I Killed the Worker Mid-Batch. Zero Tickets Paid Twice for Claude. Part-3

I killed the Claude worker mid-batch. The job did not start from zero. Same Triage pipeline as parts 1 and 2. Same tickets. Same 800-character cut. Errors reported is still zero. Misses still paint red while work runs. What changed: the remaining tickets live in Redis, not in the worker process. Claim, call Claude, checkpoint. Kill sends SIGTERM. Resume starts a new pid. DONE is skipped. In-flight is replayed. Claude is not paid twice for a ticket that already acked. Repo (this lesson is v3_resumable): https://github.com/ekb-dev-ai/claude https://github.com/ekb-dev-ai/claude/... Part 1: the job finished. Gold was wrong. Part 2: I traced every call. Misses showed up live. Part 3: the worker can die. The queue cannot. Next: I change one line in PROMPT. Quality either drops or it does not. That drop has to show up as a number on two MLflow runs, or we ship it blind. Chapters: 00:00 Intro 00:55 New queuing system 04:39 System Architecture 10:32 Code walkthrough 13:55 Part 4 Intro #claude #mlflow #redis #llmops #anthropic