To date, we have a number of hardcoded OpenAI model slug checks spread
throughout the codebase, which makes it hard to audit the various
special cases for each model. To mitigate this issue, this PR introduces
the idea of a `ModelFamily` that has fields to represent the existing
special cases, such as `supports_reasoning_summaries` and
`uses_local_shell_tool`.
There is a `find_family_for_model()` function that maps the raw model
slug to a `ModelFamily`. This function hardcodes all the knowledge about
the special attributes for each model. This PR then replaces the
hardcoded model name checks with checks against a `ModelFamily`.
Note `ModelFamily` is now available as `Config::model_family`. We should
ultimately remove `Config::model` in favor of
`Config::model_family::slug`.
Stream models thoughts and responses instead of waiting for the whole
thing to come through. Very rough right now, but I'm making the risk call to push through.
## Summary
Our recent change in #1737 can sometimes lead to the model confusing
AGENTS.md context as part of the message. But a little prompting and
formatting can help fix this!
## Testing
- Ran locally with a few different prompts to verify the model
behaves well.
- Updated unit tests
Allows users to set their experimental_instructions_file in configs.
For example the below enables experimental instructions when running
`codex -p foo`.
```
[profiles.foo]
experimental_instructions_file = "/Users/foo/.codex/prompt.md"
```
# Testing
- ✅ Running against a profile with experimental_instructions_file works.
- ✅ Running against a profile without experimental_instructions_file
works.
- ✅ Running against no profile with experimental_instructions_file
works.
- ✅ Running against no profile without experimental_instructions_file
works.
This lets us show an accumulating diff across all patches in a turn.
Refer to the docs for TurnDiffTracker for implementation details.
There are multiple ways this could have been done and this felt like the
right tradeoff between reliability and completeness:
*Pros*
* It will pick up all changes to files that the model touched including
if they prettier or another command that updates them.
* It will not pick up changes made by the user or other agents to files
it didn't modify.
*Cons*
* It will pick up changes that the user made to a file that the model
also touched
* It will not pick up changes to codegen or files that were not modified
with apply_patch
## Summary
Users frequently complain about re-approving commands that have failed
for non-sandbox reasons. We can't diagnose with complete accuracy which
errors happened because of a sandbox failure, but we can start to
eliminate some common simple cases.
This PR captures the most common case I've seen, which is a `command not
found` error.
## Testing
- [x] Added unit tests
- [x] Ran a few cases locally
The following test script fails in the codex sandbox:
```
import multiprocessing
from multiprocessing import Lock, Process
def f(lock):
with lock:
print("Lock acquired in child process")
if __name__ == '__main__':
lock = Lock()
p = Process(target=f, args=(lock,))
p.start()
p.join()
```
with
```
Traceback (most recent call last):
File "/Users/david.hao/code/codex/codex-rs/cli/test.py", line 9, in <module>
lock = Lock()
^^^^^^
File "/Users/david.hao/.local/share/uv/python/cpython-3.12.9-macos-aarch64-none/lib/python3.12/multiprocessing/context.py", line 68, in Lock
return Lock(ctx=self.get_context())
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/david.hao/.local/share/uv/python/cpython-3.12.9-macos-aarch64-none/lib/python3.12/multiprocessing/synchronize.py", line 169, in __init__
SemLock.__init__(self, SEMAPHORE, 1, 1, ctx=ctx)
File "/Users/david.hao/.local/share/uv/python/cpython-3.12.9-macos-aarch64-none/lib/python3.12/multiprocessing/synchronize.py", line 57, in __init__
sl = self._semlock = _multiprocessing.SemLock(
^^^^^^^^^^^^^^^^^^^^^^^^^
PermissionError: [Errno 1] Operation not permitted
```
After reading, adding this line to the sandbox configs fixes things -
MacOS multiprocessing appears to use sem_lock(), which opens an IPC
which is considered a disk write even though no file is created. I
interrogated ChatGPT about whether it's okay to loosen, and my
impression after reading is that it is, although would appreciate a
close look
Breadcrumb: You can run `cargo run -- debug seatbelt --full-auto <cmd>`
to test the sandbox
To make `--full-auto` safer, this PR updates the Seatbelt policy so that
a `SandboxPolicy` with a `writable_root` that contains a `.git/`
_directory_ will make `.git/` _read-only_ (though as a follow-up, we
should also consider the case where `.git` is a _file_ with a `gitdir:
/path/to/actual/repo/.git` entry that should also be protected).
The two major changes in this PR:
- Updating `SandboxPolicy::get_writable_roots_with_cwd()` to return a
`Vec<WritableRoot>` instead of a `Vec<PathBuf>` where a `WritableRoot`
can specify a list of read-only subpaths.
- Updating `create_seatbelt_command_args()` to honor the read-only
subpaths in `WritableRoot`.
The logic to update the policy is a fairly straightforward update to
`create_seatbelt_command_args()`, but perhaps the more interesting part
of this PR is the introduction of an integration test in
`tests/sandbox.rs`. Leveraging the new API in #1785, we test
`SandboxPolicy` under various conditions, including ones where `$TMPDIR`
is not readable, which is critical for verifying the new behavior.
To ensure that Codex can run its own tests, e.g.:
```
just codex debug seatbelt --full-auto -- cargo test if_git_repo_is_writable_root_then_dot_git_folder_is_read_only
```
I had to introduce the use of `CODEX_SANDBOX=sandbox`, which is
comparable to how `CODEX_SANDBOX_NETWORK_DISABLED=1` was already being
used.
Adding a comparable change for Landlock will be done in a subsequent PR.
Without this change, it is challenging to create integration tests to
verify that the folders not included in `writable_roots` in
`SandboxPolicy::WorkspaceWrite` are read-only because, by default,
`get_writable_roots_with_cwd()` includes `TMPDIR`, which is where most
integrationt
tests do their work.
This introduces a `use_exact_writable_roots` option to disable the
default
includes returned by `get_writable_roots_with_cwd()`.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/1785).
* #1765
* __->__ #1785
## Summary
- stream command stdout as `ExecCommandStdout` events
- forward streamed stdout to clients and ignore in human output
processor
- adjust call sites for new streaming API
- Add operation to summarize the context so far.
- The operation runs a compact task that summarizes the context.
- The operation clear the previous context to free the context window
- The operation didn't use `run_task` to avoid corrupting the session
- Add /compact in the tui
https://github.com/user-attachments/assets/e06c24e5-dcfb-4806-934a-564d425a919c
At 550 lines, `exec.rs` was a bit large. In particular, I found it hard
to locate the Seatbelt-related code quickly without a file with
`seatbelt` in the name, so this refactors things so:
- `spawn_command_under_seatbelt()` and dependent code moves to a new
`seatbelt.rs` file
- `spawn_child_async()` and dependent code moves to a new `spawn.rs`
file
This is a follow-up to https://github.com/openai/codex/pull/1705, as
that PR inadvertently lost the logic where `PatchApplyBeginEvent` and
`PatchApplyEndEvent` events were sent when patches were auto-approved.
Though as part of this fix, I believe this also makes an important
safety fix to `assess_patch_safety()`, as there was a case that returned
`SandboxType::None`, which arguably is the thing we were trying to avoid
in #1705.
On a high level, we want there to be only one codepath where
`apply_patch` happens, which should be unified with the patch to run
`exec`, in general, so that sandboxing is applied consistently for both
cases.
Prior to this change, `apply_patch()` in `core` would either:
* exit early, delegating to `exec()` to shell out to `apply_patch` using
the appropriate sandbox
* proceed to run the logic for `apply_patch` in memory
549846b29a/codex-rs/core/src/apply_patch.rs (L61-L63)
In this implementation, only the latter would dispatch
`PatchApplyBeginEvent` and `PatchApplyEndEvent`, though the former would
dispatch `ExecCommandBeginEvent` and `ExecCommandEndEvent` for the
`apply_patch` call (or, more specifically, the `codex
--codex-run-as-apply-patch PATCH` call).
To unify things in this PR, we:
* Eliminate the back half of the `apply_patch()` function, and instead
have it also return with `DelegateToExec`, though we add an extra field
to the return value, `user_explicitly_approved_this_action`.
* In `codex.rs` where we process `DelegateToExec`, we use
`SandboxType::None` when `user_explicitly_approved_this_action` is
`true`. This means **we no longer run the apply_patch logic in memory**,
as we always `exec()`. (Note this is what allowed us to delete so much
code in `apply_patch.rs`.)
* In `codex.rs`, we further update `notify_exec_command_begin()` and
`notify_exec_command_end()` to take additional fields to determine what
type of notification to send: `ExecCommand` or `PatchApply`.
Admittedly, this PR also drops some of the functionality about giving
the user the opportunity to expand the set of writable roots as part of
approving the `apply_patch` command. I'm not sure how much that was
used, and we should probably rethink how that works as we are currently
tidying up the protocol to the TUI, in general.
the git tests were failing on my local machine due to gpg signing config
in my ~/.gitconfig. tests should not be affected by ~/.gitconfig, so
configure them to ignore it.
Building on the work of https://github.com/openai/codex/pull/1702, this
changes how a shell call to `apply_patch` is handled.
Previously, a shell call to `apply_patch` was always handled in-process,
never leveraging a sandbox. To determine whether the `apply_patch`
operation could be auto-approved, the
`is_write_patch_constrained_to_writable_paths()` function would check if
all the paths listed in the paths were writable. If so, the agent would
apply the changes listed in the patch.
Unfortunately, this approach afforded a loophole: symlinks!
* For a soft link, we could fix this issue by tracing the link and
checking whether the target is in the set of writable paths, however...
* ...For a hard link, things are not as simple. We can run `stat FILE`
to see if the number of links is greater than 1, but then we would have
to do something potentially expensive like `find . -inum <inode_number>`
to find the other paths for `FILE`. Further, even if this worked, this
approach runs the risk of a
[TOCTOU](https://en.wikipedia.org/wiki/Time-of-check_to_time-of-use)
race condition, so it is not robust.
The solution, implemented in this PR, is to take the virtual execution
of the `apply_patch` CLI into an _actual_ execution using `codex
--codex-run-as-apply-patch PATCH`, which we can run under the sandbox
the user specified, just like any other `shell` call.
This, of course, assumes that the sandbox prevents writing through
symlinks as a mechanism to write to folders that are not in the writable
set configured by the sandbox. I verified this by testing the following
on both Mac and Linux:
```shell
#!/usr/bin/env bash
set -euo pipefail
# Can running a command in SANDBOX_DIR write a file in EXPLOIT_DIR?
# Codex is run in SANDBOX_DIR, so writes should be constrianed to this directory.
SANDBOX_DIR=$(mktemp -d -p "$HOME" sandboxtesttemp.XXXXXX)
# EXPLOIT_DIR is outside of SANDBOX_DIR, so let's see if we can write to it.
EXPLOIT_DIR=$(mktemp -d -p "$HOME" sandboxtesttemp.XXXXXX)
echo "SANDBOX_DIR: $SANDBOX_DIR"
echo "EXPLOIT_DIR: $EXPLOIT_DIR"
cleanup() {
# Only remove if it looks sane and still exists
[[ -n "${SANDBOX_DIR:-}" && -d "$SANDBOX_DIR" ]] && rm -rf -- "$SANDBOX_DIR"
[[ -n "${EXPLOIT_DIR:-}" && -d "$EXPLOIT_DIR" ]] && rm -rf -- "$EXPLOIT_DIR"
}
trap cleanup EXIT
echo "I am the original content" > "${EXPLOIT_DIR}/original.txt"
# Drop the -s to test hard links.
ln -s "${EXPLOIT_DIR}/original.txt" "${SANDBOX_DIR}/link-to-original.txt"
cat "${SANDBOX_DIR}/link-to-original.txt"
if [[ "$(uname)" == "Linux" ]]; then
SANDBOX_SUBCOMMAND=landlock
else
SANDBOX_SUBCOMMAND=seatbelt
fi
# Attempt the exploit
cd "${SANDBOX_DIR}"
codex debug "${SANDBOX_SUBCOMMAND}" bash -lc "echo pwned > ./link-to-original.txt" || true
cat "${EXPLOIT_DIR}/original.txt"
```
Admittedly, this change merits a proper integration test, but I think I
will have to do that in a follow-up PR.
Adds a `CodexAuth` type that encapsulates information about available
auth modes and logic for refreshing the token.
Changes `Responses` API to send requests to different endpoints based on
the auth type.
Updates login_with_chatgpt to support API-less mode and skip the key
exchange.
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.
This is a straight refactor, moving apply-patch-related code from
`codex.rs` and into the new `apply_patch.rs` file. The only "logical"
change is inlining `#[allow(clippy::unwrap_used)]` instead of declaring
`#![allow(clippy::unwrap_used)]` at the top of the file (which is
currently the case in `codex.rs`).
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/1703).
* #1705
* __->__ #1703
* #1702
* #1698
* #1697
Most of the time, we expect the `String` returned by
`serde_json::to_string()` to have extra capacity, so `push('\n')` is
unlikely to allocate, which seems cheaper than an extra `write(2)` call,
on average?
This update replaces the previous ratatui history widget with an
append-only log so that the terminal can handle text selection and
scrolling. It also disables streaming responses, which we'll do our best
to bring back in a later PR. It also adds a small summary of token use
after the TUI exits.
Currently, codex on start shows the value for the approval policy as
name of
[AskForApproval](2437a8d17a/codex-rs/core/src/protocol.rs (L128))
enum, which differs from
[approval_policy](2437a8d17a/codex-rs/config.md (approval_policy))
config values.
E.g. "untrusted" becomes "UnlessTrusted", "on-failure" -> "OnFailure",
"never" -> "Never".
This PR changes render names of the approval policy to match with
configuration values.
This PR updates `is_known_safe_command()` to account for "safe
operators" to expand the set of commands that can be run without
approval. This concept existed in the TypeScript CLI, and we are
[finally!] porting it to the Rust one:
c9e2def494/codex-cli/src/approvals.ts (L531-L541)
The idea is that if we have `EXPR1 SAFE_OP EXPR2` and `EXPR1` and
`EXPR2` are considered safe independently, then `EXPR1 SAFE_OP EXPR2`
should be considered safe. Currently, `SAFE_OP` includes `&&`, `||`,
`;`, and `|`.
In the TypeScript implementation, we relied on
https://www.npmjs.com/package/shell-quote to parse the string of Bash,
as it could provide a "lightweight" parse tree, parsing `'beep || boop >
/byte'` as:
```
[ 'beep', { op: '||' }, 'boop', { op: '>' }, '/byte' ]
```
Though in this PR, we introduce the use of
https://crates.io/crates/tree-sitter-bash for parsing (which
incidentally we were already using in
[`codex-apply-patch`](c9e2def494/codex-rs/apply-patch/Cargo.toml (L18))),
which gives us a richer parse tree. (Incidentally, if you have never
played with tree-sitter, try the
[playground](https://tree-sitter.github.io/tree-sitter/7-playground.html)
and select **Bash** from the dropdown to see how it parses various
expressions.)
As a concrete example, prior to this change, our implementation of
`is_known_safe_command()` could verify things like:
```
["bash", "-lc", "grep -R \"Cargo.toml\" -n"]
```
but not:
```
["bash", "-lc", "grep -R \"Cargo.toml\" -n || true"]
```
With this change, the version with `|| true` is also accepted.
Admittedly, this PR does not expand the safety check to support
subshells, so it would reject, e.g. `bash -lc 'ls || (pwd && echo hi)'`,
but that can be addressed in a subsequent PR.
`nl` is a line-numbering tool that should be on the _trusted _ list, as
there is nothing concerning on https://gtfobins.github.io/gtfobins/nl/
that would merit exclusion.
`true` and `false` are also safe, though not particularly useful given
how `is_known_safe_command()` works today, but that will change with
https://github.com/openai/codex/pull/1668.
# Summary
- Writing effective evals for codex sessions requires context of the
overall repository state at the moment the session began
- This change adds this metadata (git repository, branch, commit hash)
to the top of the rollout of the session (if available - if not it
doesn't add anything)
- Currently, this is only effective on a clean working tree, as we can't
track uncommitted/untracked changes with the current metadata set.
Ideally in the future we may want to track unclean changes somehow, or
perhaps prompt the user to stash or commit them.
# Testing
- Added unit tests
- `cargo test && cargo clippy --tests && cargo fmt -- --config
imports_granularity=Item`
### Resulting Rollout
<img width="1243" height="127" alt="Screenshot 2025-07-17 at 1 50 00 PM"
src="https://github.com/user-attachments/assets/68108941-f015-45b2-985c-ea315ce05415"
/>
1. Emit call_id to exec approval elicitations for mcp client convenience
2. Remove the `-retry` from the call id for the same reason as above but
upstream the reset behavior to the mcp client
Always store the entire conversation history.
Request encrypted COT when not storing Responses.
Send entire input context instead of sending previous_response_id
This PR adds a `load_dotenv()` helper function to the `codex-common`
crate that is available when the `cli` feature is enabled. The function
uses [`dotenvy`](https://crates.io/crates/dotenvy) to update the
environment from:
- `$CODEX_HOME/.env`
- `$(pwd)/.env`
To test:
- ran `printenv OPENAI_API_KEY` to verify the env var exists in my
environment
- ran `just codex exec hello` to verify the CLI uses my `OPENAI_API_KEY`
- ran `unset OPENAI_API_KEY`
- ran `just codex exec hello` again and got **ERROR: Missing environment
variable: `OPENAI_API_KEY`**, as expected
- created `~/.codex/.env` and added `OPENAI_API_KEY=sk-proj-...` (also
ran `chmod 400 ~/.codex/.env` for good measure)
- ran `just codex exec hello` again and it worked, verifying it picked
up `OPENAI_API_KEY` from `~/.codex/.env`
Note this functionality was available in the TypeScript CLI:
https://github.com/openai/codex/pull/122 and was recently requested over
on https://github.com/openai/codex/issues/1262#issuecomment-3093203551.
