feat(reminder): catch stale in-progress tasks, shape reminder after Claude Code

[tintinweb]

What

Reworks the task system-reminder in two parts:

1. Fix: stale in-progress tasks were never surfaced. The reminder only fired from tool_result, so a text-only turn (agent says "done" with no tool call) never triggered it — tasks left in_progress went unnoticed. Now turn_end detects this and schedules a reminder. This is the core fix from Fix stale in-progress task reminders #31 by @QuangThai — preserved intact, credited below.

2. Shape the reminder after Claude Code's todo reminders. An empty-list nudge plus a JSON state echo of the current list, matching Claude Code's wording and structure (adapted to this extension's task tool names). The echo is capped at 10 tasks (completed dropped first) so it stays bounded on large/persistent lists, and it says list truncated instead of claiming completeness when capped.

Why the trigger matters here (and not in Claude Code)

Claude Code doesn't need staleness detection because its todo list is ephemeral, single-actor and in-context — state can't outlive the agent that manages it. pi-tasks' list is persistent and multi-agent, so state can go stale with no tool call to observe it. That's why the turn_end trigger is load-bearing here and the reminder needs an out-of-band nudge Claude Code doesn't.

Notes

• Restores the "cheap-first" store.list() guard so the hot tool_result path does no disk I/O until a reminder could actually be due.
• Stops mutating the shared cadence config (interval is now passed per-call).
• Reminder size: for typical lists (≤5 tasks) it is smaller than the old fixed constant; capped at ~330 tokens for arbitrarily large lists.
• subject is echoed as content to mirror Claude Code's todo item shape.

Credit

Supersedes #31 by @QuangThai, whose turn_end staleness trigger is the load-bearing fix and is preserved here. Co-authored accordingly.

Test plan

• npx vitest run → 177 passing
• npx tsc --noEmit clean
• npx biome check clean