## Summary
`hooks/list` only consumes plugin hook declarations, but previously
loaded every enabled plugin's skills, MCP configuration, apps, and
capability summary before discarding them.
In a local benchmark, this reduced `hooks/list` latency by over 100ms
(e.g., from 594 to 467ms on startup, and 168 to 16ms when making a
`hooks/list` call later in the same TUI session). This is on the
critical path to rendering the TUI, so every 10s of ms should be eyed
skeptically (IMO).
This change adds a hook-specific plugin loading path that preserves
plugin enablement, remote/local conflict resolution, deterministic
ordering, manifest resolution, and hook-loading warnings while skipping
unrelated capabilities. (I think there's room for a more general design
here that allows you to project the capabilities you need at load-time,
but that seems unnecessary right now.)
## Summary
- skip `opentelemetry_sdk` DEBUG and TRACE events before formatting or
queueing them for the SQLite log sink
- preserve INFO, WARN, and ERROR events from the SDK, along with TRACE
events from application targets
- add a persistence-level regression test for the target and level
policy
## Why
OpenTelemetry's batch log processor emits internal
`BatchLogProcessor.ExportingDueToTimer` meta-events every second per
Codex process. In measured high-fanout `logs_2.sqlite` databases,
low-level `opentelemetry_sdk` events accounted for over 30% of retained
rows (30-60% on the machines of people I asked to check).
Persisting this SDK bookkeeping across many processes adds substantial
write volume and contention without representing application activity.
## Validation
- `just test -p codex-state` (132/132 tests passed, plus bench smoke)
- `just fix -p codex-state`
- `just fmt`
## Summary
This PR adds `memchr` for some low-hanging performance improvements
(namely, in MCP stdio, Ollama streaming, and full message-history
newline counts).
Codex produced the following release benchmarks:
| Operation | Before | After | Speedup |
| --- | ---: | ---: | ---: |
| MCP 1 MiB chunked line | 2.172 s | 3.984 ms | 545x |
| Ollama 1 MiB chunked line | 1.673 s | 2.790 ms | 600x |
| Count newlines in 10 MiB history | 132.83 ms | 20.05 ms | 6.6x |
With a "real" MCP setup (`ExecutorStdioServerLauncher` started a Python
MCP server, completed `initialize`, requested `tools/list`, and
deserialized a 1 MiB tool description over newline-delimited stdio),
it's about 16x faster end-to-end:
| Branch | 50 calls | Per call |
| --- | ---: | ---: |
| `main` | 862.53 ms | 17.25 ms |
| this branch | 53.89 ms | 1.08 ms |
`memchr` is already in our dependency tree and extremely widely used for
this kind of optimized scanning.
## Why
The skills extension needs to become the path that exposes local host
skills without losing the behavior already owned by core skill loading.
Host skill discovery is not just `$CODEX_HOME/skills`: it also includes
config layers, bundled-skill settings, plugin roots, runtime extra
roots, and the filesystem for the selected primary environment.
Rather than making the extension reload host skills and risk drifting
from that authoritative load, this PR bridges the already-loaded
per-turn skills outcome into the extension. That lets the extension
advertise host skills and inject explicit `$skill` prompts while
preserving the same roots, disabled/hidden state, rendered paths, and
environment-backed file reads that the legacy path uses.
## What Changed
- Adds `HostLoadedSkills` in `core-skills` to wrap the turn's
`SkillLoadOutcome` and read `SKILL.md` through the filesystem that
loaded that skill.
- Stores `HostLoadedSkills` in turn extension data for normal turns and
review turns, so the skills extension can consume the loaded host
catalog without reloading it.
- Adds `HostSkillProvider` under `ext/skills/src/provider/host.rs`,
mapping host-loaded skill metadata into the skills-extension
catalog/read contract.
- Registers the host provider by default from
`codex_skills_extension::install()`.
- Preserves host skill metadata such as dependencies, disabled state,
hidden-from-prompt policy, and slash-normalized display paths.
- Passes host-loaded skills through `SkillListQuery` and
`SkillReadRequest` so explicit skill invocation reads only resources
from the loaded host catalog.
- Adds integration coverage for a real legacy
`$CODEX_HOME/skills/.../SKILL.md` skill being listed and injected
through the installed extension.
## Testing
- Added `installed_extension_loads_host_skills_from_legacy_roots` in
`ext/skills/tests/skills_extension.rs`.
- `just test -p codex-skills-extension`
## Why
Goal idle continuation is extension-triggered model-visible work, so it
should follow one core-owned rule for when automatic work may start. In
particular, it should not jump ahead of queued user/client work, start
while another task is active, or inject a continuation turn while the
thread is in Plan mode.
Keeping this policy in `try_start_turn_if_idle` avoids passing
`collaboration_mode` or review-specific state through
`ThreadLifecycleContributor::on_thread_idle`. Active `/review` is
covered by the same active-task gate because Review turns are not
steerable.
## What Changed
- Teach `Session::try_start_turn_if_idle` to reject automatic idle turns
in Plan mode, both before reserving an idle turn and after building the
turn context.
- Document `CodexThread::try_start_turn_if_idle` as the extension-facing
gate for automatic idle work, including Plan-mode and active Review-task
behavior.
- Add focused coverage for Plan-mode rejection and active Review-task
rejection without queuing synthetic input.
## Testing
- `just test -p codex-core try_start_turn_if_idle`
## Why
The public Codex release workflow needs to sign and notarize macOS
binaries and DMGs without placing the Developer ID private key in
GitHub. This moves the private-key operation behind the protected
`codesigning` environment and uses GitHub OIDC with Azure Key Vault
PKCS#11, while preserving the existing external `build_unsigned` /
`promote_signed` fallback.
## What changed
- Add a reusable AKV PKCS11 setup action that authenticates to Azure
with OIDC, downloads pinned signing tools, verifies their SHA-256
digests, and loads the public signing certificate from Key Vault.
- Replace the legacy macOS signing action with scripts that support
AKV-backed `rcodesign`, notarize signed binaries and DMGs, and staple
DMG notarization tickets.
- Restructure `rust-release.yml` so macOS builds produce unsigned
artifacts first, protected jobs perform signing and notarization, macOS
runners package and verify the results, and release publishing waits for
verified artifacts.
- Preserve the manual external-signing handoff flow and make manual-mode
conditions explicit.
- Move the Codex entitlements file alongside the signing scripts and
update CODEOWNERS for the new signing surfaces.
## Verification
- [Live protected signing workflow
run](https://github.com/openai/codex/actions/runs/26903610631) completed
successfully for both macOS architectures, including binary
signing/notarization, DMG signing/notarization, and final artifact
verification.
- Downloaded both signed DMGs and independently verified their checksums
and strict signatures.
- Confirmed `xcrun stapler validate` succeeds and Gatekeeper accepts
both DMGs as `Notarized Developer ID`.
- Mounted both DMGs and confirmed the contained `codex` and
`codex-responses-api-proxy` binaries have valid Developer ID signatures
for the expected architectures.
---------
Co-authored-by: shijie-openai <shijie.rao@openai.com>
## Why
Compaction analytics need token counts that better represent the request
being compacted. The existing session snapshot can diverge from the
actual remote compaction request after output rewriting, and remote v2
can use server-side Responses usage when available.
## What changed
- Add an optional `active_context_tokens_before` override to
`CompactionAnalyticsAttempt::track(...)` for remote compaction when it
has a better before-token value than the begin-time session snapshot.
The local `/compact` path passes no override.
- For remote v1 `responses_compact`, subtract the estimated token delta
from pre-compaction output rewriting from the session snapshot, capped
by locally-added tokens since the last successful API response.
- For remote v2 `responses_compaction_v2`, use the same bounded
output-rewrite fallback as remote v1, then overwrite
`active_context_tokens_before` with server `token_usage.input_tokens`
from the `response.completed` event when present.
- Keep the existing v2 compaction-output validation while carrying the
completed response token usage through `collect_compaction_output`.
## Verification
- `just fmt`
- `just test -p codex-core
collect_compaction_output_accepts_additional_output_items`
- `git diff --check`
## Why
Codex package installs include helper binaries in `codex-path`, such as
the bundled `rg`. Package-layout launches should add that directory
before user commands run, but standalone launches were missing it while
npm launches only worked because `codex.js` had its own legacy `PATH`
rewrite. That made npm and standalone package behavior diverge.
Shell snapshot restoration can also reset `PATH` after runtime setup.
Any package-owned `PATH` prepend has to be recorded as an explicit
runtime override so shells, unified exec, and user-shell commands keep
access to `codex-path` after a snapshot is sourced.
## Repro
Before this change, a curl-installed package could contain `rg` under
`codex-path` but still fail to put it on `PATH`:
```shell
mkdir /tmp/test-codex-curl
curl -fsSL https://chatgpt.com/codex/install.sh \
| CODEX_HOME=/tmp/test-codex-curl CODEX_NON_INTERACTIVE=1 sh
/tmp/test-codex-curl/packages/standalone/current/bin/codex exec \
--skip-git-repo-check 'print `which -a rg`'
find /tmp/test-codex-curl -name rg
```
The `which -a rg` output omitted the packaged helper even though `find`
showed it under
`/tmp/test-codex-curl/packages/standalone/releases/.../codex-path/rg`.
The npm install path behaved differently only because
`codex-cli/bin/codex.js` had legacy `PATH` rewriting:
```shell
mkdir /tmp/test-codex-npm
cd /tmp/test-codex-npm
npm install @openai/codex
./node_modules/.bin/codex exec --skip-git-repo-check 'print `which -a rg`'
```
That printed the npm package's `vendor/<target>/codex-path/rg` first.
This PR moves that behavior into Rust-side package launch setup so
curl/standalone and npm/bun launches agree without JS rewriting `PATH`.
## What Changed
- `codex-rs/arg0` now uses
`InstallContext::current().package_layout.path_dir` to prepend the
package helper directory before any threads are created.
- Package helper `PATH` setup is independent from the temporary arg0
alias setup, so `codex-path` is still added even if CODEX_HOME tempdir,
lock, or symlink setup fails.
- `codex-rs/install-context` detects the canonical package layout we
ship: `bin/`, `codex-resources/`, and `codex-path/` next to
`codex-package.json`.
- Shell, local unified exec, and user-shell runtimes now record package
`codex-path` prepends in `explicit_env_overrides`, matching the existing
zsh-fork behavior so shell snapshots cannot restore over the package
helper path.
- Remote unified exec requests do not receive the local app-server
package path overlay.
- `codex-cli/bin/codex.js` no longer computes or overrides `PATH`; it
only locates the native binary in the canonical package layout and
passes npm/bun management metadata.
- Added regression tests for `PATH` ordering, package layout detection,
and shell snapshot preservation of package path prepends.
## Verification
- `node --check codex-cli/bin/codex.js`
- `just test -p codex-install-context -p codex-arg0`
- `just test -p codex-core
user_shell_snapshot_preserves_package_path_prepend`
- `just test -p codex-core tools::runtimes::tests`
- `just bazel-lock-update`
- `just bazel-lock-check`
- `just fix -p codex-install-context -p codex-arg0 -p codex-core`
# Why
When an organization requires the elevated Windows sandbox, Codex
launches an elevated helper to provision users, configure firewall and
ACL rules, and lock persistent sandbox directories.
We observed that closing the helper after setup started could leave the
machine partially initialized while the TUI still announced **Sandbox
ready**. Model-only turns continued to work, but the first shell command
retried setup and failed with Windows cancellation error `1223`.
This was not an enforcement bypass; command execution continued to fail
closed. The issue was a false readiness signal: `setup_marker.json` was
written during user provisioning, before the remaining setup stages had
completed.
# What
Treat `setup_marker.json` as the commit record for Windows sandbox
setup:
1. Before full or provisioning setup begins, remove the existing marker
and create the final marker path with a protected ACL.
2. Keep the marker empty and therefore invalid while setup is in
progress. Sandbox users cannot read, modify, or replace it.
3. Run every synchronous setup stage.
4. After setup succeeds, write the valid marker contents without
changing its ACL.
5. After the helper exits successfully, verify the existing readiness
check before enabling the sandbox.
If setup is canceled or fails, the marker remains invalid and Codex
reports setup as incomplete instead of announcing readiness.
Refresh-only and read-ACL-only helper runs continue to leave the marker
untouched. The setup version remains `5` to avoid forcing all existing
Windows users through elevated setup again.
# Verification
- Added coverage confirming sandbox users cannot read or modify the
setup marker after elevated setup.
- Added coverage confirming a successful helper exit without complete
setup artifacts is rejected.
- Ran `just test -p codex-windows-sandbox`.
## Why
PR descriptions can be visible outside the context used to generate
them. In #23710, a generated description referenced an internal
document, showing that the skill needs an explicit guardrail against
exposing confidential context.
## What changed
- Updated the `codex-pr-body` guidance to prohibit confidential
references, including codenames and OpenAI-internal URLs.
## Why
When trying to fit history under compaction limit rewrite output items
instead of removing them entirely. Otherwise we're breaking
incrementality in relation to the previous response.
## Why
Model metadata can now select multi-agent v2 even when a user has not
enabled `features.multi_agent_v2` in their config. Some existing configs
still set the legacy `agents.max_threads` knob for v1 multi-agent
behavior, so treating every v2 runtime as incompatible with
`agents.max_threads` would break users whose only v2 signal came from
the model catalog.
The incompatible configuration is specifically enabling
`features.multi_agent_v2` while also setting `agents.max_threads`.
Catalog-forced v2 should use the v2 concurrency setting and ignore the
legacy v1 cap instead of rejecting the config.
## What changed
- Split config validation from runtime concurrency calculation:
`effective_agent_max_threads` now just returns the effective cap for the
resolved multi-agent runtime.
- Added explicit validation for `features.multi_agent_v2` +
`agents.max_threads` at session startup.
- Preserved catalog-selected v2 behavior when `features.multi_agent_v2`
is disabled, so existing configs with `agents.max_threads` keep
starting.
- Updated model-runtime selector coverage so a catalog v2 model still
exposes v2 tools even when `agents.max_threads` is set and the config
flag is disabled.
## Validation
- `cargo check -p codex-core --lib`
- `just test -p codex-core --lib -E
"test(multi_agent_v2_feature_rejects_agents_max_threads) |
test(catalog_v2_allows_agents_max_threads_when_feature_disabled)"`
Michael Bolin