Add implementation for the `wait` tool.
For this we consider all status different from `PendingInit` and
`Running` as terminal. The `wait` tool call will return either after a
given timeout or when the tool reaches a non-terminal status.
A few points to note:
* The usage of a channel is preferred to prevent some races (just
looping on `get_status()` could "miss" a terminal status)
* The order of operations is very important, we need to first subscribe
and then check the last known status to prevent race conditions
* If the channel gets dropped, we return an error on purpose
Agent wouldn't "see" attached images and would instead try to use the
view_file tool:
<img width="1516" height="504" alt="image"
src="https://github.com/user-attachments/assets/68a705bb-f962-4fc1-9087-e932a6859b12"
/>
In this PR, we wrap image content items in XML tags with the name of
each image (now just a numbered name like `[Image #1]`), so that the
model can understand inline image references (based on name). We also
put the image content items above the user message which the model seems
to prefer (maybe it's more used to definitions being before references).
We also tweak the view_file tool description which seemed to help a bit
Results on a simple eval set of images:
Before
<img width="980" height="310" alt="image"
src="https://github.com/user-attachments/assets/ba838651-2565-4684-a12e-81a36641bf86"
/>
After
<img width="918" height="322" alt="image"
src="https://github.com/user-attachments/assets/10a81951-7ee6-415e-a27e-e7a3fd0aee6f"
/>
```json
[
{
"id": "single_describe",
"prompt": "Describe the attached image in one sentence.",
"images": ["image_a.png"]
},
{
"id": "single_color",
"prompt": "What is the dominant color in the image? Answer with a single color word.",
"images": ["image_b.png"]
},
{
"id": "orientation_check",
"prompt": "Is the image portrait or landscape? Answer in one sentence.",
"images": ["image_c.png"]
},
{
"id": "detail_request",
"prompt": "Look closely at the image and call out any small details you notice.",
"images": ["image_d.png"]
},
{
"id": "two_images_compare",
"prompt": "I attached two images. Are they the same or different? Briefly explain.",
"images": ["image_a.png", "image_b.png"]
},
{
"id": "two_images_captions",
"prompt": "Provide a short caption for each image (Image 1, Image 2).",
"images": ["image_c.png", "image_d.png"]
},
{
"id": "multi_image_rank",
"prompt": "Rank the attached images from most colorful to least colorful.",
"images": ["image_a.png", "image_b.png", "image_c.png"]
},
{
"id": "multi_image_choice",
"prompt": "Which image looks more vibrant? Answer with 'Image 1' or 'Image 2'.",
"images": ["image_b.png", "image_d.png"]
}
]
```
Handle null tool arguments in the MCP resource handler so optional
resource tools accept null without failing, preserving normal JSON
parsing for non-null payloads and improving robustness when models emit
null; this avoids spurious argument parse errors for list/read MCP
resource calls.
Sort list_dir entries before applying offset/limit so pagination matches
the displayed order, update pagination/truncation expectations, and add
coverage for sorted pagination. This ensures stable, predictable
directory pages when list_dir is enabled.
**Summary**
This PR makes “ApprovalDecision::AcceptForSession / don’t ask again this
session” actually work for `apply_patch` approvals by caching approvals
based on absolute file paths in codex-core, properly wiring it through
app-server v2, and exposing the choice in both TUI and TUI2.
- This brings `apply_patch` calls to be at feature-parity with general
shell commands, which also have a "Yes, and don't ask again" option.
- This also fixes VSCE's "Allow this session" button to actually work.
While we're at it, also split the app-server v2 protocol's
`ApprovalDecision` enum so execpolicy amendments are only available for
command execution approvals.
**Key changes**
- Core: per-session patch approval allowlist keyed by absolute file
paths
- Handles multi-file patches and renames/moves by recording both source
and destination paths for `Update { move_path: Some(...) }`.
- Extend the `Approvable` trait and `ApplyPatchRuntime` to work with
multiple keys, because an `apply_patch` tool call can modify multiple
files. For a request to be auto-approved, we will need to check that all
file paths have been approved previously.
- App-server v2: honor AcceptForSession for file changes
- File-change approval responses now map AcceptForSession to
ReviewDecision::ApprovedForSession (no longer downgraded to plain
Approved).
- Replace `ApprovalDecision` with two enums:
`CommandExecutionApprovalDecision` and `FileChangeApprovalDecision`
- TUI / TUI2: expose “don’t ask again for these files this session”
- Patch approval overlays now include a third option (“Yes, and don’t
ask again for these files this session (s)”).
- Snapshot updates for the approval modal.
**Tests added/updated**
- Core:
- Integration test that proves ApprovedForSession on a patch skips the
next patch prompt for the same file
- App-server:
- v2 integration test verifying
FileChangeApprovalDecision::AcceptForSession works properly
**User-visible behavior**
- When the user approves a patch “for session”, future patches touching
only those previously approved file(s) will no longer prompt gain during
that session (both via app-server v2 and TUI/TUI2).
**Manual testing**
Tested both TUI and TUI2 - see screenshots below.
TUI:
<img width="1082" height="355" alt="image"
src="https://github.com/user-attachments/assets/adcf45ad-d428-498d-92fc-1a0a420878d9"
/>
TUI2:
<img width="1089" height="438" alt="image"
src="https://github.com/user-attachments/assets/dd768b1a-2f5f-4bd6-98fd-e52c1d3abd9e"
/>
https://github.com/openai/codex/pull/8235 introduced `ConfigBuilder` and
this PR updates all call non-test call sites to use it instead of
`Config::load_from_base_config_with_overrides()`.
This is important because `load_from_base_config_with_overrides()` uses
an empty `ConfigRequirements`, which is a reasonable default for testing
so the tests are not influenced by the settings on the host. This method
is now guarded by `#[cfg(test)]` so it cannot be used by business logic.
Because `ConfigBuilder::build()` is `async`, many of the test methods
had to be migrated to be `async`, as well. On the bright side, this made
it possible to eliminate a bunch of `block_on_future()` stuff.
Previous to this PR, we used a hand-rolled PowerShell parser in
`windows_safe_commands.rs` to take a `&str` of PowerShell script see if
it is equivalent to a list of `execvp(3)` invocations, and if so, we
then test each using `is_safe_powershell_command()` to determine if the
overall command is safe:
https://github.com/openai/codex/blob/6e6338aa876bb4258abe25b02ac6417b8ea9dff0/codex-rs/core/src/command_safety/windows_safe_commands.rs#L89-L98
Unfortunately, our PowerShell parser did not recognize `@(...)` as a
special construct, so it was treated as an ordinary token. This meant
that the following would erroneously be considered "safe:"
```powershell
ls @(calc.exe)
```
The fix introduced in this PR is to do something comparable what we do
for Bash/Zsh, which is to use a "proper" parser to derive the list of
`execvp(3)` calls. For Bash/Zsh, we rely on
https://crates.io/crates/tree-sitter-bash, but there does not appear to
be a crate of comparable quality for parsing PowerShell statically
(https://github.com/airbus-cert/tree-sitter-powershell/ is the best
thing I found).
Instead, in this PR, we use a PowerShell script to parse the input
PowerShell program to produce the AST.
helpful in the future if we want more granularity for requesting
escalated permissions:
e.g when running in readonly sandbox, model can request to escalate to a
sandbox that allows writes
this PR enables TUI to approve commands and add their prefixes to an
allowlist:
<img width="708" height="605" alt="Screenshot 2025-11-21 at 4 18 07 PM"
src="https://github.com/user-attachments/assets/56a19893-4553-4770-a881-becf79eeda32"
/>
note: we only show the option to whitelist the command when
1) command is not multi-part (e.g `git add -A && git commit -m 'hello
world'`)
2) command is not already matched by an existing rule
# Unified Exec Shell Selection on Windows
## Problem
reference issue #7466
The `unified_exec` handler currently deserializes model-provided tool
calls into the `ExecCommandArgs` struct:
```rust
#[derive(Debug, Deserialize)]
struct ExecCommandArgs {
cmd: String,
#[serde(default)]
workdir: Option<String>,
#[serde(default = "default_shell")]
shell: String,
#[serde(default = "default_login")]
login: bool,
#[serde(default = "default_exec_yield_time_ms")]
yield_time_ms: u64,
#[serde(default)]
max_output_tokens: Option<usize>,
#[serde(default)]
with_escalated_permissions: Option<bool>,
#[serde(default)]
justification: Option<String>,
}
```
The `shell` field uses a hard-coded default:
```rust
fn default_shell() -> String {
"/bin/bash".to_string()
}
```
When the model returns a tool call JSON that only contains `cmd` (which
is the common case), Serde fills in `shell` with this default value.
Later, `get_command` uses that value as if it were a model-provided
shell path:
```rust
fn get_command(args: &ExecCommandArgs) -> Vec<String> {
let shell = get_shell_by_model_provided_path(&PathBuf::from(args.shell.clone()));
shell.derive_exec_args(&args.cmd, args.login)
}
```
On Unix, this usually resolves to `/bin/bash` and works as expected.
However, on Windows this behavior is problematic:
- The hard-coded `"/bin/bash"` is not a valid Windows path.
- `get_shell_by_model_provided_path` treats this as a model-specified
shell, and tries to resolve it (e.g. via `which::which("bash")`), which
may or may not exist and may not behave as intended.
- In practice, this leads to commands being executed under a non-default
or non-existent shell on Windows (for example, WSL bash), instead of the
expected Windows PowerShell or `cmd.exe`.
The core of the issue is that **"model did not specify `shell`" is
currently interpreted as "the model explicitly requested `/bin/bash`"**,
which is both Unix-specific and wrong on Windows.
## Proposed Solution
Instead of hard-coding `"/bin/bash"` into `ExecCommandArgs`, we should
distinguish between:
1. **The model explicitly specifying a shell**, e.g.:
```json
{
"cmd": "echo hello",
"shell": "pwsh"
}
```
In this case, we *do* want to respect the model’s choice and use
`get_shell_by_model_provided_path`.
2. **The model omitting the `shell` field entirely**, e.g.:
```json
{
"cmd": "echo hello"
}
```
In this case, we should *not* assume `/bin/bash`. Instead, we should use
`default_user_shell()` and let the platform decide.
To express this distinction, we can:
1. Change `shell` to be optional in `ExecCommandArgs`:
```rust
#[derive(Debug, Deserialize)]
struct ExecCommandArgs {
cmd: String,
#[serde(default)]
workdir: Option<String>,
#[serde(default)]
shell: Option<String>,
#[serde(default = "default_login")]
login: bool,
#[serde(default = "default_exec_yield_time_ms")]
yield_time_ms: u64,
#[serde(default)]
max_output_tokens: Option<usize>,
#[serde(default)]
with_escalated_permissions: Option<bool>,
#[serde(default)]
justification: Option<String>,
}
```
Here, the absence of `shell` in the JSON is represented as `shell:
None`, rather than a hard-coded string value.
I think this might help with https://github.com/openai/codex/pull/7033
because `create_approval_requirement_for_command()` will soon need
access to `Session.state`, which is a `tokio::sync::Mutex` that needs to
be accessed via `async`.
This updates `ExecParams` so that instead of taking `timeout_ms:
Option<u64>`, it now takes a more general cancellation mechanism,
`ExecExpiration`, which is an enum that includes a
`Cancellation(tokio_util::sync::CancellationToken)` variant.
If the cancellation token is fired, then `process_exec_tool_call()`
returns in the same way as if a timeout was exceeded.
This is necessary so that in #6973, we can manage the timeout logic
external to the `process_exec_tool_call()` because we want to "suspend"
the timeout when an elicitation from a human user is pending.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/6972).
* #7005
* #6973
* __->__ #6972
This PR threads execpolicy2 into codex-core.
activated via feature flag: exec_policy (on by default)
reads and parses all .codexpolicy files in `codex_home/codex`
refactored tool runtime API to integrate execpolicy logic
---------
Co-authored-by: Michael Bolin <mbolin@openai.com>
Instead of returning structured out and then re-formatting it into
freeform, return the freeform output from shell_command tool.
Keep `shell` as the default tool for GPT-5.
- Local-shell tool responses were always tagged as
`ExecCommandSource::UserShell` because handler would call
`run_exec_like` with `is_user_shell_cmd` set to true.
- Treat `ToolPayload::LocalShell` the same as other model generated
shell tool calls by deleting `is_user_shell_cmd` from `run_exec_like`
(since actual user shell commands follow a separate code path)
This adds support for a new variant of the shell tool behind a flag. To
test, run `codex` with `--enable shell_command_tool`, which will
register the tool with Codex under the name `shell_command` that accepts
the following shape:
```python
{
command: str
workdir: str | None,
timeout_ms: int | None,
with_escalated_permissions: bool | None,
justification: str | None,
}
```
This is comparable to the existing tool registered under
`shell`/`container.exec`. The primary difference is that it accepts
`command` as a `str` instead of a `str[]`. The `shell_command` tool
executes by running `execvp(["bash", "-lc", command])`, though the exact
arguments to `execvp(3)` depend on the user's default shell.
The hypothesis is that this will simplify things for the model. For
example, on Windows, instead of generating:
```json
{"command": ["pwsh.exe", "-NoLogo", "-Command", "ls -Name"]}
```
The model could simply generate:
```json
{"command": "ls -Name"}
```
As part of this change, I extracted some logic out of `user_shell.rs` as
`Shell::derive_exec_args()` so that it can be reused in
`codex-rs/core/src/tools/handlers/shell.rs`. Note the original code
generated exec arg lists like:
```javascript
["bash", "-lc", command]
["zsh", "-lc", command]
["pwsh.exe", "-NoProfile", "-Command", command]
```
Using `-l` for Bash and Zsh, but then specifying `-NoProfile` for
PowerShell seemed inconsistent to me, so I changed this in the new
implementation while also adding a `use_login_shell: bool` option to
make this explicit. If we decide to add a `login: bool` to
`ShellCommandToolCallParams` like we have for unified exec:
https://github.com/openai/codex/blob/807e2c27f0a9f2e85c50e7e6df5533f0d9b853c7/codex-rs/core/src/tools/handlers/unified_exec.rs#L33-L34
Then this should make it straightforward to support.
feature: Add "!cmd" user shell execution
This change lets users run local shell commands directly from the TUI by
prefixing their input with ! (e.g. !ls). Output is truncated to keep the
exec cell usable, and Ctrl-C cleanly
interrupts long-running commands (e.g. !sleep 10000).
**Summary of changes**
- Route Op::RunUserShellCommand through a dedicated UserShellCommandTask
(core/src/tasks/user_shell.rs), keeping the task logic out of codex.rs.
- Reuse the existing tool router: the task constructs a ToolCall for the
local_shell tool and relies on ShellHandler, so no manual MCP tool
lookup is required.
- Emit exec lifecycle events (ExecCommandBegin/ExecCommandEnd) so the
TUI can show command metadata, live output, and exit status.
**End-to-end flow**
**TUI handling**
1. ChatWidget::submit_user_message (TUI) intercepts messages starting
with !.
2. Non-empty commands dispatch Op::RunUserShellCommand { command };
empty commands surface a help hint.
3. No UserInput items are created, so nothing is enqueued for the model.
**Core submission loop**
4. The submission loop routes the op to handlers::run_user_shell_command
(core/src/codex.rs).
5. A fresh TurnContext is created and Session::spawn_user_shell_command
enqueues UserShellCommandTask.
**Task execution**
6. UserShellCommandTask::run emits TaskStartedEvent, formats the
command, and prepares a ToolCall targeting local_shell.
7. ToolCallRuntime::handle_tool_call dispatches to ShellHandler.
**Shell tool runtime**
8. ShellHandler::run_exec_like launches the process via the unified exec
runtime, honoring sandbox and shell policies, and emits
ExecCommandBegin/End.
9. Stdout/stderr are captured for the UI, but the task does not turn the
resulting ToolOutput into a model response.
**Completion**
10. After ExecCommandEnd, the task finishes without an assistant
message; the session marks it complete and the exec cell displays the
final output.
**Conversation context**
- The command and its output never enter the conversation history or the
model prompt; the flow is local-only.
- Only exec/task events are emitted for UI rendering.
**Demo video**
https://github.com/user-attachments/assets/fcd114b0-4304-4448-a367-a04c43e0b996
It's pretty amazing we have gotten here without the ability for the
model to see image content from MCP tool calls.
This PR builds off of 4391 and fixes#4819. I would like @KKcorps to get
adequete credit here but I also want to get this fix in ASAP so I gave
him a week to update it and haven't gotten a response so I'm going to
take it across the finish line.
This test highlights how absured the current situation is. I asked the
model to read this image using the Chrome MCP
<img width="2378" height="674" alt="image"
src="https://github.com/user-attachments/assets/9ef52608-72a2-4423-9f5e-7ae36b2b56e0"
/>
After this change, it correctly outputs:
> Captured the page: image dhows a dark terminal-style UI labeled
`OpenAI Codex (v0.0.0)` with prompt `model: gpt-5-codex medium` and
working directory `/codex/codex-rs`
(and more)
Before this change, it said:
> Took the full-page screenshot you asked for. It shows a long,
horizontally repeating pattern of stylized people in orange, light-blue,
and mustard clothing, holding hands in alternating poses against a white
background. No text or other graphics-just rows of flat illustration
stretching off to the right.
Without this change, the Figma, Playwright, Chrome, and other visual MCP
servers are pretty much entirely useless.
I tested this change with the openai respones api as well as a third
party completions api
jif-oai