Codex model metadata not found: what to check

How to tell this is your situation

• The CLI prints a missing-model-metadata warning followed by fallback wording.
• The selected API-key model appears as Custom in the IDE, as described in issue #34739.
• You need to check whether the request completed: that reporter received successful answers despite the warning.
• A separate rejection, capacity message or broken stream may also be present. Preserve that message rather than treating metadata as its proven cause.

The July report names Linux CLI 0.145.0 and VS Code extension 26.715.61943. Issue #12100 records similar wording on Windows CLI 0.104.0 with a custom provider. These builds and model names describe historical evidence; they are not recommendations for your current setup.

Step-by-step checks

1. Separate the warning from the result. Save the warning and exact model identifier. If practical, try one short request that does not change files. Record whether it completes, fails, or remains pending. Do not infer failure solely from the warning.
2. Identify the affected client. Record its version, operating system and sign-in method. If you already use both CLI and IDE, note whether the same situation occurs in each. You do not need to create another account or install another client for this comparison.
3. Read the selected model and provider. The configuration reference defines model as the selected identifier and model_provider as the provider ID, defaulting to openai. Use the configuration-location guide to find the relevant configuration. Inspect values without sharing credentials.
4. Record any catalog override without inventing one. Current documentation describes optional model_catalog_json, loaded at startup, with a selected profile-file override. A contributor comment says this setting was added after 0.104.0 for 0.105.0. Do not expect current configuration features in the older reported build. An override’s presence is not proof of the cause.
5. Route an actual failure separately. A response saying the model is unsupported belongs to the model-access guide. If your configuration itself produces parsing errors, use the config.toml troubleshooting guide. A model label, a warning and a rejected request are different observations.

If it is still not working

• Collect the warning, subsequent error and short-request outcome together. Reporting only the warning omits the evidence needed to investigate a failed request.
• Describe the model/provider/client combination and whether a catalog override exists. Share only relevant redacted settings, not an authentication file, complete configuration or API key.
• Check the linked issue discussion for an applicable explanation before making changes. Neither a closed report nor another user’s catalog workaround establishes a repair for every model and build. Preserve your current configuration before any later experiment.

FAQ

Does this warning mean my model request failed?

Not necessarily. The July API-key reporter received successful answers while seeing the warning and a Custom label. That observation does not guarantee your session works. Check your own request outcome and keep any separate failure message.

Should I delete the model cache or copy someone else’s catalog?

This guide does not recommend either action. The sources do not establish a universal repair. Record your existing model, provider and catalog settings first; a copied catalog may describe a different model or environment. Do not fabricate metadata just to remove a warning.

Why does model_catalog_json not help on CLI 0.104.0?

A repository contributor stated that the setting was added after that release and would be included in 0.105.0. That historical feature boundary explains why the old build cannot be assumed to support it. It does not identify a version that fixes every fallback warning.

Known limits of this guidance

The behavior is reported, not reproduced here. The sources do not measure performance loss in your session or establish a universal cause, repair or warning-free version. Model availability and account eligibility must be checked separately from these historical examples.

This guide is part of our Codex sessions troubleshooting.

First check whether this is a Codex service-side problem → View live Codex status.