Split the code-mode protocol and client from the V8-backed runtime so core and
app-server no longer link codex-code-mode in production. ThreadManager now
provisions durable code-mode sessions through a shared external host process,
while tests can still inject the in-process provider.
The IPC protocol uses a persistent stdin/stdout transport. Each frame is a
4-byte big-endian length followed by JSON, with a 16 MiB frame limit. Client
requests carry u64 request IDs so create, execute, wait, terminate, and shutdown
operations can be multiplexed over one process. Session IDs isolate durable
stored values. Execute returns an ExecutionStarted response immediately and an
asynchronous InitialResponse when the initial yield or completion is available.
Nested tool calls and notifications travel from the host back to the client as
delegate requests with their own IDs. Delegate responses, cancellation, and
cell-closed lifecycle messages use the same framed channel. Wire operations
encode errors as Result values. A dead connection fails pending operations,
cancels outstanding delegates, and lets the provider spawn a new host for later
sessions.
Build codex-code-mode-host with V8 pointer-compression sandbox support and add
it to canonical primary and app-server packages, legacy Linux and Windows
bundles, signing verification, installers, Python runtime packages, and release
CI for macOS, Linux, and Windows. The host is discovered next to the current
executable, through CODEX_CODE_MODE_HOST_PATH, or on PATH. OS-level seccomp or
seatbelt restrictions remain a follow-up to this cross-platform process split.
Benchmarks were run from release builds on Linux x86_64 with the V8 sandbox
profile and a text('ok') workload. Cold measurements used 30 samples, warm
session provisioning used 200, and warm command execution used 500. Values are
mean/p50/p95 in milliseconds:
- session startup: in-process 0.002/0.002/0.005, IPC 2.623/2.599/2.894
- fresh-session command: in-process 1.831/1.758/1.915, IPC 7.428/7.252/8.306
- warm session provisioning: in-process 0.002/0.002/0.003,
IPC 0.471/0.463/0.581
- warm command: in-process 1.759/1.757/1.940, IPC 2.005/2.001/2.166
The steady-state median command overhead is approximately 0.244 ms. The median
fresh host plus first command cost is 7.252 ms.
Validation:
- 62/62 core code-mode integration tests passed against the external host
- focused protocol, client, runtime, host, tools, and trace tests passed
- Cargo and Bazel real-process host IPC tests passed
- 11/11 package builder tests passed
- Bazel lock verification, scoped Clippy fixes, and repository formatting passed
## Why
Python files under `scripts/` were not covered by the repository
formatting recipe or the CI formatting job, so formatting drift could
merge unnoticed.
## What
- Add a dedicated `scripts/pyproject.toml` and `scripts/uv.lock` so
root-script formatting uses a locked Ruff version.
- Extend `just fmt` to format root Python scripts and add
`fmt-scripts-check` for CI.
- Run `just fmt-scripts-check` from `.github/workflows/ci.yml`,
installing `uv` through SHA-pinned `astral-sh/setup-uv` while retaining
the `uv` `0.11.3` pin.
- Apply Ruff formatting to the root Python scripts, including
`scripts/just-shell.py`, and extend
`sdk/python/tests/test_artifact_workflow_and_binaries.py` to cover the
root formatting recipe.
- Update `AGENTS.md` so agents run `just fmt` after code changes
anywhere in the repository.
## Validation
- Extended the existing Python SDK workflow test to assert that `just
fmt` includes root Python scripts.
## Why
Release packaging should be a staging step once release binaries have
already been built and signed. The Windows release job was downloading
and signing `codex-command-runner.exe` and
`codex-windows-sandbox-setup.exe`, but `scripts/build_codex_package.py`
still rebuilt those helpers while creating the package archives.
That makes the package step slower and, more importantly, risks putting
helper binaries in the archive that were produced after the signing
step. Linux had the same shape for package resources: `bwrap` could be
rebuilt by the package builder instead of being passed in as a prebuilt
release artifact.
This builds on #23752, which fixes `.tar.zst` creation when Windows
runners rely on the repository DotSlash `zstd` wrapper.
## What changed
- Add explicit prebuilt resource inputs to the Codex package builder:
- `--bwrap-bin`
- `--codex-command-runner-bin`
- `--codex-windows-sandbox-setup-bin`
- Make `.github/scripts/build-codex-package-archive.sh` pass resource
binaries from the release output directory when they are already
present.
- Build Linux `bwrap` for app-server release jobs too, so app-server
package creation does not invoke Cargo just to supply the package
resource.
- Keep macOS package creation as a no-Cargo path when `--entrypoint-bin`
is provided, since macOS packages have no resource binaries.
- Add unit coverage showing prebuilt macOS, Linux, and Windows package
inputs result in no source-built binaries.
## Verification
- `python3 -m unittest discover -s scripts/codex_package -p 'test_*.py'`
- `python3 -m py_compile scripts/codex_package/*.py`
- `bash -n .github/scripts/build-codex-package-archive.sh`
- Dry-ran Linux and Windows package builds with fake prebuilt resources
and a nonexistent Cargo path to verify the package builder did not
invoke Cargo.
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/23759).
* #23760
* __->__ #23759
