codex-cli

Official page: https://openai.com/codex/

0.153.4

It provides support for GPT-6 Astra.

The node.js counterpart works as before.

module load ceuadmin/node/22.16.0
npm i -g @openai/codex@0.153.4 --prefix=$CEUADMIN/codex-cli/0.153.4

Note that from 0.132.0 there is also a standalone distribution as with claude-code/2.1.263.

cd "$CEUADMIN/codex-cli/0.153.4"
wget -qO- https://github.com/openai/codex/releases/download/rust-v0.153.4/codex-x86_64-unknown-linux-musl.tar.gz | \
tar xfz -
ln -s codex-x86_64-unknown-linux-musl codex
ln -sf \
  lib/node_modules/@openai/codex/node_modules/@openai/codex-linux-x64/vendor/x86_64-unknown-linux-musl/bin/codex-code-mode-host \
  codex-code-mode-host
sha256sum \
   /usr/local/Cluster-Apps/ceuadmin/codex-cli/0.153.4/codex-x86_64-unknown-linux-musl \
   /usr/local/Cluster-Apps/ceuadmin/codex-cli/0.153.4/lib/node_modules/@openai/codex/node_modules/@openai/codex-linux-x64/vendor/x86_64-unknown-linux-musl/bin/codex

as codex-code-mode-host is required by agentic programming. The two versions are seen exactly the same so the standalone version is unnecessary.

56ef98ab4032d317ab26e9b5e5a175650717351edb16ed9cde0cb6d1734d62da  /usr/local/Cluster-Apps/ceuadmin/codex-cli/0.153.4/codex-x86_64-unknown-linux-musl
56ef98ab4032d317ab26e9b5e5a175650717351edb16ed9cde0cb6d1734d62da  /usr/local/Cluster-Apps/ceuadmin/codex-cli/0.153.4/lib/node_modules/@openai/codex/node_modules/@openai/codex-linux-x64/vendor/x86_64-unknown-linux-musl/bin/codex

We are also interested in a local model, e.g., in our trusted folder named codex:

module load ceuadmin/ollama/0.32.13
ollama serve > /dev/null 2>&1 &
while ! curl -s http://localhost:11434/api/tags >/dev/null; do
  sleep 1
done
ollama list | (read header; echo "$header"; sort -f -k1,1)
cd codex
codex exec \
  --oss \
  --local-provider ollama \
  --model qwen3.5:27b \
  --skip-git-repo-check \
  "Inspect AGENTS.md and tell me which R files are the primary implementations. Do not modify anything." 2>/dev/null

where a simple task is furnished on our case-cohort agenetic benchmark, giving

Based on the AGENTS.md instructions, the primary R implementation files are:

• ccsize.R
• ccsize07.R

These contain the main implementations for:

• ccsize() — Cai & Zeng (2004), rare events
• ccsize07() — Cai & Zeng (2007), non-rare events

0.120.0

Web, https://www.npmjs.com/package/@openai/codex

This following is is in line with setup for OpenClaw, Pi, etc., which is more explicit.

module load ceuadmin/node/22.16.0
npm i -g @openai/codex@0.120.0 --prefix=$CEUADMIN/codex-cli/0.120.0

A profile is with ~/.codex/config.toml.

Integration with Ollama & llama.cpp

URL, https://docs.ollama.com/integrations/codex

[model_providers.ollama-launch]
name = "Ollama"
base_url = "http://localhost:11434/v1"

[profiles.ollama-launch]
model = "gpt-oss:120b"
model_provider = "ollama-launch"

[profiles.ollama-cloud]
model = "gpt-oss:120b-cloud"
model_provider = "ollama-launch"

which enables,

codex --profile ollama-launch
codex --profile ollama-cloud

The article recommeneds use of Ollama 0.20.5,

ollama launch codex --config
codex --oss -m gpt-oss:120b
codex --oss -m gpt-oss:120b-cloud
ollama pull gemma4:31b
codex --oss -m gemma4:31b

and llama.cpp with these options.

llama-server \
  -m gemma-4-26B-A4B-it-Q4_K_M.gguf \
  --port 1234 -ngl 99 -c 32768 -np 1 --jinja \
  -ctk q8_0 -ctv q8_0

The -np 1 limits to a single slot, because multiple slots multiply KV cache memory. The -ctk q8_0 -ctv q8_0 quantises the KV cache, reducing it from 940 MB to 499 MB. The –jinja flag is required for Gemma 4's tool-calling template. And -m with a direct path avoids the -hf flag, which silently downloads a 1.1 GB vision projector that causes an out-of-memory crash. The Codex CLI config also needs web_search = "disabled", because Codex CLI sends a web_search_preview tool type that llama.cpp rejects.