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"]
}
]
```
This PR configures Codex CLI so it can be built with
[Bazel](https://bazel.build) in addition to Cargo. The `.bazelrc`
includes configuration so that remote builds can be done using
[BuildBuddy](https://www.buildbuddy.io).
If you are familiar with Bazel, things should work as you expect, e.g.,
run `bazel test //... --keep-going` to run all the tests in the repo,
but we have also added some new aliases in the `justfile` for
convenience:
- `just bazel-test` to run tests locally
- `just bazel-remote-test` to run tests remotely (currently, the remote
build is for x86_64 Linux regardless of your host platform). Note we are
currently seeing the following test failures in the remote build, so we
still need to figure out what is happening here:
```
failures:
suite::compact::manual_compact_twice_preserves_latest_user_messages
suite::compact_resume_fork::compact_resume_after_second_compaction_preserves_history
suite::compact_resume_fork::compact_resume_and_fork_preserve_model_history_view
```
- `just build-for-release` to build release binaries for all
platforms/architectures remotely
To setup remote execution:
- [Create a buildbuddy account](https://app.buildbuddy.io/) (OpenAI
employees should also request org access at
https://openai.buildbuddy.io/join/ with their `@openai.com` email
address.)
- [Copy your API key](https://app.buildbuddy.io/docs/setup/) to
`~/.bazelrc` (add the line `build
--remote_header=x-buildbuddy-api-key=YOUR_KEY`)
- Use `--config=remote` in your `bazel` invocations (or add `common
--config=remote` to your `~/.bazelrc`, or use the `just` commands)
## CI
In terms of CI, this PR introduces `.github/workflows/bazel.yml`, which
uses Bazel to run the tests _locally_ on Mac and Linux GitHub runners
(we are working on supporting Windows, but that is not ready yet). Note
that the failures we are seeing in `just bazel-remote-test` do not occur
on these GitHub CI jobs, so everything in `.github/workflows/bazel.yml`
is green right now.
The `bazel.yml` uses extra config in `.github/workflows/ci.bazelrc` so
that macOS CI jobs build _remotely_ on Linux hosts (using the
`docker://docker.io/mbolin491/codex-bazel` Docker image declared in the
root `BUILD.bazel`) using cross-compilation to build the macOS
artifacts. Then these artifacts are downloaded locally to GitHub's macOS
runner so the tests can be executed natively. This is the relevant
config that enables this:
```
common:macos --config=remote
common:macos --strategy=remote
common:macos --strategy=TestRunner=darwin-sandbox,local
```
Because of the remote caching benefits we get from BuildBuddy, these new
CI jobs can be extremely fast! For example, consider these two jobs that
ran all the tests on Linux x86_64:
- Bazel 1m37s
https://github.com/openai/codex/actions/runs/20861063212/job/59940545209?pr=8875
- Cargo 9m20s
https://github.com/openai/codex/actions/runs/20861063192/job/59940559592?pr=8875
For now, we will continue to run both the Bazel and Cargo jobs for PRs,
but once we add support for Windows and running Clippy, we should be
able to cutover to using Bazel exclusively for PRs, which should still
speed things up considerably. We will probably continue to run the Cargo
jobs post-merge for commits that land on `main` as a sanity check.
Release builds will also continue to be done by Cargo for now.
Earlier attempt at this PR: https://github.com/openai/codex/pull/8832
Earlier attempt to add support for Buck2, now abandoned:
https://github.com/openai/codex/pull/8504
---------
Co-authored-by: David Zbarsky <dzbarsky@gmail.com>
Co-authored-by: Michael Bolin <mbolin@openai.com>
This seems to be necessary to get the Bazel builds on ARM Linux to go
green on https://github.com/openai/codex/pull/8875.
I don't feel great about timeout-whack-a-mole, but we're still learning
here...
I have seen this test flake out sometimes when running the macOS build
using Bazel in CI: https://github.com/openai/codex/pull/8875. Perhaps
Bazel runs with greater parallelism, inducing a heavier load, causing an
issue?
Historically we started with a CodexAuth that knew how to refresh it's
own tokens and then added AuthManager that did a different kind of
refresh (re-reading from disk).
I don't think it makes sense for both `CodexAuth` and `AuthManager` to
be mutable and contain behaviors.
Move all refresh logic into `AuthManager` and keep `CodexAuth` as a data
object.
Add metrics capabilities to Codex. The `README.md` is up to date.
This will not be merged with the metrics before this PR of course:
https://github.com/openai/codex/pull/8350
To support Bazelification in https://github.com/openai/codex/pull/8875,
this PR introduces a new `find_resource!` macro that we use in place of
our existing logic in tests that looks for resources relative to the
compile-time `CARGO_MANIFEST_DIR` env var.
To make this work, we plan to add the following to all `rust_library()`
and `rust_test()` Bazel rules in the project:
```
rustc_env = {
"BAZEL_PACKAGE": native.package_name(),
},
```
Our new `find_resource!` macro reads this value via
`option_env!("BAZEL_PACKAGE")` so that the Bazel package _of the code
using `find_resource!`_ is injected into the code expanded from the
macro. (If `find_resource()` were a function, then
`option_env!("BAZEL_PACKAGE")` would always be
`codex-rs/utils/cargo-bin`, which is not what we want.)
Note we only consider the `BAZEL_PACKAGE` value when the `RUNFILES_DIR`
environment variable is set at runtime, indicating that the test is
being run by Bazel. In this case, we have to concatenate the runtime
`RUNFILES_DIR` with the compile-time `BAZEL_PACKAGE` value to build the
path to the resource.
In testing this change, I discovered one funky edge case in
`codex-rs/exec-server/tests/common/lib.rs` where we have to _normalize_
(but not canonicalize!) the result from `find_resource!` because the
path contains a `common/..` component that does not exist on disk when
the test is run under Bazel, so it must be semantically normalized using
the [`path-absolutize`](https://crates.io/crates/path-absolutize) crate
before it is passed to `dotslash fetch`.
Because this new behavior may be non-obvious, this PR also updates
`AGENTS.md` to make humans/Codex aware that this API is preferred.
Adds a new feature
`enable_request_compression` that will compress using zstd requests to
the codex-backend. Currently only enabled for codex-backend so only enabled for openai providers when using chatgpt::auth even when the feature is enabled
Added a new info log line too for evaluating the compression ratio and
overhead off compressing before requesting. You can enable with
`RUST_LOG=$RUST_LOG,codex_client::transport=info`
```
2026-01-06T00:09:48.272113Z INFO codex_client::transport: Compressed request body with zstd pre_compression_bytes=28914 post_compression_bytes=11485 compression_duration_ms=0
```
We used to override truncation policy by comparing model info vs config
value in context manager. A better way to do it is to construct model
info using the config value
**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"
/>
- Merge ModelFamily into ModelInfo
- Remove logic for adding instructions to apply patch
- Add compaction limit and visible context window to `ModelInfo`
Set login=false for the shell tool in the timing-based parallelism test
so it does not depend on slow user login shells, making the test
deterministic without user-facing changes. This prevents occasional
flakes when running locally.
With `config.toml`:
```
model = "gpt-5.1-codex"
```
(where `gpt-5.1-codex` has `show_in_picker: false` in
[`model_presets.rs`](https://github.com/openai/codex/blob/main/codex-rs/core/src/models_manager/model_presets.rs);
this happens if the user hasn't used codex in a while so they didn't see
the popup before their model was changed to `show_in_picker: false`)
The upgrade picker used to not show (because `gpt-5.1-codex` was
filtered out of the model list in code). Now, the filtering is done
downstream in tui and app-server, so the model upgrade popup shows:
<img width="1503" height="227" alt="Screenshot 2026-01-06 at 5 04 37 PM"
src="https://github.com/user-attachments/assets/26144cc2-0b3f-4674-ac17-e476781ec548"
/>
Add `web_search_cached` feature to config. Enables `web_search` tool
with access only to cached/indexed results (see
[docs](https://platform.openai.com/docs/guides/tools-web-search#live-internet-access)).
This takes precedence over the existing `web_search_request`, which
continues to enable `web_search` over live results as it did before.
`web_search_cached` is disabled for review mode, as `web_search_request`
is.
Added an agent control plane that lets sessions spawn or message other
conversations via `AgentControl`.
`AgentBus` (core/src/agent/bus.rs) keeps track of the last known status
of a conversation.
ConversationManager now holds shared state behind an Arc so AgentControl
keeps only a weak back-reference, the goal is just to avoid explicit
cycle reference.
Follow-ups:
* Build a small tool in the TUI to be able to see every agent and send
manual message to each of them
* Handle approval requests in this TUI
* Add tools to spawn/communicate between agents (see related design)
* Define agent types
Adds an optional `justification` parameter to the `prefix_rule()`
execpolicy DSL so policy authors can attach human-readable rationale to
a rule. That justification is propagated through parsing/matching and
can be surfaced to the model (or approval UI) when a command is blocked
or requires approval.
When a command is rejected (or gated behind approval) due to policy, a
generic message makes it hard for the model/user to understand what went
wrong and what to do instead. Allowing policy authors to supply a short
justification improves debuggability and helps guide the model toward
compliant alternatives.
Example:
```python
prefix_rule(
pattern = ["git", "push"],
decision = "forbidden",
justification = "pushing is blocked in this repo",
)
```
If Codex tried to run `git push origin main`, now the failure would
include:
```
`git push origin main` rejected: pushing is blocked in this repo
```
whereas previously, all it was told was:
```
execpolicy forbids this command
```
What changed
- Added `outputSchema` support to the app-server APIs, mirroring `codex
exec --output-schema` behavior.
- V1 `sendUserTurn` now accepts `outputSchema` and constrains the final
assistant message for that turn.
- V2 `turn/start` now accepts `outputSchema` and constrains the final
assistant message for that turn (explicitly per-turn only).
Core behavior
- `Op::UserTurn` already supported `final_output_json_schema`; now V1
`sendUserTurn` forwards `outputSchema` into that field.
- `Op::UserInput` now carries `final_output_json_schema` for per-turn
settings updates; core maps it into
`SessionSettingsUpdate.final_output_json_schema` so it applies to the
created turn context.
- V2 `turn/start` does NOT persist the schema via `OverrideTurnContext`
(it’s applied only for the current turn). Other overrides
(cwd/model/etc) keep their existing persistent behavior.
API / docs
- `codex-rs/app-server-protocol/src/protocol/v1.rs`: add `output_schema:
Option<serde_json::Value>` to `SendUserTurnParams` (serialized as
`outputSchema`).
- `codex-rs/app-server-protocol/src/protocol/v2.rs`: add `output_schema:
Option<JsonValue>` to `TurnStartParams` (serialized as `outputSchema`).
- `codex-rs/app-server/README.md`: document `outputSchema` for
`turn/start` and clarify it applies only to the current turn.
- `codex-rs/docs/codex_mcp_interface.md`: document `outputSchema` for v1
`sendUserTurn` and v2 `turn/start`.
Tests added/updated
- New app-server integration tests asserting `outputSchema` is forwarded
into outbound `/responses` requests as `text.format`:
- `codex-rs/app-server/tests/suite/output_schema.rs`
- `codex-rs/app-server/tests/suite/v2/output_schema.rs`
- Added per-turn semantics tests (schema does not leak to the next
turn):
- `send_user_turn_output_schema_is_per_turn_v1`
- `turn_start_output_schema_is_per_turn_v2`
- Added protocol wire-compat tests for the merged op:
- serialize omits `final_output_json_schema` when `None`
- deserialize works when field is missing
- serialize includes `final_output_json_schema` when `Some(schema)`
Call site updates (high level)
- Updated all `Op::UserInput { .. }` constructions to include
`final_output_json_schema`:
- `codex-rs/app-server/src/codex_message_processor.rs`
- `codex-rs/core/src/codex_delegate.rs`
- `codex-rs/mcp-server/src/codex_tool_runner.rs`
- `codex-rs/tui/src/chatwidget.rs`
- `codex-rs/tui2/src/chatwidget.rs`
- plus impacted core tests.
Validation
- `just fmt`
- `cargo test -p codex-core`
- `cargo test -p codex-app-server`
- `cargo test -p codex-mcp-server`
- `cargo test -p codex-tui`
- `cargo test -p codex-tui2`
- `cargo test -p codex-protocol`
- `cargo clippy --all-features --tests --profile dev --fix -- -D
warnings`
Fixes /review base-branch prompt resolution to use the session/turn cwd
(respecting runtime cwd overrides) so merge-base/diff guidance is
computed from the intended repo; adds a regression test for cwd
overrides; tested with cargo test -p codex-core --test all
review_uses_overridden_cwd_for_base_branch_merge_base.
last token count in context manager is initialized to 0. Gets populated
only on events from server.
This PR populates it on resume so we can decide if we need to compact or
not.
This PR introduces a `codex-utils-cargo-bin` utility crate that
wraps/replaces our use of `assert_cmd::Command` and
`escargot::CargoBuild`.
As you can infer from the introduction of `buck_project_root()` in this
PR, I am attempting to make it possible to build Codex under
[Buck2](https://buck2.build) as well as `cargo`. With Buck2, I hope to
achieve faster incremental local builds (largely due to Buck2's
[dice](https://buck2.build/docs/insights_and_knowledge/modern_dice/)
build strategy, as well as benefits from its local build daemon) as well
as faster CI builds if we invest in remote execution and caching.
See
https://buck2.build/docs/getting_started/what_is_buck2/#why-use-buck2-key-advantages
for more details about the performance advantages of Buck2.
Buck2 enforces stronger requirements in terms of build and test
isolation. It discourages assumptions about absolute paths (which is key
to enabling remote execution). Because the `CARGO_BIN_EXE_*` environment
variables that Cargo provides are absolute paths (which
`assert_cmd::Command` reads), this is a problem for Buck2, which is why
we need this `codex-utils-cargo-bin` utility.
My WIP-Buck2 setup sets the `CARGO_BIN_EXE_*` environment variables
passed to a `rust_test()` build rule as relative paths.
`codex-utils-cargo-bin` will resolve these values to absolute paths,
when necessary.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/8496).
* #8498
* __->__ #8496
This isn't very useful parameter.
logic:
```
if model puts `**` in their reasoning, trim it and visualize the header.
if couldn't trim: don't render
if model doesn't support: don't render
```
We can simplify to:
```
if could trim, visualize header.
if not, don't render
```
### Motivation
- Persist richer per-turn configuration in rollouts so resumed/forked
sessions and tooling can reason about the exact instruction inputs and
output constraints used for a turn.
### Description
- Extend `TurnContextItem` to include optional `base_instructions`,
`user_instructions`, and `developer_instructions`.
- Record the optional `final_output_json_schema` associated with a turn.
- Add an optional `truncation_policy` to `TurnContextItem` and populate
it when writing turn-context rollout items.
- Introduce a protocol-level `TruncationPolicy` representation and
convert from core truncation policy when recording.
### Testing
- `cargo test -p codex-protocol` (pass)
## Summary
Adds a FeatureFlag to enforce UTF8 encoding in powershell, particularly
Windows Powershell v5. This should help address issues like #7290.
Notably, this PR does not include the ability to parse `apply_patch`
invocations within UTF8 shell commands (calls to the freeform tool
should not be impacted). I am leaving this out of scope for now. We
should address before this feature becomes Stable, but those cases are
not the default behavior at this time so we're okay for experimentation
phase. We should continue cleaning up the `apply_patch::invocation`
logic and then can handle it more cleanly.
## Testing
- [x] Adds additional testing
### Summary
With codesigning on Mac, Windows and Linux, we should be able to safely
remove `features.rmcp_client` and `use_experimental_use_rmcp_client`
check from the codebase now.
# External (non-OpenAI) Pull Request Requirements
Before opening this Pull Request, please read the dedicated
"Contributing" markdown file or your PR may be closed:
https://github.com/openai/codex/blob/main/docs/contributing.md
If your PR conforms to our contribution guidelines, replace this text
with a detailed and high quality description of your changes.
Include a link to a bug report or enhancement request.
This adds support for `allowed_sandbox_modes` in `requirements.toml` and
provides legacy support for constraining sandbox modes in
`managed_config.toml`. This is converted to `Constrained<SandboxPolicy>`
in `ConfigRequirements` and applied to `Config` such that constraints
are enforced throughout the harness.
Note that, because `managed_config.toml` is deprecated, we do not add
support for the new `external-sandbox` variant recently introduced in
https://github.com/openai/codex/pull/8290. As noted, that variant is not
supported in `config.toml` today, but can be configured programmatically
via app server.
Fixes#8214 by removing the '--staged' flag from the undo git restore
command. This ensures that while the working tree is reverted to the
snapshot state, the user's staged changes (index) are preserved,
preventing data loss. Also adds a regression test.
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.
# External (non-OpenAI) Pull Request Requirements
Before opening this Pull Request, please read the dedicated
"Contributing" markdown file or your PR may be closed:
https://github.com/openai/codex/blob/main/docs/contributing.md
If your PR conforms to our contribution guidelines, replace this text
with a detailed and high quality description of your changes.
Include a link to a bug report or enhancement request.
# External (non-OpenAI) Pull Request Requirements
Before opening this Pull Request, please read the dedicated
"Contributing" markdown file or your PR may be closed:
https://github.com/openai/codex/blob/main/docs/contributing.md
If your PR conforms to our contribution guidelines, replace this text
with a detailed and high quality description of your changes.
Include a link to a bug report or enhancement request.
