gpt-5.6-sol does not receive tool_search, leaving deferred MCP tools unreachable

What version of the Codex App are you using (From “About Codex” dialog)?

Version 26.707.91948

What subscription do you have?

Pro plan 20x

What platform is your computer?

Darwin 24.6.0 arm64 arm

What issue are you seeing?

Codex App bug report: gpt-5.6-sol cannot load deferred MCP tools

What issue are you seeing?

Codex Desktop can connect to a configured MCP server, but a fresh
gpt-5.6-sol task is not given the native tool-loading path needed to call the
server's deferred tools.

I reproduced this with OpenAI's public Developer Docs MCP server, so the
reproduction does not depend on a private server, OAuth, a third-party plugin,
or project-specific code.

The affected task's telemetry indicates that MCP tools are deferred. In the
gpt-5.6-sol task:

• The configured MCP server is available to Codex.
• Its callable functions are not included directly in the model-facing tool

list.

• tool_search is not included in the session.
• The model therefore cannot discover or load the deferred MCP function tools.
• Generic list_mcp_resources and list_mcp_resource_templates calls do not

provide access to a tool-only server's functions.

• The model eventually bypasses Codex's native MCP tool interface by sending

raw MCP JSON-RPC requests to the public endpoint with shell curl commands.

There is no explicit runtime error message. The failure is the absence of both
the deferred MCP functions and the tool_search loader from the model-facing
toolbox.

The equivalent gpt-5.5 task works normally:

• It receives tool_search.
• It searches for the OpenAI Developer Docs tools.
• Codex loads the mcp__openaiDeveloperDocs namespace.
• The model calls search_openai_docs and fetch_openai_doc through the

native MCP interface.

The prompt, workspace, Codex installation, MCP configuration, and server were
the same. Only the selected model changed.

Environment:

• Embedded Codex runtime: 0.144.5
• Platform: Darwin 24.6.0 arm64 arm
• MCP transport: Streamable HTTP
• MCP server: https://developers.openai.com/mcp

This appears to be a Codex Desktop task/tool-exposure defect affecting
gpt-5.6-sol, rather than an MCP server defect.

Related reports:

• https://github.com/openai/codex/issues/19425
• https://github.com/openai/codex/issues/19649

What steps can reproduce the bug?

1. Configure OpenAI's public Developer Docs MCP server:

``toml
[mcp_servers.openaiDeveloperDocs]
url = "https://developers.openai.com/mcp"
``

The equivalent CLI command is:

``sh
codex mcp add openaiDeveloperDocs --url https://developers.openai.com/mcp
``

1. Restart Codex Desktop.

1. Start a fresh task using gpt-5.5.

1. Submit this exact prompt:

``text
Use the OpenAI Developer Docs MCP server to search for Codex MCP configuration documentation.
``

1. Observe the native tool path:

``text
tool_search
-> loads mcp__openaiDeveloperDocs
-> search_openai_docs
-> fetch_openai_doc
-> successful response
``

1. Start another fresh task in the same workspace using gpt-5.6-sol.

1. Submit the identical prompt.

1. Observe that the task receives neither tool_search nor the deferred

mcp__openaiDeveloperDocs function tools.

1. In the reproduced task, the model tries generic MCP resource calls and then

works around the missing native tool path with commands equivalent to:

``sh
curl -sS -X POST https://developers.openai.com/mcp \
-H 'Content-Type: application/json' \
-H 'Accept: application/json, text/event-stream' \
--data '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'
``

It then sends a raw tools/call request through curl. This proves that
the public MCP endpoint is reachable, but it is not a successful native
Codex MCP tool invocation.

Session evidence

Working gpt-5.5 session:

019f6b3e-86fc-73a2-ad6a-6b658338822d

Affected gpt-5.6-sol session:

019f6b3f-2a0e-7c70-9814-b4c288fee944

Token and context usage

Both were fresh, single-prompt tasks with a model context window of 258,400
tokens.

First recorded token-count event:

• gpt-5.5: 21,300 / 258,400 tokens, approximately 8.2%
• gpt-5.6-sol: 17,979 / 258,400 tokens, approximately 7.0%

Final model request:

• gpt-5.5: 65,602 / 258,400 tokens, approximately 25.4%
• gpt-5.6-sol: 61,620 / 258,400 tokens, approximately 23.8%

No context compaction, context-window exhaustion, or token-limit failure
occurred in either task.

What is the expected behavior?

When Codex defers MCP function tools, every supported model session should
receive a working mechanism for discovering and loading those tools.

For gpt-5.6-sol, Codex should do one of the following:

• Provide tool_search, allowing the model to discover and load the deferred

MCP namespace; or

• Expose the applicable MCP function tools directly when tool_search is not

available.

With the same configuration and prompt, gpt-5.6-sol should follow the same
native path demonstrated by gpt-5.5:

tool_search
  -> load mcp__openaiDeveloperDocs
  -> call search_openai_docs
  -> call fetch_openai_doc
  -> return the documentation result

The model should not need to inspect MCP configuration, call generic resource
APIs, or manually recreate the MCP protocol with shell curl commands.

Additional information

<img width="874" height="1152" alt="Image" src="https://github.com/user-attachments/assets/efa8d2e7-7a99-430c-ba13-181559f04e1a" />

<img width="935" height="1156" alt="Image" src="https://github.com/user-attachments/assets/0b637c3d-da17-4611-a203-91c187b79753" />

First discover when building a mcp plugin for our platform:

<img width="910" height="975" alt="Image" src="https://github.com/user-attachments/assets/b4b7a5eb-8c86-463c-96c7-c00239ddc5b8" />

<img width="965" height="1031" alt="Image" src="https://github.com/user-attachments/assets/eb551926-584f-4f42-95ca-abb7eca186d6" />