2.6 KiB
2.6 KiB
name, description
| name | description |
|---|---|
| diagnose-codex-relay | Diagnose Codex relay connectivity and compatibility without exposing credentials. Use when a user asks to test a custom API base URL, list available models, verify the Responses API or SSE streaming, investigate first-token latency, or check whether web search, image generation, and function calling are available through a relay. |
Diagnose Codex Relay
Diagnose one layer at a time and report evidence instead of assuming a capability works.
Safety
- Never print, quote, save, or commit an API key.
- Redact authorization headers, query tokens, cookies, passwords, and signed URLs.
- Prefer read-only health checks and minimal requests.
- Ask before running sustained concurrency or bandwidth tests.
- Do not modify Codex configuration unless the user explicitly asks for a change.
Workflow
- Identify the configured API base URL without displaying its secret.
- Check DNS resolution, TLS validity, and basic HTTP reachability.
- Request
/v1/modelsand record status, elapsed time, and returned model IDs. - Send a minimal synchronous request to the requested model.
- Test
/v1/responsesseparately from Chat Completions. - Test SSE streaming and measure time to the first meaningful event.
- Test function calling, web search, or image generation only when the user requests it.
- Separate client, reverse-proxy, relay, and upstream latency when diagnosing delays.
- Report results as passed, failed, unsupported, or not tested.
Evidence
For each test, record:
- Endpoint category, without secret query parameters.
- HTTP status or transport error.
- Total elapsed time and first-event latency where applicable.
- Whether valid response content was received.
- Whether the result proves the claimed capability.
Do not claim web search succeeded without a returned search result or source. Do not claim image generation succeeded without a native image result or generated artifact.
Diagnosis Rules
/v1/modelssucceeding proves authentication and catalog access only.- A synchronous response does not prove SSE streaming works.
- Chat Completions success does not prove the Responses API works.
- A model name appearing in the catalog does not prove tool support.
- Compare direct relay access with the public domain when isolating reverse-proxy delay.
- When context length affects latency, compare short and long prompts separately.
Output
Return a compact table of tests, followed by:
- The most likely fault layer.
- Tests that were not performed.
- The smallest next action.
- A reminder to rotate any credential accidentally exposed during testing.