Use Codex

Codex CLI speaks the OpenAI Responses API. Current releases accept only wire_api = "responses". The server translates POST /v1/responses onto chat completions, including client function tools. input_image blocks become chat image_url parts, including images returned by a tool. Streamed turns send each token as it is generated. The server does not run server-side tools such as web search.

Use the recommended sequence length. Codex sends its instructions, skills, and tool list on every turn, about 14,000 tokens before the reply, and that profile leaves room for the prompt and a short completion. Change either flag if your prompt is larger or you want a smaller engine.

edgellm serve Qwen/Qwen3.5-4B \
  --max-input-len 32768 \
  --max-kv-cache-capacity 40960 \
  --enable-auto-tool-choice \
  --served-model-name qwen35-local \
  --port 8000

Point Codex at the server from ~/.codex/config.toml. Provider settings belong in that user config, not in a project .codex/config.toml.

model = "qwen35-local"
model_provider = "edgellm"

[model_providers.edgellm]
name = "TensorRT Edge-LLM"
base_url = "http://127.0.0.1:8000/v1"
env_key = "EDGE_LLM_API_KEY"
wire_api = "responses"

export EDGE_LLM_API_KEY=local
codex exec "Reply with the single word ready."

model must match --served-model-name. Set EDGE_LLM_API_KEY to any non-empty value when the server omits --api-key, and to the same value when the server sets --api-key. --enable-auto-tool-choice is required because Codex sends tool definitions.