uri) => {
const base64Image = await getAdaPngBase64();
return {
contents: [
{
uri: uri.href,
mimeType: "image/png",
blob: base64Image,
},
],
};
});
async function getAdaPngBase64() {
const __dirname = new URL(".", import.meta.url).pathname;
// From 9705ce2c59/assets/Ada.png
const filePath = join(__dirname, "Ada.png");
const imageData = await readFile(filePath);
const base64Image = imageData.toString("base64");
return base64Image;
}
const transport = new StdioServerTransport();
await server.connect(transport);
```
With the local changes from this PR, I added the following to my
`config.toml`:
```toml
[mcp_servers.kitty]
command = "node"
args = ["/Users/mbolin/code/kitty-mcp/index.js"]
```
Running the TUI from source:
```
cargo run --bin codex -- --model o3 'I need a picture of a cat'
```
I get:
<img width="732" alt="image"
src="https://github.com/user-attachments/assets/bf80b721-9ca0-4d81-aec7-77d6899e2869"
/>
Now, that said, I have only tested in iTerm and there is definitely some
funny business with getting an accurate character-to-pixel ratio
(sometimes the `CompletedMcpToolCallWithImageOutput` thinks it needs 10
rows to render instead of 4), so there is still work to be done here.
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.
Historically, we spawned the Seatbelt and Landlock sandboxes in
substantially different ways:
For **Seatbelt**, we would run `/usr/bin/sandbox-exec` with our policy
specified as an arg followed by the original command:
d1de7bb383/codex-rs/core/src/exec.rs (L147-L219)
For **Landlock/Seccomp**, we would do
`tokio::runtime::Builder::new_current_thread()`, _invoke
Landlock/Seccomp APIs to modify the permissions of that new thread_, and
then spawn the command:
d1de7bb383/codex-rs/core/src/exec_linux.rs (L28-L49)
While it is neat that Landlock/Seccomp supports applying a policy to
only one thread without having to apply it to the entire process, it
requires us to maintain two different codepaths and is a bit harder to
reason about. The tipping point was
https://github.com/openai/codex/pull/1061, in which we had to start
building up the `env` in an unexpected way for the existing
Landlock/Seccomp approach to continue to work.
This PR overhauls things so that we do similar things for Mac and Linux.
It turned out that we were already building our own "helper binary"
comparable to Mac's `sandbox-exec` as part of the `cli` crate:
d1de7bb383/codex-rs/cli/Cargo.toml (L10-L12)
We originally created this to build a small binary to include with the
Node.js version of the Codex CLI to provide support for Linux
sandboxing.
Though the sticky bit is that, at this point, we still want to deploy
the Rust version of Codex as a single, standalone binary rather than a
CLI and a supporting sandboxing binary. To satisfy this goal, we use
"the arg0 trick," in which we:
* use `std::env::current_exe()` to get the path to the CLI that is
currently running
* use the CLI as the `program` for the `Command`
* set `"codex-linux-sandbox"` as arg0 for the `Command`
A CLI that supports sandboxing should check arg0 at the start of the
program. If it is `"codex-linux-sandbox"`, it must invoke
`codex_linux_sandbox::run_main()`, which runs the CLI as if it were
`codex-linux-sandbox`. When acting as `codex-linux-sandbox`, we make the
appropriate Landlock/Seccomp API calls and then use `execvp(3)` to spawn
the original command, so do _replace_ the process rather than spawn a
subprocess. Incidentally, we do this before starting the Tokio runtime,
so the process should only have one thread when `execvp(3)` is called.
Because the `core` crate that needs to spawn the Linux sandboxing is not
a CLI in its own right, this means that every CLI that includes `core`
and relies on this behavior has to (1) implement it and (2) provide the
path to the sandboxing executable. While the path is almost always
`std::env::current_exe()`, we needed to make this configurable for
integration tests, so `Config` now has a `codex_linux_sandbox_exe:
Option<PathBuf>` property to facilitate threading this through,
introduced in https://github.com/openai/codex/pull/1089.
This common pattern is now captured in
`codex_linux_sandbox::run_with_sandbox()` and all of the `main.rs`
functions that should use it have been updated as part of this PR.
The `codex-linux-sandbox` crate added to the Cargo workspace as part of
this PR now has the bulk of the Landlock/Seccomp logic, which makes
`core` a bit simpler. Indeed, `core/src/exec_linux.rs` and
`core/src/landlock.rs` were removed/ported as part of this PR. I also
moved the unit tests for this code into an integration test,
`linux-sandbox/tests/landlock.rs`, in which I use
`env!("CARGO_BIN_EXE_codex-linux-sandbox")` as the value for
`codex_linux_sandbox_exe` since `std::env::current_exe()` is not
appropriate in that case.
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.
`config.rs` is already quite long without these definitions. Since they
have no real dependencies of their own, let's move them to their own
file so `config.rs` can focus on the business logic of loading a config.
This introduces an experimental `--output-last-message` flag that can be
used to identify a file where the final message from the agent will be
written. Two use cases:
- Ultimately, we will likely add a `--quiet` option to `exec`, but even
if the user does not want any output written to the terminal, they
probably want to know what the agent did. Writing the output to a file
makes it possible to get that information in a clean way.
- Relatedly, when using `exec` in CI, it is easier to review the
transcript written "normally," (i.e., not as JSON or something with
extra escapes), but getting programmatic access to the last message is
likely helpful, so writing the last message to a file gets the best of
both worlds.
I am calling this "experimental" because it is possible that we are
overfitting and will want a more general solution to this problem that
would justify removing this flag.
I did a bit of research to understand why I could not use my mouse to
drag to select text to copy to the clipboard in iTerm.
Apparently https://github.com/openai/codex/pull/641 to enable mousewheel
scrolling broke this functionality. It seems that, unless we put in a
bit of effort, we can have drag-to-select or scrolling, but not both.
Though if you know the trick to hold down `Option` will dragging with
the mouse in iTerm, you can probably get by with this. (I did not know
about this option prior to researching this issue.)
Nevertheless, users may still prefer to disable mouse capture
altogether, so this PR introduces:
* the ability to set `tui.disable_mouse_capture = true` in `config.toml`
to disable mouse capture
* a new command, `/toggle-mouse-mode` to toggle mouse capture
The new `codex-mini-latest` model expects a new tool with `{"type":
"local_shell"}`. Its contract is similar to the existing `function` tool
with `"name": "shell"`, so this takes the `local_shell` tool call into
`ExecParams` and sends it through the existing
`handle_container_exec_with_params()` code path.
This also adds the following logic when adding the default set of tools
to a request:
```rust
let default_tools = if self.model.starts_with("codex") {
&DEFAULT_CODEX_MODEL_TOOLS
} else {
&DEFAULT_TOOLS
};
```
That is, if the model name starts with `"codex"`, we add `{"type":
"local_shell"}` to the list of tools; otherwise, we add the
aforementioned `shell` tool.
To test this, I ran the TUI with `-m codex-mini-latest` and verified
that it used the `local_shell` tool. Though I also had some entries in
`[mcp_servers]` in my personal `config.toml`. The `codex-mini-latest`
model seemed eager to try the tools from the MCP servers first, so I
have personally commented them out for now, so keep an eye out if you're
testing `codex-mini-latest`!
Perhaps we should include more details with `{"type": "local_shell"}` or
update the following:
fd0b1b0208/codex-rs/core/prompt.md
For reference, the corresponding change in the TypeScript CLI is
https://github.com/openai/codex/pull/951.
