The observability gap is absolutely something that I have faced. I've resorted to --verbose and reading terminal scrollback to understand why Claude made certain decisions. The graph view sounds like the right abstraction - tool calls form a natural DAG. Does it capture the reasoning between tool calls, or just the calls themselves? The reasoning chain is usually where I lose my thread!
Right now reasoning shows up as its own nodes in the graph alongside the tool calls so you can follow the full chain what claude was thinking then what it did. Still early and there's room to improve how it's presented, but the core is there.