Addresses #15527
Problem: Nested `codex exec` commands could source a shell snapshot that
re-exported the parent `CODEX_THREAD_ID`, so commands inside the nested
session were attributed to the wrong thread.
Solution: Reapply the live command env's `CODEX_THREAD_ID` after
sourcing the snapshot.
Addresses #15532
Problem: Nested read-only `apply_patch` rejections report in-project
files as outside the project.
Solution: Choose the rejection message based on sandbox mode so
read-only sessions report a read-only-specific reason, and add focused
safety coverage.
Problem: The multi-agent followup interrupt test polled history before
interrupt cleanup and mailbox wakeup were guaranteed to settle, which
made it flaky under CI scheduling variance.
Solution: Wait for the child turn's `TurnAborted(Interrupted)` event
before asserting that the redirected assistant envelope is recorded and
no plain user message is left behind.
## Summary
- reduce public module visibility across Rust crates, preferring private
or crate-private modules with explicit crate-root public exports
- update external call sites and tests to use the intended public crate
APIs instead of reaching through module trees
- add the module visibility guideline to AGENTS.md
## Validation
- `cargo check --workspace --all-targets --message-format=short` passed
before the final fix/format pass
- `just fix` completed successfully
- `just fmt` completed successfully
- `git diff --check` passed
# 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.
## Summary
- make AGENTS.md discovery and loading fully FS-aware and remove the
non-FS discover helper
- migrate remote-aware codex-core tests to use TestEnv workspace setup
instead of syncing a local workspace copy
- add AGENTS.md corner-case coverage, including directory fallbacks and
remote-aware integration coverage
## Testing
- cargo test -p codex-core project_doc -- --nocapture
- cargo test -p codex-core hierarchical_agents -- --nocapture
- cargo test -p codex-core agents_md -- --nocapture
- cargo test -p codex-tui status -- --nocapture
- cargo test -p codex-tui-app-server status -- --nocapture
- just fix
- just fmt
- just bazel-lock-update
- just bazel-lock-check
- just argument-comment-lint
- remote Linux executor tests in progress via scripts/test-remote-env.sh
## Summary
This adds `experimental_network.danger_full_access_denylist_only` for
orgs that want yolo / danger-full-access sessions to keep full network
access while still enforcing centrally managed deny rules.
When the flag is true and the session sandbox is `danger-full-access`,
the network proxy starts with:
- domain allowlist set to `*`
- managed domain `deny` entries enforced
- upstream proxy use allowed
- all Unix sockets allowed
- local/private binding allowed
Caveat: the denylist is best effort only. In yolo / danger-full-access
mode, Codex or the model can use an allowed socket or other
local/private network path to bypass the proxy denylist, so this should
not be treated as a hard security boundary.
The flag is intentionally scoped to `SandboxPolicy::DangerFullAccess`.
Read-only and workspace-write modes keep the existing managed/user
allowlist, denylist, Unix socket, and local-binding behavior. This does
not enable the non-loopback proxy listener setting; that still requires
its own explicit config.
This also threads the new field through config requirements parsing,
app-server protocol/schema output, config API mapping, and the TUI debug
config output.
## How to use
Add the flag under `[experimental_network]` in the network policy config
that is delivered to Codex. The setting is not under `[permissions]`.
```toml
[experimental_network]
enabled = true
danger_full_access_denylist_only = true
[experimental_network.domains]
"blocked.example.com" = "deny"
"*.blocked.example.com" = "deny"
```
With that configuration, yolo / danger-full-access sessions get broad
network access except for the managed denied domains above. The denylist
remains a best-effort proxy policy because the session may still use
allowed sockets to bypass it. Other sandbox modes do not get the
wildcard domain allowlist or the socket/local-binding relaxations from
this flag.
## Verification
- `cargo test -p codex-config network_requirements`
- `cargo test -p codex-core network_proxy_spec`
- `cargo test -p codex-app-server map_requirements_toml_to_api`
- `cargo test -p codex-tui debug_config_output`
- `cargo test -p codex-app-server-protocol`
- `just write-app-server-schema`
- `just fmt`
- `just fix -p codex-config -p codex-core -p codex-app-server-protocol
-p codex-app-server -p codex-tui`
- `just fix -p codex-core -p codex-config`
- `git diff --check`
- `cargo clean`
## Summary
- make `CODEX_EXEC_SERVER_URL=none` map to an explicit disabled
environment mode instead of inferring from a missing URL
- expose environment capabilities (`exec_enabled`, `filesystem_enabled`)
so tool building can gate behavior explicitly and future
multi-environment work has a clearer seam
- suppress env-backed tools when the relevant capability is unavailable,
including exec tools, `js_repl`, `apply_patch`, `list_dir`, and
`view_image`
- keep handler/runtime backstops so disabled environments still reject
execution if a tool path somehow bypasses registration
## Testing
- `just fmt`
- `cargo test -p codex-exec-server`
- `cargo test -p codex-tools
disabled_environment_omits_environment_backed_tools`
- `cargo test -p codex-tools
environment_capabilities_gate_exec_and_filesystem_tools_independently`
- remote devbox Bazel build via `codex-applied-devbox`:
`//codex-rs/cli:cli`
Extract a shared helper that builds AuthManager from Config and applies
the forced ChatGPT workspace override in one place.
Create the shared AuthManager at MessageProcessor call sites so that
upcoming new transport's initialization can reuse the same handle, and
keep only external auth refresher wiring inside `MessageProcessor`.
Remove the now-unused `AuthManager::shared_with_external_auth` helper.
## Description
Add requirements.toml support for `allowed_approvals_reviewers =
["user", "guardian_subagent"]`, so admins can now restrict the use of
guardian mode.
Note: If a user sets a reviewer that isn’t allowed by requirements.toml,
config loading falls back to the first allowed reviewer and emits a
startup warning.
The table below describes the possible admin controls.
| Admin intent | `requirements.toml` | User `config.toml` | End result |
|---|---|---|---|
| Leave Guardian optional | omit `allowed_approvals_reviewers` or set
`["user", "guardian_subagent"]` | user chooses `approvals_reviewer =
"user"` or `"guardian_subagent"` | Guardian off for `user`, on for
`guardian_subagent` + `approval_policy = "on-request"` |
| Force Guardian off | `allowed_approvals_reviewers = ["user"]` | any
user value | Effective reviewer is `user`; Guardian off |
| Force Guardian on | `allowed_approvals_reviewers =
["guardian_subagent"]` and usually `allowed_approval_policies =
["on-request"]` | any user reviewer value; user should also have
`approval_policy = "on-request"` unless policy is forced | Effective
reviewer is `guardian_subagent`; Guardian on when effective approval
policy is `on-request` |
| Allow both, but default to manual if user does nothing |
`allowed_approvals_reviewers = ["user", "guardian_subagent"]` | omit
`approvals_reviewer` | Effective reviewer is `user`; Guardian off |
| Allow both, and user explicitly opts into Guardian |
`allowed_approvals_reviewers = ["user", "guardian_subagent"]` |
`approvals_reviewer = "guardian_subagent"` and `approval_policy =
"on-request"` | Guardian on |
| Invalid admin config | `allowed_approvals_reviewers = []` | anything |
Config load error |
Problem: `rejects_escalated_permissions_when_policy_not_on_request`
retried a real shell command after asserting the escalation rejection,
so Windows CI could fail on command startup timing instead of approval
behavior.
Solution: Keep the rejection assertion, verify no turn permissions were
granted, and assert through exec-policy evaluation that the same command
would be allowed without escalation instead of timing a subprocess.
Send pending mailbox mail after completed reasoning or commentary items
so follow-up requests can pick it up mid-turn.
---------
Co-authored-by: Codex <noreply@openai.com>
This adds an `include_environment_context` config/profile flag that
defaults on, and guards both initial injection and later environment
updates to allow skipping injection of `<environment_context>`.
This PR adds root and profile config switches to omit the generated
`<permissions instructions>` and `<apps_instructions>` prompt blocks
while keeping both enabled by default, and it gates both the initial
developer-context injection and later permissions diff injection so
turning the permissions block off stays effective across turn-context
overrides.
Also added a prompt debug tool that can be used as `codex debug
prompt-input "hello"` and dumps the constructed items list.
The `OPENAI_BASE_URL` environment variable has been a significant
support issue, so we decided to deprecate it in favor of an
`openai_base_url` config key. We've had the deprecation warning in place
for about a month, so users have had time to migrate to the new
mechanism. This PR removes support for `OPENAI_BASE_URL` entirely.
## Why
Extracted from [#16528](https://github.com/openai/codex/pull/16528) so
the Windows Bazel app-server test failures can be reviewed independently
from the rest of that PR.
This PR targets:
-
`suite::v2::thread_shell_command::thread_shell_command_runs_as_standalone_turn_and_persists_history`
-
`suite::v2::thread_start::thread_start_with_elevated_sandbox_trusts_project_and_followup_loads_project_config`
-
`suite::v2::thread_start::thread_start_with_nested_git_cwd_trusts_repo_root`
There were two Windows-specific assumptions baked into those tests and
the underlying trust lookup:
- project trust keys were persisted and looked up using raw path
strings, but Bazel's Windows test environment can surface canonicalized
paths with `\\?\` / UNC prefixes or normalized symlink/junction targets,
so follow-up `thread/start` requests no longer matched the project entry
that had just been written
- `item/commandExecution/outputDelta` assertions compared exact trailing
line endings even though shell output chunk boundaries and CRLF handling
can differ on Windows, and Bazel made that timing-sensitive mismatch
visible
There was also one behavior bug separate from the assertion cleanup:
`thread/start` decided whether to persist trust from the final resolved
sandbox policy, but on Windows an explicit `workspace-write` request may
be downgraded to `read-only`. That incorrectly skipped writing trust
even though the request had asked to elevate the project, so the new
logic also keys off the requested sandbox mode.
## What
- Canonicalize project trust keys when persisting/loading `[projects]`
entries, while still accepting legacy raw keys for existing configs.
- Persist project trust when `thread/start` explicitly requests
`workspace-write` or `danger-full-access`, even if the resolved policy
is later downgraded on Windows.
- Make the Windows app-server tests compare persisted trust paths and
command output deltas in a path/newline-normalized way.
## Verification
- Existing app-server v2 tests cover the three failing Windows Bazel
cases above.
- Keep only parent system/developer/user messages plus assistant
final-answer messages in forked child history.
- Strip parent tool/reasoning items and remove the unmatched synthetic
spawn output.
## Why
We were seeing failures in the following tests as part of trying to get
all the tests running under Bazel on Windows in CI
(https://github.com/openai/codex/pull/16528):
```
suite::shell_command::unicode_output::with_login
suite::shell_command::unicode_output::without_login
```
Certainly `PATHEXT` should have been included in the extra `CORE_VARS`
list, so we fix that up here, but also take things a step further for
now by forcibly ensuring it is set on Windows in the return value of
`create_env()`. Once we get the Windows Bazel build working reliably
(i.e., after #16528 is merged), we should come back to this and confirm
we can remove the special case in `create_env()`.
## What
- Split core env inheritance into `COMMON_CORE_VARS` plus
platform-specific allowlists for Windows and Unix in
[`exec_env.rs`](https://github.com/openai/codex/blob/1b55c88fbf585b32cd553cb9d02ec817f2ad6ebc/codex-rs/core/src/exec_env.rs#L45-L81).
- Preserve `PATHEXT`, `USERNAME`, and `USERPROFILE` on Windows, and
`HOME` / locale vars on Unix.
- Backfill a default `PATHEXT` in `create_env()` on Windows if the
parent env does not provide one, so child process launch still works in
stripped-down Bazel environments.
- Extend the Windows exec-env test to assert mixed-case `PathExt`
survives case-insensitive core filtering, and document why the
shell-command Unicode test goes through a child process.
## Verification
- `cargo test -p codex-core exec_env::tests`
Stacked on #16508.
This removes the temporary `codex-core` / `codex-login` re-export shims
from the ownership split and rewrites callsites to import directly from
`codex-model-provider-info`, `codex-models-manager`, `codex-api`,
`codex-protocol`, `codex-feedback`, and `codex-response-debug-context`.
No behavior change intended; this is the mechanical import cleanup layer
split out from the ownership move.
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
- split `models-manager` out of `core` and add `ModelsManagerConfig`
plus `Config::to_models_manager_config()` so model metadata paths stop
depending on `core::Config`
- move login-owned/auth-owned code out of `core` into `codex-login`,
move model provider config into `codex-model-provider-info`, move API
bridge mapping into `codex-api`, move protocol-owned types/impls into
`codex-protocol`, and move response debug helpers into a dedicated
`response-debug-context` crate
- move feedback tag emission into `codex-feedback`, relocate tests to
the crates that now own the code, and keep broad temporary re-exports so
this PR avoids a giant import-only rewrite
## Major moves and decisions
- created `codex-models-manager` as the owner for model
cache/catalog/config/model info logic, including the new
`ModelsManagerConfig` struct
- created `codex-model-provider-info` as the owner for provider config
parsing/defaults and kept temporary `codex-login`/`codex-core`
re-exports for old import paths
- moved `api_bridge` error mapping + `CoreAuthProvider` into
`codex-api`, while `codex-login::api_bridge` temporarily re-exports
those symbols and keeps the `auth_provider_from_auth` wrapper
- moved `auth_env_telemetry` and `provider_auth` ownership to
`codex-login`
- moved `CodexErr` ownership to `codex-protocol::error`, plus
`StreamOutput`, `bytes_to_string_smart`, and network policy helpers to
protocol-owned modules
- created `codex-response-debug-context` for
`extract_response_debug_context`, `telemetry_transport_error_message`,
and related response-debug plumbing instead of leaving that behavior in
`core`
- moved `FeedbackRequestTags`, `emit_feedback_request_tags`, and
`emit_feedback_request_tags_with_auth_env` to `codex-feedback`
- deferred removal of temporary re-exports and the mechanical import
rewrites to a stacked follow-up PR so this PR stays reviewable
## Test moves
- moved auth refresh coverage from `core/tests/suite/auth_refresh.rs` to
`login/tests/suite/auth_refresh.rs`
- moved text encoding coverage from
`core/tests/suite/text_encoding_fix.rs` to
`protocol/src/exec_output_tests.rs`
- moved model info override coverage from
`core/tests/suite/model_info_overrides.rs` to
`models-manager/src/model_info_overrides_tests.rs`
---------
Co-authored-by: Codex <noreply@openai.com>
## Why
The Windows `ProviderAuthScript` test helpers do not need PowerShell.
Running them through `cmd.exe` is enough to emit the next fixture token
and rotate `tokens.txt`, and it avoids a PowerShell-specific dependency
in these tests.
## What changed
- Replaced the Windows `print-token.ps1` fixtures with `print-token.cmd`
in `codex-rs/core/src/models_manager/manager_tests.rs` and
`codex-rs/login/src/auth/auth_tests.rs`.
- Switched the failing external-auth helper in
`codex-rs/login/src/auth/auth_tests.rs` from `powershell.exe -Command
'exit 1'` to `cmd.exe /d /s /c 'exit /b 1'`.
- Updated Windows timeout comments so they no longer call out PowerShell
specifically.
## Verification
- `cargo test -p codex-login`
- `cargo test -p codex-core` (fails in unrelated
`core/src/config/config_tests.rs` assertions in this checkout)
## Why
This continues the compile-time cleanup from #16630. `SessionTask`
implementations are monomorphized, but `Session` stores the task behind
a `dyn` boundary so it can drive and abort heterogenous turn tasks
uniformly. That means we can move the `#[async_trait]` expansion off the
implementation trait, keep a small boxed adapter only at the storage
boundary, and preserve the existing task lifecycle semantics while
reducing the amount of generated async-trait glue in `codex-core`.
One measurement caveat showed up while exploring this: a warm
incremental benchmark based on `touch core/src/tasks/mod.rs && cargo
check -p codex-core --lib` was basically flat, but that was the wrong
benchmark for this change. Using package-clean `codex-core` rebuilds,
like #16630, shows the real win.
Relevant pre-change code:
- [`SessionTask` with
`#[async_trait]`](https://github.com/openai/codex/blob/3c7f013f9735e67796c70d95f75f436b7f97e3ec/codex-rs/core/src/tasks/mod.rs#L129-L182)
- [`RunningTask` storing `Arc<dyn
SessionTask>`](https://github.com/openai/codex/blob/3c7f013f9735e67796c70d95f75f436b7f97e3ec/codex-rs/core/src/state/turn.rs#L69-L77)
## What changed
- Switched `SessionTask::{run, abort}` to native RPITIT futures with
explicit `Send` bounds.
- Added a private `AnySessionTask` adapter that boxes those futures only
at the `Arc<dyn ...>` storage boundary.
- Updated `RunningTask` to store `Arc<dyn AnySessionTask>` and removed
`#[async_trait]` from the concrete task impls plus test-only
`SessionTask` impls.
## Timing
Benchmarked package-clean `codex-core` rebuilds with dependencies left
warm:
```shell
cargo check -p codex-core --lib >/dev/null
cargo clean -p codex-core >/dev/null
/usr/bin/time -p cargo +nightly rustc -p codex-core --lib -- \
-Z time-passes \
-Z time-passes-format=json >/dev/null
```
| revision | rustc `total` | process `real` | `generate_crate_metadata`
| `MIR_borrow_checking` | `monomorphization_collector_graph_walk` |
| --- | ---: | ---: | ---: | ---: | ---: |
| parent `3c7f013f9735` | 67.21s | 67.71s | 24.61s | 23.43s | 22.43s |
| this PR `2cafd783ac22` | 35.08s | 35.60s | 8.01s | 7.25s | 7.15s |
| delta | -47.8% | -47.4% | -67.5% | -69.1% | -68.1% |
For completeness, the warm touched-file benchmark stayed flat (`1.96s`
parent vs `1.97s` this PR), which is why that benchmark should not be
used to evaluate this refactor.
## Verification
- Ran `cargo test -p codex-core`; this change compiled and task-related
tests passed before hitting the same unrelated 5
`config::tests::*guardian*` failures already present on the parent
stack.
Recently, I merged a number of PRs to increase startup timeouts for
scripts that ran under PowerShell, but in the failure for
`suite::codex_tool::test_shell_command_approval_triggers_elicitation`, I
found this in the error logs when running on Bazel with BuildBuddy:
```
[mcp stderr] 2026-04-02T19:54:10.758951Z ERROR codex_core::tools::router: error=Exit code: 1
[mcp stderr] Wall time: 0.2 seconds
[mcp stderr] Output:
[mcp stderr] 'New-Item' is not recognized as an internal or external command,
[mcp stderr] operable program or batch file.
[mcp stderr]
```
This error implies that the command was run under `cmd.exe` instead of
`pwsh.exe`. Under GitHub Actions, I suspect that the `%PATH%` that is
passed to our Bazel builder is scrubbed such that our tests cannot find
PowerShell where GitHub installs it. Having these explicit fallback
paths should help.
While we could enable these only for tests, I don't see any harm in
keeping them in production, as well.
Ahmed Ibrahim ·
Eric Traut ·
jif-oai ·
pakrym-oai ·
starr-openai ·
Won Park ·
viyatb-oai ·
Matthew Zeng ·
rhan-oai ·
xl-openai ·
Ruslan Nigmatullin ·
Owen Lin ·
Thibault Sottiaux ·
Michael Bolin ·