[Bug]: Codex per-thread mcp_servers override replaces MCP servers added with -c launch args #15828

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

1. Add a Codex provider instance whose launch args add a stdio MCP server, e.g. -c mcp_servers.extra.command="/path/to/server" (plus any args).
2. Start a thread on that instance.
3. Ask the agent to list its tools, or check which MCP servers Codex started.

Expected behavior

The thread gets T3's t3-code MCP server and the extra server from the instance's launch args, the same way a Claude instance with an extra --mcp-config gets both.

Actual behavior

Only t3-code is available. The extra server is never started.

codexThreadRuntimeParams (apps/server, Codex provider) sends:

config: {
  ...CODEX_THREAD_CONFIG,                       // { "tools.update_plan.enabled": true }
  mcp_servers: { "t3-code": { url, http_headers } },
}

Codex applies per-thread overrides after the process's -c flags, in the same session-flags layer. A single-segment key (mcp_servers) replaces the whole table those -c mcp_servers.extra.* flags built. Servers in ~/.codex/config.toml survive, because that is a lower layer and layers are deep-merged. This happens on every thread/start, thread/resume and thread/fork, so no launch-arg quoting can work around it.

Suggested fix (not tested)

Use a dotted key, as CODEX_THREAD_CONFIG already does for tools.update_plan.enabled:

"mcp_servers.t3-code": { url, http_headers },

That should set only T3's entry and keep other servers.

Impact

A Codex provider instance can't add its own MCP server through launch args, though Claude instances can with --mcp-config. The remaining options are editing the global ~/.codex/config.toml (which affects every Codex use on the machine) or wrapping the Codex binary to re-add the server.

Version or commit

T3 Code 0.0.46-nightly.20261003.2638 (Linux AppImage)

Environment

Fedora Linux, codex-cli 0.160.0, provider instance with custom launch args.

Workaround

A wrapper executable as the instance's binaryPath that re-adds the server, or a global ~/.codex/config.toml entry.