OpenAI GPT-5 verification fails in openshell inference set due to max_tokens request parameter #517

Summary

openshell inference set works for older OpenAI chat-completions models like gpt-4.1, but verification fails for GPT-5-family models like gpt-5.4 because the validation request appears to send max_tokens, which OpenAI rejects for that model family.

Environment

• OpenShell: 0.0.12
• Host: macOS (Apple Silicon workstation)
• Provider type: openai
• Base URL: https://api.openai.com/v1

Reproduction

Provider setup succeeds:

export OPENAI_API_KEY=...
openshell provider create --name openai-api --type openai --credential OPENAI_API_KEY --config OPENAI_BASE_URL=https://api.openai.com/v1

This succeeds:

openshell inference set --provider openai-api --model gpt-4.1

Output:

Gateway inference configured:

  Route: inference.local
  Provider: openai-api
  Model: gpt-4.1
  Version: 1
  Validated Endpoints:
    - https://api.openai.com/v1/chat/completions (openai_chat_completions)

This fails:

openshell inference set --provider openai-api --model gpt-5.4

Output:

Error:   × failed to verify inference endpoint for provider 'openai-api' and model 'gpt-5.4' at 'https://api.openai.com/v1': upstream rejected the validation request with HTTP 400 Bad Request. Response body: {
  "error": {
    "message": "Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.",
    "type": "invalid_request_error",
    "param": "max_token"

Expected behavior

openshell inference set --provider openai-api --model gpt-5.4 should validate successfully for supported OpenAI GPT-5-family models, or OpenShell should adapt the verification request to the model family requirements.

Actual behavior

Verification appears to send a request shaped for older chat-completions semantics (max_tokens), which GPT-5-family models reject.

Notes

• This does not appear to be a provider creation problem.
• The same provider and base URL work for gpt-4.1.
• As a workaround, --no-verify can be used, but that is not ideal for normal onboarding flows.

Likely fix area

OpenShell inference endpoint verification for OpenAI provider/model combinations likely needs model-aware request shaping for GPT-5-family models, specifically around max_completion_tokens vs max_tokens.