## Summary
Fixes#11678 by removing the Windows-specific
`PASTE_BURST_CHAR_INTERVAL` override. Windows now uses the same `8ms`
paste-burst character interval as macOS and Linux, which removes the
extra per-character hold that made fast typing and key repeat feel
delayed on Windows.
The paste-burst heuristic itself is unchanged, and the Windows-specific
active idle timeout remains in place. This PR only restores the shared
character-to-character burst threshold that decides whether adjacent
plain character events are part of a paste.
## Motivation
PR #9348 raised the Windows character interval from `8ms` to `30ms` to
protect the multiline paste behavior tracked in #2137, where pasted
newlines could be interpreted as submits in Windows terminals. That
fixed the paste failure, but it also made ordinary typing visibly laggy
because the TUI waits briefly before flushing a single typed character
while it checks whether a paste burst is forming.
The deployed behavior here is to remove that Windows-only delay and
return to the cross-platform threshold. Manual Windows validation of the
critical VS Code integrated terminal path shows multiline paste still
works with the final `8ms` value, including testing on VS Code
`1.107.0`.
## Testing
- `cargo test -p codex-tui`
- Manual Windows validation in VS Code integrated PowerShell with the
final `8ms` interval
## Why
Bazel CI was not actually exercising some sharded Rust integration-test
targets on macOS. The `rules_rust` sharding wrapper expects a symlink
runfiles tree, but this repo runs Bazel with `--noenable_runfiles`. In
that configuration the wrapper could fail to find the generated test
binary, produce an empty test list, and exit successfully. That made
targets such as `//codex-rs/core:core-all-test` look green even when
Cargo CI could still catch failures in the same Rust tests.
The coverage gap appears to have been introduced by
[#18082](https://github.com/openai/codex/pull/18082), which enabled
rules_rust native sharding on `//codex-rs/core:core-all-test` and the
other large Rust test labels. The manifest-runfiles setup itself
predates that change in
[#10098](https://github.com/openai/codex/pull/10098), but #18082 is
where the affected integration tests started running through the
incompatible rules_rust sharding wrapper.
[#18913](https://github.com/openai/codex/pull/18913) fixed the same
class of issue for wrapped unit-test shards, but integration-test shards
were still going through the rules_rust wrapper until this PR.
We still do not have the V8/code-mode pieces stable under the Bazel CI
cross-compile setup, so this keeps those tests out of Bazel while
restoring coverage for the rest of the sharded Rust integration suites.
Cargo CI remains responsible for V8/code-mode coverage for now.
This change did uncover a real failing core test on `main`:
`approved_folder_write_request_permissions_unblocks_later_apply_patch`.
That fix is split into
[#21060](https://github.com/openai/codex/pull/21060), which enables the
`apply_patch` tool in the test, teaches the aggregate core test binary
to dispatch the sandboxed filesystem helper, canonicalizes the macOS
temp patch target, and isolates the core test harness from managed
local/enterprise config. Keeping that fix separate lets this PR stay
focused on restoring Bazel coverage while documenting the first failure
it exposed.
## What changed
- Build sharded Rust integration tests as manual `*-bin` binaries and
run them through the existing manifest-aware `workspace_root_test`
launcher.
- Keep Bazel sharding on the launcher target so Rust test cases are
still distributed by stable test-name hashing.
- Configure Bazel CI to skip Rust tests whose names contain
`suite::code_mode::`.
- Exclude the standalone `codex-rs/code-mode` and `codex-rs/v8-poc`
unit-test targets from `bazel.yml`.
## Verification
- `bazel query --output=build //codex-rs/core:core-all-test` now shows
`workspace_root_test` wrapping `//codex-rs/core:core-all-test-bin`.
- `bazel test --test_output=all --nocache_test_results
--test_sharding_strategy=disabled //codex-rs/core:core-all-test
--test_filter=suite::request_permissions_tool::approved_folder_write_request_permissions_unblocks_later_apply_patch`
runs the actual Rust test body and passes.
- `bazel test --test_output=errors --nocache_test_results
--test_env=CODEX_BAZEL_TEST_SKIP_FILTERS=suite::code_mode::
//codex-rs/core:core-all-test` runs the sharded target with code-mode
skipped and passes overall locally, with one flaky attempt retried by
the existing `flaky = True` setting.
## Why
Fixes#21046.
Codex TUI 0.128.0 can show Backspace/Delete-related editor shortcuts in
`/keymap`, but Windows-style modified Backspace/Delete events were still
dropped by the composer because the default editor keymap did not
include those modified special-key variants. On Windows/CMD this meant
`Shift+Backspace` and `Shift+Delete` did not fall through to normal
character deletion, and `Ctrl+Backspace` / `Ctrl+Delete` did not perform
the word deletion users expect from Windows text inputs.
## What Changed
- Added default editor bindings for `shift-backspace` and `shift-delete`
so shifted delete keys keep normal grapheme deletion behavior.
- Added default editor bindings for `ctrl-backspace`,
`ctrl-shift-backspace`, `ctrl-delete`, and `ctrl-shift-delete` so
Windows-style word deletion works when terminals preserve those
modifiers.
- Added regression coverage for the resolved default keymap and textarea
behavior.
## How to Test
1. Start Codex in the TUI on Windows CMD or another terminal that
reports modified Backspace/Delete keys distinctly.
2. Type `hello world` in the composer.
3. Press `Ctrl+Backspace`; confirm `world` is removed and `hello `
remains.
4. Type `world` again, move the cursor before it, then press
`Ctrl+Delete`; confirm the next word is removed.
5. Type a few characters and press `Shift+Backspace` and `Shift+Delete`;
confirm they delete one character in the expected direction instead of
doing nothing.
6. Open `/keymap`, inspect the Editor deletion actions, and confirm the
modified Backspace/Delete aliases are visible as configurable defaults.
Targeted tests:
- `cargo test -p codex-tui keymap::tests`
- `cargo test -p codex-tui bottom_pane::textarea::tests`
- `cargo test -p codex-tui keymap_setup::tests`
## Why
The Bazel test coverage change exposed
`approved_folder_write_request_permissions_unblocks_later_apply_patch`,
and `rust-ci-full.yml` showed the same test failing on `main` on macOS.
There were two separate classes of problems here.
### Clean CI failure
The test emits an `apply_patch` tool call, but its config did not enable
the `apply_patch` tool, so the mocked response completed without an
`apply-patch-call` output. After enabling the tool, the same path also
needs the aggregate `codex-core` test binary to dispatch
`--codex-run-as-fs-helper`; sandboxed `apply_patch` uses that helper
under macOS Seatbelt.
The test now also canonicalizes the temporary patch target before
building the patch payload so the path matches normalized grants on
macOS, where `/var` paths often normalize to `/private/var`.
### Local/enterprise config isolation
The core test harness now builds its default test config with managed
config disabled, so host-managed enterprise config cannot alter these
tests. The request-permissions turns in this test also explicitly use
the user reviewer path, keeping the assertions focused on
`request_permissions` behavior rather than reviewer defaults from the
host.
## What Changed
- Enable `apply_patch` in
`approved_folder_write_request_permissions_unblocks_later_apply_patch`.
- Teach the core integration test binary to dispatch
`CODEX_FS_HELPER_ARG1`, matching the existing apply-patch and
linux-sandbox dispatch paths.
- Canonicalize the tempdir-backed patch target before creating the
patch.
- Ignore managed config in default core test configs and explicitly pin
this test to `ApprovalsReviewer::User`.
## Verification
Run outside the Codex app sandbox because these macOS tests
intentionally spawn Seatbelt:
- `cargo test -p codex-core
approved_folder_write_request_permissions_unblocks_later_apply_patch`
- `cargo test -p codex-core
approved_folder_write_request_permissions_unblocks_later_exec_without_sandbox_args`
## Why
Feedback reports do not currently surface a direct pointer to the last
model call, so investigations may require searching through many
requests in a session to find the bad response. Preserve the last
model-side IDs at response-stream time so immediate feedback reports
carry that breadcrumb.
## What changed
- Record `last_model_request_id` when a Responses stream exposes an
upstream request ID.
- Record `last_model_response_id` when the model response completes.
- Add unit coverage for the emitted feedback tags.
## Verification
- `cargo test -p codex-core
client::tests::response_stream_records_last_model_feedback_ids`
## Why
MCP servers can provide `instructions` that explain what their tools are
for. Directly exposed MCP namespaces already use those instructions when
a connector description is not available, but deferred `tool_search`
results did not preserve that fallback. The direct path falls back from
connector metadata to server instructions, while the deferred path only
carried `connector_description` and otherwise fell back to generic
namespace text.
That meant a plain MCP server could provide useful model-facing guidance
and still appear as `Tools in the X namespace.` whenever it was
discovered lazily through `tool_search`.
## What changed
- Store one model-facing `namespace_description` on `ToolInfo`, using
connector descriptions for connector-backed tools and server
instructions for plain MCP servers.
- Thread that namespace description through the `tool_search` source
list, search indexing, and returned namespace metadata.
- Add an end-to-end regression test for deferred non-app MCP search
results exposing server instructions as the namespace description.
## Verification
- `cargo test -p codex-tools
search_tool_description_lists_each_mcp_source_once --lib`
- `cargo test -p codex-core --test all
tool_search_uses_non_app_mcp_server_instructions_as_namespace_description`
## Summary
- normalize terminal-emitted C0 control characters through configurable
editor keymaps, covering raw control-key fallbacks like
Shift+Enter-as-LF in terminals from #20555 and #20898, plus part of the
modified-Enter behavior in #20580
- add default-unbound keymap actions for toggling Fast mode and killing
the current composer line, giving #20698 users a configurable zsh-style
Ctrl+U option without changing the existing default Ctrl+U behavior
- wire the new actions through gated /keymap picker entries, schema
generation, and snapshot coverage
Fixes#20555.
Fixes#20898.
## Testing
- just write-config-schema
- just fmt
- cargo test -p codex-config
- cargo test -p codex-tui keymap::tests
- cargo test -p codex-tui bottom_pane::textarea::tests
- cargo test -p codex-tui keymap_setup::tests
- cargo insta pending-snapshots
- just fix -p codex-tui
- git diff --check
- just argument-comment-lint
## Why?
The Codex App already exposes branch and PR context in its
branch-details UI. This brings the same context into the CLI footer as
opt-in statusline items, so users can choose the extra signal without
making the default footer busier.
## What?
Add optional `pull-request-number` and `branch-changes` items to the
configurable TUI status line.
- `pull-request-number` shows the open PR for the current checkout and
renders as a clickable terminal hyperlink when OSC 8 links are
supported.
- `branch-changes` shows committed additions/deletions against the
repository default branch, or `No changes` when the branch has no
committed diff.
<img width="1257" height="261" alt="CleanShot 2026-05-03 at 20 44 15"
src="https://github.com/user-attachments/assets/10b4380b-c3e9-4729-9ee1-3f742068fa47"
/>
## Architecture
This follows the same client/app-server split as the Codex App: the TUI
owns presentation, caching, and optional rendering, while
workspace-sensitive `git` and `gh` discovery runs through app-server.
The new TUI-local `workspace_command` layer sends bounded,
non-interactive `command/exec` requests to the active app-server. That
makes the implementation remote-friendly: the TUI does not decide
whether commands run in an embedded local workspace or a remote
workspace, and it does not bypass app-server sandbox or permission
policy.
The branch summary logic stays internal to `codex-tui` because this PR
only needs TUI statusline behavior. The command boundary is still
isolated behind `WorkspaceCommandExecutor`, so the lookup code can be
lifted or reused later without changing statusline rendering.
## How?
- Add a TUI `WorkspaceCommandExecutor` abstraction backed by app-server
`command/exec`.
- Add branch summary probes for:
- current branch name,
- open PR metadata,
- committed branch diff stats against the default branch.
- Prefer remote-tracking default branch refs for diff stats, avoiding
stale or absent local `main` branches.
- Resolve PRs with `gh pr view` first, then fall back to
commit-associated PR lookup across parent/fork repos.
- Add `/statusline` picker entries, preview values, rendering, and OSC 8
clickable PR links.
- Keep all probes best-effort so missing `git`, missing `gh`, auth
failures, or non-git directories hide optional items instead of
surfacing footer errors.
## Validation
- `cargo test -p codex-tui branch_summary -- --nocapture`
- Snapshot coverage for the `/statusline` preview/setup rendering paths
- Hyperlink rendering coverage for clickable PR statusline cells
## Why
SQLite state was still being opened from consumer paths, including lazy
`OnceCell`-backed thread-store call sites. That let one process
construct multiple state DB connections for the same Codex home, which
makes SQLite lock contention and `database is locked` failures much
easier to hit.
State DB lifetime should be chosen by main-like entrypoints and tests,
then passed through explicitly. Consumers should use the supplied
`Option<StateDbHandle>` or `StateDbHandle` and keep their existing
filesystem fallback or error behavior when no handle is available.
The startup path also needs to keep the rollout crate in charge of
SQLite state initialization. Opening `codex_state::StateRuntime`
directly bypasses rollout metadata backfill, so entrypoints should
initialize through `codex_rollout::state_db` and receive a handle only
after required rollout backfills have completed.
## What Changed
- Initialize the state DB in main-like entrypoints for CLI, TUI,
app-server, exec, MCP server, and the thread-manager sample.
- Pass `Option<StateDbHandle>` through `ThreadManager`,
`LocalThreadStore`, app-server processors, TUI app wiring, rollout
listing/recording, personality migration, shell snapshot cleanup,
session-name lookup, and memory/device-key consumers.
- Remove the lazy local state DB wrapper from the thread store so
non-test consumers use only the supplied handle or their existing
fallback path.
- Make `codex_rollout::state_db::init` the local state startup path: it
opens/migrates SQLite, runs rollout metadata backfill when needed, waits
for concurrent backfill workers up to a bounded timeout, verifies
completion, and then returns the initialized handle.
- Keep optional/non-owning SQLite helpers, such as remote TUI local
reads, as open-only paths that do not run startup backfill.
- Switch app-server startup from direct
`codex_state::StateRuntime::init` to the rollout state initializer so
app-server cannot skip rollout backfill.
- Collapse split rollout lookup/list APIs so callers use the normal
methods with an optional state handle instead of `_with_state_db`
variants.
- Restore `getConversationSummary(ThreadId)` to delegate through
`ThreadStore::read_thread` instead of a LocalThreadStore-specific
rollout path special case.
- Keep DB-backed rollout path lookup keyed on the DB row and file
existence, without imposing the filesystem filename convention on
existing DB rows.
- Verify readable DB-backed rollout paths against `session_meta.id`
before returning them, so a stale SQLite row that points at another
thread's JSONL falls back to filesystem search and read-repairs the DB
row.
- Keep `debug prompt-input` filesystem-only so a one-off debug command
does not initialize or backfill SQLite state just to print prompt input.
- Keep goal-session test Codex homes alive only in the goal-specific
helper, rather than leaking tempdirs from the shared session test
helper.
- Update tests and call sites to pass explicit state handles where DB
behavior is expected and explicit `None` where filesystem-only behavior
is intended.
## Validation
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo check -p
codex-rollout -p codex-thread-store -p codex-app-server -p codex-core -p
codex-tui -p codex-exec -p codex-cli --tests`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p
codex-rollout state_db_`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p
codex-rollout find_thread_path`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p
codex-rollout find_thread_path -- --nocapture`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p
codex-rollout try_init_ -- --nocapture`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p
codex-rollout`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo clippy -p
codex-rollout --lib -- -D warnings`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p
codex-thread-store
read_thread_falls_back_when_sqlite_path_points_to_another_thread --
--nocapture`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p
codex-thread-store`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p codex-core
shell_snapshot`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p codex-core
--test all personality_migration`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p codex-core
--test all rollout_list_find`
- `RUST_MIN_STACK=8388608 CODEX_SKIP_VENDORED_BWRAP=1
CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p codex-core
--test all rollout_list_find::find_prefers_sqlite_path_by_id --
--nocapture`
- `RUST_MIN_STACK=8388608 CODEX_SKIP_VENDORED_BWRAP=1
CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p codex-core
--test all rollout_list_find -- --nocapture`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p codex-core
interrupt_accounts_active_goal_before_pausing`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p
codex-app-server get_auth_status -- --test-threads=1`
- `CODEX_SKIP_VENDORED_BWRAP=1
CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo test -p
codex-app-server --lib`
- `CODEX_SKIP_VENDORED_BWRAP=1
CARGO_TARGET_DIR=/tmp/codex-target-state-db cargo check -p codex-rollout
-p codex-app-server --tests`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db just fix -p codex-rollout
-p codex-thread-store -p codex-core -p codex-app-server -p codex-tui -p
codex-exec -p codex-cli`
- `CODEX_SKIP_VENDORED_BWRAP=1
CARGO_TARGET_DIR=/tmp/codex-target-state-db just fix -p codex-rollout -p
codex-app-server`
- `CARGO_TARGET_DIR=/tmp/codex-target-state-db just fix -p
codex-rollout`
- `CODEX_SKIP_VENDORED_BWRAP=1
CARGO_TARGET_DIR=/tmp/codex-target-state-db just fix -p codex-core`
- `just argument-comment-lint -p codex-core`
- `just argument-comment-lint -p codex-rollout`
Focused coverage added in `codex-rollout`:
- `recorder::tests::state_db_init_backfills_before_returning` verifies
the rollout metadata row exists before startup init returns.
- `state_db::tests::try_init_waits_for_concurrent_startup_backfill`
verifies startup waits for another worker to finish backfill instead of
disabling the handle for the process.
-
`state_db::tests::try_init_times_out_waiting_for_stuck_startup_backfill`
verifies startup does not hang indefinitely on a stuck backfill lease.
-
`tests::find_thread_path_accepts_existing_state_db_path_without_canonical_filename`
verifies DB-backed lookup accepts valid existing rollout paths even when
the filename does not include the thread UUID.
-
`tests::find_thread_path_falls_back_when_db_path_points_to_another_thread`
verifies DB-backed lookup ignores a stale row whose existing path
belongs to another thread and read-repairs the row after filesystem
fallback.
Focused coverage updated in `codex-core`:
- `rollout_list_find::find_prefers_sqlite_path_by_id` now uses a
DB-preferred rollout file with matching `session_meta.id`, so it still
verifies that valid SQLite paths win without depending on stale/empty
rollout contents.
`cargo test -p codex-app-server thread_list_respects_search_term_filter
-- --test-threads=1 --nocapture` was attempted locally but timed out
waiting for the app-server test harness `initialize` response before
reaching the changed thread-list code path.
`bazel test //codex-rs/thread-store:thread-store-unit-tests
--test_output=errors` was attempted locally after the thread-store fix,
but this container failed before target analysis while fetching `v8+`
through BuildBuddy/direct GitHub. The equivalent local crate coverage,
including `cargo test -p codex-thread-store`, passes.
A plain local `cargo check -p codex-rollout -p codex-app-server --tests`
also requires system `libcap.pc` for `codex-linux-sandbox`; the
follow-up app-server check above used `CODEX_SKIP_VENDORED_BWRAP=1` in
this container.
## Why
This stack adds configured exec-server environments, including
environments reached over stdio. Before client-side stdio transports or
config can use that path, the exec-server binary itself needs a
first-class stdio listen mode so it can speak the same JSON-RPC protocol
over stdin/stdout that it already speaks over websockets.
**Stack position:** this is PR 1 of 5. It is the server-side transport
foundation for the stack.
## What Changed
- Accept `stdio` and `stdio://` for `codex exec-server --listen`.
- Promote the existing stdio `JsonRpcConnection` helper from test-only
code into normal exec-server transport code.
- Add parse coverage for stdio listen URLs while preserving the existing
websocket default.
## Stack
- **1. This PR:** https://github.com/openai/codex/pull/20663 - Add stdio
exec-server listener
- 2. https://github.com/openai/codex/pull/20664 - Add stdio exec-server
client transport
- 3. https://github.com/openai/codex/pull/20665 - Make environment
providers own default selection
- 4. https://github.com/openai/codex/pull/20666 - Add CODEX_HOME
environments TOML provider
- 5. https://github.com/openai/codex/pull/20667 - Load configured
environments from CODEX_HOME
Split from original draft: https://github.com/openai/codex/pull/20508
## Validation
Not run locally; this was split out of the original draft stack.
---------
Co-authored-by: Codex <noreply@openai.com>
## Why
On Windows, background terminals could stay visible after their shell
process had already exited. The elevated runner waits for the PTY output
reader to reach EOF before it sends the final exit message, but the
ConPTY helper was reducing ownership down to raw handles too early. That
left the pseudoconsole's borrowed pipe handles alive past teardown, so
EOF never propagated and the session stayed `running`.
## What changed
- change `utils/pty/src/win/conpty.rs` to hand off owned ConPTY
resources instead of leaking only raw handles
- make `windows-sandbox-rs/src/conpty/mod.rs` keep the pseudoconsole
owner and the backing pipe handles together until teardown
- update the elevated runner and the legacy unified-exec backend to keep
that `ConptyInstance` alive, take only the specific pipe handles they
need, and drop the owner at teardown instead of trying to close a
detached pseudoconsole handle later
## Testing
- desktop app in `Auto-review`: 11 x `cmd /c "ping -n 3 google.com"` all
exited cleanly and did not accumulate in the UI
- desktop app in `Auto-review`: 5 x `cmd /c "ping -n 30 google.com"`
appeared in the UI and drained back out on their own
## Why
This is a prep PR in the multi-environment process-tool stack. It
separates ownership/config cleanup from the behavior change that teaches
process tools to route by selected environment, so the follow-up PR can
focus on model-facing `environment_id` behavior.
## Stack
1. https://github.com/openai/codex/pull/20646 - `EnvironmentContext`
rendering for selected environments
2. https://github.com/openai/codex/pull/20669 - selected-environment
ownership and tool config prep (this PR)
3. https://github.com/openai/codex/pull/20647 - process-tool
`environment_id` routing
## What Changed
- keep the resolved turn environment list wrapped in
`ResolvedTurnEnvironments` through `TurnContext` instead of unwrapping
it back to a raw `Vec`
- add `TurnContext::resolve_path_against` so cwd-relative path
resolution has one shared helper
- replace the old tool config boolean with `ToolEnvironmentMode::{None,
Single, Multiple}`
## Testing
- Tests not run locally; this prep refactor is covered by GitHub CI for
the stack.
Co-authored-by: Codex <noreply@openai.com>
## Why
The TUI currently exposes overlapping command names for the same
permissions flow: `/permissions` and the older `/approvals` alias. It
also uses `/autoreview` for the manual retry flow, even though the
action users take there is approving one denied auto-review request.
This change makes the command surface consistent with the hard rebrand:
- `/permissions` is the only command for permission settings.
- `/approve` is the command for approving a recent auto-review denial.
## What changed
- Removed the legacy `/approvals` slash command and its dispatch path.
- Kept `/permissions` as the single permissions command shown and
accepted by the TUI.
- Renamed the auto-review denial command from `/autoreview` to
`/approve`.
- Updated nearby comments so they refer to `/permissions` rather than
the retired `/approvals` name.
## Verification
- Updated the slash-command unit test to assert that `AutoReview` now
renders and parses as `approve`.
## Why
We constantly get bug reports about keys not being recognized by Codex
when the terminal is not handling the key press. Running `/keymap debug`
or `/keymap` and going to the Debug tab, we can allow the user to either
understand that the key being pressed is not being recognized or to
check what it's being recognized as and report or reassign that key.
| Menu | Inspector | Hint |
|---|---|---|
| <img width="1369" height="796" alt="CleanShot 2026-05-02 at 12 57 12"
src="https://github.com/user-attachments/assets/512b6faa-344e-4aee-9c00-b4bdc633a662"
/> | <img width="1261" height="754" alt="CleanShot 2026-05-02 at 12 56
36"
src="https://github.com/user-attachments/assets/a6ddae7d-e174-4ee4-893f-e6bec4fff4ab"
/> | <img width="1369" height="796" alt="CleanShot 2026-05-02 at 12 57
30"
src="https://github.com/user-attachments/assets/db507784-f40a-4cff-ac23-a61d9703769b"
/> |
## Summary
- add a Debug tab to `/keymap` and support `/keymap debug` for direct
access
- show what key Codex receives, the config key representation, raw event
details, and matching actions
- add a progressive missing-key hint that escalates after a few seconds
with no detected keypress
## Validation
- `just fmt`
- `cargo test -p codex-tui keymap_setup::tests::debug_view`
- `cargo test -p codex-tui keymap_setup::tests`
- `cargo test -p codex-tui slash_keymap`
- `cargo test -p codex-tui` (unit tests passed; integration test
`suite::model_availability_nux::resume_startup_does_not_consume_model_availability_nux_count`
failed locally by itself with `codex resume` exiting 1 and terminal
probe escape output)
- `just fix -p codex-tui`
- `just argument-comment-lint`
- `cargo insta pending-snapshots`
- `git diff --check`
## Why
Codex `0.128` started using `--perms` in more routine Linux sandbox
construction when protected workspace metadata mounts landed in #19852.
Upstream bubblewrap added `--perms` in `v0.5.0`, so system `bwrap`
versions older than that, including the `v0.4.0` and `v0.4.1` family, do
not support the flag. The launcher still selected those binaries as long
as they existed on `PATH`.
That means affected hosts can fail every sandboxed command up front
with:
```text
bwrap: Unknown option --perms
```
The reports in #20590 and duplicate #20623 match that compatibility gap;
#20623 explicitly shows system bubblewrap `0.4.0`.
## What changed
- Replace the single `--argv0` probe with a small system-bwrap
capability probe in `codex-rs/linux-sandbox/src/launcher.rs`.
- Continue using the old-system `--argv0` compatibility path when
needed, but only select a system `bwrap` if it also advertises
`--perms`.
- Fall back to the vendored `bwrap` when the system binary is too old
for the flags Codex now requires.
- Add regression coverage for the old-system-bwrap case so binaries
without `--perms` stay on the vendored path.
## Verification
- Added `falls_back_to_vendored_when_system_bwrap_lacks_perms` to cover
the reported compatibility gap.
- Ran `cargo test -p codex-linux-sandbox` and `cargo clippy -p
codex-linux-sandbox --tests` locally. On macOS, the crate builds but its
Linux-only tests are cfg-gated out, so the new regression test still
needs Linux CI or a Linux devbox run for real execution coverage.
## Related issues
- Fixes#20590
- Duplicate report: #20623
## Why
Whenever we return a thread's history (turns and items) over app-server,
always return the limited form as specified by the rollout policy
`EventPersistenceMode::Limited`, even if the thread was previously
started with `EventPersistenceMode::Extended`.
We're finding it is quite unscalable to be returning the extended
history, so let's apply the same filtering logic of the rollout policy
when we load and return the thread's history.
## What Changed
- Reuse the rollout persistence policy when reconstructing app-server
`ThreadItem` history so only `EventPersistenceMode::Limited` rollout
items are replayed into API turns.
- Route `thread/read`, `thread/resume`, `thread/fork`,
`thread/turns/list`, and rollback responses through the same filtered
app-server history projection.
- Keep live active turns intact when composing a response for a
currently running thread.
- Update command execution coverage so persisted extended command events
are excluded from returned history for `thread/read`, `thread/fork`, and
`thread/turns/list`.
## Test Plan
- `cargo test -p codex-app-server limited`
- `cargo test -p codex-app-server thread_shell_command`
- `cargo test -p codex-app-server thread_read`
- `cargo test -p codex-app-server thread_rollback`
- `cargo test -p codex-app-server thread_fork`
- `cargo test -p codex-app-server-protocol`
## Summary
- Treat `approval_mode = "approve"` as skip-review across all permission
modes.
- Remove the mode-specific split in the MCP auto-approval gate so
approved tools bypass review consistently.
- Expand regression coverage in the shared MCP helper and the core
tool-call flow.
## Testing
- `just fmt`
- `cargo test -p codex-mcp`
- `cargo test -p codex-core
approve_mode_skips_arc_and_guardian_in_every_permission_mode`
- `git diff --check`
- Full `cargo test -p codex-core` was also attempted, but the suite hit
an unrelated pre-existing stack overflow in an existing multi-agent test
- [x] Persist a special type of MCP tool calls for triggering MCP App,
this type of mcp tool calls has 'mcpAppResourceUri` set. These events
are needed so that the Codex App can correctly render the MCP App after
resume.
## Why
`ModelClientSession` and `compact_conversation_history()` were still
rebuilding the same `ResponsesApiRequest` fields separately. That
duplication makes it easy for normal `/responses` turns and compact
requests to drift when request-shape changes land later, which is
exactly the kind of cache-affecting divergence we want to avoid.
This follow-up keeps the scope small by extracting the shared
request-construction logic into one helper and using it from both paths.
## What changed
- move `ResponsesApiRequest` construction into a shared
`ModelClient::build_responses_request(...)` helper in
`core/src/client.rs`
- update the normal `/responses` streaming path to call that helper
instead of the old `ModelClientSession`-local implementation
- update `compact_conversation_history()` to derive its compact payload
from the same helper so `model`, `instructions`, `input`, `tools`,
`parallel_tool_calls`, `reasoning`, and `text` stay aligned with normal
request building
- add a unit test covering the shared helper's prompt cache key,
installation metadata, and `service_tier` behavior
## Verification
- `cargo test -p codex-core
build_responses_request_sets_shared_cache_and_metadata_fields`
- `cargo test -p codex-core --test all
remote_compact_v2_reuses_context_compaction_for_followups`
## Docs
No docs update needed.
## Why
The local memories MCP backend only rejected symlinks after resolving
the final path. That left room for scoped requests like
`skills/secret.md` to walk through a symlinked ancestor directory and
escape the configured memories root.
This change also makes missing scoped paths fail explicitly instead of
looking like an empty `list` / `search` result or a `NotFile` read
error.
## What Changed
- walk each scoped path component in
`LocalMemoriesBackend::resolve_scoped_path` and reject symlinked
ancestors before accessing the target
- reject scoped paths that traverse through a non-directory intermediate
component
- add a `NotFound` backend error for missing `read`, `list`, and
`search` paths and map it through the MCP server error conversion
- add coverage for missing paths and symlinked ancestor directories in
`codex-rs/memories/mcp/src/local_tests.rs`
## Testing
- added unit coverage in `codex-rs/memories/mcp/src/local_tests.rs` for
missing paths and symlinked ancestor directories across `read`, `list`,
and `search`
## Why
The app-server request path had grown around a large
`CodexMessageProcessor` plus separate API wrapper/helper modules. That
made the dependency graph hard to see and forced unrelated request
families to share broad processor state.
This PR makes the split mechanical and command-prefix oriented so
request families own only the dependencies they use.
## What changed
- Replaced `CodexMessageProcessor` with command-prefix request
processors under `app-server/src/request_processors/`.
- Removed the old config, device-key, external-agent-config, and fs API
wrapper files by moving their API handling into processors.
- Split apps, plugins, marketplace, catalog, account, MCP, command exec,
fs, git, feedback, thread, turn, thread goals, and Windows sandbox
handling into dedicated processors.
- Kept shared lifecycle, summary conversion, token usage replay, and
shared error mapping only where multiple processors use them; single-use
helpers were inlined into their owning processor.
- Removed the fallback processor path and moved processor tests to
`_tests` files.
## Validation
- `cargo test -p codex-app-server`
- `cargo check -p codex-app-server`
- `just fix -p codex-app-server`
## Summary
Early adopters of the `/goal` feature have provided feedback that they
expect a goal they explicitly paused to remain paused when they resume a
thread. Previously, resuming a thread would reactivate a paused goal.
This PR keeps persisted goal status unchanged during thread resume. This
honors the user feedback while also simplifying the core goal logic.
Rather than have the core logic automatically resume a paused goal, that
responsibility is transferred to the client. The TUI now detects a
resumed thread with a paused goal and asks the user whether to `Resume
goal` or `Leave paused`. The prompt appears only for quiet resume flows,
so users who resume with an immediate prompt are not interrupted.
<img width="544" height="111" alt="image"
src="https://github.com/user-attachments/assets/0ac9de1c-6ee6-47ba-b223-c03c8eb4c192"
/>
## Why
Returning from a `/side` conversation restores the parent thread by
replaying its snapshot into the TUI. For very long parent threads,
replaying every transcript row can take noticeable time even though most
rows immediately scroll out of terminal history.
## What Changed
- Buffer thread-switch replay for parent restores when terminal resize
reflow is enabled.
- Reuse the existing resize-reflow tail renderer so only the retained
transcript tail is written back to scrollback when a row cap is
configured.
## Summary
Early adopters of the `/goal` feature have provided feedback that they
expect a goal they explicitly paused to remain paused when they resume a
thread. Previously, resuming a thread would reactivate a paused goal.
This PR keeps persisted goal status unchanged during thread resume. This
honors the user feedback while also simplifying the core goal logic.
Rather than have the core logic automatically resume a paused goal, that
responsibility is transferred to the client. The TUI now detects a
resumed thread with a paused goal and asks the user whether to `Resume
goal` or `Leave paused`. The prompt appears only for quiet resume flows,
so users who resume with an immediate prompt are not interrupted.
<img width="544" height="111" alt="image"
src="https://github.com/user-attachments/assets/0ac9de1c-6ee6-47ba-b223-c03c8eb4c192"
/>
## Why
The memories MCP `search` tool only accepts a single substring today,
which makes it hard for clients to express combined queries or explain
why a line matched. This change adds the richer search shape needed for
the next client iteration while keeping the legacy single-`query` call
working.
## What changed
- accept either the legacy `query` field or a new `queries` array, plus
`match_mode: any|all`
- teach the local memories backend to evaluate multi-query line matches
and return `matched_queries` on each hit
- update the MCP input/output schema and add coverage for parser
behavior, ordering, pagination, case sensitivity, and match modes
## Testing
- added unit coverage in `memories/mcp/src/local_tests.rs` and
`memories/mcp/src/server.rs`
## Why
The paginated memories MCP `search` tool still returned only the
matching line text, which made it harder for clients to present useful
search results or decide whether they needed to follow up with a
separate `read` call. Adding a small amount of surrounding context makes
individual hits much more usable while keeping the search response
deterministic and line-addressable.
## What changed
- add an optional `context_lines` search argument and thread it through
the MCP server into the local memories backend
- change search matches to return the matched `line_number` plus a
`start_line_number` and multi-line `content` block for the requested
context window
- update the search tool schema and description to document the new
request/response shape
- extend the local backend tests to cover zero-context matches,
contextual results, pagination, and invalid cursors that point past the
end of the result set
## Testing
- Added targeted unit coverage in `memories/mcp/src/local_tests.rs`
- GitHub Actions are running for the branch
---------
Co-authored-by: Codex <noreply@openai.com>
## Why
The memories MCP `search` tool previously stopped once it hit
`max_results`, so callers could tell there were more matches via
`truncated` but had no way to fetch the rest of the result set. That
made large searches awkward for clients that need to keep paging through
a stable, deterministic view of the matches.
## What changed
- add an optional `cursor` field to `SearchMemoriesRequest` / tool input
and return `next_cursor` in `SearchMemoriesResponse`
- update the MCP schemas and tool wiring so clients can request
subsequent pages explicitly
- change the local memories backend to collect and sort the full scoped
match list, then slice the requested page and reject invalid cursors
- add unit coverage for paginated search results and invalid cursor
handling in `memories/mcp/src/local_tests.rs`
## Testing
- Added targeted unit coverage in `memories/mcp/src/local_tests.rs`
- GitHub Actions are running for the branch
## Why
The memories MCP `list` tool should behave like a directory listing, not
a recursive tree walk. Recursive results make pagination harder to
reason about, return unexpectedly deep paths for scoped requests, and no
longer match the intended tool contract.
## What Changed
- Changed the local memories backend so `list` returns only the
immediate children of the requested path.
- Preserved file-scoped requests by returning the file itself, and
missing paths by returning an empty result.
- Updated cursor handling to paginate over the shallow sibling set and
reject cursors past the available results.
- Updated the MCP tool description to say it lists immediate files and
directories under a path.
- Reworked the local backend tests to cover shallow top-level listing,
shallow scoped listing, sibling ordering, and pagination.
## Testing
- `cargo test -p codex-memories-mcp`
## Why
Large memories trees do not fit well into a single MCP `list` response.
This change makes the memories MCP server page `list` results so callers
can continue walking the tree without overfetching or relying on
ambiguous truncation.
## What changed
- add an optional `cursor` input to the memories MCP `list` API and
return `next_cursor` alongside `truncated` in the response
- paginate recursive local-memory traversal while preserving
lexicographic path order across directories
- reject malformed and out-of-range cursors as invalid MCP requests
- update the server/schema wiring and add coverage for pagination,
ordering, and cursor validation in `memories/mcp/src/local_tests.rs`
## Testing
- `cargo test -p codex-memories-mcp`
## Why
The memories MCP `read` tool already supports `line_offset`, but it
cannot return a bounded line range. That makes it awkward to page
through large memory files or request a small slice without relying on
token truncation.
## What changed
- add an optional `max_lines` parameter to the memories MCP `read` tool
schema and request parsing
- cap local backend reads to the requested number of lines before token
truncation
- treat `max_lines = 0` as an invalid request and surface it as
`invalid_params`
- add backend tests for bounded reads and invalid line request
validation
## Testing
- added coverage in `memories/mcp/src/local_tests.rs` for `max_lines`
reads and invalid `max_lines` / `line_offset` requests
## Why
Memory clients sometimes need to continue reading a file from a known
line instead of starting over from the top. Adding a line offset to the
`read` MCP keeps that resume logic simple and avoids re-reading
already-consumed content.
## What changed
- Added an optional `line_offset` argument to the memory `read` tool,
defaulting to `1`.
- Read content starting at the requested 1-indexed line before token
truncation, and return `start_line_number` in the response.
- Treat invalid offsets as invalid params errors and cover the new
behavior in `codex-rs/memories/mcp/src/local_tests.rs`.
## Testing
- Added unit tests for reading from a non-default starting line.
- Added unit tests for rejecting `0` and past-end line offsets.
## Why
This adds the `remote_compaction_v2` client path so remote compaction
can run through the normal Responses stream and install a
`context_compaction` item that trigger a compaction.
The goal is to migrate some of the compaction logic on the client side
We keeps the v2 transport behind a feature flag while letting follow-up
requests reuse the compacted context instead of falling back to the
legacy compaction item shape.
## What changed
- add `ResponseItem::ContextCompaction` and refresh the generated
app-server / schema / TypeScript fixtures that expose response items on
the wire
- add `core/src/compact_remote_v2.rs` to send compaction through the
standard streamed Responses client, require exactly one
`context_compaction` output item, and install that item into compacted
history
- route manual compact and auto-compaction through the v2 path when
`remote_compaction_v2` is enabled, while keeping the existing remote
compaction path as the fallback
- preserve the new item type across history retention, follow-up request
construction, telemetry, rollout persistence, and rollout-trace
normalization
- add targeted coverage for the feature flag, `context_compaction`
serialization, rollout-trace normalization, and remote-compaction
follow-up behavior
## Verification
- added protocol tests for `context_compaction`
serialization/deserialization in `protocol/src/models.rs`
- added rollout-trace coverage for `context_compaction` normalization in
`rollout-trace/src/reducer/conversation_tests.rs`
- added remote compaction integration coverage for v2 follow-up reuse
and mixed compaction output streams in
`core/tests/suite/compact_remote.rs`
---------
Co-authored-by: Codex <noreply@openai.com>
## Why
`multi_agent_v2` usage hints sometimes need to reference resolved config
values such as the effective thread limit. Those values only exist after
config layering, defaulting, and feature materialization, so the raw
TOML alone was not enough to render them.
## What changed
- allow
`features.multi_agent_v2.{usage_hint_text,root_agent_usage_hint_text,subagent_usage_hint_text}`
to use `{{ ... }}` placeholders backed by the materialized effective
config
- fail config loading with a targeted error when a referenced
placeholder does not exist or does not resolve to a scalar value
- move resolved-config materialization into a shared helper so config
interpolation and config-lock export/replay both serialize the same
resolved feature, memory, and agent settings
## Example
```
[features.multi_agent_v2]
enabled = true
usage_hint_text = "lorem {{ features.multi_agent_v2.max_concurrent_threads_per_session }} ipsum"
```
gets rendered as
```
"description": String("... \lorem 4 ipsum"),
```
## Why
`McpToolCall` was still an app-server item synthesized from deprecated
legacy begin/end events. Recent item migrations moved this ownership
into core `TurnItem`s, so MCP tool calls now follow the same canonical
lifecycle and leave legacy events as compatibility fanout.
Keeping the core item close to the v2 `ThreadItem::McpToolCall` shape
also avoids spreading MCP result semantics across app-server conversion
code. Core now owns whether a completed call is `completed` or `failed`,
and whether the payload is a tool result or an error.
## What changed
- Added core `TurnItem::McpToolCall` with flattened `server`, `tool`,
`arguments`, `status`, `result`, and `error` fields.
- Updated MCP tool call emitters, including MCP resource tools, to emit
`ItemStarted`/`ItemCompleted` around directly constructed core MCP
items.
- Updated app-server v2 conversion to project the core MCP item into
`ThreadItem::McpToolCall` without deriving status or splitting `Result`
locally.
- Ignored live deprecated MCP legacy fanout in app-server v2 to avoid
duplicate item notifications, while keeping thread history replay on the
legacy event path.
## Verification
- `cargo test -p codex-protocol`
- `cargo test -p codex-app-server-protocol`
- `cargo test -p codex-core --lib mcp_tool_call`
- `cargo check -p codex-app-server`
- `cargo test -p codex-app-server
mcp_tool_call_completion_notification_contains_truncated_large_result`
## Why
App-server request handling had response sending spread across many
individual handlers, which made it harder to see which requests return
payloads, which methods send their own delayed response, and which
branches emit notifications after a response.
## What changed
- Centralized normal `ClientResponsePayload` sending in the dispatch
path.
- Kept explicit-response methods explicit where they need custom
ordering or delayed delivery.
- Removed forward-only handler wrappers and immediate `async { ...
}.await` bodies where they were not needed.
- Moved branch-specific post-response notifications into the branches
that own the response ordering.
- Replaced unreachable delegated request-family error arms with explicit
`unreachable!` cases.
## Verification
- `cargo check -p codex-app-server`
- `cargo test -p codex-app-server thread_goal`
- `just fix -p codex-app-server`
## Summary
We should not check local-only docs or planning specs into this
repository. Keeping those files here duplicates the canonical Codex
documentation surface and makes transient implementation notes look like
supported docs.
This PR removes the local-only docs/spec files from `docs/` and trims
`docs/config.md` back to links for the maintained configuration
documentation on developers.openai.com.
## Why
The automated issue labeler needs more precise area labels for newly
opened GitHub issues so triage can distinguish new Codex app and agent
feature surfaces without falling back to broad labels.
## What Changed
- Added labeler prompt entries for `computer-use`, `browser`, `memory`,
`imagen`, `remote`, `performance`, `automations`, and `pets` in
`.github/workflows/issue-labeler.yml`.
- Updated the agent-area guidance so `memory` is used for agentic memory
storage/retrieval and `performance` is used for slow behavior, high
memory utilization, and leaks.
- Expanded the fallback `agent` guidance so Codex prefers the new
specific labels when applicable.
## Verification
- Parsed `.github/workflows/issue-labeler.yml` with `yq e '.'`.
- Ran `git diff --check` for the workflow change.
## Why
We saw Responses websocket sessions recover only after a long quiet
period when the server had already logged the websocket as disconnected.
The normal connect path is already bounded by
`websocket_connect_timeout_ms`, but the first request send on an
established websocket reused only the receive-side idle timeout after
the write completed. If the socket write/pump stalls, the client can sit
in `ws_stream.send(...)` without reaching the existing receive timeout.
Tool suggest still misfires when model needs tool_search, updating the
prompts to further disambiguate it:
- [x] rename it from `tool_suggest` to `request_plugin_install`
- [x] rephrase "suggestion" to "install" in the tool descriptions.
- [x] disambiguate "the tool" vs "the plugin/connector".
Tested with the Codex App and verified it still works.
## Summary
Bound TUI startup terminal response probes so unsupported terminals
cannot stall startup for multiple seconds.
This replaces the Unix startup uses of crossterm's blocking response
probes with short `/dev/tty` probes that use nonblocking reads and
`poll` with a 100ms timeout. It covers the initial cursor-position
query, keyboard enhancement support detection, and OSC 10/11
default-color detection. The default-color probe uses one shared
deadline for foreground and background instead of allowing two
independent full waits.
The diagnostic mode/trace env vars from the investigation branch are
intentionally not included. The shipped behavior is simply bounded
probing by default, while non-Unix keeps the existing crossterm fallback
path.
## Details
- Add a private `terminal_probe` module for bounded Unix terminal probes
and response parsers.
- Let `custom_terminal::Terminal` accept a caller-provided initial
cursor position so startup can compute it before constructing the
terminal.
- Use bounded cursor, keyboard enhancement, and default-color probes on
Unix startup.
- Preserve default-color cache behavior so a failed attempted query does
not retry forever.
## Validation
- `cd codex-rs && just fmt`
- `cd codex-rs && cargo test -p codex-tui terminal_probe`
- `cd codex-rs && just fix -p codex-tui`
- `cd codex-rs && just argument-comment-lint`
- `git diff --check`
- `git diff --cached --check`
`cd codex-rs && cargo test -p codex-tui` still aborts on the
pre-existing local stack overflow in
`app::tests::discard_side_thread_keeps_local_state_when_server_close_fails`;
I reproduced that same focused failure on `main` before this PR work, so
it is not introduced by this change.
Manual validation in the VM showed the original crossterm path taking
about 2s per unanswered probe, while bounded probing returned in about
100ms per probe.
Refs:
https://linear.app/openai/issue/SE-6311/login-fails-for-experian-users-behind-tls-inspecting-proxy
## Summary
- When a custom CA bundle is configured, force the shared `codex-client`
reqwest builder onto rustls before registering custom roots.
- Add the `rustls-tls-native-roots` reqwest feature so the rustls client
preserves native roots plus the enterprise CA bundle.
- Add subprocess TLS coverage for both a direct local TLS 1.3 server and
a hermetic local CONNECT TLS-intercepting proxy that forwards a
token-exchange-shaped POST to a local origin.
## Plain-language explanation
Experian users are behind a TLS-inspecting proxy, so the login token
exchange needs to trust the enterprise CA bundle from
`CODEX_CA_CERTIFICATE` or `SSL_CERT_FILE`. Before this change, that
custom-CA branch still used reqwest default TLS selection, which could
fail in the proxy environment. Now, only when a custom CA is configured,
Codex selects rustls first and then adds the custom CA roots, matching
the validated behavior from the Experian test build while leaving normal
system-root clients unchanged.
The new regression test recreates the enterprise-proxy shape locally:
the probe client sends an HTTPS `POST /oauth/token` through an explicit
HTTP CONNECT proxy, the proxy presents a leaf certificate signed by a
runtime-generated test CA, decrypts the request, forwards it to a local
origin, and relays the `ok` response back.
## Scope note
- The actual production fix is the first commit: `8368119282 Fix custom
CA reqwest clients to use rustls`.
- The second commit is integration-test coverage only. It generates all
test CA and localhost certificate material at runtime.
## Validation
- `cd codex-rs && cargo test -p codex-client --test ca_env
posts_to_token_origin_through_tls_intercepting_proxy_with_custom_ca_bundle
-- --nocapture`
- `cd codex-rs && cargo test -p codex-client`
- `cd codex-rs && cargo test -p codex-login`
- `cd codex-rs && just fmt`
- `cd codex-rs && just bazel-lock-update`
- `cd codex-rs && just bazel-lock-check`
- `cd codex-rs && just fix -p codex-client`
## Why
#20585 moved the Windows Bazel test job to the cross-compile path, but
the Windows Bazel clippy and verify-release-build jobs were still using
the native Windows/MSVC-host fallback. Those two jobs became the slowest
Windows PR legs, even though both are build-only signal and do not need
to execute the resulting binaries.
## What Changed
- Switches the Windows Bazel clippy job from
`--windows-msvc-host-platform` to `--windows-cross-compile`, so clippy
build actions use Linux RBE while still targeting
`x86_64-pc-windows-gnullvm`.
- Switches the Windows Bazel verify-release-build job to
`--windows-cross-compile` as well. This job only compiles
`cfg(not(debug_assertions))` Rust code under `fastbuild`, so it does not
need a native Windows build host.
- Keeps the old `--skip_incompatible_explicit_targets` behavior only for
fork/community PRs without `BUILDBUDDY_API_KEY`, where `run-bazel-ci.sh`
falls back to the local Windows MSVC-host shape.
- Adds `--windows-cross-compile` support to
`.github/scripts/run-bazel-query-ci.sh`, so target-discovery queries
select the same `ci-windows-cross` config as the subsequent build.
- Threads that option through `scripts/list-bazel-clippy-targets.sh` so
the Windows clippy job discovers targets under the same platform shape
as the subsequent clippy build.
## Verification
Local checks:
```shell
bash -n .github/scripts/run-bazel-query-ci.sh
bash -n scripts/list-bazel-clippy-targets.sh
ruby -e 'require "yaml"; YAML.load_file(".github/workflows/bazel.yml"); puts "ok"'
RUNNER_OS=Linux ./scripts/list-bazel-clippy-targets.sh | grep -c -- '-windows-cross-bin$'
RUNNER_OS=Windows ./scripts/list-bazel-clippy-targets.sh --windows-cross-compile | grep -c -- '-windows-cross-bin$'
```
The Linux target-list check reported `0` Windows-cross internal test
binaries, while the Windows cross target-list check reported `47`,
preserving the test-code clippy coverage shape from the existing Windows
job.
## Status
This is the Bazel PR-CI cross-compilation follow-up to #20485. It is
intentionally split from the Cargo/cargo-xwin release-build PoC so
#20485 can stay as the historical release-build exploration. The
unrelated async-utils test cleanup has been moved to #20686, so this PR
is focused on the Windows Bazel CI path.
The intended tradeoff is now explicit in `.github/workflows/bazel.yml`:
pull requests get the fast Windows cross-compiled Bazel test leg, while
post-merge pushes to `main` run both that fast cross leg and a fully
native Windows Bazel test leg. The native main-only job keeps full
V8/code-mode coverage and gets a 40-minute timeout because it is less
latency-sensitive than PR CI. All other Bazel jobs remain at 30 minutes.
## Why
Windows Bazel PR CI currently does the expensive part of the build on
Windows. A native Windows Bazel test job on `main` completed in about
28m12s, leaving very little headroom under the 30-minute job timeout and
making Windows the slowest PR signal.
#20485 showed that Windows cross-compilation can be materially faster
for Cargo release builds, but PR CI needs Bazel because Bazel owns our
test sharding, flaky-test retries, and integration-test layout. This PR
applies the same high-level shape we already use for macOS Bazel CI:
compile with remote Linux execution, then run platform-specific tests on
the platform runner.
The compromise is deliberately signal-aware: code-mode/V8 changes are
rare enough that PR CI can accept losing the direct V8/code-mode
smoke-test signal temporarily, while `main` still runs the native
Windows job post-merge to catch that class of regression. A follow-up PR
should investigate making the cross-built Windows gnullvm V8 archive
pass the direct V8/code-mode tests so this tradeoff can eventually go
away.
## What Changed
- Adds a `ci-windows-cross` Bazel config that targets
`x86_64-pc-windows-gnullvm`, uses Linux RBE for build actions, and keeps
`TestRunner` actions local on the Windows runner.
- Adds explicit Windows platform definitions for
`windows_x86_64_gnullvm`, `windows_x86_64_msvc`, and a bridge toolchain
that lets gnullvm test targets execute under the Windows MSVC host
platform.
- Updates the Windows Bazel PR test leg to opt into the cross-compile
path via `--windows-cross-compile` and `--remote-download-toplevel`.
- Adds a `test-windows-native-main` job that runs only for `push` events
on `refs/heads/main`, uses the native Windows Bazel path, includes
V8/code-mode smoke tests, and has `timeout-minutes: 40`.
- Keeps fork/community PRs without `BUILDBUDDY_API_KEY` on the previous
local Windows MSVC-host fallback, including
`--host_platform=//:local_windows_msvc` and `--jobs=8`.
- Preserves the existing integration-test shape on non-gnullvm
platforms, while generating Windows-cross wrapper targets only for
`windows_gnullvm`.
- Resolves `CARGO_BIN_EXE_*` values from runfiles at test runtime,
avoiding hard-coded Cargo paths and duplicate test runfiles.
- Extends the V8 Bazel patches enough for the
`x86_64-pc-windows-gnullvm` target and Linux remote execution path.
- Makes the Windows sandbox test cwd derive from `INSTA_WORKSPACE_ROOT`
at runtime when Bazel provides it, because cross-compiled binaries may
contain Linux compile-time paths.
- Keeps the direct V8/code-mode unit smoke tests out of the Windows
cross PR path for now while native Windows CI continues to cover them
post-merge.
## Command Shape
The fast Windows PR test leg invokes the normal Bazel CI wrapper like
this:
```shell
./.github/scripts/run-bazel-ci.sh \
--print-failed-action-summary \
--print-failed-test-logs \
--windows-cross-compile \
--remote-download-toplevel \
-- \
test \
--test_tag_filters=-argument-comment-lint \
--test_verbose_timeout_warnings \
--build_metadata=COMMIT_SHA=${GITHUB_SHA} \
-- \
//... \
-//third_party/v8:all \
-//codex-rs/code-mode:code-mode-unit-tests \
-//codex-rs/v8-poc:v8-poc-unit-tests
```
With the BuildBuddy secret available on Windows, the wrapper selects
`--config=ci-windows-cross` and appends the important Windows-cross
overrides after rc expansion:
```shell
--host_platform=//:rbe
--shell_executable=/bin/bash
--action_env=PATH=/usr/bin:/bin
--host_action_env=PATH=/usr/bin:/bin
--test_env=PATH=${CODEX_BAZEL_WINDOWS_PATH}
```
The native post-merge Windows job intentionally omits
`--windows-cross-compile` and does not exclude the V8/code-mode unit
targets:
```shell
./.github/scripts/run-bazel-ci.sh \
--print-failed-action-summary \
--print-failed-test-logs \
-- \
test \
--test_tag_filters=-argument-comment-lint \
--test_verbose_timeout_warnings \
--build_metadata=COMMIT_SHA=${GITHUB_SHA} \
--build_metadata=TAG_windows_native_main=true \
-- \
//... \
-//third_party/v8:all
```
## Research Notes
The existing macOS Bazel CI config already uses the model we want here:
build actions run remotely with `--strategy=remote`, but `TestRunner`
actions execute on the macOS runner. This PR mirrors that pattern for
Windows with `--strategy=TestRunner=local`.
The important Bazel detail is that `rules_rs` is already targeting
`x86_64-pc-windows-gnullvm` for Windows Bazel PR tests. This PR changes
where the build actions execute; it does not switch the Bazel PR test
target to Cargo, `cargo-nextest`, or the MSVC release target.
Cargo release builds differ from this Bazel path for V8: the normal
Windows Cargo release target is MSVC, and `rusty_v8` publishes prebuilt
Windows MSVC `.lib.gz` archives. The Bazel PR path targets
`windows-gnullvm`; `rusty_v8` does not publish a prebuilt Windows
GNU/gnullvm archive, so this PR builds that archive in-tree. That
Linux-RBE-built gnullvm archive currently crashes in direct V8/code-mode
smoke tests, which is why the workflow keeps native Windows coverage on
`main`.
The less obvious Bazel detail is test wrapper selection. Bazel chooses
the Windows test wrapper (`tw.exe`) from the test action execution
platform, not merely from the Rust target triple. The outer
`workspace_root_test` therefore declares the default test toolchain and
uses the bridge toolchain above so the test action executes on Windows
while its inner Rust binary is built for gnullvm.
The V8 investigation exposed a Windows-client gotcha: even when an
action execution platform is Linux RBE, Bazel can still derive the
genrule shell path from the Windows client. That produced remote
commands trying to run `C:\Program Files\Git\usr\bin\bash.exe` on Linux
workers. The wrapper now passes `--shell_executable=/bin/bash` with
`--host_platform=//:rbe` for the Windows cross path.
The same Windows-client/Linux-RBE boundary also affected
`third_party/v8:binding_cc`: a multiline genrule command can carry CRLF
line endings into Linux remote bash, which failed as `$'\r'`. That
genrule now keeps the `sed` command on one physical shell line while
using an explicit Starlark join so the shell arguments stay readable.
## Verification
Local checks included:
```shell
bash -n .github/scripts/run-bazel-ci.sh
bash -n workspace_root_test_launcher.sh.tpl
ruby -e "require %q{yaml}; YAML.load_file(%q{.github/workflows/bazel.yml}); puts %q{ok}"
RUNNER_OS=Linux ./scripts/list-bazel-clippy-targets.sh
RUNNER_OS=Windows ./scripts/list-bazel-clippy-targets.sh
RUNNER_OS=Linux ./tools/argument-comment-lint/list-bazel-targets.sh
RUNNER_OS=Windows ./tools/argument-comment-lint/list-bazel-targets.sh
```
The Linux clippy and argument-comment target lists contain zero
`*-windows-cross-bin` labels, while the Windows lists still include 47
Windows-cross internal test binaries.
CI evidence:
- Baseline native Windows Bazel test on `main`: success in about 28m12s,
https://github.com/openai/codex/actions/runs/25206257208/job/73907325959
- Green Windows-cross Bazel run on the split PR before adding the
main-only native leg: Windows test 9m16s, Windows release verify 5m10s,
Windows clippy 4m43s,
https://github.com/openai/codex/actions/runs/25231890068
- The latest SHA adds the explicit PR-vs-main tradeoff in `bazel.yml`;
CI is rerunning on that focused diff.
## Follow-Up
A subsequent PR should investigate making a cross-built Windows binary
work with V8/code-mode enabled. Likely options are either making the
Linux-RBE-built `windows-gnullvm` V8 archive correct at runtime, or
evaluating whether a Bazel MSVC target/toolchain can reuse the same
prebuilt MSVC `rusty_v8` archive shape that Cargo release builds already
use.
Felipe Coury ·
Michael Bolin ·
charley-openai ·
sayan-oai ·
Owen Lin ·
Ruslan Nigmatullin ·
starr-openai ·
iceweasel-oai ·
Won Park ·
viyatb-oai ·
Matthew Zeng ·
jif-oai ·
pakrym-oai ·
Eric Traut ·
jgershen-oai ·