kwongfuk/codex-skills/codex-custom-provider-pack

Codex Custom Provider Pack

Use this skill when the user wants to:

• configure Codex CLI, desktop, or editor plugins against https://api.mol.us.ci/v1
• keep a maintained table of available hosted models
• remember local Hugging Face model cache paths under /scratch/gguo/.hf_cache/hub
• validate custom-provider credentials, list models, or generate small evaluation datasets

Guardrails

• Never store usernames, passwords, or API keys inside the skill, repo, or history log.
• If a secret was pasted into chat, treat it as exposed and recommend rotating it before routine use.
• Keep SKILL.md short. Read the reference files only when needed.

Workflow

1. Read references/package-guide.md for the package activation flow, token creation flow, and Codex configuration snippets.
2. Read references/model-registry.md for the hosted-model table and current local cache registry.
3. Use the helper script for automation:

python /home/gguo/code/.codex/skills/codex-custom-provider-pack/scripts/codex_provider_helper.py local-models
python /home/gguo/code/.codex/skills/codex-custom-provider-pack/scripts/codex_provider_helper.py list-models
python /home/gguo/code/.codex/skills/codex-custom-provider-pack/scripts/codex_provider_helper.py smoke-test --model gpt-5.4
python /home/gguo/code/.codex/skills/codex-custom-provider-pack/scripts/codex_provider_helper.py generate-data \
  --model gpt-5.4 \
  --input-file prompts.txt \
  --output-file outputs.jsonl

Notes

• local-models scans Hugging Face cache directories such as models--Qwen--Qwen3-VL-32B-Instruct.
• For local inference, prefer the resolved snapshots/<commit> directory instead of the cache root when a framework expects actual model files.
• For remote calls, the script searches for OPENAI_API_KEY in this order: CLI flag, environment variable, workspace .codex/auth.json, then ~/.codex/auth.json.
• Refresh references/model-registry.md when the local cache materially changes.