Interviewer stops responding mid-session; ending the interview manually produces no evaluation (Gemini 503)

[hengcsie]

Description

During a "Coding + behavioral" interview, the AI interviewer stopped responding after I finished speaking. I waited, but no reply came, so I ended the interview myself. The resulting report contained no evaluation and was marked INCOMPLETE.

Steps to reproduce

1. Start an interview: Coding + behavioral, problem implement-trie-prefix-tree
2. Work through the coding round in Python while talking with the interviewer
3. After I finished speaking (~22 min into the coding round), the interviewer never responded
4. End the interview manually
5. Report shows "No evaluation" / INCOMPLETE

Error message (from the report)

The automatic evaluation could not be completed: Rust LiveKit runner ended (candidate_ended) but Gemini report generation failed for model gemini-3.1-flash-lite on implement-trie-prefix-tree. Final editor state: 388 bytes of python. Error: HTTP status server error (503 Service Unavailable) for url (https://generativelanguage.googleapis.com/v1beta/models/gemini-3.1-flash-lite:generateContent).

Report details

• Contract bundle 12 · rubric 1 · report schema 2
• Interview rounds:
  • coding · 22 min · incomplete
  • behavioral · 8 min · skipped
• Integrity evidence:
  • SESSION_START 2026-09-24T14:52:30.237Z
  • MEDIA_PREFLIGHT_PASSED 2026-09-24T14:52:30.240Z

Expected behavior

• The interviewer responds after the candidate finishes speaking.
• If the interviewer cannot respond (e.g. upstream API error), the UI shows an error or retry option instead of silently waiting.
• Ending the interview early still produces an evaluation of the work done so far.

Actual behavior

• The interviewer went silent with no error or status shown in the UI.
• After ending the session manually, report generation failed with 503 and no scores were produced.

Environment

• OS: Windows 11 Home 25H2 (OS Build 26200.9457)
• Browser: Google Chrome 153.0.8010.53 (Official Build, 64-bit)

Possible improvement

• Detect when the interviewer fails to respond (timeout or upstream error) and notify the candidate.
• Retry transient Gemini 5xx errors with backoff, both during the conversation and for report generation.
• Allow report generation to be re-triggered from the saved session without rerunning the interview.

[scbsns123-ctrl]

Update: Replacing the API key resolved the issue in my case

I encountered a similar issue( failed with 503) and was able to resume using CodeTrial after replacing my Google Gemini API key.

What I did:

1. Deleted the old API key and created a new one.
2. Opened config/codetrial.env.local inside the CodeTrial folder.
3. Replaced the existing API key value with the new key and saved the file.

After replacing the key, CodeTrial worked again in my subsequent attempt.

This appears to be a workaround in my case, but I have not confirmed the root cause. Since the reported error was 503 Service Unavailable, the recovery could also have coincided with the upstream service becoming available again.

This does not confirm that the original incomplete report can be recovered. An in-app error message and a retry option would still help users understand and recover from this situation.

[jserv]

#95 reports the same thing: the interviewer stops replying mid-session and doesn't come back. Open PR #202 targets it. When a socket stops answering, it is replaced, the room shows "Thinking...", and the interview ends with a report instead of nothing. The 503 on report generation is #97, which already has several reproductions and tracks retrying a report without rerunning the interview. Please continue in #95, and if you can, try the PR #202 branch the next time the interviewer goes quiet.