Repository navigation
[Bug]: [Presets] A repeatedly dying agent retries forever instead of failing #4163
Description
Steps to reproduce
- Make the agent process die at startup without submitting a report, in a way that still prints one
assistantline. For example, with subscription auth supplied only by user-scope settings while the native OAuth session is expired: dstack spawnsclaudewith--setting-sources project,local(agent.py:356-357), which excludes the user scope, so the agent starts and dies immediately. - Start any preset creation:
dstack apply -f preset.dstack.yml -y.
Actual behaviour
Agent process exited without a report; resuming in 30s.
repeats every 30 seconds, indefinitely, before any instance is provisioned. Never 60s or 120s: agent.py:551-552 sets made_progress on any assistant line without inspecting its content, and agent.py:238-239 resets the retry budget (agent.py:52) before the pop at agent.py:246, so the budget never drains and the while True at agent.py:216 has no attempt cap. The agent's own error is captured at agent.py:555-556 but the loop message omits it, so the cause is only visible in .agent-stdout.jsonl. A report-less death that prints no assistant line does bound correctly and exits after 30/60/120s.
Expected behaviour
The creation fails immediately, printing the agent's error:
Failed to authenticate: OAuth session expired and could not be refreshed
Retries stay only for a transient death after real progress, bounded, printing the error each attempt.
Activity
[-][Presets] Agent startup and auth failures retry forever instead of failing[/-][+][Bug]: [Presets] Agent startup and auth failures retry forever instead of failing[/+][-][Bug]: [Presets] Agent startup and auth failures retry forever instead of failing[/-][+][Bug]: [Presets] A repeatedly dying agent retries forever instead of failing[/+]breken-ai commentedon Sep 9, 2026
Reproduced on current HEAD. The retry budget still resets on any assistant output, so the crash loop you described is still unbounded: an agent that prints one assistant line then dies (your expired-OAuth case) makes identical progress every attempt. The loop's
made_progressboolean can't distinguish "got further" from "got to the same place again". This changes it to an assistant-line count and only resets the budget when an attempt exceeds the best previous attempt, so a dying agent drains the 30/60/120s retries and returns the error while a genuinely progressing one still earns resets. Includes an end-to-end regression test with a fake claude process that dies after one assistant line: unbounded on main, exactly 4 attempts with the fix. PR: #4277Built by breken, your AI support engineer - breken.ai finds the bug, proves the fix with unmatched RCA, answers customers in your voice, and keeps your code and support docs clean - this one's on us.
github-actions commentedon Oct 10, 2026
This issue is stale because it has been open for 30 days with no activity.