Adds new events for streaming apply_patch changes from responses api.
This is to enable clients to show progress during file writes.
Caveat: This does not work with apply_patch in function call mode, since
that required adding streaming json parsing.
## Summary
- mark `features.use_legacy_landlock` as a deprecated feature flag
- emit a startup deprecation notice when the flag is configured
- add feature- and core-level regression coverage for the notice
<img width="1288" height="93" alt="Screenshot 2026-04-15 at 11 14 00 PM"
src="https://github.com/user-attachments/assets/fffc628b-614c-4521-9374-64e50a269252"
/>
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
- adds macOS Seatbelt deny rules for unreadable glob patterns
- expands unreadable glob matches on Linux and masks them in bwrap,
including canonical symlink targets
- keeps Linux glob expansion robust when `rg` is unavailable in minimal
or Bazel test environments
- adds sandbox integration coverage that runs `shell` and `exec_command`
with a `**/*.env = none` policy and verifies the secret contents do not
reach the model
## Linux glob expansion
```text
Prefer: rg --files --hidden --no-ignore --glob <pattern> -- <search-root>
Fallback: internal globset walker when rg is not installed
Failure: any other rg failure aborts sandbox construction
```
```
[permissions.workspace.filesystem]
glob_scan_max_depth = 2
[permissions.workspace.filesystem.":project_roots"]
"**/*.env" = "none"
```
This keeps the common path fast without making sandbox construction
depend on an ambient `rg` binary. If `rg` is present but fails for
another reason, the sandbox setup fails closed instead of silently
omitting deny-read masks.
## Platform support
- macOS: subprocess sandbox enforcement is handled by Seatbelt regex
deny rules
- Linux: subprocess sandbox enforcement is handled by expanding existing
glob matches and masking them in bwrap
- Windows: policy/config/direct-tool glob support is already on `main`
from #15979; Windows subprocess sandbox paths continue to fail closed
when unreadable split filesystem carveouts require runtime enforcement,
rather than silently running unsandboxed
## Stack
1. #15979 - merged: cross-platform glob deny-read
policy/config/direct-tool support for macOS, Linux, and Windows
2. This PR - macOS/Linux subprocess sandbox enforcement plus Windows
fail-closed clarification
3. #17740 - managed deny-read requirements
## Verification
- Added integration coverage for `shell` and `exec_command` glob
deny-read enforcement
- `cargo check -p codex-sandboxing -p codex-linux-sandbox --tests`
- `cargo check -p codex-core --test all`
- `cargo clippy -p codex-linux-sandbox -p codex-sandboxing --tests`
- `just bazel-lock-check`
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
This is the minimal client-side follow-up for the Codex Auto Review
model slug rollout. It updates the guardian reviewer preferred model
from `gpt-5.4` to `codex-auto-review`, so the client can rely on the
backend catalog + Statsig mapping instead of hardcoding the GPT-5.4
slug.
Context:
https://openai.slack.com/archives/C0AF9328RL0/p1775777479388369?thread_ts=1775773094.071629&cid=C0AF9328RL0
## Testing
- `cargo fmt --package codex-core --check`
- `cargo test -p codex-core guardian::`
- `bazel test --experimental_remote_downloader= --test_output=errors
//codex-rs/core:core-unit-tests --test_arg=guardian`
## Summary
- set the realtime v2 server VAD silence delay to 500ms
- update the default realtime 1.5 backend prompt to the v4 text
- keep the session payload and prompt rendering tests aligned with those
changes
## Why
- the VAD change gives the voice path a longer pause before ending the
user's turn
- the prompt change makes the default bundled realtime prompt match the
current v4 content
## Validation
- `cargo +1.93.0 test -p codex-core realtime_prompt --manifest-path
/tmp/codex-realtime-v2-vad-prompt-v4/codex-rs/Cargo.toml`
- `CARGO_TARGET_DIR=/tmp/codex-pr-v4-target cargo +1.93.0 test -p
codex-api
realtime_v2_session_update_includes_background_agent_tool_and_handoff_output_item
--manifest-path
/tmp/codex-realtime-v2-vad-prompt-v4/codex-rs/Cargo.toml`
- `CARGO_TARGET_DIR=/tmp/codex-pr-v4-target cargo +1.93.0 test -p
codex-app-server --test all
'suite::v2::realtime_conversation::realtime_webrtc_start_emits_sdp_notification'
--manifest-path /tmp/codex-realtime-v2-vad-prompt-v4/codex-rs/Cargo.toml
-- --exact`
# Why
We already emit analytics for completed hook runs, but we don't have
matching OTEL metrics to track hook volume and latency.
# What
- add `codex.hooks.run` and `codex.hooks.run.duration_ms`
- tag both metrics with `hook_name`, `source`, and `status`
- emit the metrics from the completed hook path
Verified locally against a dummy OTLP collector
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
Stack PR3 for feature-gated agent identity support.
This PR adds per-thread agent task registration behind
`features.use_agent_identity`. Tasks are minted on the first real user
turn and cached in thread runtime state for later turns.
## Stack
- PR1: https://github.com/openai/codex/pull/17385 - add
`features.use_agent_identity`
- PR2: https://github.com/openai/codex/pull/17386 - register agent
identities when enabled
- PR3: https://github.com/openai/codex/pull/17387 - this PR, original
task registration slice
- PR3.1: https://github.com/openai/codex/pull/17978 - persist and
prewarm registered tasks per thread
- PR4: https://github.com/openai/codex/pull/17980 - use `AgentAssertion`
downstream when enabled
## Validation
Covered as part of the local stack validation pass:
- `just fmt`
- `cargo test -p codex-core --lib agent_identity`
- `cargo test -p codex-core --lib agent_assertion`
- `cargo test -p codex-core --lib websocket_agent_task`
- `cargo test -p codex-api api_bridge`
- `cargo build -p codex-cli --bin codex`
## Notes
The full local app-server E2E path is still being debugged after PR
creation. The current branch stack is directionally ready for review
while that follow-up continues.
## Summary
- remove the final whole-blob truncation pass from realtime
startup-context assembly
- enforce fixed per-section budgets, including each section heading
- keep the existing per-section caps and raise the overall realtime
startup-context budget to `5300`, matching the sum of those section
budgets
- add focused tests for the new wrapping and section-budget behavior
## Why
The previous flow truncated each section and then middle-truncated the
final combined startup-context blob again. Small input changes could
shift that combined cut point, which made retained context unstable and
caused nondeterministic tests.
## Impact
Startup context now preserves section boundaries and ordering
deterministically. Each section is still budgeted independently, but the
final assembled blob is no longer truncated again as a single opaque
string. To match that design, the overall startup-context token budget
is updated to the sum of the existing section budgets rather than
lowering the section caps.
## Validation
- `cargo +1.93.0 test -p codex-core realtime_context`
- `cargo +1.93.0 test -p codex-core --test all
suite::realtime_conversation::conversation_start_injects_startup_context_from_thread_history
-- --exact`
- `cargo +1.93.0 test -p codex-core --test all
suite::realtime_conversation::conversation_startup_context_current_thread_selects_many_turns_by_budget
-- --exact`
- `cargo +1.93.0 test -p codex-core --test all
suite::realtime_conversation::conversation_startup_context_falls_back_to_workspace_map
-- --exact`
- `cargo +1.93.0 test -p codex-core --test all
suite::realtime_conversation::conversation_startup_context_is_truncated_and_sent_once_per_start
-- --exact`
## Problem
When a user resumed or forked a session, the TUI could render the
restored thread history immediately, but it did not receive token usage
until a later model turn emitted a fresh usage event. That left the
context/status UI blank or stale during the exact window where the user
expects resumed state to look complete. Core already reconstructed token
usage from the rollout; the missing behavior was app-server lifecycle
replay to the client that just attached.
## Mental model
Token usage has two representations. The rollout is the durable source
of historical `TokenCount` events, and the core session cache is the
in-memory snapshot reconstructed from that rollout on resume or fork.
App-server v2 clients do not read core state directly; they learn about
usage through `thread/tokenUsage/updated`. The fix keeps those roles
separate: core exposes the restored `TokenUsageInfo`, and app-server
sends one targeted notification after a successful `thread/resume` or
`thread/fork` response when that restored snapshot exists.
This notification is not a new model event. It is a replay of
already-persisted state for the client that just attached. That
distinction matters because using the normal core event path here would
risk duplicating `TokenCount` entries in the rollout and making future
resumes count historical usage twice.
## Non-goals
This change does not add a new protocol method or payload shape. It
reuses the existing v2 `thread/tokenUsage/updated` notification and the
TUI’s existing handler for that notification.
This change does not alter how token usage is computed, accumulated,
compacted, or written during turns. It only exposes the token usage that
resume and fork reconstruction already restored.
This change does not broadcast historical usage replay to every
subscribed client. The replay is intentionally scoped to the connection
that requested resume or fork so already-attached clients are not
surprised by an old usage update while they may be rendering live
activity.
## Tradeoffs
Sending the usage notification after the JSON-RPC response preserves a
clear lifecycle order: the client first receives the thread object, then
receives restored usage for that thread. The tradeoff is that usage is
still a notification rather than part of the `thread/resume` or
`thread/fork` response. That keeps the protocol shape stable and avoids
duplicating usage fields across response types, but clients must
continue listening for notifications after receiving the response.
The helper selects the latest non-in-progress turn id for the replayed
usage notification. This is conservative because restored usage belongs
to completed persisted accounting, not to newly attached in-flight work.
The fallback to the last turn preserves a stable wire payload for
unusual histories, but histories with no meaningful completed turn still
have a weak attribution story.
## Architecture
Core already seeds `Session` token state from the last persisted rollout
`TokenCount` during `InitialHistory::Resumed` and
`InitialHistory::Forked`. The new core accessor exposes the complete
`TokenUsageInfo` through `CodexThread` without giving app-server direct
session mutation authority.
App-server calls that accessor from three lifecycle paths: cold
`thread/resume`, running-thread resume/rejoin, and `thread/fork`. In
each path, the server sends the normal response first, then calls a
shared helper that converts core usage into
`ThreadTokenUsageUpdatedNotification` and sends it only to the
requesting connection.
The tests build fake rollouts with a user turn plus a persisted token
usage event. They then exercise `thread/resume` and `thread/fork`
without starting another model turn, proving that restored usage arrives
before any next-turn token event could be produced.
## Observability
The primary debug path is the app-server JSON-RPC stream. After
`thread/resume` or `thread/fork`, a client should see the response
followed by `thread/tokenUsage/updated` when the source rollout includes
token usage. If the notification is absent, check whether the rollout
contains an `event_msg` payload of type `token_count`, whether core
reconstruction seeded `Session::token_usage_info`, and whether the
connection stayed attached long enough to receive the targeted
notification.
The notification is sent through the existing
`OutgoingMessageSender::send_server_notification_to_connections` path,
so existing app-server tracing around server notifications still
applies. Because this is a replay, not a model turn event, debugging
should start at the resume/fork handlers rather than the turn event
translation in `bespoke_event_handling`.
## Tests
The focused regression coverage is `cargo test -p codex-app-server
emits_restored_token_usage`, which covers both resume and fork. The core
reconstruction guard is `cargo test -p codex-core
record_initial_history_seeds_token_info_from_rollout`.
Formatting and lint/fix passes were run with `just fmt`, `just fix -p
codex-core`, and `just fix -p codex-app-server`. Full crate test runs
surfaced pre-existing unrelated failures in command execution and plugin
marketplace tests; the new token usage tests passed in focused runs and
within the app-server suite before the unrelated command execution
failure.
# Why
Add product analytics for hook handler executions so we can understand
which hooks are running, where they came from, and whether they
completed, failed, stopped, or blocked work.
# What
- add the new `codex_hook_run` analytics event and payload plumbing in
`codex-rs/analytics`
- emit hook-run analytics from the shared hook completion path in
`codex-rs/core`
- classify hook source from the loaded hook path as `system`, `user`,
`project`, or `unknown`
```
{
"event_type": "codex_hook_run",
"event_params": {
"thread_id": "string",
"turn_id": "string",
"model_slug": "string",
"hook_name": "string, // any HookEventName
"hook_source": "system | user | project | unknown",
"status": "completed | failed | stopped | blocked"
}
}
```
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
- replace the unsubscribe-during-turn test's sleep/polling flow with a
gated streaming SSE response
- add request-count notification support to the streaming SSE test
server so the test can wait for the in-flight Responses request
deterministically
## Scope
- codex-rs/app-server/tests/suite/v2/thread_unsubscribe.rs
- codex-rs/core/tests/common/streaming_sse.rs
## Validation
- Not run locally; this is a narrow extraction from the prior CI-green
branch.
---------
Co-authored-by: Codex <noreply@openai.com>
This was flagged by the Codex Security tool: the `state` lock was held
longer than necessary, which included being held across an `async` call,
increasing the potential for deadlock.
While this was flagged by the Codex Security tool, I will look into
enabling
https://rust-lang.github.io/rust-clippy/stable/index.html#await_holding_lock
in a follow-up PR (though unfortunately, that Clippy rule claims it
reports false positives when `drop()` is used to drop a guard instead of
using the end of block scope to drop). Though I can't seem to find a
Clippy rule that checks for opportunities to drop a guard as soon as it
is no longer referenced, in general.
## Summary
- Add best-effort auto-upgrade for user-configured Git marketplaces
recorded in `config.toml`.
- Track the last activated Git revision with `last_revision` so
unchanged marketplace sources skip clone work.
- Trigger the upgrade from plugin startup and `plugin/list`, while
preserving existing fail-open plugin behavior with warning logs rather
than new user-visible errors.
## Details
- Remote configured marketplaces use `git ls-remote` to compare the
source/ref against the recorded revision.
- Upgrades clone into a staging directory, validate that
`.agents/plugins/marketplace.json` exists and that the manifest name
matches the configured marketplace key, then atomically activate the new
root.
- Local `.agents/plugins/marketplace.json` marketplaces remain live
filesystem state and are not auto-pulled.
- Existing non-curated plugin cache refresh is kicked after successful
marketplace root upgrades.
## Validation
- `just write-config-schema`
- `cargo test -p codex-core marketplace_upgrade`
- `cargo check -p codex-cli -p codex-app-server`
- `just fix -p codex-core`
Did not run the complete `cargo test` suite because the repo
instructions require asking before a full core workspace run.
## Summary
- Add an explicit stdin mode to process/start.
- Keep normal non-interactive exec stdin closed while allowing
pipe-backed processes.
## Stack
```text
o #18027 [8/8] Fail exec client operations after disconnect
│
o #18025 [7/8] Cover MCP stdio tests with executor placement
│
o #18089 [6/8] Wire remote MCP stdio through executor
│
o #18088 [5/8] Add executor process transport for MCP stdio
│
o #18087 [4/8] Abstract MCP stdio server launching
│
o #18020 [3/8] Add pushed exec process events
│
@ #18086 [2/8] Support piped stdin in exec process API
│
o #18085 [1/8] Add MCP server environment config
│
o main
```
Co-authored-by: Codex <noreply@openai.com>
## Summary
This makes `DangerFullAccess` / yolo tool execution fully opt out of
managed-network enforcement.
Previously, yolo turns could have `turn.network` stripped while tool
orchestration still derived `enforce_managed_network=true` from
`requirements.toml.network`. That created an inconsistent state where
the turn had no managed proxy attached, but tool execution still behaved
like managed networking was active.
This updates the tool orchestration and JS REPL paths to treat managed
networking as active only when the current turn actually has
`turn.network`.
## Behavior
- Yolo / `DangerFullAccess`: no managed proxy, no managed-network
enforcement.
- Guardian / workspace-write with managed proxy: managed-network
enforcement still applies.
- Avoids the half-state where yolo has no proxy but still gets
managed-network sandbox behavior.
## Tests
- `just fmt`
- `cargo test -p codex-core
danger_full_access_tool_attempts_do_not_enforce_managed_network --
--nocapture`
- `cargo test -p codex-core danger_full_access -- --nocapture`
- `just fix -p codex-core`
Co-authored-by: jgershen-oai <jgershen@openai.com>
## Summary
- Promote `image_generation` from under-development to stable
- Enable image generation by default in the feature registry
- Update feature coverage for the new launch-state expectation
- Add the missing image-generation auth fixture field in a tool registry
test
## Testing
- `just fmt`
- `cargo test -p codex-features`
- `cargo test -p codex-tools` currently fails:
`test_full_toolset_specs_for_gpt5_codex_unified_exec_web_search` needs
its expected default tool list updated for `image_generation`
## Summary
- hide deferred MCP/app nested tool descriptions from the `exec` prompt
in code mode
- add short guidance that omitted nested tools are still available
through `ALL_TOOLS`
- cover the code_mode_only path with an integration test that discovers
and calls a deferred app tool
## Motivation
`code_mode_only` exposes only top-level `exec`/`wait`, but the `exec`
description could still include a large nested-tool reference. This
keeps deferred nested tools callable while avoiding that prompt bloat.
## Tests
- `just fmt`
- `just fix -p codex-code-mode`
- `just fix -p codex-tools`
- `cargo test -p codex-code-mode
exec_description_mentions_deferred_nested_tools_when_available`
- `cargo test -p codex-tools
create_code_mode_tool_matches_expected_spec`
- `cargo test -p codex-core
code_mode_only_guides_all_tools_search_and_calls_deferred_app_tools`
## Summary
- Add an MCP server environment setting with local as the default.
- Thread the default through config serialization, schema generation,
and existing config fixtures.
## Stack
```text
o #18027 [8/8] Fail exec client operations after disconnect
│
o #18025 [7/8] Cover MCP stdio tests with executor placement
│
o #18089 [6/8] Wire remote MCP stdio through executor
│
o #18088 [5/8] Add executor process transport for MCP stdio
│
o #18087 [4/8] Abstract MCP stdio server launching
│
o #18020 [3/8] Add pushed exec process events
│
o #18086 [2/8] Support piped stdin in exec process API
│
@ #18085 [1/8] Add MCP server environment config
│
o main
```
Co-authored-by: Codex <noreply@openai.com>
Stabilizes the Responses API proxy header test by splitting the coverage
at the right boundary:
- Core integration test now verifies parent/subagent identity headers
directly from captured `/responses` requests.
- Proxy dump unit test now verifies those identity headers are preserved
in dumped request JSON.
- Removes the flaky real proxy process + temp-file dump polling path
from the core test.
## Summary
- parse chatgpt_account_is_fedramp from signed ChatGPT auth metadata
- add _account_is_fedramp=true to ChatGPT backend-api requests only for
FedRAMP ChatGPT-auth accounts
Split plugin loading, marketplace, and related infrastructure out of
core into codex-core-plugins, while keeping the core-facing
configuration and orchestration flow in codex-core.
---------
Co-authored-by: Codex <noreply@openai.com>
- [x] Add resource uri meta to tool call item so that the app-server
client can start prefetching resources immediately without loading mcp
server status.
## Summary
- Promote `Feature::ToolSearch` to `Stable` and enable it in the default
feature set
- Update feature tests and tool registry coverage to match the new
default
- Adjust the search-tool integration test to assert the default-on path
and explicit disable fallback
## Testing
- `just fmt`
- `cargo test -p codex-features`
- `cargo test -p codex-core --test all search_tool`
- `cargo test -p codex-tools`
**Summary**
- prevent managed requirements.toml network settings from leaking into
DangerFullAccess / yolo turns by gating managed proxy attachment on
sandbox mode
- keep guardian/sandboxed modes on the managed proxy path, while making
true yolo bypass the proxy entirely, including /shell full-access
commands
## Summary
- wrap realtime startup context in
`<startup_context>...</startup_context>` tags
- prefix V2 mirrored user text and relayed backend text with `[USER]` /
`[BACKEND]`
- remove the V2 progress suffix and replace the final V2 handoff output
with a short completion acknowledgement while preserving the existing V1
wrapper
## Testing
- cargo test -p codex-api
realtime_v2_session_update_includes_background_agent_tool_and_handoff_output_item
-- --exact
- cargo test -p codex-app-server webrtc_v2_background_agent_
- cargo test -p codex-app-server webrtc_v2_text_input_is_
- cargo test -p codex-core conversation_user_text_turn_is_
## Summary
- honor `_meta["codex/imageDetail"] == "original"` on MCP image content
and map it to `detail: "original"` where supported
- strip that detail back out when the active model does not support
original-detail image inputs
- update code-mode `image(...)` to accept individual MCP image blocks
- teach `js_repl` / `codex.emitImage(...)` to preserve the same hint
from raw MCP image outputs
- document the new `_meta` contract and add generic RMCP-backed coverage
across protocol, core, code-mode, and js_repl paths
## Summary
1. Revert https://github.com/openai/codex/pull/17848 so the Bazel and
`BUILD` file changes leave `main`.
2. Prepare for a narrower follow up that restores only `SECURITY.md`.
## Validation
1. Reviewed the revert diff against `main`.
2. Ran a clean diff check before push.
- Discover marketplace manifests from different supported layout paths
instead of only .agents/plugins/marketplace.json.
- Accept local plugin sources written either as { source: "local", path:
... } or as a direct string path.
- Skip unsupported or invalid plugin source entries without failing the
entire marketplace, and keep valid local plugins loadable.
# Summary
- implement local ThreadStore archive/unarchive operations
- implement local ThreadStore read_thread operation
- break up the various ThreadStore local method implementations into
separate files
- migrate app-server archive/unarchive and core archive fixture to use
ThreadStore (but not all read operations yet!)
- use the ThreadStore's read operation as a proxy check for thread
persistence/existence in the app server code
- move all other filesystem operations related to archive (path
validation etc) into the local thread store.
# Tests
- add dedicated local store archive/unarchive tests
## Summary
1. Add a Security Boundaries section to `SECURITY.md`.
2. Point readers to the Codex Agent approvals and security documentation
for sandboxing, approvals, and network controls.
## Validation
1. Reviewed the `SECURITY.md` diff in a clean worktree.
2. No tests run. Docs only change.
Azure Responses providers were still falling back to local compaction
because the compaction gate only checked
`ModelProviderInfo::is_openai()`.
Move the capability check onto `ModelProviderInfo` with
`supports_remote_compaction()`, backed by the existing Azure Responses
endpoint detection used in `codex-api`, and have `core::compact`
delegate to that helper.
Add regression coverage for:
- OpenAI providers using remote compaction
- Azure providers using remote compaction
- non-OpenAI/non-Azure providers staying on the local path
resolves#17773
---------
Co-authored-by: Michael Bolin <mbolin@openai.com>
## Why
#17763 moved sandbox-state delivery for MCP tool calls to request
`_meta` via the `codex/sandbox-state-meta` experimental capability.
Keeping the older `codex/sandbox-state` capability meant Codex still
maintained a second transport that pushed updates with the custom
`codex/sandbox-state/update` request at server startup and when the
session sandbox policy changed.
That duplicate MCP path is redundant with the per-tool-call metadata
path and makes the sandbox-state contract larger than needed. The
existing managed network proxy refresh on sandbox-policy changes is
still needed, so this keeps that behavior separate from the removed MCP
notification.
## What Changed
- Removed the exported `MCP_SANDBOX_STATE_CAPABILITY` and
`MCP_SANDBOX_STATE_METHOD` constants.
- Removed detection of `codex/sandbox-state` during MCP initialization
and stopped sending `codex/sandbox-state/update` at server startup.
- Removed the `McpConnectionManager::notify_sandbox_state_change`
plumbing while preserving the managed network proxy refresh when a user
turn changes sandbox policy.
- Slimmed `McpConnectionManager::new` so startup paths pass only the
initial `SandboxPolicy` needed for MCP elicitation state.
- Kept `codex/sandbox-state-meta` support intact; servers that opt in
still receive the current `SandboxState` on tool-call request `_meta`
([remaining call
path](ff2d3c1e72/codex-rs/core/src/mcp_tool_call.rs (L487-L526))).
- Added regression coverage for refreshing the live managed network
proxy on a per-turn sandbox-policy change.
## Verification
- `cargo test -p codex-core
new_turn_refreshes_managed_network_proxy_for_sandbox_change`
- `cargo test -p codex-mcp`
## Summary
- Move auth header construction into the
`AuthProvider::add_auth_headers` contract.
- Inline `CoreAuthProvider` header mutation in its provider impl and
remove the shared header-map helper.
- Update HTTP, websocket, file upload, sideband websocket, and test auth
callsites to use the provider method.
- Add direct coverage for `CoreAuthProvider` auth header mutation.
## Testing
- `just fmt`
- `cargo test -p codex-api`
- `cargo test -p codex-core
client::tests::auth_request_telemetry_context_tracks_attached_auth_and_retry_phase`
- `cargo test -p codex-core` failed on unrelated/reproducible
`tools::handlers::multi_agents::tests::multi_agent_v2_followup_task_interrupts_busy_child_without_losing_message`
---------
Co-authored-by: Celia Chen <celia@openai.com>
