Claude Code chat ends, Celery workers die silent

The developer is building ALPHA FORGE, a Japanese-stock AI screening and continuous-learning tool, as a solo project alongside Claude Code. The article is a reconstruction of the 14th installment in his development-log series, and its three traps all share a feature: none of them produced an error.

The first trap was structural. Because the AI launched the worker processes as its own children, ending the chat session also ended the workers. The developer chose CLAUDE.md, the file the AI reads before each task, and wrote a WMI-based launch command so the parent would be the OS rather than the AI session.

The second trap was a stale, empty PID file dating from September 15. The script read its presence as proof the system was running and exited silently, while the auto-start task reported success. The fix, in code this time, was to check whether the recorded process was actually alive, and to return a real failure to the caller.

The third trap appeared on October 1, when the developer set up a cloud work environment. The API key the app uses to call an LLM and the key Claude Code uses for itself share the same standard environment variable name, so the developer feared the development tool's charges could land on the app's key. His answer was to store the app key under a different name and copy it only into the app's config file at launch.

The piece closes with a four-question framework for deciding where a lesson belongs: code and tests if a machine can enforce it, CLAUDE.md if the AI must know it before every task, skills for steps in a fixed procedure, and memory for the developer's own judgment habits, which he describes as something to maintain rather than grow. What the outcome hinges on, by his own account, is the human noticing at all — he caught the first failure the next morning and set the rule himself, while the AI traced the causes and proposed where to write each fix.