Codex¶
Codex speaks the Responses API, which the server answers at /v1/responses.
npm install -g @openai/codex # install
superfluid launch codex --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_M
superfluid launch codex -- exec "fix the failing test"
Aliases: codex-cli.
What launch gives it¶
| given | what for |
|---|---|
-c overrides |
a superfluid provider: base_url the server's /v1, wire_api = "responses", the key read from SUPERFLUID_LAUNCH_KEY; web_search = "disabled", since the server runs function tools only |
-m <model> |
the served model |
$SUPERFLUID_HOME/launch/codex-models.json |
a one-model catalog (model_catalog_json) carrying the server's context window |
Codex sends its sub-agent and MCP tools as namespace tools, which the Responses route flattens to namespace.member functions; see Responses API. The catalog names no freeform apply_patch tool, so Codex sends it as a function.