Plan history: show the latest finished slot, and stop the false 'stale' warning - #5414

Posted by Claude on behalf of @chalfontchubby.

Summary

The web page's Plan → History view had two problems:

1. The most recent slots were missing. History is rebuilt from what actually happened, at most once an hour, while the live Plan view starts at the current slot. So the slots finished since the last rebuild appeared in neither view. For example, at 21:35 the History stopped at the 20:30 slot and the Plan started at 21:30, so 21:00–21:30 was nowhere.
2. It said "Plan data is stale" most of the time. The page judged every view against the live plan's 15-minute limit. History is only rebuilt hourly, so the warning showed for about 45 minutes of every hour although nothing was wrong.

With this change History is rebuilt once per completed slot, so the last finished slot shows within one or two re-plans of ending. Each view is judged against how often it is actually republished, so the warning only appears when something has really stopped.

Detail

Rebuild cadence: calculate_yesterday()

• It now rebuilds once per completed plan slot (plan_interval_minutes), on the first re-plan at least one run (5 minutes) after the slot ends. The wait is so that the boundary run's own predbat.cost_today write is already recorded when its history is read back. The key is history_slot(): a (date, slot number, UTC offset), stepped back one run in UTC. The offset keeps the hour that repeats when the clocks go back as two separate slots.
• A change of local date also rebuilds straight away, so the forced midnight re-plan still moves cost_yesterday and savings_yesterday_* on to the day just ended at 00:00.
• How often that is:
  • before Fix potential lost octopus slots, fix saving calc fluctuations #2913 it ran on every re-plan (about 144 times a day at the 10-minute default)
  • Fix potential lost octopus slots, fix saving calc fluctuations #2913 limited it to hourly, with the comment "we don't need to re-compute this one every iteration"
  • this PR: once per slot, 48 times a day at 30-minute slots
• A rebuild takes 0.4–0.9 s on a live install.
• The one-run wait uses RUN_EVERY (the run cadence), not PREDICT_STEP.

Stale warning: get_plan_renderer_js()

• Each dataset the page reads carries refresh_minutes, the longest Predbat should take to republish it:
  • the live plan: calculate_plan_every, set only when the plan is published
  • History and "Yesterday without Predbat": a slot, plus the run it waits, plus a re-plan interval
• The page warns a run (5 minutes) after that, never under 15 minutes. At the defaults that's Plan 15 minutes (as before) and History 50.
• The check uses the data the current view shows. Before, the "unchanged" poll always checked the plan's timestamp. It hides the warning when the view has no data. The view-to-data mapping is now a single helper (currentViewData()), and the "no data yet" text no longer says "about once an hour".

Trade-offs, accepted

• With calculate_plan_every at or above the slot length, re-plans can fall on slot boundaries, and the rebuild then happens on the boundary run.
• Just after midnight the history is rebuilt twice: once on the 00:00 re-plan (new day), and again on the first re-plan after 00:05 (first slot of the new day).
• cost_yesterday and savings_yesterday_predbat, which have large html/json attributes, are written twice as often.
• The 1 am savings totals (savings_total_*) add the figures from the last rebuild before them. That is now about 00:40 at the defaults, not about 01:00, so late cloud load/PV data has about 40 minutes rather than an hour to land before it is counted.

Tests

• test_calculate_yesterday:
  • early exit within a slot
  • no rebuild on the boundary run, then a rebuild on the next run (this test fails on main)
  • 15-minute slots and the midnight key
  • the repeated hour at the autumn clock change, under both pytz and zoneinfo (fails without the UTC offset in the key)
  • the midnight re-plan rebuilds, although its slot key still counts towards the day before
  • refresh_minutes on the published History and baseline data
• Test doubles: the stubs returned "{}" (a string) where publish_html_plan and plan_write_debug return a dict. They now match the real functions.
• New test_plan_staleness_js: structural checks on the page script (the per-view data, the limit arithmetic, hiding the warning without data, no remaining bare-timestamp checks). It fails on main's script.
• Pre-commit and ./run_all --quick pass. The rendered page script passes node --check.

🤖 Generated with Claude Code