v0.4.1

The agent now keeps its own task list, the TodoWrite pattern from Claude Code distilled into one 79-line file. todo_write takes the full checklist each call (pending / in_progress / done), keeps it in session memory, and the agent loop re-injects the current list into the system context every round, so the model always plans against live state instead of digging through old tool results. Sub-agents share the parent's list, a deliberate simplification documented in the module docstring. 103 tests pass, including the state transitions, the context injection, and the validation paths. The README code map and tool count are updated in both languages.