Skip to content

0056 — A png or summary reads its indicators' warmup

Status: Accepted Date: 2026-11-09 Amends: ADR-0009 (png, summary and both no longer compute over exactly bars) and ADR-0032 (the indicators’ need now applies to them too, without a warmup argument)

ADR-0009 gave warmup to format=series alone: png, summary and both computed their indicators over exactly bars bars. An agent asking for the default 200 bars therefore got values the web chart never shows (#125, #131):

  • EMA 200 and SMA 200 had one value. The image drew no line, but the legend and the summary printed that value as if it were the average.
  • kalman_volume needs about 285 bars before its first value, so its panel was empty and its summary values were null, while format=series with any warmup had them (ADR-0032 raises the warmup to what an indicator declares).
  • CCI, RSI and the rest started 14 to 20 bars into the image, and every EMA-seeded value differed from the web chart’s, which always loads with warmup: 300.

The same call should give the same numbers in every format and in the web app.

  • One value per bar: png, summary, series and the web chart agree.
  • The bars shown stay exactly bars, so drawings, signals and indices mean what they did.
  • Bounded cost: no more upstream reads than a series call may already make.
  1. Read the indicators’ need before the bars shown, at least the web chart’s 300, capped at 1000, then trim (png, summary and both).
  2. Read only what the indicators declare (warmup_bars), without the 300 floor.
  3. Keep exactly bars and mark unconverged values (— in the legend, null in the summary).
  4. Honour the warmup argument in png and summary as series does.

Chosen option: 1. A png, summary or both call with indicators reads max(300, the largest warmup_bars_at of its indicators) older bars, at most MAX_WARMUP (1000), computes over all of them and trims every result to the bars it shows, the same way a series is trimmed (trim_panel_result). Patterns, levels and events are re-based onto the bars shown, and an index before the first of them is negative, as in a series. Volume and time profiles that aggregate over their window (windowed_profile) are recomputed over the bars shown, so warmup bars do not leak into them. A call without indicators reads nothing more, and neither does the demo series (a call without a ticker): it is generated from its first bar on, so a warmup would only move the bars it shows. Price-based styles read their fixed source window, as before (ADR-0047); their indicators run on every brick built from it and are trimmed to the bars shown in every format, as a series’ are. The warmup argument still applies to series alone.

Option 2 leaves indicators that declare no need (CCI, Williams %R, Bollinger Bands) starting mid-image and slightly different from the web chart. Option 3 hides the problem instead of fixing it, though the image keeps it as a safety net: a line with a single value shows — in its legend, for a symbol with too little history. Option 4 puts the burden on every agent, and the defaults stay wrong.

  • Good: an image’s legend, a summary’s last_value and the web chart show the same number. kalman_volume, EMA 200 and SMA 200 have values on 200 bars.
  • Good: the image draws every indicator from its first bar.
  • Bad: a call reads up to 1000 more bars. 200 bars with ema_stack read 1,091 (two Binance requests instead of one), and chart_patterns or auto_fib read 600 more. Series calls could already ask for this much, and call_cost (ADR-0052) counts the bars shown, not those read, for every format alike. The extra compute is transient: only the trimmed results stay.
  • Neutral: the demo series (no ticker) shows the same bars and values as before, unconverged.
  • Neutral: cumulative indicators (OBV, A/D, CVD) start from a different bar, as in the web chart. Their level was never meaningful, only its changes.