Developer describes daily workflow using Claude Code agent with manual review for open-source Go projects

The author runs Claude Code with broad permissions inside a virtual machine for isolation, manages up to 3 parallel tmux sessions for different tasks, and uses a tool called caveman to reduce Claude's verbose narration and token count.

For feature development on MetalLB and OpenPerouter projects, the author writes detailed specs upfront, uses diffity to review generated code with annotations, and for larger features enables CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 to run 3 coding agents in parallel alongside 1 reviewer and 1 QE agent — but always reviews the initial output and iterates through CI failures and GitHub review feedback before merging.

To handle CI test failures, the author built artifactsdownloader to fetch artifacts from failed PR runs and uses Claude to correlate logs across multiple files and identify probable root causes, even when the proposed explanation proves wrong; for PR triage, the author uses CI classification to summarize incoming PRs and flag structural issues rather than reading each one cold.