Building a Linear-Driven Agent Loop with Claude Code

Running agent loops inside Claude Code causes context rot. This bash-based approach starts each iteration with a fresh context window, pulling issues from Linear, building features, running code review with subagents, and opening pull requests autonomously.