## Summary
Stack PR 2 of 4 for feature-gated agent identity support.
This PR adds agent identity registration behind
`features.use_agent_identity`. It keeps the app-server protocol
unchanged and starts registration after ChatGPT auth exists rather than
requiring a client restart.
## Stack
- PR1: https://github.com/openai/codex/pull/17385 - add
`features.use_agent_identity`
- PR2: https://github.com/openai/codex/pull/17386 - this PR
- PR3: https://github.com/openai/codex/pull/17387 - register agent tasks
when enabled
- PR4: https://github.com/openai/codex/pull/17388 - 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.
stacked on #17402.
MCP tools returned by `tool_search` (deferred tools) get registered in
our `ToolRegistry` with a different format than directly available
tools. this leads to two different ways of accessing MCP tools from our
tool catalog, only one of which works for each. fix this by registering
all MCP tools with the namespace format, since this info is already
available.
also, direct MCP tools are registered to responsesapi without a
namespace, while deferred MCP tools have a namespace. this means we can
receive MCP `FunctionCall`s in both formats from namespaces. fix this by
always registering MCP tools with namespace, regardless of deferral
status.
make code mode track `ToolName` provenance of tools so it can map the
literal JS function name string to the correct `ToolName` for
invocation, rather than supporting both in core.
this lets us unify to a single canonical `ToolName` representation for
each MCP tool and force everywhere to use that one, without supporting
fallbacks.
## Summary
- Fix marketplace-add local path detection on Windows by using
`Path::is_absolute()`.
- Make marketplace-add local-source tests parse/write TOML through the
same helpers instead of raw string matching.
- Update `rand` 0.9.x to 0.9.3 and document the remaining audited `rand`
0.8.5 advisory exception.
- Refresh `MODULE.bazel.lock` after the Cargo.lock update.
## Why
Latest `main` had two independent CI blockers: marketplace-add tests
were not portable to Windows path/TOML escaping, and cargo-deny still
reported `RUSTSEC-2026-0097` after the recent rustls-webpki fix.
## Validation
- `cargo test -p codex-core marketplace_add -- --nocapture`
- `cargo deny --all-features check`
- `just bazel-lock-check`
- `just fix -p codex-core`
- `just fmt`
- `git diff --check`
## Changes
Allows MCPs to opt in to receiving sandbox config info through `_meta`
on model-initiated tool calls. This lets MCPs adhere to the thread's
sandbox if they choose to.
## Details
- Adds the `codex/sandbox-state-meta` experimental MCP capability.
- Tracks whether each MCP server advertises that capability.
- When a server opts in, `codex-core` injects the current `SandboxState`
into model-initiated MCP tool-call request `_meta`.
## Verification
- added an integration test for the capability
`exec()` had a number of arguments that were unused, making the function
signature misleading. This PR aims to clean things up to clarify the
role of this function and to clarify which fields of `ExecParams` are
unused and why.
## Why
`spawn_command_under_seatbelt()` in `codex-rs/core/src/seatbelt.rs` had
fallen out of production use and was only referenced by test-only
wrappers. That left us with sandbox tests that could stay green even if
the actual seatbelt exec path regressed, because production shell
execution now flows through `SandboxManager::transform()` and
`ExecRequest::from_sandbox_exec_request()` instead of that helper.
Removing the dead helper also exposed one downstream `codex-exec`
integration test that still imported it, which broke `just clippy`.
## What Changed
- Removed `codex-rs/core/src/seatbelt.rs` and stopped exporting
`codex_core::seatbelt`.
- Removed the redundant `codex-rs/core/tests/suite/seatbelt.rs` coverage
that only exercised the dead helper.
- Kept the `openpty` regression check, but moved it into
`codex-rs/core/tests/suite/exec.rs` so it now runs through
`process_exec_tool_call()`.
- Fixed the seatbelt denial test in `codex-rs/core/tests/suite/exec.rs`
to use `/usr/bin/touch`, so it actually exercises the sandbox instead of
a nonexistent path.
- Updated `codex-rs/exec/tests/suite/sandbox.rs` on macOS to build the
sandboxed command through `build_exec_request()` and spawn the
transformed command, instead of importing the removed helper.
- Left the lower-level seatbelt policy coverage in
`codex-rs/sandboxing/src/seatbelt_tests.rs`, where the policy generator
is still covered directly.
## Verification
- `cargo test -p codex-core suite::exec::`
- `cargo test -p codex-exec`
- `cargo clippy -p codex-exec --tests -- -D warnings`
Fix clippy warnings in external agent config migration
```
error: this expression creates a reference which is immediately dereferenced by the compiler
--> core/src/external_agent_config.rs:188:55
|
188 | let migrated = build_config_from_external(&settings)?;
| ^^^^^^^^^ help: change this to: `settings`
|
= help: for further information visit https://rust-lang.github.io/rust-clippy/rust-1.93.0/index.html#needless_borrow
= note: requested on the command line with `-D clippy::needless-borrow`
error: useless conversion to the same type: `codex_utils_absolute_path::AbsolutePathBuf`
--> core/src/external_agent_config.rs:355:27
|
355 | match AbsolutePathBuf::try_from(
| ___________________________^
356 | | add_marketplace_outcome
357 | | .installed_root
358 | | .join(INSTALLED_MARKETPLACE_MANIFEST_RELATIVE_PATH),
359 | | ) {
| |_____________________^
|
= help: consider removing `AbsolutePathBuf::try_from()`
= help: for further information visit https://rust-lang.github.io/rust-clippy/rust-1.93.0/index.html#useless_conversion
= note: `-D clippy::useless-conversion` implied by `-D warnings`
= help: to override `-D warnings` add `#[allow(clippy::useless_conversion)]`
error: aborting due to 2 previous errors
```
## Summary
- wrap routed delegation text in a small XML envelope before submitting
it as a user turn
- escape XML text content so the envelope stays well formed
- update focused coverage for the wrapper and the affected routed-turn
expectations
## What
Disable `Feature::CodexHooks` when building guardian review session
config
## Why
Guardian review sessions were respecting the Stop hook and could ingest
synthetic `<hook_prompt>` user turns Guardian should ignore hooks, while
the main session and regular subagents continue to respect them
In other words Guardian was getting ralph-looped
Co-authored-by: Codex <noreply@openai.com>
### **Issue**
guardian_parallel_reviews_fork_from_last_committed_trunk_history was
failing on Windows/Bazel with a stack overflow:
`thread
'guardian::tests::guardian_parallel_reviews_fork_from_last_committed_trunk_history'
has overflowed its stack`
- This problem was a stack-headroom problem
### **Solution**
Reduced stack pressure in the guardian async path by boxing thin wrapper
futures, and run the affected test on a dedicated 2 MiB thread stack.
Concretely:
- added Box::pin(...) around thin async wrapper hops in the guardian
review/delegate path
- changed
guardian_parallel_reviews_fork_from_last_committed_trunk_history to run
inside an explicitly sized thread stack so it has enough headroom in
low-stack environments
## Summary
- Port marketplace source support into the shared core marketplace-add
flow
- Support local marketplace directory sources
- Support direct `marketplace.json` URL sources
- Persist the new source types in config/schema and cover them in CLI
and app-server tests
## Validation
- `cargo test -p codex-core marketplace_add`
- `cargo test -p codex-cli marketplace_add`
- `cargo test -p codex-app-server marketplace_add`
- `just write-config-schema`
- `just fmt`
- `just fix -p codex-core`
- `just fix -p codex-cli`
## Context
Current `main` moved marketplace-add behavior into shared core code and
still assumed only git-backed sources. This change keeps that structure
but restores support for local directories and direct manifest URLs in
the shared path.
## Summary
- Allows selected MCP results to return a larger default result set.
- Keeps the existing default cap for other MCP results.
- Applies the cap consistently when higher explicit limits are
requested.
## Testing
- `cargo test -p codex-core tool_search`
- Ran a local CLI smoke test with two stdio MCP servers exposing 100
tools each; the selected-server query returned 20 tools and the
regular-server query returned 8.
## Summary
- route apply_patch runtime execution through the selected Environment
filesystem instead of the local self-exec path
- keep the standalone apply_patch command surface intact while restoring
its launcher/test/docs contract
- add focused apply_patch filesystem sandbox regression coverage
## Validation
- remote devbox Bazel run in progress
- passed: //codex-rs/apply-patch:apply-patch-unit-tests
--test_filter=test_read_file_utf8_with_context_reports_invalid_utf8
- in progress / follow-up: focused core and exec Bazel test slices on
dev
## Follow-up under review
- remote pre-verification and approval/retry behavior still need
explicit scrutiny for delete/update flows
- runtime sandbox-denial classification may need a tighter assertion
path than rendered stderr matching
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
Adds `thread_source` field to the existing Codex turn metadata sent to
Responses API
- Sends `thread_source: "user"` for user-initiated sessions: CLI, VS
Code, and Exec
- Sends `thread_source: "subagent"` for subagent sessions
- Omits `thread_source` for MCP, custom, and unknown session sources
- Uses the existing turn metadata transport:
- HTTP requests send through the `x-codex-turn-metadata` header
- WebSocket `response.create` requests send through
`client_metadata["x-codex-turn-metadata"]`
## Testing
- `cargo test -p codex-protocol
session_source_thread_source_name_classifies_user_and_subagent_sources`
- `cargo test -p codex-core turn_metadata_state`
- `cargo test -p codex-core --test responses_headers
responses_stream_includes_turn_metadata_header_for_git_workspace_e2e --
--nocapture`
## Summary
This PR removes `image_detail_original` as a runtime experiment and
makes original image detail available whenever the selected model
supports it.
Concretely, this change:
- drops the `image_detail_original` feature flag from the feature
registry and generated config schema
- makes tool-emitted image detail depend only on
`ModelInfo.supports_image_detail_original`
- updates `view_image` and `code_mode`/`js_repl` image emission to use
that capability check directly
- removes now-redundant experiment-specific tests and instruction
coverage
- keeps backward compatibility for existing configs by silently ignoring
a stale `features.image_detail_original` entry
The net effect is that `detail: "original"` is always available on
supported models, without requiring an experiment toggle.
- Add outputModality to thread/realtime/start and wire text/audio output
selection through app-server, core, API, and TUI.\n- Rename the realtime
transcript delta notification and add a separate transcript done
notification that forwards final text from item done without correlating
it with deltas.
This changes multi-agent v2 mailbox handling so incoming inter-agent
messages no longer preempt an in-flight sampling stream at reasoning or
commentary output-item boundaries.
## Summary
Move `codex marketplace add` onto a shared core implementation so the
CLI and app-server path can use one source of truth.
This change:
- adds shared marketplace-add orchestration in `codex-core`
- switches the CLI command to call that shared implementation
- removes duplicated CLI-only marketplace add helpers
- preserves focused parser and add-path coverage while moving the shared
behavior into core tests
## Why
The new `marketplace/add` RPC should reuse the same underlying
marketplace-add flow as the CLI. This refactor lands that consolidation
first so the follow-up app-server PR can be mostly protocol and handler
wiring.
## Validation
- `cargo test -p codex-core marketplace_add`
- `cargo test -p codex-cli marketplace_cmd`
- `just fix -p codex-core`
- `just fix -p codex-cli`
- `just fmt`
## Summary
- Add `turn/inject_items` app-server v2 request support for appending
raw Responses API items to a loaded thread history without starting a
turn.
- Generate JSON schema and TypeScript protocol artifacts for the new
params and empty response.
- Document the new endpoint and include a request/response example.
- Preserve compatibility with the typo alias `turn/injet_items` while
returning the canonical method name.
## Testing
- Not run (not requested)
## Why
For more advanced MCP usage, we want the model to be able to emit
parallel MCP tool calls and have Codex execute eligible ones
concurrently, instead of forcing all MCP calls through the serial block.
The main design choice was where to thread the config. I made this
server-level because parallel safety depends on the MCP server
implementation. Codex reads the flag from `mcp_servers`, threads the
opted-in server names into `ToolRouter`, and checks the parsed
`ToolPayload::Mcp { server, .. }` at execution time. That avoids relying
on model-visible tool names, which can be incomplete in
deferred/search-tool paths or ambiguous for similarly named
servers/tools.
## What was added
Added `supports_parallel_tool_calls` for MCP servers.
Before:
```toml
[mcp_servers.docs]
command = "docs-server"
```
After:
```toml
[mcp_servers.docs]
command = "docs-server"
supports_parallel_tool_calls = true
```
MCP calls remain serial by default. Only tools from opted-in servers are
eligible to run in parallel. Docs also now warn to enable this only when
the server’s tools are safe to run concurrently, especially around
shared state or read/write races.
## Testing
Tested with a local stdio MCP server exposing real delay tools. The
model/Responses side was mocked only to deterministically emit two MCP
calls in the same turn.
Each test called `query_with_delay` and `query_with_delay_2` with `{
"seconds": 25 }`.
| Build/config | Observed | Wall time |
| --- | --- | --- |
| main with flag enabled | serial | `58.79s` |
| PR with flag enabled | parallel | `31.73s` |
| PR without flag | serial | `56.70s` |
PR with flag enabled showed both tools start before either completed;
main and PR-without-flag completed the first delay before starting the
second.
Also added an integration test.
Additional checks:
- `cargo test -p codex-tools` passed
- `cargo test -p codex-core
mcp_parallel_support_uses_exact_payload_server` passed
- `git diff --check` passed
Cap mirrored user text sent to realtime with the existing 300-token turn
budget while preserving the full model turn.
Adds integration coverage for capped realtime mirror payloads.
---------
Co-authored-by: Codex <noreply@openai.com>
# External (non-OpenAI) Pull Request Requirements
Before opening this Pull Request, please read the dedicated
"Contributing" markdown file or your PR may be closed:
https://github.com/openai/codex/blob/main/docs/contributing.md
If your PR conforms to our contribution guidelines, replace this text
with a detailed and high quality description of your changes.
Include a link to a bug report or enhancement request.
- stop `list_tool_suggest_discoverable_plugins()` from reloading the
curated marketplace for each discoverable plugin
- reuse a direct plugin-detail loader against the already-resolved
marketplace entry
The trigger was to stop those logs spamming:
```
d=019d81cf-6f69-7230-98aa-74294ff2dc5a}:submission_dispatch{otel.name="op.dispatch.user_input" submission.id="019d86c8-0a8e-7013-b442-109aabbf75c9" codex.op="user_input"}:turn{otel.name="session_task.turn" thread.id=019d81cf-6f69-7230-98aa-74294ff2dc5a turn.id=019d86c8-0a8e-7013-b442-109aabbf75c9 model=gpt-5.4}: ignoring interface.defaultPrompt: prompt must be at most 128 characters path=/Users/jif/.codex/.tmp/plugins/plugins/life-science-research/.codex-plugin/plugin.json
2026-04-13T12:27:30.402Z WARN [019d81cf-6f69-7230-98aa-74294ff2dc5a] codex_core::plugins::manifest - session_loop{thread_id=019d81cf-6f69-7230-98aa-74294ff2dc5a}:submission_dispatch{otel.name="op.dispatch.user_input" submission.id="019d86c8-0a8e-7013-b442-109aabbf75c9" codex.op="user_input"}:turn{otel.name="session_task.turn" thread.id=019d81cf-6f69-7230-98aa-74294ff2dc5a turn.id=019d86c8-0a8e-7013-b442-109aabbf75c9 model=gpt-5.4}: ignoring interface.defaultPrompt: prompt must be at most 128 characters path=/Users/jif/.codex/.tmp/plugins/plugins/build-ios-apps/.codex-plugin/plugin.json
2026-04-13T12:27:30.402Z WARN [019d81cf-6f69-7230-98aa-74294ff2dc5a] codex_core::plugins::manifest - session_loop{thread_id=019d81cf-6f69-7230-98aa-74294ff2dc5a}:submission_dispatch{otel.name="op.dispatch.user_input" submission.id="019d86c8-0a8e-7013-b442-109aabbf75c9" codex.op="user_input"}:turn{otel.name="session_task.turn" thread.id=019d81cf-6f69-7230-98aa-74294ff2dc5a turn.id=019d86c8-0a8e-7013-b442-109aabbf75c9 model=gpt-5.4}: ignoring interface.defaultPrompt: prompt must be at most 128 characters path=/Users/jif/.codex/.tmp/plugins/plugins/life-science-research/.codex-plugin/plugin.json
2026-04-13T12:27:30.405Z WARN [019d81cf-6f69-7230-98aa-74294ff2dc5a] codex_core::plugins::manifest - session_loop{thread_id=019d81cf-6f69-7230-98aa-74294ff2dc5a}:submission_dispatch{otel.name="op.dispatch.user_input" submission.id="019d86c8-0a8e-7013-b442-109aabbf75c9" codex.op="user_input"}:turn{otel.name="session_task.turn" thread.id=019d81cf-6f69-7230-98aa-74294ff2dc5a turn.id=019d86c8-0a8e-7013-b442-109aabbf75c9 model=gpt-5.4}: ignoring interface.defaultPrompt: prompt must be at most 128 characters path=/Users/jif/.codex/.tmp/plugins/plugins/build-ios-apps/.codex-plugin/plugin.json
2026-04-13T12:27:30.406Z WARN [019d81cf-6f69-7230-98aa-74294ff2dc5a] codex_core::plugins::manifest - session_loop{thread_id=019d81cf-6f69-7230-98aa-74294ff2dc5a}:submission_dispatch{otel.name="op.dispatch.user_input" submission.id="019d86c8-0a8e-7013-b442-109aabbf75c9" codex.op="user_input"}:turn{otel.name="session_task.turn" thread.id=019d81cf-6f69-7230-98aa-74294ff2dc5a turn.id=019d86c8-0a8e-7013-b442-109aabbf75c9 model=gpt-5.4}: ignoring interface.defaultPrompt: prompt must be at most 128 characters path=/Users/jif/.codex/.tmp/plugins/plugins/life-science-research/.codex-plugin/plugin.json
2026-04-13T12:27:30.408Z WARN [019d81cf-6f69-7230-98aa-74294ff2dc5a] codex_core::plugins::manifest - session_loop{thread_id=019d81cf-6f69-7230-98aa-74294ff2dc5a}:submission_dispatch{otel.name="op.dispatch.user_input" submission.id="019d86c8-0a8e-7013-b442-109aabbf75c9" codex.op="user_input"}:turn{otel.name="session_task.turn" thread.id=019d81cf-6f69-7230-98aa-74294ff2dc5a turn.id=019d86c8-0a8e-7013-b442-109aabbf75c9 model=gpt-5.4}: ignoring interface.defaultPrompt: prompt must be at most 128 characters path=/Users/jif/.codex/.tmp/plugins/plugins/build-ios-apps/.codex-plugin/plugin.json
```
Addresses #16255
Problem: Incomplete Responses streams could leave completed custom tool
outputs out of cleanup and retry prompts, making persisted history
inconsistent and retries stale.
Solution: Route stream and output-item errors through shared cleanup,
and rebuild retry prompts from fresh session history after the first
attempt.
## Summary
When a `spawn_agent` call does a full-history fork, keep the parent's
effective agent type and model configuration instead of applying child
role/model overrides.
This is the minimal config-inheritance slice of #16055. Prompt-cache key
inheritance and MCP tool-surface stability are split into follow-up PRs.
## Design
- Reject `agent_type`, `model`, and `reasoning_effort` for v1
`fork_context` spawns.
- Reject `agent_type`, `model`, and `reasoning_effort` for v2
`fork_turns = "all"` spawns.
- Keep v2 partial-history forks (`fork_turns = "N"`) configurable;
requested model/reasoning overrides and role config still apply there.
- Keep non-forked spawn behavior unchanged.
## Tests
- `cargo +1.93.1 test -p codex-core spawn_agent_fork_context --lib`
- `cargo +1.93.1 test -p codex-core multi_agent_v2_spawn_fork_turns
--lib`
- `cargo +1.93.1 test -p codex-core
multi_agent_v2_spawn_partial_fork_turns_allows_agent_type_override
--lib`
## Summary
- add an exec-server `envPolicy` field; when present, the server starts
from its own process env and applies the shell environment policy there
- keep `env` as the exact environment for local/embedded starts, but
make it an overlay for remote unified-exec starts
- move the shell-environment-policy builder into `codex-config` so Core
and exec-server share the inherit/filter/set/include behavior
- overlay only runtime/sandbox/network deltas from Core onto the
exec-server-derived env
## Why
Remote unified exec was materializing the shell env inside Core and
forwarding the whole map to exec-server, so remote processes could
inherit the orchestrator machine's `HOME`, `PATH`, etc. This keeps the
base env on the executor while preserving Core-owned runtime additions
like `CODEX_THREAD_ID`, unified-exec defaults, network proxy env, and
sandbox marker env.
## Validation
- `just fmt`
- `git diff --check`
- `cargo test -p codex-exec-server --lib`
- `cargo test -p codex-core --lib unified_exec::process_manager::tests`
- `cargo test -p codex-core --lib exec_env::tests`
- `cargo test -p codex-core --lib exec_env_tests` (compile-only; filter
matched 0 tests)
- `cargo test -p codex-config --lib shell_environment` (compile-only;
filter matched 0 tests)
- `just bazel-lock-update`
## Known local validation issue
- `just bazel-lock-check` is not runnable in this checkout: it invokes
`./scripts/check-module-bazel-lock.sh`, which is missing.
---------
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: pakrym-oai <pakrym@openai.com>
Addresses #17313
Problem: The visual context meter in the status line was confusing and
continued to draw negative feedback, and context reporting should remain
an explicit opt-in rather than part of the default footer.
Solution: Remove the visual meter, restore opt-in context remaining/used
percentage items that explicitly say "Context", keep existing
context-usage configs working as a hidden alias, and update the setup
text and snapshots.
- Let typed user messages submit while realtime is active and mirror
accepted text into the realtime text stream.
- Add integration coverage and snapshot for outbound realtime text.
## Summary
- register flattened handler aliases for deferred MCP tools
- cover the node_repl-shaped deferred MCP call path in tool registry
tests
## Root Cause
Deferred MCP tools were registered only under their namespaced handler
key, e.g. `mcp__node_repl__:js`. If the model/bridge emitted the
flattened qualified name `mcp__node_repl__js`, core parsed it as an MCP
payload but dispatch looked up the flattened handler key and returned
`unsupported call` before reaching the MCP handler.
## Validation
- `just fmt`
- `cargo test -p codex-tools
search_tool_registers_deferred_mcp_flattened_handlers`
- `cargo test -p codex-core
search_tool_registers_namespaced_mcp_tool_aliases`
- `git diff --check`
Select Current Thread startup context by budget from newest turns, cap
each rendered turn at 300 approximate tokens, and add formatter plus
integration snapshot coverage.
## Summary
- update the guardian timeout guidance to say permission approval review
timed out
- simplify the retry guidance to say retry once or ask the user for
guidance or explicit approval
## Testing
- cargo test -p codex-core
guardian_timeout_message_distinguishes_timeout_from_policy_denial
- cargo test -p codex-core
guardian_review_decision_maps_to_mcp_tool_decision
**Summary**
This PR treats Guardian timeouts as distinct from explicit denials in
the core approval paths.
Timeouts now return timeout-specific guidance instead of Guardian
policy-rejection messaging.
It updates the command, shell, network, and MCP approval flows and adds
focused test coverage.
# External (non-OpenAI) Pull Request Requirements
Before opening this Pull Request, please read the dedicated
"Contributing" markdown file or your PR may be closed:
https://github.com/openai/codex/blob/main/docs/contributing.md
If your PR conforms to our contribution guidelines, replace this text
with a detailed and high quality description of your changes.
Include a link to a bug report or enhancement request.
Addresses #17302
Problem: `thread/list` compared cwd filters with raw path equality, so
`resume --last` could miss Windows sessions when the saved cwd used a
verbatim path form and the current cwd did not.
Solution: Normalize cwd comparisons through the existing path comparison
utilities before falling back to direct equality, and add Windows
regression coverage for verbatim paths. I made this a general utility
function and replaced all of the duplicated instance of it across the
code base.
