Codex CLI 0.147.0 Sends Empty Tool Description to Azure Responses API

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #37422
• #37380
• #37432

Powered by Codex Action

Duplicate of #37425

downgrade 0.146.1 and working fine

npm install -g @openai/codex@0.146.1

Same issue, what's the solution?:

{
  "error": {
    "message": "Invalid 'input[0].tools[0].description': empty string. Expected a string with minimum length 1, but got an empty string
instead.",
    "type": "invalid_request_error",
    "param": "input[0].tools[0].description",
    "code": "empty_string"
  }
}

facing the same issue here

+1 Same issue, reproducible on 0.147 by sending any message

I just checked codex v0.148.0-alpha.9 : the problem still exists there. For now, use npm install -g @openai/codex@0.146.1

I had faced same issue and the older version did help.
However, according to the error message which can be address to the parameter in models.json. The parameter is called "use_responses_lite". The default value is true, and all you have to change is to false.
This works, now I successfully update openai codex to the newer version(>=0.147). Hope this would help!

can confirm that turning off "use_responses_lite" works on 0.147

@oferlaor Thanks. To spell out what I did:

• Identify your Codex directory
• By default the directory will be ~/.codex/.
• However, like some others, I prefer to change it to ~/.local/share/codex by using the CODEX_HOME environment variable. Your choice.
• Clone the codex source code
• Make sure you checkout the right tag (e.g. rust-v0.147.0)
• copy its models.json to your local Codex directory
• Edit your models.json
• change use_responses_lite from true to false for the various gpt-5.6 models you use
• Edit your config.toml
• Change model_catalog_json to point to your local copy of models.json

```toml
model = "gpt-5.6-sol"
model_provider = "azure"
model_catalog_json = "~/.codex/models.json"

[model_providers.azure]
name = "azure"
base_url = "https://<resource>.cognitiveservices.azure.com/openai/v1"
wire_api = "responses"
env_key = "AZURE_OPENAI_KEY"
```

• To help future you, add a note to your Codex directory explaining what you did

One big request: Please sit down with a pen and paper for _at least_ 15 minutes and consider each of the following claims to the best of your ability:

• It is wiser and more ethical to _not_ use AI to improve AI _capabilities_.
• Given some degree of agency and ability, it is our responsibility to help improve ourselves, our families, and/or our local communities, as best we can.
• The above point is a great start. If enough of us do this, we can help nudge the world in a better direction.
• We want humans to better understand each other and to solve problems peacefully.
• We want to make better decisions.
• Can you responsibly use AI to help more than hurt? I think you can find a way. If you can't, stop using AI and do something else.
• The current AI labs _do not have things under control_. Anthropic might have a better safety track record at present, and this is good, but no current labs are taking safety seriously enough.
• Don't bash any company for disclosing problems.
• Don't become "that guy" who simply says everything an AI lab does "is for marketing purposes". Doing this is intellectually lazy. The world is not that simple. Trust me, I'm no optimist, and I have seen dysfunction over and over again, but there _are indeed_ people there trying to do the right thing, even if their company as a whole is off-track. Listen to them.
• Learning about what _alignment_ means is essential. Learning about why is it _hard_ will help us to avoid seeking half-assed "solutions".
• The onset of superintelligence has at least a good chance of ending humanity, one way or another, with or without malevolence. You don't have to agree with Yudkowsky and Soares (who put the probability above 90%) to take their concerns seriously.
• I agree with Yudkowsky and Soares. I've been following the arguments closely for several years. I have training and experience in electrical engineering, liberal arts, statistics, economics, computer science, machine learning, and my recent specialization is AI safety.
• According to the Future of Life Institute, "Three‑quarters of U.S. adults want strong regulations on AI development, preferring oversight akin to pharmaceuticals rather than industry "self‑regulation." PDF.
• Read Tech employees call for US-backed global effort to manage risks of advanced AI
• Can I pick something specific to do? (If not _me_, then who?)

P.S. I took the time to write up these detail, so please forgive me if you disagree with my points at the end. It is irresponsible for people to just "shut up and do their work" now. Just like it was unethical for physicists to work on the Manhattan Project when their ethics said it was wrong.

+1，When will this be fixed? v26.810.41047 cannot
After downgrading my Windows plugin to 26.803.61601, it works. The CLI version should be v0.147.0-alpha.6.5

<img width="877" height="159" alt="Image" src="https://github.com/user-attachments/assets/7286c60c-b320-4c60-81da-8b283e56cabe" />

@wssf812 This is just a tooling comment, apologies if it is not helpful or I'm missing something... If that screenshot reflects what you see, perhaps your terminal emulator is causing your glyphs to not line up. It doesn't have to look that bad... :) Have you considered using a modern terminal emulator like Ghostty (my favorite) or Alacritty (also good)?

Same root cause as #37952 (also #37692, #37675): default_namespace_description deliberately returns an empty string for the default functions namespace (https://github.com/openai/codex/blob/1f41cc5d92/codex-rs/tools/src/responses_api.rs#L64-L70); OpenAI's backend tolerates it, Azure's validator rejects both the empty string and the field's absence. One-line fix discussed in #37952 — this looks like a duplicate.

0.148 still has this issue.

same here