MCP sub-tools not callable - server injected under bare namespaced name and marked unsupported

Summary

The node_repl MCP server is injected to the model as a top-level tool under its bare namespaced name mcp__node_repl and marked "unsupported". The model cannot reach the server's actual sub-tools (js, js_reset, js_add_node_module_dir). Calling the bare name with empty args {} returns unsupported call.

Repro

• Reproduced on desktop build OpenAI.Codex_26.721.4979.0 (ChatGPT.exe -> codex.exe -c features.code_mode_host=true app-server) and on the standalone CLI codex.exe.
• Two different models (session model + TUI claude-opus-4-8) both fail the same way (bare name mcp__node_repl + empty args {} -> unsupported call).

Server is healthy

Connecting directly to node_repl.exe's MCP js tool executes real JS:

• nodeRepl.write(1+1) -> 2

So the process / pipe / handshake / schema are all fine. The defect is in the client-side tool injection / name-mapping layer.

Expected

The client should expose an MCP server's sub-tools (e.g. js) as callable mcp__server__tool entries, instead of injecting the server's bare name and marking it unsupported.

Environment

• OS: Windows
• Desktop build: OpenAI.Codex_26.721.4979.0
• CLI: codex.exe
• CODEX_HOME: G:\AI-Global\codex-home