@myx I honestly don’t do near enough debug of log messages for that. With my simple requirements and low use, I just rely on my /skill to steer the agent. I do get it document everything it finds though, and I use human-in-the-loop (me) to review its theories. If it is complicated then I get the agent to dump logs around the error (say +5/-50 lines) into a log file too. Probably worth noting too that I have a decent CLAUDE.md for base info, and always use “rules” to guardrail it. The Anthropic Opus model is smart enough to not make craziness happen, but having the whole thing spit out ideas first is huge, by not trusting it I never get frustrated (just sometimes LOL at its naivety.) My baseline investigate skill: https://github.com/OttawaCloudConsulting/agentic-ai/blob/main/commands/investigate.md