intractable diseases
When encountering a call failure, first determine the error type and then troubleshoot according to the path.
401 or 403
- Whether the API Key is copied completely.
- Whether the Key is disabled.
- Whether the account has available quota or permissions.
Model does not exist
- Is the model name spelled correctly?
- Whether the current account supports this model.
- Whether the client caches old model names.
Request timeout
- Is the local network stable?
- Whether proxy settings affect connections.
- Whether images or long context tasks require longer timeouts.
Codex has been working and has no output.
Common reasons:
- The command is still running. For example, the recursive search scanned large directories such as
node_modulesandbuild. - The background process has not ended, such as dev server, watch, test or interactive commands are still waiting.
- The command output is too large and the terminal is still processing a lot of logs.
- Network or package manager stuck, such as installing dependencies, accessing GitHub or registry.
Processing method: directly press Esc to interrupt, and then enter "continue". This won't hurt the code, at best it will interrupt the current command.
Common upstream errors
Claude Code 500 errors in a row
You can try adding in settings.json:
CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1
Selected model is at capacity
This type of error usually indicates a high load on the official service. You can enter "continue" or try again later.
AGENT_CRAFT_CORE_FINISH_REASON_ERROR or terminated
If there is any problem of not being able to connect to the service, please first go back to the official website User Center - [API Key] page and confirm that the Base URL filled in by the client is consistent with the page display.
No available accounts
Check the group icon corresponding to the API Key. If it is a group as shown in the figure below, please use the Anthropic protocol request.
502 or 503 service unavailable
There may be an account balance that runs out, and the Codex number will be temporarily stuck. Generally, it will detach after about ten minutes.
Too many pending requests
Too much concurrency, please reduce concurrency or try again later.
Upstream stream ended without a response
The return stream is truncated, usually due to network problems. In a few cases, it may be truncated by the large model security mechanism.
Zhipu prompt violates fair use policy
Don't worry, you can continue to use it, and we will continue to pay GLM.

Deduction question
Save the request time, model name and Key name, and go to the console to view usage records. If it's still unclear, contact support again.