Add a Plan/Todo System to Your Agent

PROJECT CONTEXT:
My agent handles single-step requests fine but falls apart on multi-step tasks: it forgets steps, repeats work, and declares victory early. Source analyses of Claude Code show a first-class todo/plan tool is core to its reliability.

GOAL:
Add a plan/todo subsystem: the model creates and updates a structured task list, and the harness renders it back into context every turn.

REQUIREMENTS:
- A todo tool the model can call: create, update status, and rewrite items
- Todo state stored by the harness and re-rendered into context each turn
- Statuses: pending, in_progress, completed - with exactly one in_progress at a time
- A rule that non-trivial tasks must start with a plan before any other tool call
- Completion gate: the agent cannot finish while items remain pending

CONSTRAINTS:
- Todos are structured data, not freeform markdown scraped by regex
- The harness owns todo state; the model only mutates it via the tool

DEFINITION OF DONE:
- A 5-step task produces a plan first, then executes items in order
- The rendered todo list appears in context every turn and matches state
- Finishing with pending items is blocked with a corrective message
- Interrupting and resuming preserves the plan and current item

COMMON FAILURES TO AVOID:
- Letting the plan live only in the model's head where it evaporates under context pressure
- Multiple items in_progress so progress becomes untrackable
- Marking items complete without evidence the work happened
- Plans created but never updated - a dead artifact instead of a working memory