GPT-5.6 Luna missing from Codex model catalog — working Ubuntu workaround using

[C-NHXT-N]

GPT-5.6 Luna missing from Codex: findings and a working Ubuntu workaround

I want to document a workaround that successfully restored GPT-5.6 Luna on my Ubuntu Codex CLI setup, and also correct an earlier conclusion I made while troubleshooting this issue.

Summary

Several users, including me, have recently seen GPT-5.6 Luna disappear from the Codex model selector.

In my case:

• GPT-5.6 Luna disappeared from the normal /model picker.
• The same thing happened on both my remote Ubuntu server and my untouched Windows 11 PC.
• Both machines use the same account.
• I only modified the Ubuntu server.
• I was able to restore Luna locally on Ubuntu by creating a custom model_catalog_json.
• Restarting an old Codex app-server was necessary for my local workaround to fully take effect.

Important

I originally thought the stale app-server was the cause of Luna disappearing.

I no longer believe that is correct.

The stale app-server only prevented my custom local model catalog from taking effect immediately. The original disappearance of Luna appears to be a separate catalog/account-side issue.

---

What I observed

The normal Codex model picker stopped showing Luna.

My current model list looked like this:

GPT-6 Astra
GPT-5.6 Sol
GPT-5.6 Terra
GPT-5.5

GPT-5.6 Luna was missing.

I also checked a completely separate Windows 11 machine using the same account.

I made no changes on that machine:

• no custom model catalog
• no cache modifications
• no config.toml changes
• no app-server changes

Luna was missing there as well.

Because two independent environments showed the same catalog change, I do not think the original disappearance can be explained by a Linux-only cache or stale-process issue.

---

Environment used for the workaround

The workaround below was successfully tested on:

OS: Ubuntu / Linux
Codex CLI: 0.155.1

I have not yet tested this workaround on Windows 11.

---

Workaround

1. Check the current downloaded model catalog

Run:

jq -r '.models[] | [.slug, .display_name] | @tsv' \
~/.codex/models_cache.json

In my case, the catalog contained models such as:

gpt-6-astra
gpt-reserve
gpt-5.6-sol
gpt-5.6-terra
gpt-5.5
codex-auto-review

but did not contain:

gpt-5.6-luna

---

2. Back up your configuration

Before changing anything, make backups:

cp ~/.codex/models_cache.json \
   ~/.codex/models_cache.json.before_luna_fix.bak

cp ~/.codex/config.toml \
   ~/.codex/config.toml.before_luna_fix.bak

---

3. Do not modify models_cache.json directly

I first tried manually adding Luna to:

~/.codex/models_cache.json

That worked temporarily, but Codex later refreshed the model cache and overwrote the modification.

So editing models_cache.json directly is not persistent.

Instead, I created a separate custom catalog.

---

4. Create a custom model catalog containing Luna

I generated a new catalog from the current cache and added a Luna entry based on the existing GPT-5.6 Terra entry:

jq '
if any(.models[]; .slug=="gpt-5.6-luna") then
  .
else
  .models += [
    (
      .models[]
      | select(.slug=="gpt-5.6-terra")
      | .slug="gpt-5.6-luna"
      | .display_name="GPT-5.6-Luna"
      | .description="Fast and affordable agentic coding model."
      | .priority=8
      | .supported_reasoning_levels |= map(select(.effort!="ultra"))
    )
  ]
end
' ~/.codex/models_cache.json > ~/.codex/models_with_luna.json

Validate the generated JSON:

jq empty ~/.codex/models_with_luna.json && echo "JSON OK"

Then verify the Luna entry:

jq '.models[] | select(.slug=="gpt-5.6-luna") |
{
  slug,
  display_name,
  description,
  default_reasoning_level,
  supported_reasoning_levels,
  visibility,
  priority,
  context_window,
  max_context_window
}' ~/.codex/models_with_luna.json

The resulting entry should include:

slug: gpt-5.6-luna
display_name: GPT-5.6-Luna
visibility: list
priority: 8

and reasoning levels:

low
medium
high
xhigh
max

---

5. Configure Codex to use the custom catalog

Edit:

~/.codex/config.toml

and add the following as a top-level setting:

model_catalog_json = "/home/YOUR_USERNAME/.codex/models_with_luna.json"

Replace YOUR_USERNAME with your actual Linux username.

For example:

model_provider = "openai"
model = "gpt-5.6-luna"
model_reasoning_effort = "medium"

model_catalog_json = "/home/YOUR_USERNAME/.codex/models_with_luna.json"

Setting:

model = "gpt-5.6-luna"

is optional, but useful if you want Luna to be the default model while testing.

---

6. Verify that Codex loaded the custom catalog

Run:

codex debug models | jq -r \
'.models[] |
[.slug, .display_name, .visibility, (.priority|tostring)] |
@tsv'

After configuring the custom catalog, my output included:

gpt-5.6-luna    GPT-5.6-Luna    list    8

This is an important checkpoint.

• models_with_luna.json contains Luna
• config.toml points to the custom catalog
• codex debug models reports gpt-5.6-luna
• Restart any stale Codex app-server
• Start Codex again and verify Luna is available

If codex debug models does not show Luna, the custom catalog has not been loaded correctly yet.

---

7. Check for an old Codex app-server

In my environment, codex debug models could already see Luna, but an older Codex app-server was still running from before I changed the configuration.

Check only processes belonging to your own user:

ps -u "$USER" -o pid,ppid,lstart,cmd \
| grep -E '[c]odex( |$)|[c]odex-code-mode-host'

I found a process similar to:

codex -c features.code_mode_host=true app-server --listen unix://

with a corresponding:

codex-code-mode-host

If the app-server was started before the catalog change, terminate your own old process:

kill <PID>

Then verify that it exited:

ps -u "$USER" -o pid,ppid,lstart,cmd \
| grep -E '[c]odex( |$)|[c]odex-code-mode-host'

Warning

On a shared server, do not blindly run:

pkill codex

That may terminate Codex processes belonging to other users. Check the process owner and terminate only your own process.

---

8. Start Codex again

Start a fresh Codex process:

codex

After restarting Codex with the new catalog loaded, GPT-5.6 Luna became available again on my Ubuntu server.

That is the workaround that worked for me.

---

What I believe is happening

My current understanding is:

GPT-5.6 Luna disappears from the current model catalog
                         ↓
      Luna disappears from the normal model picker
                         ↓
       Create a custom local model catalog
                         ↓
             Add Luna to that catalog
                         ↓
      Configure model_catalog_json in config.toml
                         ↓
         codex debug models can see Luna
                         ↓
      Restart the old app-server if necessary
                         ↓
          Start a fresh Codex process
                         ↓
             Luna is available locally

There are therefore two separate issues.

1. The original disappearance

Luna appears to be missing from the current model catalog received by some users/accounts.

My untouched Windows 11 machine shows the same absence as my Ubuntu server, which makes a Linux-specific local cache problem unlikely.

2. The Ubuntu workaround

A custom model_catalog_json can locally add Luna back to the model catalog.

If an older app-server is still running, it may need to be restarted before the new local catalog is fully reflected.

---

Correction to my earlier conclusion

My earlier troubleshooting gave too much importance to the stale app-server.

The more accurate conclusion is:

The stale app-server did not cause GPT-5.6 Luna to disappear.

It only prevented my locally modified model catalog from taking effect immediately.

The upstream catalog issue and the local workaround should be treated as separate things.

---

Important caveat

This is an unofficial local workaround.

The Luna entry in this workaround is constructed locally using the current GPT-5.6 Terra catalog entry as a base, with Luna-specific fields changed.

It is not an official Luna catalog entry downloaded from OpenAI.

There may therefore be metadata differences that this workaround does not reproduce.

Please keep backups before trying it.

To restore the original configuration:

cp ~/.codex/config.toml.before_luna_fix.bak \
   ~/.codex/config.toml

---

Information that may help investigate the upstream issue

If you are experiencing the same problem, it would be useful to share:

OS:
Codex version:
Does Luna appear in /model?
Does Luna appear in model/list?
Does codex -m gpt-5.6-luna work?
Can an older existing Luna session still use the model?

If anyone from OpenAI can clarify whether GPT-5.6 Luna was intentionally removed from the current Codex model catalog for some accounts, or whether this is a temporary catalog/rollout issue, that would be very helpful.

Hopefully this workaround helps others who are seeing the same issue.