Incomplete-todo nudge repeats every turn with no backoff, even when the agent is blocked on await_members

Summary

The end-of-turn reminder You have N incomplete todos. Continue working, or update the todo tool. repeats on every single turn with no backoff and no awareness of whether the agent is actually able to make progress. When a coordinator is legitimately blocked waiting on background swarm workers, this fires indefinitely.

Observed

In one session today it fired ~25 times in ~12 minutes, in unbroken succession, while the coordinator was blocked on swarm await_members waiting for a worker to finish a browser scrape. The two remaining todos were:

1. assigned to a running worker, with a wake already registered via await_members
2. explicitly waiting on a human decision

Neither could be advanced by the model. Each nudge still consumed a full turn: the model either re-polled swarm list (pure token burn, no new information) or emitted "still waiting" prose. That was hundreds of thousands of tokens of pure overhead, and it crowds the transcript so real worker reports are harder to spot.

Why it's wrong in this case

The harness already knows the coordinator is blocked — await_members / bg wait registered a wake, and the session is in a waiting state. Nudging an agent to "continue working" when it has explicitly parked on a wake is asking it to do the one thing that cannot help.

Suggested fixes (any one would fix the pathology)

• Suppress while a wake is registered. If the session has an outstanding await_members / bg wait / ScheduleWakeup, skip the nudge — that's the strongest signal that waiting is intentional.
• Exponential backoff. Turn 1, then 2, 4, 8… rather than every turn.
• Ownership awareness. If every incomplete todo is delegated (or the session has live owned workers), suppress or soften to an informational line.
• Cap per session. Stop after N consecutive nudges with no change in todo state; resume only when the todo list actually changes.

Contrast with a user-side gate

For reference, a pre_tool discipline gate in my own config gets this right: it counts only actionable tool calls (bash/write/edit/multiedit/apply_patch), blocks once, and resets its counter so it cannot deadlock. The built-in turn-end nudge has neither the actionable-only filter nor the block-once/reset behaviour.

Environment

jcode v0.67.1 (88a19f38e), macOS aarch64. Coordinator session with 5 spawned swarm workers; [display] pin_todos = true.

Trace the end-of-turn incomplete-todo nudge and its interaction with await_members, bg wait, and ScheduleWakeup. Reproduce a coordinator blocked on a registered wake with delegated or human-blocked todos, then verify the chosen suppression, backoff, or cap behavior prevents repeated nudges while preserving useful reminders after todo state changes.