[robdevops]

• "What was the humidity each hour for the last 5 hours" got "the tool can't give me an hour-by-hour table", because weather_history only sends a period's lows and highs. The new readings tool (lib/history.py) returns a row per interval, read straight from ecobot.sqlite.
• Inputs: a metric (outdoor.humidity, rainfall.rain, air.pm2_5 …), a start, an optional end, and an interval: 1min, 5min, 30min, 1hour, 4hour or 1day.
• Intervals:
  • An interval a sensor stores is read as it is. Any other is worked out from finer readings, never coarser ones (the air sensor's older hourly history is not split into half hours).
  • Intervals are blocks of local time: an hour runs from :00 to :00, a day from midnight.
  • The exception is 4hour, which follows the station's own 4-hour records. Ecowitt cuts those on UTC hours, so in AEDT they start at 3am, 7am and so on.
• Rows: means per 30-minute slot, as the charts take them. Gusts give the records' own highest. Rain gives the mm that fell in the interval, with the midnight reset handled. Air particles have glitches removed, as air_quality does. Days come from the day rows, with the times of their lows and highs.
• Gaps: a gap is a row with no value. Where a period is held only at a coarser interval (or only as day rows), the result says which interval to ask for (fits).
• Size limit: at most 300 rows a request. A longer request is refused, and the refusal names the finest interval that fits.
• Metrics offered: only those each sensor has actually sent. The air sensor's own temperature and humidity are offered only when the question names them.
• Routing (lib/intent.py): "each hour", "every 30 minutes", "every ten minutes", "hourly" and "in 15 minute steps" reach the model, which the prompt points at readings. Before, some of these took a code shortcut to a highs-and-lows summary.
  • A chart of hourly values ("hourly rain chart 7d") is still drawn in code.
  • "miles per hour" and "km/h per hour" are units, not intervals.
• scripts/show_request.py can run the tool against a file.
• A review (three reviewers, each finding checked by an agent trying to refute it) confirmed 18 problems in the first commit; the second commit fixes them:
  • Rain at the seam between two resolutions read as a midnight reset: 11.8 mm for an hour that had 1.5 mm.
  • 4-hour records were filed into the wrong blocks.
  • Gusts ignored each record's own highest.
  • The tool suggested a coarser interval when only the latest, still-filling row was empty.
  • Several smaller ones: glitches at the edge of the window, start and end times sent with a time-zone offset, off-by-one row limits, and the metric lookup falling back to another sensor.
• Checks: pytest passed (1,123 tests, 20 of them for the new tool) and the offline evals passed (92 cases, 2 new). The live evals, which ask the real model, were not run because they need the OpenRouter key.

---

Generated by Claude Code