Dissecting Claude Code- The Reflexes
"What happens when things break. Exponential backoff with jitter, cascade protection that treats retry traffic as a shared resource, and the risks of resuming after an unclean shutdown

"What happens when things break. Exponential backoff with jitter, cascade protection that treats retry traffic as a shared resource, and the risks of resuming after an unclean shutdown

Three memory layers, three timescales. How Claude Code forgets gracefully, survives compaction, and the context engineering patterns that make it all work
Nine-step permission cascade, a classifier that uses Claude to police itself, and a 5-way race to decide who answers the permission prompt first

41 tools, one interface. How Claude Code keeps file reads, shell commands, and web searches coherent through a single polymorphic contract with input-dependent safety classification

"No hard depth limit. No explicit synthesis step. How Claude Code spawns autonomous sub-agents, isolates their state, and lets the model be its own coordinator.

Inside the while(true) loop that makes Claude Code an agent. Five exit conditions, seven re-entry points, and the architecture that separates a tool that answers from one that works.