fix: improve todo completion reliability

[trungutt]

Problem

Agents create todos (e.g. 5 items), do the work, but only mark a subset as completed (e.g. 3 out of 5). The remaining items stay "pending" in the sidebar even though the underlying tasks were actually done.

This is an LLM attention problem. The CRUD model (update_todos with just an ID) lets the LLM update one item without ever seeing the rest. As the conversation grows and old tool results get truncated by the 40k-token budget, the LLM loses awareness of what it originally planned.

Changes

Full-state responses on every tool call. create_todo, create_todos, and update_todos now include an all_todos field containing the complete current todo list, plus a reminder listing any incomplete items. This means the LLM sees the full state every time it touches its todos — it cannot mark todo_3 as completed without also seeing that todo_1 and todo_5 are still pending. This mirrors the approach used by OpenCode, where every todo write forces the full list into the response as a cognitive forcing function.

Stronger instructions. The Instructions() prompt now explicitly requires the LLM to call list_todos before its final response and to never leave items in a pending/in-progress state.

Remove auto-clear on all-completed. Previously, marking all items as completed would call storage.Clear(), wiping the list. This destroyed the audit trail and made list_todos return empty, which could confuse the LLM if it tried to verify its work. Completed items now remain visible in storage.

New CreateTodoOutput type. create_todo previously returned a bare Todo. It now returns CreateTodoOutput with created, all_todos, and reminder fields, consistent with the other tool responses.

[docker-agent]

Review Summary

Assessment: 🟢 APPROVE

This PR successfully addresses the LLM attention problem by adding reminder strings that alert the agent about incomplete todos. The implementation is sound with proper control flow:

✅ Clear separation between completed/pending/in-progress states
✅ Correct conditional logic (reminders only shown when needed)
✅ Storage clearing happens at the right time (after all completed)
✅ Test updated to reflect new output format

The changes effectively solve the problem of LLMs forgetting to mark todos as completed by bringing incomplete items back into their attention window.

Minor optimization opportunities (informational only, not blocking):

• Two small inefficiencies noted below could be addressed in a future cleanup

---

Posted via docker-agent PR review

[docker-agent]

[LOW] Minor inefficiency: storage.All() called twice

When incompleteReminder() is called, allCompleted() has already called h.storage.All() to check completion status. This results in iterating the storage twice. Consider refactoring to pass the slice or cache the result.

This doesn't affect correctness, just performance (minimal impact in practice given todo lists are typically small).

[docker-agent]

[LOW] Redundant counter: total is always len(all)

The total variable is incremented for every item but will always equal len(all). You can simplify this:

var completed int
for _, todo := range all {
    if todo.Status == "completed" {
        completed++
    }
}
fmt.Fprintf(&output, "Current todos (%d/%d completed):\n", completed, len(all))

This is a minor code clarity improvement with no functional impact.