## Summary
In collaboration with @gpeal: upgrade the onboarding flow, and persist
user settings.
---------
Co-authored-by: Gabriel Peal <gabriel@openai.com>
Trying to use `core` as the default has been "too clever." Users can
always take responsibility for controlling the env without this setting
at all by specifying the `env` they use when calling `codex` in the
first place.
See https://github.com/openai/codex/issues/1249.
- For absolute, use non-cached input + output.
- For estimating what % of the model's context window is used, we need
to account for reasoning output tokens from prior turns being dropped
from the context window. We approximate this here by subtracting
reasoning output tokens from the total. This will be off for the current
turn and pending function calls. We can improve it later.
Replaces the `include_default_writable_roots` option on
`sandbox_workspace_write` (that defaulted to `true`, which was slightly
weird/annoying) with `exclude_tmpdir_env_var`, which defaults to
`false`.
Though perhaps more importantly `/tmp` is now enabled by default as part
of `sandbox_mode = "workspace-write"`, though `exclude_slash_tmp =
false` can be used to disable this.
This sets up the scaffolding and basic flow for a TUI onboarding
experience. It covers sign in with ChatGPT, env auth, as well as some
safety guidance.
Next up:
1. Replace the git warning screen
2. Use this to configure default approval/sandbox modes
Note the shimmer flashes are from me slicing the video, not jank.
https://github.com/user-attachments/assets/0fbe3479-fdde-41f3-87fb-a7a83ab895b8
## Summary
- Prioritize provider-specific API keys over default Codex auth when
building requests
- Add test to ensure provider env var auth overrides default auth
## Testing
- `just fmt`
- `just fix` *(fails: `let` expressions in this position are unstable)*
- `cargo test --all-features` *(fails: `let` expressions in this
position are unstable)*
------
https://chatgpt.com/codex/tasks/task_i_68926a104f7483208f2c8fd36763e0e3
The docs and code do not match. It turns out the docs are "right" in
they are what we have been meaning to support, so this PR updates the
code:
ae88b69b09/README.md (L298-L302)
Support for `instructions.md` is a holdover from the TypeScript CLI, so
we are just going to drop support for it altogether rather than maintain
it in perpetuity.
## Summary
Forgot to remove this in #1869 last night! Too much of a performance hit
on the main thread. We can bring it back via an async thread on startup.
## Summary
Includes a new user message in the api payload which provides useful
environment context for the model, so it knows about things like the
current working directory and the sandbox.
## Testing
Updated unit tests
## Summary
A split-up PR of #1763 , stacked on top of a tools refactor #1858 to
make the change clearer. From the previous summary:
> Let's try something new: tell the model about the sandbox, and let it
decide when it will need to break the sandbox. Some local testing
suggests that it works pretty well with zero iteration on the prompt!
## Testing
- [x] Added unit tests
- [x] Tested locally and it appears to work smoothly!
## Summary
In an effort to make tools easier to work with and more configurable,
I'm introducing `ToolConfig` and updating `Prompt` to take in a general
list of Tools. I think this is simpler and better for a few reasons:
- We can easily assemble tools from various sources (our own harness,
mcp servers, etc.) and we can consolidate the logic for constructing the
logic in one place that is separate from serialization.
- client.rs no longer needs arbitrary config values, it just takes in a
list of tools to serialize
A hefty portion of the PR is now updating our conversion of
`mcp_types::Tool` to `OpenAITool`, but considering that @bolinfest
accurately called this out as a TODO long ago, I think it's time we
tackled it.
## Testing
- [x] Experimented locally, no changes, as expected
- [x] Added additional unit tests
- [x] Responded to rust-review
## Summary
Escalating out of sandbox is (almost always) not going to fix
long-running commands timing out - therefore we should just pass the
failure back to the model instead of asking the user to re-run a command
that took a long time anyway.
## Testing
- [x] Ran locally with a timeout and confirmed this worked as expected
This PR started as an investigation with the goal of eliminating the use
of `unsafe { std::env::set_var() }` in `ollama/src/client.rs`, as
setting environment variables in a multithreaded context is indeed
unsafe and these tests were observed to be flaky, as a result.
Though as I dug deeper into the issue, I discovered that the logic for
instantiating `OllamaClient` under test scenarios was not quite right.
In this PR, I aimed to:
- share more code between the two creation codepaths,
`try_from_oss_provider()` and `try_from_provider_with_base_url()`
- use the values from `Config` when setting up Ollama, as we have
various mechanisms for overriding config values, so we should be sure
that we are always using the ultimate `Config` for things such as the
`ModelProviderInfo` associated with the `oss` id
Once this was in place,
`OllamaClient::try_from_provider_with_base_url()` could be used in unit
tests for `OllamaClient` so it was possible to create a properly
configured client without having to set environment variables.
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>
https://github.com/openai/codex/pull/1835 has some messed up history.
This adds support for streaming chat completions, which is useful for ollama. We should probably take a very skeptical eye to the code introduced in this PR.
---------
Co-authored-by: Ahmed Ibrahim <aibrahim@openai.com>
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.
