Session-scoped active todo reminders for update_plan

What variant of Codex are you using?

CLI / App core session engine (codex-core)

What feature would you like to see?

I would like update_plan to maintain a session-scoped active todo list and use it to remind the model at task-continuation boundaries.

Since external PRs are not currently accepted for this repository, I opened this issue with a reference implementation in a fork PR:

• Reference PR: https://github.com/toddwyl/codex/pull/1

Problem

update_plan is already useful as a structured TODO/checklist tool that clients can render, but the core session does not currently treat the latest unfinished plan as active execution state.

In practice this creates a few gaps:

• The UI can receive and render a PlanUpdate, but the model still relies on normal transcript context to remember the checklist later.
• A later user turn such as "continue" may not reliably steer the model back to the unfinished in_progress / pending items.
• A resumed session can restore transcript history, but there is no explicit active todo state recovered from the latest unfinished update_plan call.
• A naive reminder on every model request would be too noisy, especially during tool continuations and retries.

The intended behavior is closer to Claude Code's TodoWrite-style workflow: the todo list is not just UI decoration; it is active session state that helps prevent task drift during long-running work.

Proposed behavior

When update_plan has unfinished items, Codex should keep that plan as the session's active todo list and remind the model only when that reminder is useful:

• after update_plan updates the active todo list
• at the beginning of a later user turn while unfinished todos exist
• after resume restores an unfinished todo list from history

The reminder should be one-shot: once it is injected into the next sampling request, it should be consumed so the same todo list is not repeated on every model request or tool continuation.

If update_plan is empty or every item is completed, the active todo list should be cleared.

Reference implementation

The fork PR implements the behavior as follows:

• Adds active_todo_list to SessionState.
• Updates PlanHandler so parsed update_plan arguments are stored in session state before the existing PlanUpdate event is emitted.
• Clears active todo state when the plan is empty or all steps are completed.
• Adds a pending-reminder flag so reminders are queued and consumed once.
• Queues reminders after update_plan, on new user turns with active todos, and after resume restores active todos.
• Appends the reminder as an ephemeral developer message in run_sampling_request only when a reminder is pending.
• Restores active todos on resume by scanning rollout items for the latest unfinished update_plan function call.

The reference PR uses the name active_todo_list rather than latest_todo_list because the state represents the current unfinished checklist, not just the most recent plan payload.

Verification in the reference PR

The fork PR includes tests covering:

• reminder formatting
• skipping empty / completed todo lists
• restoring the latest unfinished update_plan from rollout history
• storing and clearing active todo state through the update_plan handler
• injecting and consuming the reminder in model request construction

Commands run locally:

cargo test -p codex-core todo_reminder
cargo test -p codex-core update_plan_

Related issues

This appears related to, but narrower than, these existing issues:

• #2966 asks for TODO support for long-running tasks.
• #18920 asks for the update_plan task list to remain visibly rendered in the TUI.

This issue is specifically about the core session/model-reminder behavior: preserving the unfinished update_plan checklist as active execution state and injecting a bounded reminder to reduce task drift.