detectkit 0.31.1__tar.gz → 0.33.0__tar.gz

{detectkit-0.31.1/detectkit.egg-info → detectkit-0.33.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: detectkit
-Version: 0.31.1
+Version: 0.33.0
 Summary: Metric monitoring with automatic anomaly detection
 Author: detectkit team
 License: MIT

{detectkit-0.31.1 → detectkit-0.33.0}/detectkit/__init__.py

@@ -4,7 +4,7 @@ detectk - Anomaly Detection for Time-Series Metrics
 A Python library for data analysts and engineers to monitor metrics with automatic anomaly detection.
 """
 
-__version__ = "0.31.1"
+__version__ = "0.33.0"
 
 from detectkit.core.interval import Interval
 from detectkit.core.models import ColumnDefinition, TableModel

{detectkit-0.31.1 → detectkit-0.33.0}/detectkit/autotune/grid_search.py

@@ -21,6 +21,7 @@ from detectkit.autotune.window_select import (
     detect_level_shift,
     half_life_grid,
     min_samples_for,
+    seasonal_fill_window,
     select_window,
     trend_present,
 )
@@ -45,6 +46,21 @@ def grid_search(
     base: dict[str, Any] = {}
     if seasonality:
         base["seasonality_components"] = seasonality
+        # Per-group stats engage only when the window holds min_samples_per_group
+        # points of the current seasonal key (~min_samples_per_group * cardinality
+        # points). If even the largest fold-feasible window can't reach that, the
+        # chosen seasonality will silently fall back to global at runtime — flag it
+        # in the decision log so the tuned config isn't trusted to be seasonal.
+        fill = seasonal_fill_window(tuner)
+        if fill and grid and max(grid) < fill:
+            tuner.log(
+                "window",
+                f"seasonality {seasonality} needs window_size >= {fill} to engage per-group "
+                f"stats, but the fold budget caps the window at {max(grid)} on this history "
+                "— the band will use global statistics (seasonality has no effect). Tune on "
+                "more history, or use `dtk tune` to set a larger window manually.",
+                seasonal_fill_window=fill,
+            )
 
     has_trend = trend_present(tuner)
     if not has_trend:

{detectkit-0.31.1 → detectkit-0.33.0}/detectkit/autotune/html_labeler.py

@@ -295,7 +295,11 @@ const TICK_STEPS=[MS_S,2*MS_S,5*MS_S,10*MS_S,15*MS_S,30*MS_S,
   MS_H,2*MS_H,3*MS_H,6*MS_H,12*MS_H, MS_D,2*MS_D,3*MS_D,5*MS_D,7*MS_D,14*MS_D];
 function niceTimeTicks(lo, hi, target) {
   const span=Math.max(hi-lo, 1);
-  if (span > target*28*MS_D) {  // month/year calendar ticks for long spans
+  // Escalate to calendar months/years once even the coarsest sub-monthly step
+  // (14d) would exceed the target count (span > target*14d). The old threshold
+  // (target*28d) left a gap where no sub-monthly step fit and the loop fell
+  // through to a 14d step, packing ~13 labels in and overlapping.
+  if (span > target*14*MS_D) {  // month/year calendar ticks for long spans
     const mSteps=[1,2,3,6,12,24,36,60,120,240]; let stepM=mSteps[mSteps.length-1];
     for (const m of mSteps) { if (span/(m*30.44*MS_D) <= target) { stepM=m; break; } }
     const d=new Date(lo); let y=d.getUTCFullYear(), mo=d.getUTCMonth();
@@ -322,6 +326,17 @@ const fmtAxis = (ts, step) => { const s=new Date(ts).toISOString();
   if (step >= MS_D) return s.slice(5,10);
   if (step >= MS_H) return s.slice(5,16).replace('T',' ');
   return s.slice(5,19).replace('T',' '); };
+// Which ticks may show a TEXT label without colliding: keep the first, then keep
+// each only if its x clears the previous kept label by its width + a gutter.
+// Gridlines/ticks still draw for every entry; only text is thinned. Measures with
+// the passed context's CURRENT font, so set it before calling.
+function labelMask(g, ticks, xOf, step) {
+  const show=new Array(ticks.length).fill(false); if (!ticks.length) return show;
+  const minGap=g.measureText(fmtAxis(ticks[0], step)).width + 16*dpr; let lastX=-Infinity;
+  for (let i=0;i<ticks.length;i++) { const xx=xOf(ticks[i]);
+    if (xx-lastX >= minGap) { show[i]=true; lastX=xx; } }
+  return show;
+};
 function fmtDur(ms) { const m = Math.round(ms/60000); if (m<60) return m+'m';
   const h = Math.floor(m/60), mm = m%60; if (h<24) return h+'h'+(mm?(' '+mm+'m'):'');
   const d = Math.floor(h/24), hh = h%24; return d+'d'+(hh?(' '+hh+'h'):''); }
@@ -439,9 +454,11 @@ function draw() {
   // in real time reads off the grid (not only by chasing the cursor).
   ctx.textBaseline = 'top';
   const xtk=niceTimeTicks(viewMin, viewMax, 7);
-  xtk.ticks.forEach(ts => { const xx=px(ts);
+  const showXL=labelMask(ctx, xtk.ticks, px, xtk.step);
+  xtk.ticks.forEach((ts,ti) => { const xx=px(ts);
     ctx.strokeStyle='rgba(255,255,255,0.05)'; ctx.lineWidth=1*dpr;
     ctx.beginPath(); ctx.moveTo(xx, M.t*dpr); ctx.lineTo(xx, c.height-M.b*dpr); ctx.stroke();
+    if (!showXL[ti]) return;
     ctx.fillStyle='#6e675b';
     ctx.textAlign = xx < (M.l+24)*dpr ? 'left' : xx > c.width-(M.r+24)*dpr ? 'right' : 'center';
     ctx.fillText(fmtAxis(ts, xtk.step), xx, c.height-(M.b-8)*dpr); });
@@ -578,9 +595,11 @@ function drawOverview() {
   // absolute-time anchors under the strip, so a point's real time is readable
   // across the whole series without hovering.
   octx.font=(10*dpr)+'px ui-monospace, monospace'; octx.textBaseline='top';
-  otk.ticks.forEach(ts => { const xx=ovpx(ts);
+  const showOL=labelMask(octx, otk.ticks, ovpx, otk.step);
+  otk.ticks.forEach((ts,ti) => { const xx=ovpx(ts);
     octx.strokeStyle='rgba(201,194,180,0.25)'; octx.lineWidth=1*dpr;
     octx.beginPath(); octx.moveTo(xx, plotBot); octx.lineTo(xx, plotBot+3*dpr); octx.stroke();
+    if (!showOL[ti]) return;
     octx.fillStyle='#6e675b';
     octx.textAlign = xx < (OM.l+26)*dpr ? 'left' : xx > ov.width-(OM.r+26)*dpr ? 'right' : 'center';
     octx.fillText(fmtAxis(ts, otk.step), xx, plotBot+5*dpr); });

{detectkit-0.31.1 → detectkit-0.33.0}/detectkit/autotune/window_select.py

@@ -16,6 +16,15 @@ import numpy as np
 from detectkit.autotune._base import _AutoTuneBase
 from detectkit.autotune._types import CandidateEval
 from detectkit.detectors.factory import DetectorFactory
+from detectkit.detectors.seasonality import parse_seasonality_data
+from detectkit.detectors.statistical._windowed import WindowedStatDetector
+
+# Reference per-group floor used to size a seasonality-fill window candidate. The
+# windowed detectors share this default (MAD 10 is the largest); using it makes
+# the fill window cover every windowed type so cross-validation can actually
+# discover whether conditioning on a seasonal key helps. See
+# ``seasonal_fill_window``.
+_MSPG_REF = int(WindowedStatDetector.MIN_SAMPLES_PER_GROUP_DEFAULT)
 
 
 def min_samples_for(window_size: int, floor: int) -> int:
@@ -23,8 +32,54 @@ def min_samples_for(window_size: int, floor: int) -> int:
     return min(window_size, max(floor, round(window_size / 4)))
 
 
+def max_seasonal_cardinality(tuner: _AutoTuneBase) -> int:
+    """Largest distinct-key count among the available single seasonality columns.
+
+    Per-group statistics only engage once the window holds
+    ``min_samples_per_group`` points sharing the current point's key, and same-key
+    points recur every *cardinality* grid positions — so this is the recurrence
+    period the window must cover. We use the most granular single column (e.g.
+    ``hour_of_day`` → 24) as the representative key; conjunctive groupings are
+    rarer and are backstopped by the detector's runtime under-fill warning.
+    Returns 0 when no seasonality columns are present.
+    """
+    columns = [c for c in tuner.data.get("seasonality_columns", []) if c != "is_holiday"]
+    if not columns:
+        return 0
+    season = parse_seasonality_data(tuner.data.get("seasonality_data", np.array([])), columns)
+    card = 0
+    for col in columns:
+        vals = season.get(col)
+        if vals is None or len(vals) == 0:
+            continue
+        distinct = {v for v in vals.tolist() if v is not None}
+        card = max(card, len(distinct))
+    return card
+
+
+def seasonal_fill_window(tuner: _AutoTuneBase) -> int:
+    """Smallest window that can fill a single-column seasonal group, else 0.
+
+    ``min_samples_per_group * cardinality``. None of the natural-unit candidates
+    (≈1 day, ≈1 week) reach this for hourly ``hour_of_day`` data (24 keys → 240),
+    so without this the chosen seasonality silently never engages at the tuned
+    window. Returns 0 when there is no seasonality to fill.
+    """
+    card = max_seasonal_cardinality(tuner)
+    return _MSPG_REF * card if card > 0 else 0
+
+
 def window_grid(tuner: _AutoTuneBase) -> list[int]:
-    """Candidate window sizes (≈1 day, ≈1 week, default 100), clamped to fit folds."""
+    """Candidate window sizes (≈1 day, ≈1 week, default 100), clamped to fit folds.
+
+    When the data carries seasonality columns, also offer a window large enough to
+    fill the most granular seasonal group (``min_samples_per_group * cardinality``)
+    so that — if seasonality is chosen downstream — cross-validation can evaluate a
+    window where the per-group band actually engages instead of silently falling
+    back to global statistics. Capped (like the other candidates) so it never
+    exceeds the fold budget; if it doesn't fit, seasonality simply can't fill on
+    this much history and the detector's runtime warning will say so.
+    """
     fixed = tuner.settings.fixed_params.get("window_size")
     if isinstance(fixed, int):
         return [fixed]
@@ -34,6 +89,9 @@ def window_grid(tuner: _AutoTuneBase) -> list[int]:
     cap = max(20, n // (fold_count + 1))
     pts_per_day = max(1, round(86400 / tuner.interval_seconds))
     candidates = {100, pts_per_day, 7 * pts_per_day}
+    fill = seasonal_fill_window(tuner)
+    if fill:
+        candidates.add(fill)
     grid = sorted({w for w in candidates if 20 <= w <= cap})
     if not grid:
         grid = [max(2, min(cap, 100))]

{detectkit-0.31.1 → detectkit-0.33.0}/detectkit/cli/assets/claude/CLAUDE.section.md

@@ -37,6 +37,7 @@ version — **read the relevant one on demand** instead of guessing:
 | Choosing/tuning detectors, preprocessing, trends, seasonality | `.claude/rules/detectkit/detectors.md` |
 | Alert rules (quorum/direction/consecutive), cooldown, recovery, templates | `.claude/rules/detectkit/alerting.md` |
 | Auto-tuning a detector (`dtk autotune`), labels file, the `autotune:` block, `_dtk_autotune_runs` | `.claude/rules/detectkit/autotune.md` |
+| Manually/interactively tuning a detector in the browser and writing it back in place (`dtk tune`) | `.claude/rules/detectkit/cli.md` |
 
 ### Skills
 
@@ -48,7 +49,10 @@ version — **read the relevant one on demand** instead of guessing:
   to a YAML file that validates and is ready to run.
 - **Tune / pick a detector, or build an alert from scratch and tune it** — use
   the **`dtk-autotune`** skill; it gathers seasonality + incidents, runs
-  `dtk autotune`, and explains the chosen, annotated config.
+  `dtk autotune`, and explains the chosen, annotated config (an automatic,
+  cross-validated search). For **hands-on** tuning instead, `dtk tune` opens a
+  browser view where you drag the detector's knobs against the real series and
+  write the result back in place (see `cli.md`).
 - **Hit a detectkit bug, or have feedback** — once you've ruled out a local
   config fix (see the gotchas below), use the **`dtk-feedback`** skill to file a
   redacted bug report, feature request, or comment as a GitHub issue on the

{detectkit-0.31.1 → detectkit-0.33.0}/detectkit/cli/assets/claude/rules/cli.md

@@ -77,9 +77,10 @@ Full reference: `autotune.md`.
 
 The **manual, interactive** sibling of `dtk autotune`. Opens a localhost browser
 view of the metric's **real** persisted series and lets you turn the detector's
-knobs (type, threshold, window, recency weighting + half-life, detrend, smoothing,
-**seasonality groups**, alert `consecutive_anomalies`) while the confidence band
-and flagged anomalies **recompute live**. The chart is **zoomable** (scroll/drag +
+knobs (type — including **Manual bounds** with lower/upper sliders — threshold,
+window, recency weighting + half-life, detrend, smoothing, **seasonality groups**,
+**direction** (both/up/down), alert `consecutive_anomalies`) while the confidence
+band and flagged anomalies **recompute live**. The chart is **zoomable** (scroll/drag +
 a navigator strip) and a **"Points shown"** slider trims the active sample to speed
 up recompute on a long metric. Clicking **Apply** writes the chosen
 config back into the metric YAML **in place** (autotune, by contrast, writes a new

{detectkit-0.31.1 → detectkit-0.33.0}/detectkit/cli/assets/claude/rules/detectors.md

@@ -101,7 +101,15 @@ seasonality_components:
 - `["hour", "day_of_week"]` — two *separate* adjustments.
 - `[["hour", "day_of_week"]]` — one *combined* group per pair.
 
-`min_samples_per_group` is the floor of points required per bucket.
+`min_samples_per_group` is the floor of points required per bucket. **Size the
+window to fill a bucket:** a group engages only when the window holds
+`min_samples_per_group` points of the current key, and same-key points recur once
+per *cardinality*, so `window_size ≳ min_samples_per_group × distinct_keys`. With
+hourly `["hour"]` (24 keys, mad default 10) that's `≳ 240` — the default
+`window_size = 100` fills no bucket, so seasonality silently falls back to the
+global band (the detector logs a one-time warning). Raise `window_size`, lower
+`min_samples_per_group`, or use a coarser grouping. `dtk autotune` now offers a
+fill-sized window candidate automatically.
 
 ## Preprocessing — `input_type`
 

{detectkit-0.31.1 → detectkit-0.33.0}/detectkit/detectors/statistical/_windowed.py

@@ -26,6 +26,7 @@ Subclasses define:
 
 from __future__ import annotations
 
+import logging
 import math
 from abc import abstractmethod
 from typing import Any
@@ -40,6 +41,8 @@ from detectkit.detectors.seasonality import (
 )
 from detectkit.utils.stats import effective_sample_size, weighted_median
 
+logger = logging.getLogger(__name__)
+
 _INPUT_TYPES = {"values", "changes", "absolute_changes", "log_changes"}
 _CHANGE_INPUT_TYPES = {"changes", "absolute_changes", "log_changes"}
 _SMOOTHING_METHODS = {None, "ema", "sma"}
@@ -141,6 +144,10 @@ class WindowedStatDetector(BaseDetector):
             weight_decay=weight_decay,
             detrend=detrend,
         )
+        # Whether the one-time "seasonality groups can't fill this window" check
+        # has run for this instance (so the warning fires at most once per run,
+        # not once per detect() batch).
+        self._underfill_checked: bool = False
 
     # ------------------------------------------------------------------
     # Subclass hooks
@@ -386,6 +393,56 @@ class WindowedStatDetector(BaseDetector):
         # (age 0) is values + slope * age
         return (med_new - med_old) / (age_old - age_new)
 
+    def _warn_if_groups_cannot_fill(
+        self,
+        seasonality_components: list[str | list[str]],
+        seasonality_dict: dict[str, np.ndarray],
+        window_size: int,
+        min_samples_per_group: int,
+    ) -> None:
+        """Warn once if the window is too small to ever fill a seasonality group.
+
+        Per-group statistics only engage when the trailing window holds at least
+        ``min_samples_per_group`` points sharing the current point's seasonal key.
+        Same-key points recur every *cardinality* positions on the grid, so the
+        window must hold roughly ``min_samples_per_group * cardinality`` points
+        before any group fills — otherwise EVERY point silently falls back to the
+        global (un-conditioned) band and the configured seasonality has no effect.
+        This is easy to hit by accident (e.g. the default ``window_size=100`` with
+        24-key hourly seasonality needs ``>= 240``), so surface it loudly once.
+        """
+        if self._underfill_checked:
+            return
+        self._underfill_checked = True
+
+        for group in seasonality_components:
+            cols = [group] if isinstance(group, str) else list(group)
+            present = [c for c in cols if c in seasonality_dict]
+            if not present:
+                continue
+            # Distinct seasonal keys actually present in the data (conjunction of
+            # the group's columns) — the per-key recurrence period on the grid.
+            keys = set(zip(*[seasonality_dict[c].tolist() for c in present], strict=False))
+            cardinality = len(keys)
+            if cardinality <= 0:
+                continue
+            needed = min_samples_per_group * cardinality
+            if window_size < needed:
+                logger.warning(
+                    "%s: seasonality group %s has %d distinct key(s) but window_size=%d "
+                    "holds only ~%d same-key point(s) (< min_samples_per_group=%d), so this "
+                    "group falls back to global statistics and the seasonality has no effect. "
+                    "Increase window_size to >= %d (min_samples_per_group * distinct keys), "
+                    "lower min_samples_per_group, or use a coarser grouping.",
+                    self.__class__.__name__,
+                    cols,
+                    cardinality,
+                    window_size,
+                    window_size // cardinality,
+                    min_samples_per_group,
+                    needed,
+                )
+
     # ------------------------------------------------------------------
     # Detection pipeline
     # ------------------------------------------------------------------
@@ -415,6 +472,9 @@ class WindowedStatDetector(BaseDetector):
         seasonality_dict = {}
         if seasonality_components and len(seasonality_data) > 0 and seasonality_columns:
             seasonality_dict = parse_seasonality_data(seasonality_data, seasonality_columns)
+            self._warn_if_groups_cannot_fill(
+                seasonality_components, seasonality_dict, window_size, min_samples_per_group
+            )
 
         weight_lut = self._build_weight_lut(timestamps)
 

detectkit-0.33.0/detectkit/tuning/assets/tune.js

@@ -0,0 +1,82 @@
+"use strict";(()=>{var bt={"--term-bg":"#211e1a","--clay":"#d15b36","--st-anomaly":"#d63232","--st-recovery":"#36a64f","--st-nodata":"#f0ad4e","--st-error":"#5a7a8c","--faint":"#9a9384","--muted":"#6e675b","--border":"#332f29","--term-border":"#332f29"};function X(t){return getComputedStyle(document.documentElement).getPropertyValue(t).trim()||bt[t]||"#888"}function gt(t){let n=t.replace("#","").trim();n.length===3&&(n=n[0]+n[0]+n[1]+n[1]+n[2]+n[2]);let l=parseInt(n,16);return n.length!==6||Number.isNaN(l)?[209,91,54]:[l>>16&255,l>>8&255,l&255]}function F(t,n){let[l,e,m]=gt(t);return`rgba(${l},${e},${m},${n})`}function ot(t){let n=Math.max(1,window.devicePixelRatio||1),l=t.clientWidth||t.offsetWidth||0,e=t.clientHeight||t.offsetHeight||0;return t.width=Math.round(l*n),t.height=Math.round(e*n),n}var Ke=Number.isFinite;function Xe(t,n,l,e,m,S,o,r,u,M,h,w){let z=n.length,H=Math.max(1,Math.round(o)),A=m-e||1,W=0;for(let I=0;I<z;I++){let Y=l[I];!Ke(Y)||n[I]<e||n[I]>m||W++}if(t.strokeStyle=M,t.lineWidth=h*w,t.lineJoin="round",t.beginPath(),W<=H){let I=!1;for(let Y=0;Y<z;Y++){let re=l[Y],L=n[Y];if(!Ke(re)||L<e||L>m){I=!1;continue}let E=r(L),j=u(re);I?t.lineTo(E,j):(t.moveTo(E,j),I=!0)}}else{let I=new Array(H).fill(null),Y=new Array(H).fill(null);for(let L=0;L<z;L++){let E=l[L],j=n[L];if(!Ke(E)||j<e||j>m)continue;let U=Math.floor((j-e)/A*(H-1));U=U<0?0:U>H-1?H-1:U,(I[U]===null||E<I[U])&&(I[U]=E),(Y[U]===null||E>Y[U])&&(Y[U]=E)}let re=!1;for(let L=0;L<H;L++){if(Y[L]===null){re=!1;continue}let E=S+L,j=u(Y[L]),U=u(I[L]);re?t.lineTo(E,j):(t.moveTo(E,j),re=!0),t.lineTo(E,U)}}t.stroke()}function rt(t){let n=Math.abs(t);return n>=1e3?t.toFixed(0):n>=10?t.toFixed(1):n>=1?t.toFixed(2):t.toFixed(3)}function it(t,n){let l=new Date(t).toISOString();return n<2*864e5?l.slice(5,16).replace("T"," "):l.slice(5,10)}function st(t,n,l){return{left:n.l*l,top:n.t*l,right:t.width-n.r*l,bottom:t.height-n.b*l}}function at(t,n,l,e,m,S,o){let r=st(n,l,e),u=Math.max(r.left,Math.min(m(S),r.right));u<=r.left+.5||(t.save(),t.fillStyle="rgba(17,15,13,0.42)",t.fillRect(r.left,r.top,u-r.left,r.bottom-r.top),t.strokeStyle=F(X("--faint"),.7),t.lineWidth=1*e,t.setLineDash([4*e,4*e]),t.beginPath(),t.moveTo(u,r.top),t.lineTo(u,r.bottom),t.stroke(),t.setLineDash([]),t.fillStyle=F(X("--faint"),.95),t.font=`${10*e}px ui-monospace, monospace`,t.textAlign="left",t.textBaseline="top",t.fillText(o,u+6*e,r.top+5*e),t.restore())}function lt(t,n,l,e,m,S,o){let r=st(n,l,e),u=5*e;t.save();for(let M=0;M<S.length;M++){let h=S[M],w=m(h.t);if(w<r.left-1||w>r.right+1)continue;let z=o(h.kind);t.strokeStyle=F(z,.45),t.lineWidth=1*e,t.beginPath(),t.moveTo(w,r.top),t.lineTo(w,r.bottom),t.stroke(),t.fillStyle=z,t.beginPath(),t.moveTo(w-u,r.top),t.lineTo(w+u,r.top),t.lineTo(w,r.top+u*1.4),t.closePath(),t.fill()}t.restore()}var vt={mad:1,zscore:2,iqr:4};function xt(t,n){let l=t.seasonalityData;if(!l||l.length===0)return 0;let e=0;for(let m of n){let S=new Set;for(let o of l)S.add(m.map(r=>{var u;return String((u=o==null?void 0:o[r])!=null?u:"")}).join("|"));e=Math.max(e,S.size)}return e}function ut(t,n){let l=t.timestamps.length;if(n.type==="manual_bounds")return Math.min(n.inputType!=="values"?1:0,l);let e=Math.max(n.minSamples,vt[n.type]);if(n.smoothing==="sma"&&(e=Math.max(e,n.smoothingWindow-1)),n.smoothing==="ema"&&(e=Math.max(e,Math.ceil(5/n.smoothingAlpha))),n.inputType!=="values"&&(e=Math.max(e,1)),n.seasonalityComponents!==null&&n.seasonalityComponents.length>0&&Array.isArray(t.seasonalityData)&&t.seasonalityData.length>0){let S=xt(t,n.seasonalityComponents);if(S>0){let o=n.minSamplesPerGroup*S;n.windowSize>=o&&(e=Math.max(e,o))}}return Math.min(e,l)}var T={l:52,r:14,t:14,b:26},Je=46,ct=10,wt=13,ae=Number.isFinite,Ce=1e3,Te=60*Ce,Ae=60*Te,Z=24*Ae,Qe=[Ce,2*Ce,5*Ce,10*Ce,15*Ce,30*Ce,Te,2*Te,5*Te,10*Te,15*Te,30*Te,Ae,2*Ae,3*Ae,6*Ae,12*Ae,Z,2*Z,3*Z,5*Z,7*Z,14*Z];function dt(t,n,l){let e=Math.max(n-t,1);if(e>l*14*Z){let r=[1,2,3,6,12,24,36,60,120,240],u=r[r.length-1];for(let A of r)if(e/(A*30.44*Z)<=l){u=A;break}let M=new Date(t),h=M.getUTCFullYear(),w=M.getUTCMonth();if(u>=12){let A=u/12;h=Math.floor(h/A)*A,w=0}else w=Math.floor(w/u)*u;let z=[],H=Date.UTC(h,w,1);for(;H<=n;)H>=t&&z.push(H),w+=u,h+=Math.floor(w/12),w%=12,H=Date.UTC(h,w,1);return{ticks:z,step:u*30*Z}}let m=Qe[Qe.length-1];for(let r of Qe)if(e/r<=l){m=r;break}let S;if(m%Z===0){let r=new Date(t);S=Date.UTC(r.getUTCFullYear(),r.getUTCMonth(),r.getUTCDate());let u=m/Z,M=Math.round(S/Z);S+=(u-M%u)%u*Z}else S=Math.ceil(t/m)*m;let o=[];for(let r=S;r<=n;r+=m)r>=t&&o.push(r);return{ticks:o,step:m}}function Ze(t,n){let l=new Date(t).toISOString();return n>=320*Z?l.slice(0,4):n>=26*Z?l.slice(0,7):n>=Z?l.slice(5,10):n>=Ae?l.slice(5,16).replace("T"," "):l.slice(5,19).replace("T"," ")}function mt(t,n={}){let l=t.getContext("2d");if(!l)throw new Error("chart: 2D context unavailable");let e=l,m=!!n.navigable,S=n.yFit==="data",o=1,r=null,u=-1,M=0,h=0,w=1,z=0,H=1,A=0,W=1,I="";function Y(){o=ot(t)}let re=()=>m?(ct+Je)*o:0,L=()=>t.width-(T.l+T.r)*o,E=()=>t.height-(T.t+T.b)*o-re(),j=()=>m?A:h,U=()=>m?W:w,le=()=>U()-j()||1,de=()=>w-h||1,O=i=>T.l*o+(i-j())/le()*L(),ie=i=>t.height-T.b*o-re()-(i-z)/(H-z||1)*E(),he=()=>t.height-Je*o,ee=()=>(Je-wt)*o,V=()=>t.width-(T.l+T.r)*o,B=i=>T.l*o+(i-h)/de()*V(),Re=i=>he()+ee()-(i-z)/(H-z||1)*ee(),De=(i,a,d)=>{let c=new Array(i.length).fill(!1);if(i.length===0)return c;let f=e.measureText(Ze(i[0],d)).width+16*o,v=-1/0;for(let y=0;y<i.length;y++){let _=a(i[y]);_-v>=f&&(c[y]=!0,v=_)}return c},me=(i,a,d)=>Math.max(a,Math.min(d,i)),te=()=>{let i=r==null?void 0:r.series,a=i?i.timestamps.length:0,d=a>1?de()/(a-1):1e3;return Math.max(d*8,1e3)};function se(i,a){let d=a-i,c=te();if(d<c){let f=(i+a)/2;i=f-c/2,a=f+c/2,d=c}d>=de()&&(i=h,a=w),i<h&&(a+=h-i,i=h),a>w&&(i-=a-w,a=w),A=me(i,h,w),W=me(a,h,w),fe()}let Be=i=>{let a=(i-T.l*o)/(L()||1);return j()+me(a,0,1)*le()},ue=i=>{let a=(i-T.l*o)/(V()||1);return h+me(a,0,1)*de()};function Oe(i){let a=r==null?void 0:r.series;if(!a||a.timestamps.length===0)return-1;let d=(i-T.l*o)/(L()||1),c=j()+d*le(),f=a.timestamps,v=0,y=f.length-1;for(;v<y;){let _=v+y>>1;f[_]<c?v=_+1:y=_}return v>0&&c-f[v-1]<f[v]-c&&(v-=1),v}function be(i,a){let d=i.timestamps;if(h=d[0],w=d[d.length-1],m){let y=`${d.length}:${h}:${w}`;y!==I?(I=y,A=h,W=w):(A=me(A,h,w),W=me(Math.max(W,A+te()),h,w))}let c=1/0,f=-1/0;for(let y of i.values)ae(y)&&(y<c&&(c=y),y>f&&(f=y));if(!S)for(let y of a)y.scored&&(ae(y.lower)&&y.lower<c&&(c=y.lower),ae(y.upper)&&y.upper>f&&(f=y.upper));(!ae(c)||!ae(f))&&(c=0,f=1),f<=c&&(f=c+1);let v=(f-c)*.06;z=c-v,H=f+v}function we(i,a,d){let c=r.series.timestamps;Xe(e,c,i,j(),U(),T.l*o,L(),O,ie,a,d,o)}function ye(i,a){let d=[],c=-1;for(let f=Math.max(0,a);f<i.length;f++){let v=i[f];v.scored&&ae(v.lower)&&ae(v.upper)?c===-1&&(c=f):c!==-1&&(d.push([c,f-1]),c=-1)}return c!==-1&&d.push([c,i.length-1]),d}function Ue(){if(M=0,!r||t.width===0||t.height===0)return;let{series:i,scored:a,params:d,alerts:c}=r;if(i.timestamps.length===0){e.fillStyle=X("--term-bg"),e.fillRect(0,0,t.width,t.height);return}let f=X("--clay"),v=X("--st-anomaly"),y=X("--faint"),_=X("--muted");e.fillStyle=X("--term-bg"),e.fillRect(0,0,t.width,t.height),e.font=`${11*o}px ui-monospace, 'JetBrains Mono', monospace`,e.textBaseline="middle";for(let g=0;g<=4;g++){let k=z+(H-z)*g/4,x=ie(k);e.strokeStyle=F(y,.1),e.lineWidth=1*o,e.beginPath(),e.moveTo(T.l*o,x),e.lineTo(t.width-T.r*o,x),e.stroke(),e.fillStyle=_,e.textAlign="right",e.fillText(rt(k),(T.l-8)*o,x)}e.textBaseline="top";let Q=t.height-T.b*o-re();if(m){let g=dt(j(),U(),7),k=De(g.ticks,O,g.step);for(let x=0;x<g.ticks.length;x++){let D=g.ticks[x],ne=O(D);e.strokeStyle=F(y,.1),e.lineWidth=1*o,e.beginPath(),e.moveTo(ne,T.t*o),e.lineTo(ne,Q),e.stroke(),k[x]&&(e.fillStyle=_,e.textAlign=ne<(T.l+24)*o?"left":ne>t.width-(T.r+24)*o?"right":"center",e.fillText(Ze(D,g.step),ne,Q+7*o))}}else{let g=le();for(let k=0;k<=5;k++){let x=h+g*k/5,D=O(x);e.fillStyle=_,e.textAlign=k===0?"left":k===5?"right":"center",e.fillText(it(x,g),D,Q+7*o)}}e.save(),e.beginPath(),e.rect(T.l*o,T.t*o,L(),E()),e.clip();let K=a.length,R=Math.min(ut(i,d),K),$=R<K?i.timestamps[R]:void 0,q=ye(a,R);e.fillStyle=F(f,.13);for(let[g,k]of q){e.beginPath(),e.moveTo(O(a[g].timestamp),ie(a[g].upper));for(let x=g+1;x<=k;x++)e.lineTo(O(a[x].timestamp),ie(a[x].upper));for(let x=k;x>=g;x--)e.lineTo(O(a[x].timestamp),ie(a[x].lower));e.closePath(),e.fill()}e.strokeStyle=F(f,.4),e.lineWidth=1*o;for(let[g,k]of q)for(let x of["upper","lower"]){e.beginPath();for(let D=g;D<=k;D++){let ne=O(a[D].timestamp),xe=ie(a[D][x]);D===g?e.moveTo(ne,xe):e.lineTo(ne,xe)}e.stroke()}e.strokeStyle=F(y,.55),e.lineWidth=1*o,e.setLineDash([3*o,3*o]);for(let[g,k]of q){e.beginPath();for(let x=g;x<=k;x++){let D=a[x].center;if(!ae(D))continue;let ne=O(a[x].timestamp),xe=ie(D);x===g?e.moveTo(ne,xe):e.lineTo(ne,xe)}e.stroke()}if(e.setLineDash([]),d.smoothing!=="none"){we(i.values,F(f,.28),1.25);let g=a.map(k=>k.processedValue);we(g,f,1.6)}else we(i.values,f,1.5);for(let g=R;g<K;g++){let k=a[g];if(!k.scored||!ae(k.value))continue;let x=O(k.timestamp),D=ie(k.value);k.isAnomaly?(e.fillStyle=F(v,.18),e.beginPath(),e.arc(x,D,6*o,0,Math.PI*2),e.fill(),e.fillStyle=v,e.beginPath(),e.arc(x,D,3*o,0,Math.PI*2),e.fill()):i.truthAnomaly[g]&&(e.strokeStyle=F(_,.7),e.lineWidth=1.25*o,e.beginPath(),e.arc(x,D,3.5*o,0,Math.PI*2),e.stroke())}$!==void 0&&at(e,t,T,o,O,$,"detection at full power \u2192"),u>=0&&u<a.length&&ke(u,d.windowSize,a,i,y),c&&c.length&&lt(e,t,T,o,O,c,g=>g==="anomaly"?X("--st-anomaly"):g==="recovery"?X("--st-recovery"):X("--st-nodata")),e.restore(),m&&qe(i,c,f,y,_)}function qe(i,a,d,c,f){let v=he(),y=v+ee(),_=T.l*o,Q=t.width-T.r*o,K=dt(h,w,5);e.save(),e.beginPath(),e.rect(_,v,V(),ee()),e.clip(),e.strokeStyle=F(c,.1),e.lineWidth=1*o;for(let k of K.ticks){let x=B(k);e.beginPath(),e.moveTo(x,v),e.lineTo(x,y),e.stroke()}Xe(e,i.timestamps,i.values,h,w,_,V(),B,Re,F(d,.7),1.1,o),e.restore();let R=B(A),$=B(W);if(e.fillStyle="rgba(27,25,22,0.55)",e.fillRect(_,v,R-_,ee()),e.fillRect($,v,Q-$,ee()),a&&a.length){let k=X("--st-anomaly");e.fillStyle=F(k,.85);for(let x of a){let D=B(x.t);e.fillRect(D-1*o,v,2*o,ee())}}e.fillStyle="rgba(245,241,232,0.06)",e.fillRect(R,v,$-R,ee()),e.strokeStyle=d,e.lineWidth=1.5*o,e.strokeRect(R,v+1,$-R,ee()-2),e.fillStyle=d;let q=v+ee()/2-8*o;e.fillRect(R-2*o,q,4*o,16*o),e.fillRect($-2*o,q,4*o,16*o),e.font=`${10*o}px ui-monospace, monospace`,e.textBaseline="top",e.fillStyle=f;let g=De(K.ticks,B,K.step);for(let k=0;k<K.ticks.length;k++){let x=K.ticks[k],D=B(x);e.strokeStyle=F(c,.25),e.lineWidth=1*o,e.beginPath(),e.moveTo(D,y),e.lineTo(D,y+3*o),e.stroke(),g[k]&&(e.textAlign=D<(T.l+26)*o?"left":D>t.width-(T.r+26)*o?"right":"center",e.fillText(Ze(x,K.step),D,y+5*o))}}function ke(i,a,d,c,f){let v=c.timestamps,y=Math.max(0,i-a),_=i-1,Q=T.t*o,K=E();if(_>=y){let q=c.intervalSeconds*1e3/le()*L()*.5,g=O(v[y])-q,k=O(v[_])+q;e.fillStyle="rgba(255,255,255,0.05)",e.fillRect(g,Q,k-g,K),e.strokeStyle=F(f,.5),e.lineWidth=1*o,e.beginPath(),e.moveTo(g,Q),e.lineTo(g,Q+K),e.moveTo(k,Q),e.lineTo(k,Q+K),e.stroke()}let R=O(v[i]);e.strokeStyle=F(f,.85),e.lineWidth=1*o,e.setLineDash([2*o,2*o]),e.beginPath(),e.moveTo(R,Q),e.lineTo(R,Q+K),e.stroke(),e.setLineDash([]);let $=d[i];if($.scored&&ae($.lower)&&ae($.upper)){let q=5*o;e.strokeStyle=F(X("--clay"),.85),e.lineWidth=1.5*o;for(let g of[$.lower,$.upper]){let k=ie(g);e.beginPath(),e.moveTo(R-q,k),e.lineTo(R+q,k),e.stroke()}}if(ae($.value)){let q=ie($.value);e.fillStyle=X("--term-bg"),e.beginPath(),e.arc(R,q,4*o,0,Math.PI*2),e.fill(),e.strokeStyle=$.isAnomaly?X("--st-anomaly"):X("--clay"),e.lineWidth=2*o,e.beginPath(),e.arc(R,q,4*o,0,Math.PI*2),e.stroke()}}function fe(){M===0&&(M=requestAnimationFrame(Ue))}function ge(){var d;if(!n.onHover||!r)return;if(u<0){n.onHover(null);return}let i=(d=r.scored[u])!=null?d:null,a={index:u,point:i,windowStart:Math.max(0,u-r.params.windowSize),windowEnd:u-1};n.onHover(a)}let oe=null,G=null,Se=i=>{let a=t.getBoundingClientRect();return{x:(i.clientX-a.left)*o,y:(i.clientY-a.top)*o}},Ve=i=>m&&i>=he(),Pe=i=>{let a=B(A),d=B(W),c=8*o;return Math.abs(i-a)<=c?"l":Math.abs(i-d)<=c?"r":i>a&&i<d?"move":"out"};function Le(i){if(!r)return;let{x:a,y:d}=Se(i);if(oe||G)return;if(Ve(d)){u!==-1&&(u=-1,ge(),fe());let f=Pe(a);t.style.cursor=f==="l"||f==="r"?"ew-resize":f==="move"?"grab":"pointer";return}m&&(t.style.cursor="grab");let c=Oe(a);c!==u&&(u=c,ge(),fe())}function J(){u!==-1&&(u=-1,ge(),fe())}function Me(i){if(!m||!r)return;let{x:a,y:d}=Se(i);if(d>=he()){let c=Pe(a);if(c==="l")G={type:"l",grab:0,vMin:A,vMax:W};else if(c==="r")G={type:"r",grab:0,vMin:A,vMax:W};else if(c==="move")G={type:"move",grab:ue(a),vMin:A,vMax:W};else{let f=ue(a),v=le();se(f-v/2,f+v/2),G={type:"move",grab:f,vMin:A,vMax:W}}t.style.cursor="grabbing",i.preventDefault()}else d>T.t*o&&d<he()-ct*o&&(oe={x:a,vMin:A,vMax:W},t.style.cursor="grabbing",i.preventDefault())}function _e(i){if(!r)return;let{x:a}=Se(i);if(oe){let d=(a-oe.x)*(oe.vMax-oe.vMin)/(L()||1);se(oe.vMin-d,oe.vMax-d)}else if(G){let d=ue(a);G.type==="l"?se(Math.min(d,W-te()),W):G.type==="r"?se(A,Math.max(d,A+te())):se(G.vMin+(d-G.grab),G.vMax+(d-G.grab))}}function Ee(){(oe||G)&&(oe=null,G=null,t.style.cursor="grab")}function ve(i){if(!m||!r)return;i.preventDefault();let{x:a,y:d}=Se(i),c=me(le()*Math.pow(1.0015,i.deltaY),te(),de());if(d>=he()){let f=ue(a);se(f-c/2,f+c/2)}else{let f=Be(a),v=(f-A)/(le()||1);se(f-v*c,f-v*c+c)}}function $e(){m&&se(h,w)}t.addEventListener("mousemove",Le),t.addEventListener("mouseleave",J),m&&(t.addEventListener("mousedown",Me),t.addEventListener("wheel",ve,{passive:!1}),t.addEventListener("dblclick",$e),window.addEventListener("mousemove",_e),window.addEventListener("mouseup",Ee));function Fe(i){r=i,be(i.series,i.scored),u>=i.series.timestamps.length&&(u=-1),fe()}function Ye(){Y(),r&&(be(r.series,r.scored),fe())}function Ge(){t.removeEventListener("mousemove",Le),t.removeEventListener("mouseleave",J),m&&(t.removeEventListener("mousedown",Me),t.removeEventListener("wheel",ve),t.removeEventListener("dblclick",$e),window.removeEventListener("mousemove",_e),window.removeEventListener("mouseup",Ee)),M!==0&&cancelAnimationFrame(M),M=0,r=null}return Y(),{render:Fe,resize:Ye,destroy:Ge}}var yt={mad:3,zscore:3,iqr:1.5},kt={mad:10,zscore:3,iqr:4},St="dtk-tune";function b(t,n,l){let e=document.createElement(t);return n&&(e.className=n),l!=null&&(e.textContent=l),e}function je(t,n){let l=b("label","dtk-ctl-label",t);if(n){l.title=n;let e=b("span","dtk-ctl-info","\u24D8");e.title=n,l.appendChild(document.createTextNode(" ")),l.appendChild(e)}return l}function He(t,n,l,e,m){let S=b("div","dtk-ctl");S.appendChild(je(t,m));let o=b("div","dtk-seg"),r=l,u=[],M=()=>{u.forEach(h=>h.classList.toggle("on",h.dataset.v===r))};return n.forEach(h=>{let w=b("button","dtk-seg-btn",h.label);w.type="button",w.dataset.v=h.value,w.onclick=()=>{r=h.value,M(),e(r)},u.push(w),o.appendChild(w)}),M(),S.appendChild(o),{row:S,get:()=>r,set:h=>{r=h,M()}}}function We(t,n,l){var M;let e=b("div","dtk-ctl"),m=b("div","dtk-ctl-head"),S=je(t,n.hint),o=b("span","dtk-ctl-val"),r=(M=n.fmt)!=null?M:(h=>String(h));m.appendChild(S),m.appendChild(o),e.appendChild(m);let u=b("input","dtk-range");return u.type="range",u.min=String(n.min),u.max=String(n.max),u.step=String(n.step),u.value=String(n.value),o.textContent=r(n.value),u.oninput=()=>{let h=Number(u.value);o.textContent=r(h),l(h)},e.appendChild(u),{row:e,get:()=>Number(u.value),setMax:h=>{u.max=String(h),Number(u.value)>h&&(u.value=String(h),o.textContent=r(h))}}}function pt(t){if(t.type==="manual_bounds"){let l={};return t.lowerBound!=null&&(l.lower_bound=t.lowerBound),t.upperBound!=null&&(l.upper_bound=t.upperBound),t.inputType!=="values"&&(l.input_type=t.inputType),l}let n={threshold:t.threshold,window_size:t.windowSize};return t.windowWeights!=="none"&&(n.window_weights=t.windowWeights,t.windowWeights==="exponential"&&t.halfLife!=null&&(n.half_life=t.halfLife)),t.detrend!=="none"&&(n.detrend=t.detrend),t.smoothing!=="none"&&(n.smoothing=t.smoothing),t.inputType!=="values"&&(n.input_type=t.inputType),t.seasonalityComponents&&t.seasonalityComponents.length&&(n.seasonality_components=t.seasonalityComponents,n.min_samples_per_group=t.minSamplesPerGroup),n}function Mt(t,n){let l=pt(t),e=[`type: ${t.type}`];for(let[m,S]of Object.entries(l))e.push(`${m}=${typeof S=="object"?JSON.stringify(S):S}`);return t.direction&&t.direction!=="any"&&e.push(`direction=${t.direction}`),e.push(`consecutive_anomalies=${n}`),e.join("  \xB7  ")}function Ct(t,n){var D,ne,xe;Dt(),n.classList.add(St),n.innerHTML="";let l=b("div","dtk-tune-root");n.appendChild(l);let e=t.points.length,m={timestamps:t.points.map(s=>s.t),values:t.points.map(s=>s.v==null?NaN:s.v),intervalSeconds:t.interval_seconds,truthAnomaly:new Array(e).fill(!1),seasonalityData:t.seasonality_columns.length?t.seasonality:void 0,seasonalityColumns:t.seasonality_columns.length?t.seasonality_columns:void 0},S=m,o=s=>{let p=Math.max(0,e-s);return p<=0?m:{timestamps:m.timestamps.slice(p),values:m.values.slice(p),intervalSeconds:m.intervalSeconds,truthAnomaly:m.truthAnomaly.slice(p),seasonalityData:m.seasonalityData?m.seasonalityData.slice(p):void 0,seasonalityColumns:m.seasonalityColumns}},r=t.detector,u=t.consecutive_anomalies,M=m.values.filter(s=>Number.isFinite(s)).sort((s,p)=>s-p),h=M.length?M[0]:0,w=M.length?M[M.length-1]:1,z=s=>M.length?M[Math.min(M.length-1,Math.max(0,Math.round(s*(M.length-1))))]:0,H=Math.max((w-h)*.05,1e-9),A=h-H,W=w+H,I=Math.max((W-A)/400,1e-9),Y=r.lowerBound!=null?r.lowerBound:z(.05),re=r.upperBound!=null?r.upperBound:z(.95),L=(D=r.seasonalityComponents)!=null?D:[],E=new Map;L.forEach((s,p)=>s.forEach(C=>E.set(C,p+1)));let j=()=>{let s=0;E.forEach(C=>{C>s&&(s=C)});let p=[];for(let C=1;C<=s;C++){let N=t.seasonality_columns.filter(P=>E.get(P)===C);N.length&&p.push(N)}return p.length?p:null},U=s=>{if(!s||!t.seasonality.length)return 0;let p=0;for(let C of s){let N=new Set;for(let P of t.seasonality)N.add(C.map(pe=>{var ze;return String((ze=P==null?void 0:P[pe])!=null?ze:"")}).join("|"));p=Math.max(p,N.size)}return p},le=()=>{var s;return{type:Me.get(),threshold:ve.get(),windowSize:a.get(),minSamples:r.minSamples,inputType:r.inputType,smoothing:y.get(),smoothingAlpha:r.smoothingAlpha,smoothingWindow:r.smoothingWindow,windowWeights:d.get(),halfLife:d.get()==="exponential"?c.get():null,detrend:v.get(),seasonalityComponents:j(),minSamplesPerGroup:(s=kt[Me.get()])!=null?s:r.minSamplesPerGroup,consecutiveAnomalies:u,direction:Q.get(),lowerBound:_e.get(),upperBound:Ee.get()}},de=b("div","dtk-tune-header"),O=b("div","dtk-tune-titlerow");O.appendChild(b("h1","dtk-tune-title",t.metric));let ie=b("span","dtk-tune-badge",t.save_url?"manual tuning":"preview");O.appendChild(ie),de.appendChild(O);let he=t.project?`${t.project} \xB7 `:"";de.appendChild(b("div","dtk-tune-sub",`${he}${e} points \xB7 ${At(t.interval_seconds)} grid`)),t.description&&de.appendChild(b("div","dtk-tune-desc",t.description)),l.appendChild(de);let ee=b("div","dtk-tune-grid"),V=b("div","dtk-tune-controls"),B=b("div","dtk-tune-main");ee.appendChild(V),ee.appendChild(B),l.appendChild(ee);let Re=b("div","dtk-tune-trim"),De=b("div","dtk-tune-trim-head");De.appendChild(je("Points shown","Trim the active sample to the most-recent N points. Fewer points recompute faster (cost grows with points \xD7 window) and make a shorter period easier to read \u2014 handy once you can see a smaller window/period is enough."));let me=b("span","dtk-tune-trim-val");De.appendChild(me),Re.appendChild(De);let te=b("input","dtk-range");te.type="range",te.min=String(Math.min(e,200)),te.max=String(e),te.step=String(Math.max(1,Math.round(e/200))),te.value=String(e),Re.appendChild(te),B.appendChild(Re);let se=b("div","dtk-tune-chart"),Be=b("canvas");se.appendChild(Be);let ue=b("div","dtk-tune-spin");ue.appendChild(b("span","dtk-spin-ring")),ue.appendChild(b("span","dtk-spin-txt","computing\u2026")),se.appendChild(ue),B.appendChild(se);let Oe=b("div","dtk-tune-legend"),be=(s,p,C)=>{let N=b("span","dtk-leg-item");N.title=C,N.appendChild(b("span",`dtk-leg-sw ${s}`)),N.appendChild(b("span","dtk-leg-txt",p)),Oe.appendChild(N)};be("line","metric","The metric value over time."),be("band","expected range","The detector's confidence band \u2014 values inside it read as normal."),be("center","band center","The expected value at the middle of the band."),be("dot","anomaly","A point the detector flagged as anomalous (outside the band)."),be("alert","alert","Where an alert fired \u2014 enough consecutive anomalies to meet the rule."),B.appendChild(Oe);let we=b("div","dtk-tune-readout");B.appendChild(we);let ye=b("div","dtk-tune-warn");ye.style.display="none",B.appendChild(ye);let Ue=s=>{let p=s.seasonalityComponents,C=U(p),N=s.minSamplesPerGroup*C;p&&C>0&&s.windowSize<N?(ye.textContent=`\u26A0 Seasonality inactive at this window: ${s.windowSize} < ${N} (min_samples_per_group ${s.minSamplesPerGroup} \xD7 ${C} key${C===1?"":"s"}). Each point keeps only ~${Math.floor(s.windowSize/C)} same-key point(s), so the band falls back to global statistics (seasonality has no effect). Raise the window to \u2265 ${N}.`,ye.style.display=""):ye.style.display="none"},qe=mt(Be,{navigable:!0,yFit:"data",onHover:s=>{if(!s||!s.point||!s.point.scored){we.textContent="";return}let p=s.point;we.textContent=`t=${Tt(p.timestamp)}  value=${Ie(p.value)}  band=[${Ie(p.lower)}, ${Ie(p.upper)}]`+(p.isAnomaly?`  \u26A0 ${p.direction} (sev ${p.severity.toFixed(2)})`:"")}}),ke=new Worker(URL.createObjectURL(new Blob([`"use strict";(()=>{var X={mad:1,zscore:2,iqr:4};function C(e){let n=0;for(let t of e)n+=t;return n}function un(e){let n=e.map((t,o)=>o);return n.sort((t,o)=>e[t]<e[o]?-1:e[t]>e[o]?1:t-o),n}function cn(e,n,t){let o=n.length;if(o===0)return NaN;if(e<=n[0])return t[0];if(e>=n[o-1])return t[o-1];for(let i=1;i<o;i++)if(e<=n[i]){let r=n[i-1],s=n[i],c=t[i-1],u=t[i];if(s===r)return c;let d=(e-r)/(s-r);return c+d*(u-c)}return t[o-1]}function j(e,n,t){let o=un(e),i=o.map(b=>e[b]),r=o.map(b=>n[b]),s=C(r),c=new Array(r.length),u=0;for(let b=0;b<r.length;b++)u+=r[b],c[b]=(u-.5*r[b])/s;let d=t/100;return cn(d,c,i)}function P(e,n){return j(e,n,50)}function ln(e,n,t){let o=t===void 0?P(e,n):t,i=e.map(r=>Math.abs(r-o));return P(i,n)}function mn(e,n){let t=C(n),o=0;for(let i=0;i<e.length;i++)o+=e[i]*(n[i]/t);return o}function an(e,n,t,o=0){let i=C(n),r=n.map(u=>u/i),s=t===void 0?C(e.map((u,d)=>u*r[d])):t,c=0;for(let u=0;u<e.length;u++)c+=r[u]*(e[u]-s)**2;if(o===1){let u=1-C(r.map(d=>d*d));u>1e-12&&(c/=u)}return Math.sqrt(c)}function fn(e,n){return n.smoothing==="none"?e.slice():n.smoothing==="ema"?dn(e,n.smoothingAlpha):bn(e,n.smoothingWindow)}function dn(e,n){let t=e.length,o=new Array(t).fill(NaN);if(t===0)return o;let i=-1;for(let r=0;r<t;r++)if(!Number.isNaN(e[r])){i=r;break}if(i===-1)return o;o[i]=e[i];for(let r=i+1;r<t;r++)Number.isNaN(e[r])?o[r]=o[r-1]:o[r]=n*e[r]+(1-n)*o[r-1];return o}function bn(e,n){let t=e.length,o=new Array(t).fill(NaN);for(let i=0;i<t;i++){let r=Math.max(0,i-n+1),s=0,c=0;for(let u=r;u<=i;u++)Number.isNaN(e[u])||(s+=e[u],c+=1);o[i]=c>0?s/c:NaN}return o}function Y(e,n){let t=e.length;if(n.inputType==="values")return e.slice();let o=new Array(t).fill(NaN);for(let i=1;i<t;i++){let r=e[i-1],s=e[i];n.inputType==="changes"?o[i]=(s-r)/r:n.inputType==="absolute_changes"?o[i]=s-r:o[i]=Math.log(s+1)-Math.log(r+1)}return o}function pn(e,n){return e.halfLife!==null?e.halfLife:Math.max(e.windowSize/20,n/2,1)}function yn(e,n){if(e.windowWeights==="none")return null;let t=e.windowSize,o=new Array(t);if(e.windowWeights==="exponential"){let i=pn(e,n);for(let r=1;r<=t;r++){let s=Math.min(r/i,1e3);o[r-1]=Math.pow(.5,s)}return o}for(let i=1;i<=t;i++)o[i-1]=(t+1-i)/t;return o}function Q(e,n){return n===null?e.map(()=>1):e.map(t=>n[t-1])}function gn(e){let n=e[0];for(let t of e)t>n&&(n=t);return n}function hn(e){let n=e[0];for(let t of e)t<n&&(n=t);return n}function Sn(e,n,t){if(e.length<4)return 0;let o=(gn(n)+hn(n))/2,i=n.map(a=>a>o),r=i.map(a=>!a),s=0,c=0;for(let a=0;a<n.length;a++)i[a]?s++:c++;if(s<2||c<2)return 0;let u=(a,M)=>M.filter((x,w)=>a[w]),d=u(r,e),b=u(i,e),N=u(r,n),y=u(i,n),h=u(r,t),m=u(i,t),A=P(d,h),p=P(b,m),S=P(N,h),g=P(y,m);return g===S?0:(A-p)/(g-S)}var wn={mad:[["median","center"],["mad","spread"]],zscore:[["mean","center"],["std","spread"]],iqr:[["q1","center"],["q3","center"],["iqr","spread"]]};function U(e,n,t){if(e==="mad"){let r=P(n,t);return{median:r,mad:ln(n,t,r)}}if(e==="zscore"){let r=mn(n,t);return{mean:r,std:an(n,t,r,1)}}let o=j(n,t,25),i=j(n,t,75);return{q1:o,q3:i,iqr:i-o}}function Nn(e,n,t){if(e==="mad"){if(n.mad===0)return[n.median-1e-10,n.median+1e-10];let o=t*1.4826*n.mad;return[n.median-o,n.median+o]}return e==="zscore"?n.std===0?[n.mean-1e-10,n.mean+1e-10]:[n.mean-t*n.std,n.mean+t*n.std]:n.iqr===0?[n.q1-1e-10,n.q3+1e-10]:[n.q1-t*n.iqr,n.q3+t*n.iqr]}function An(e,n,t){if(e==="mad"){let o=1.4826*n.mad;return o>0?t/o:1/0}return e==="zscore"?n.std>0?t/n.std:1/0:n.iqr>0?t/n.iqr:1/0}function Mn(e,n){return e==="mad"?n.median:e==="zscore"?n.mean:(n.q1+n.q3)/2}function F(e,n,t,o,i){return{index:e,timestamp:n,value:t,processedValue:o,scored:!1,isAnomaly:!1,lower:NaN,upper:NaN,center:NaN,direction:null,severity:0,reason:i}}function Dn(e,n){var d,b,N;let{timestamps:t,values:o}=e,i=t.length,r=Y(o,n),s=(d=n.lowerBound)!=null?d:null,c=(b=n.upperBound)!=null?b:null,u=[];for(let y=0;y<i;y++){let h=o[y],m=r[y],A=t[y];if(Number.isNaN(m)){u.push(F(y,A,h,m,"missing_data"));continue}let p=!1,S=null,g=0;s!==null&&m<s&&(p=!0,S="below",g=s-m),c!==null&&m>c&&(p=!0,S="above",g=m-c);let a=0;if(p)if(s!==null&&c!==null){let v=c-s;a=v>0?g/v:1/0}else a=g;let M=s!=null?s:-1/0,x=c!=null?c:1/0,w=s!==null&&c!==null?(s+c)/2:(N=s!=null?s:c)!=null?N:NaN;u.push({index:y,timestamp:A,value:h,processedValue:m,scored:!0,isAnomaly:p,lower:M,upper:x,center:w,direction:S,severity:a,reason:"ok"})}return u}function Z(e,n){if(n.type==="manual_bounds")return Dn(e,n);let t=n.type,o=Math.max(n.minSamples,X[t]),i=wn[t],{timestamps:r,values:s}=e,c=r.length,u=fn(s,n),d=Y(u,n),b=n.seasonalityComponents!==null&&n.seasonalityComponents.length>0&&Array.isArray(e.seasonalityData)&&e.seasonalityData.length>0,N=e.seasonalityData,y=yn(n,o),h=[];for(let m=0;m<c;m++){let A=s[m],p=d[m],S=r[m];if(Number.isNaN(p)){h.push(F(m,S,A,p,"missing_data"));continue}let g=Math.max(0,m-n.windowSize),a=m-g,M=d.slice(g,m),x=M.map(l=>!Number.isNaN(l)),w=[];for(let l=0;l<M.length;l++)x[l]&&w.push(M[l]);if(w.length<o){h.push(F(m,S,A,p,"insufficient_data"));continue}let v=new Array(a);for(let l=0;l<a;l++)v[l]=a-l;let L=[];for(let l=0;l<a;l++)x[l]&&L.push(v[l]);let B=Q(L,y),W=0;n.detrend==="linear"&&(W=Sn(w,L,B));let nn=W!==0?w.map((l,R)=>l+W*L[R]):w,G=U(t,nn,B),_={...G};if(b&&N){let l=N[m];for(let R of n.seasonalityComponents){let K=new Array(a).fill(!0);for(let f=0;f<a;f++){let q=N[g+f],T=!0;for(let V of R)if(!q||q[V]!==(l==null?void 0:l[V])){T=!1;break}K[f]=T}let z=[],E=[];for(let f=0;f<a;f++)x[f]&&K[f]&&(z.push(M[f]),E.push(v[f]));if(z.length<n.minSamplesPerGroup)continue;let tn=Q(E,y),on=W!==0?z.map((f,q)=>f+W*E[q]):z,rn=U(t,on,tn);for(let[f,q]of i){let T=G[f],sn=(q==="spread"?T>0:T!==0)?rn[f]/T:1;_[f]*=sn}}}let[D,k]=Nn(t,_,n.threshold);D>k&&([D,k]=[k,D]);let H=p<D||p>k,O=null,J=0;if(H){let l;p<D?(O="below",l=D-p):(O="above",l=p-k),J=An(t,_,l)}let en=Mn(t,_);h.push({index:m,timestamp:S,value:A,processedValue:p,scored:!0,isAnomaly:H,lower:D,upper:k,center:en,direction:O,severity:J,reason:"ok"})}return h}function Pn(e,n){let t=e.seasonalityData;if(!t||t.length===0)return 0;let o=0;for(let i of n){let r=new Set;for(let s of t)r.add(i.map(c=>{var u;return String((u=s==null?void 0:s[c])!=null?u:"")}).join("|"));o=Math.max(o,r.size)}return o}function $(e,n){let t=e.timestamps.length;if(n.type==="manual_bounds")return Math.min(n.inputType!=="values"?1:0,t);let o=Math.max(n.minSamples,X[n.type]);if(n.smoothing==="sma"&&(o=Math.max(o,n.smoothingWindow-1)),n.smoothing==="ema"&&(o=Math.max(o,Math.ceil(5/n.smoothingAlpha))),n.inputType!=="values"&&(o=Math.max(o,1)),n.seasonalityComponents!==null&&n.seasonalityComponents.length>0&&Array.isArray(e.seasonalityData)&&e.seasonalityData.length>0){let r=Pn(e,n.seasonalityComponents);if(r>0){let s=n.minSamplesPerGroup*r;n.windowSize>=s&&(o=Math.max(o,s))}}return Math.min(o,t)}var I=null;function xn(e,n){if(!n||n==="any")return;let t=n==="up"?"above":"below";for(let o of e)o.isAnomaly&&o.direction!==t&&(o.isAnomaly=!1)}function vn(e,n,t){let o=[],i=0;for(let r=0;r<e.length;r++){if(!(e[r].scored&&e[r].isAnomaly)){i=0;continue}let c=r>0&&e[r-1].scored&&e[r-1].isAnomaly,u=r>0&&e[r].timestamp-e[r-1].timestamp===n;i=c&&u?i+1:1,i===t&&o.push(r)}return o}self.onmessage=e=>{let n=e.data;if(n.type==="series"){I=n.series;return}if(n.type==="run"&&I){let t=n.params,o=Z(I,t);xn(o,t.direction);let i=I.intervalSeconds*1e3,r=vn(o,i,t.consecutiveAnomalies),s=$(I,t),c=0;for(let u of o)u.scored&&u.isAnomaly&&c++;self.postMessage({type:"result",id:n.id,scored:o,fires:r,eff:s,flagged:c})}};})();
+`],{type:"text/javascript"})));ke.postMessage({type:"series",series:S});let fe=0,ge=null;ke.onmessage=s=>{let p=s.data;if(p.type!=="result"||p.id!==fe||!ge)return;ue.classList.remove("on");let C=ge,N=p.fires.map(P=>({t:S.timestamps[P],kind:"anomaly"}));qe.render({series:S,scored:p.scored,params:C,alerts:N}),q.textContent=`${p.flagged} flagged \xB7 ${p.fires.length} alert${p.fires.length===1?"":"s"} \xB7 warm-up ${p.eff} pts`,Ue(C),k.textContent=Mt(C,u)},ke.onerror=()=>{ue.classList.remove("on"),q.textContent="recompute failed \u2014 see the browser console"};let oe=()=>{ge=le(),fe+=1,ue.classList.add("on"),ke.postMessage({type:"run",id:fe,params:ge})},G=s=>s>1?et(m.timestamps[e-1]-m.timestamps[e-s]):"\u2014",Se=s=>{me.textContent=s>=e?`${e} pts \xB7 full (${G(e)})`:`${s} pts \xB7 ${G(s)}`};Se(e);function Ve(s){S=o(s),ke.postMessage({type:"series",series:S}),J()}let Pe=0;te.oninput=()=>{let s=Number(te.value);Se(s),Pe&&window.clearTimeout(Pe),Pe=window.setTimeout(()=>Ve(s),200)};let Le=0,J=()=>{Le&&window.clearTimeout(Le),Le=window.setTimeout(oe,130)},Me=He("Detector",[{label:"MAD",value:"mad"},{label:"Z-Score",value:"zscore"},{label:"IQR",value:"iqr"},{label:"Manual",value:"manual_bounds"}],r.type,s=>{var p;Ye((p=yt[s])!=null?p:3),$(),J()},"The statistic for the band: MAD (robust median, default), Z-Score (mean/std) or IQR (quartiles) \u2014 all windowed \u2014 or Manual (fixed lower/upper thresholds, no window/history).");V.appendChild(Me.row);let _e=We("Lower bound",{min:A,max:W,step:I,value:Y,fmt:s=>Ie(s),hint:"Manual bounds: values below this read as anomalous. Drag in from the data range to see how many points fall outside (and how many alerts that yields)."},J);V.appendChild(_e.row);let Ee=We("Upper bound",{min:A,max:W,step:I,value:re,fmt:s=>Ie(s),hint:"Manual bounds: values above this read as anomalous."},J);V.appendChild(Ee.row);let ve=We("Threshold (\u03C3-equivalent)",{min:.5,max:10,step:.1,value:r.threshold,fmt:s=>s.toFixed(1),hint:"Band half-width in \u03C3-equivalents. Lower = tighter band = more flags; higher = wider band = fewer flags."},J);V.appendChild(ve.row);let $e=ve.row.querySelector("input"),Fe=ve.row.querySelector(".dtk-ctl-val"),Ye=s=>{$e&&($e.value=String(s)),Fe&&(Fe.textContent=s.toFixed(1))},Ge=Math.max(50,Math.min(2e3,Math.floor(e/2))),i=Math.max(Ge,r.windowSize),a=We("Window size (points)",{min:Math.max(1,Math.min(10,r.windowSize)),max:i,step:1,value:r.windowSize,fmt:s=>`${s} \xB7 ${et(s*t.interval_seconds*1e3)}`,hint:"How many trailing points form the baseline window for each scored point. Larger = steadier baseline (more history); smaller = adapts faster to shifts."},J);V.appendChild(a.row);let d=He("Recency weighting",[{label:"none",value:"none"},{label:"exponential",value:"exponential"},{label:"linear",value:"linear"}],r.windowWeights,()=>{$(),J()},"Weight recent points in the window more heavily: none (flat), exponential (half-life decay) or linear. Helps the baseline track a drifting level.");V.appendChild(d.row);let c=We("Half-life (points)",{min:1,max:i,step:1,value:(ne=r.halfLife)!=null?ne:Math.max(5,Math.round(r.windowSize/20)),fmt:s=>`${s} \xB7 ${et(s*t.interval_seconds*1e3)}`,hint:"Exponential weighting only: the age (in points) at which a point counts half as much as the newest. Smaller = faster decay = fresher baseline."},J),f=c.row;f.style.display=r.windowWeights==="exponential"?"":"none",V.appendChild(f);let v=He("Detrend",[{label:"none",value:"none"},{label:"linear",value:"linear"}],r.detrend,J,"Remove a robust linear trend from each window before computing the band, so a steadily rising/falling metric is not flagged for the trend itself.");V.appendChild(v.row);let y=He("Smoothing",[{label:"none",value:"none"},{label:"EMA",value:"ema"},{label:"SMA",value:"sma"}],r.smoothing,J,"Smooth the series before detection (EMA or SMA) so single-point jitter does not flag. The detector judges the smoothed line; the raw values show as a faint ghost.");V.appendChild(y.row);let _=null;if(t.seasonality_columns.length){let s=t.seasonality_columns,p=b("div","dtk-ctl");_=p,p.appendChild(je("Seasonality groups","Condition the band on seasonal keys. Pick a group per column: columns in the SAME group are combined into one key (e.g. dow\xD7hour); separate groups each apply their own correction. Off = ignore that column."));let C=Math.min(s.length,6),N=[{label:"\u2014",value:"0"}];for(let P=1;P<=C;P++)N.push({label:`G${P}`,value:String(P)});s.forEach(P=>{let pe=b("div","dtk-season-row");pe.appendChild(b("span","dtk-season-col",P));let ze=b("div","dtk-seg dtk-season-seg"),tt=[],ht=()=>{var ce;return(ce=E.get(P))!=null?ce:0},nt=()=>tt.forEach(ce=>ce.classList.toggle("on",Number(ce.dataset.v)===ht()));N.forEach(ce=>{let Ne=b("button","dtk-seg-btn",ce.label);Ne.type="button",Ne.dataset.v=ce.value,Ne.title=ce.value==="0"?`ignore ${P}`:`put ${P} in group ${ce.value}`,Ne.onclick=()=>{E.set(P,Number(ce.value)),nt(),J()},tt.push(Ne),ze.appendChild(Ne)}),nt(),pe.appendChild(ze),p.appendChild(pe)}),V.appendChild(p)}let Q=He("Direction",[{label:"both",value:"any"},{label:"up",value:"up"},{label:"down",value:"down"}],(xe=t.direction)!=null?xe:"any",J,"Which anomalies to show and count toward alerts: both directions, only spikes ABOVE the band (up) or only drops BELOW it (down). A preview filter mirroring the alert direction policy \u2014 it never changes the band itself.");V.appendChild(Q.row);let K=We("Alert: consecutive anomalies",{min:1,max:10,step:1,value:u,fmt:s=>String(s),hint:"How many anomalies in a row are required before an alert fires. Higher = fewer, more-confident alerts (the \u25BC markers on the chart)."},s=>{u=s,J()});V.appendChild(K.row);let R=[ve.row,a.row,d.row,v.row,y.row];_&&R.push(_);function $(){let s=Me.get()==="manual_bounds";for(let p of R)p.style.display=s?"none":"";_e.row.style.display=s?"":"none",Ee.row.style.display=s?"":"none",f.style.display=!s&&d.get()==="exponential"?"":"none"}$();let q=b("div","dtk-tune-stat");B.appendChild(q);let g=b("div","dtk-tune-cfg");g.appendChild(b("span","dtk-tune-cfg-k","// effective config"));let k=b("code","dtk-tune-cfg-v");if(g.appendChild(k),B.appendChild(g),t.save_url){let s=b("div","dtk-tune-apply"),p=b("button","dtk-apply-btn","Apply to metric");p.type="button",p.title="Write this detector config back into the metric YAML (the previous version is archived under metrics/.history/). Trimming the sample does not change what is written.";let C=b("span","dtk-apply-msg");p.onclick=()=>{let N=le();p.disabled=!0,C.className="dtk-apply-msg info",C.textContent="Applying\u2026",fetch(t.save_url,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({detector:{type:N.type,params:pt(N)},consecutive_anomalies:u})}).then(P=>P.ok?P.json():P.text().then(pe=>{throw new Error(pe||`HTTP ${P.status}`)})).then(P=>{var pe;C.className="dtk-apply-msg ok",C.textContent=`Applied \u2192 ${(pe=P.saved)!=null?pe:"metric"} (previous archived). You can close this tab.`}).catch(P=>{p.disabled=!1,C.className="dtk-apply-msg err",C.textContent=`Apply failed: ${P.message}`})},s.appendChild(p),s.appendChild(C),B.appendChild(s)}else B.appendChild(b("div","dtk-tune-note","Static preview \u2014 sliders recompute live, but there is no write-back. Run `dtk tune` (without --no-serve) to apply a config."));oe();let x=0;window.addEventListener("resize",()=>{x&&cancelAnimationFrame(x),x=requestAnimationFrame(()=>qe.resize())})}function Ie(t){if(!Number.isFinite(t))return"\u2014";let n=Math.abs(t);return n!==0&&(n<.01||n>=1e6)?t.toExponential(2):Number(t.toFixed(n<1?4:2)).toString()}function Tt(t){return new Date(t).toISOString().slice(0,16).replace("T"," ")}function et(t){let n=Math.round(t/6e4);if(n<60)return`${n}m`;let l=Math.floor(n/60);if(l<24)return`${l}h`;let e=Math.floor(l/24),m=l%24;return m?`${e}d ${m}h`:`${e}d`}function At(t){return t%86400===0?`${t/86400}d`:t%3600===0?`${t/3600}h`:t%60===0?`${t/60}min`:`${t}s`}var ft=!1;function Dt(){if(ft)return;ft=!0;let t=`
+.dtk-tune{--c:#d15b36;--c7:#b4471f;--ink:#1b1916;--muted:#6e675b;--faint:#9a9384;
+  --paper:#f5f1e8;--surface:#fbf9f3;--border:#e6e0d4;--green:#2e9e73;--anom:#d63232;
+  --mono:'JetBrains Mono',ui-monospace,Menlo,monospace;
+  --sans:'Schibsted Grotesk',system-ui,-apple-system,Segoe UI,Roboto,sans-serif;}
+.dtk-tune-root{max-width:1200px;margin:0 auto;padding:24px 20px 56px;font-family:var(--sans);color:var(--ink);}
+.dtk-tune-titlerow{display:flex;align-items:center;gap:12px;}
+.dtk-tune-title{font-size:24px;margin:0;font-weight:700;}
+.dtk-tune-badge{font-family:var(--mono);font-size:11px;text-transform:uppercase;letter-spacing:.06em;
+  color:#fff;background:var(--c);border-radius:999px;padding:3px 10px;}
+.dtk-tune-sub{color:var(--muted);font-size:13px;margin-top:4px;font-family:var(--mono);}
+.dtk-tune-desc{color:var(--muted);font-size:13px;margin-top:8px;white-space:pre-wrap;}
+.dtk-tune-grid{display:grid;grid-template-columns:280px 1fr;gap:24px;margin-top:20px;align-items:start;}
+@media(max-width:820px){.dtk-tune-grid{grid-template-columns:1fr;}}
+.dtk-tune-controls{display:flex;flex-direction:column;gap:16px;background:var(--surface);
+  border:1px solid var(--border);border-radius:12px;padding:16px;}
+.dtk-ctl{display:flex;flex-direction:column;gap:6px;}
+.dtk-ctl-head{display:flex;justify-content:space-between;align-items:baseline;}
+.dtk-ctl-label{font-size:12px;font-weight:600;color:var(--ink);}
+.dtk-ctl-val{font-family:var(--mono);font-size:12px;color:var(--c7);}
+.dtk-seg{display:flex;gap:4px;background:var(--paper);border:1px solid var(--border);border-radius:8px;padding:3px;}
+.dtk-seg.dtk-wrap{flex-wrap:wrap;}
+.dtk-seg-btn{flex:1 1 auto;border:0;background:transparent;color:var(--muted);font-family:var(--sans);
+  font-size:12px;padding:5px 8px;border-radius:6px;cursor:pointer;white-space:nowrap;}
+.dtk-seg-btn:hover{color:var(--ink);}
+.dtk-seg-btn.on{background:var(--c);color:#fff;font-weight:600;}
+.dtk-range{width:100%;accent-color:var(--c);cursor:pointer;}
+.dtk-check{display:flex;align-items:center;gap:6px;font-size:12px;color:var(--muted);margin-top:2px;cursor:pointer;}
+.dtk-tune-main{display:flex;flex-direction:column;gap:10px;min-width:0;}
+.dtk-tune-chart{position:relative;width:100%;height:470px;background:var(--surface);
+  border:1px solid var(--border);border-radius:12px;overflow:hidden;}
+.dtk-tune-chart canvas{width:100%;height:100%;display:block;}
+.dtk-tune-readout{font-family:var(--mono);font-size:12px;color:var(--muted);min-height:18px;}
+.dtk-tune-stat{font-family:var(--mono);font-size:12px;color:var(--ink);}
+.dtk-tune-warn{font-family:var(--mono);font-size:12px;line-height:1.5;color:var(--c7);
+  background:rgba(240,173,78,0.13);border:1px solid rgba(240,173,78,0.5);border-radius:8px;padding:8px 11px;}
+.dtk-tune-cfg{background:var(--ink);color:#c9c2b4;border-radius:8px;padding:10px 12px;font-family:var(--mono);
+  font-size:12px;overflow-x:auto;}
+.dtk-tune-cfg-k{color:var(--faint);display:block;margin-bottom:4px;}
+.dtk-tune-cfg-v{color:#e6e0d4;white-space:pre-wrap;word-break:break-word;}
+.dtk-tune-apply{display:flex;align-items:center;gap:12px;flex-wrap:wrap;margin-top:6px;}
+.dtk-apply-btn{background:var(--c);color:#fff;border:0;border-radius:8px;padding:10px 18px;font-family:var(--sans);
+  font-size:14px;font-weight:600;cursor:pointer;}
+.dtk-apply-btn:hover{background:var(--c7);}
+.dtk-apply-btn:disabled{opacity:.55;cursor:default;}
+.dtk-apply-msg{font-size:13px;}
+.dtk-apply-msg.ok{color:var(--green);}
+.dtk-apply-msg.err{color:var(--anom);}
+.dtk-apply-msg.info{color:var(--muted);}
+.dtk-tune-note{font-size:13px;color:var(--muted);background:var(--surface);border:1px dashed var(--border);
+  border-radius:8px;padding:10px 12px;}
+.dtk-ctl-info{color:var(--faint);font-size:10px;cursor:help;vertical-align:super;}
+.dtk-tune-trim{display:flex;flex-direction:column;gap:6px;background:var(--surface);
+  border:1px solid var(--border);border-radius:10px;padding:9px 12px;}
+.dtk-tune-trim-head{display:flex;justify-content:space-between;align-items:baseline;}
+.dtk-tune-trim-val{font-family:var(--mono);font-size:12px;color:var(--c7);}
+.dtk-tune-spin{position:absolute;top:10px;right:12px;display:none;align-items:center;gap:7px;
+  background:rgba(27,25,22,0.78);color:#e6e0d4;border:1px solid #332f29;border-radius:999px;
+  padding:4px 11px 4px 8px;font-family:var(--mono);font-size:11px;pointer-events:none;}
+.dtk-tune-spin.on{display:inline-flex;}
+.dtk-spin-ring{width:12px;height:12px;border-radius:50%;border:2px solid rgba(245,241,232,0.25);
+  border-top-color:var(--c);animation:dtk-spin .7s linear infinite;}
+@keyframes dtk-spin{to{transform:rotate(360deg);}}
+.dtk-tune-legend{display:flex;flex-wrap:wrap;gap:14px;font-size:12px;color:var(--muted);padding:2px 2px 0;}
+.dtk-leg-item{display:inline-flex;align-items:center;gap:6px;cursor:help;}
+.dtk-leg-sw{display:inline-block;flex:0 0 auto;}
+.dtk-leg-sw.line{width:16px;height:3px;background:var(--c);border-radius:2px;}
+.dtk-leg-sw.band{width:16px;height:11px;background:rgba(209,91,54,0.18);
+  border:1px solid rgba(209,91,54,0.5);border-radius:2px;}
+.dtk-leg-sw.center{width:16px;height:2px;
+  background:repeating-linear-gradient(90deg,var(--faint) 0 4px,transparent 4px 7px);}
+.dtk-leg-sw.dot{width:9px;height:9px;border-radius:50%;background:var(--anom);}
+.dtk-leg-sw.alert{width:0;height:0;border-left:5px solid transparent;border-right:5px solid transparent;
+  border-top:7px solid var(--anom);}
+.dtk-leg-txt{white-space:nowrap;}
+.dtk-season-row{display:flex;align-items:center;justify-content:space-between;gap:8px;}
+.dtk-season-col{font-family:var(--mono);font-size:11.5px;color:var(--muted);
+  overflow:hidden;text-overflow:ellipsis;white-space:nowrap;}
+.dtk-season-seg{flex:0 0 auto;padding:2px;}
+.dtk-season-seg .dtk-seg-btn{flex:0 0 auto;padding:3px 7px;font-family:var(--mono);font-size:11px;}
+`,n=document.createElement("style");n.textContent=t,document.head.appendChild(n)}window.__DTK_TUNE__={render:Ct};})();

{detectkit-0.31.1 → detectkit-0.33.0}/detectkit/tuning/config_writer.py

@@ -20,9 +20,10 @@ import yaml
 from detectkit.config.metric_config import MetricConfig
 from detectkit.detectors.factory import DetectorFactory
 
-# The detector types the interactive tuner can emit (windowed statistical). Other
-# types (e.g. manual_bounds) need domain thresholds, not interactive tuning.
-_TUNABLE_TYPES = {"mad", "zscore", "iqr"}
+# The detector types the interactive tuner can emit: the windowed statistical
+# detectors plus the stateless manual_bounds (lower/upper threshold) detector,
+# whose bounds the tuner lets you drag against the real series.
+_TUNABLE_TYPES = {"mad", "zscore", "iqr", "manual_bounds"}
 
 _RULE = "# " + "─" * 61
 

{detectkit-0.31.1 → detectkit-0.33.0}/detectkit/tuning/payload.py

@@ -30,12 +30,25 @@ from detectkit.reporting.builder import _ms, _num_or_none, _parse_seasonality
 _TUNE_COMPUTE_BUDGET = 20_000_000  # ~points x window per recompute (off-thread)
 _TUNE_MIN_POINTS = 3000
 _TUNE_MAX_POINTS = 15000
+# A point only gets a band where the trailing window can fill, so showing fewer
+# than a few windows of history would leave a large-window metric with almost no
+# scored region (the band never reaches its real width). Floor the shown points
+# at this many windows so the seeded window is actually exercised in the preview.
+_TUNE_WINDOW_COVERAGE = 3
 
 
 def default_window_points(seed_window: int) -> int:
-    """Smart default point count for a seeded window size (budget-bound + clamped)."""
+    """Smart default point count for a seeded window size.
+
+    Inversely sized to the window under a recompute budget (small windows can
+    afford far more points), but floored at a few windows' worth of history so a
+    large-window metric still has a meaningful scored region, and clamped to a
+    render/payload-comfortable range.
+    """
     w = max(int(seed_window or 0), 1)
-    return max(_TUNE_MIN_POINTS, min(_TUNE_MAX_POINTS, round(_TUNE_COMPUTE_BUDGET / w)))
+    budget_points = round(_TUNE_COMPUTE_BUDGET / w)
+    fill_points = w * _TUNE_WINDOW_COVERAGE
+    return max(_TUNE_MIN_POINTS, min(_TUNE_MAX_POINTS, max(budget_points, fill_points)))
 
 
 # Per-type interval-width defaults (mirror the detector class defaults and the
@@ -43,7 +56,9 @@ def default_window_points(seed_window: int) -> int:
 _THRESHOLD_DEFAULT = {"mad": 3.0, "zscore": 3.0, "iqr": 1.5}
 # Per-type min-samples-per-group defaults (mirror the detector subclasses).
 _MIN_SAMPLES_PER_GROUP_DEFAULT = {"mad": 10, "zscore": 3, "iqr": 4}
-_TUNABLE_TYPES = ("mad", "zscore", "iqr")
+# Detector types the interactive tuner can seed/emit: the windowed statistical
+# detectors plus the stateless manual_bounds (lower/upper threshold) detector.
+_TUNABLE_TYPES = ("mad", "zscore", "iqr", "manual_bounds")
 
 
 def _normalize_seasonality_components(
@@ -65,9 +80,12 @@ def _normalize_seasonality_components(
 def _seed_detector(metric_config: MetricConfig) -> dict[str, Any]:
     """Map the metric's current detector config to the TS ``DetectorParams`` shape.
 
-    Picks the first tunable (mad/zscore/iqr) detector to seed the sliders; falls
-    back to MAD defaults when the metric has none (e.g. only ``manual_bounds``).
-    Param names are camelCased to match the client port.
+    Picks the first tunable (mad/zscore/iqr/manual_bounds) detector to seed the
+    sliders; falls back to MAD defaults when the metric has none. Param names are
+    camelCased to match the client port. The windowed knobs always carry sane
+    defaults (even for a manual_bounds seed) so switching detector type in the UI
+    never hits an empty slider; ``lowerBound``/``upperBound`` carry the manual
+    thresholds (``None`` for a windowed metric — the client picks a data default).
     """
     chosen = next(
         (d for d in metric_config.detectors if d.type in _TUNABLE_TYPES),
@@ -99,10 +117,26 @@ def _seed_detector(metric_config: MetricConfig) -> dict[str, Any]:
         "minSamplesPerGroup": params.get(
             "min_samples_per_group", _MIN_SAMPLES_PER_GROUP_DEFAULT.get(dtype, 10)
         ),
+        # manual_bounds thresholds (None for a windowed metric).
+        "lowerBound": params.get("lower_bound"),
+        "upperBound": params.get("upper_bound"),
     }
     return seed
 
 
+def _seed_direction(metric_config: MetricConfig) -> str:
+    """Seed the tuner's direction view-filter from the first alerting block.
+
+    The tuner offers ``any``/``up``/``down``; the alerting policy may also be
+    ``same`` (a multi-detector agreement rule). Tuning is single-detector, so
+    ``same`` reads as ``any`` for the preview. Defaults to ``any``.
+    """
+    if not metric_config.alerting:
+        return "any"
+    raw = (metric_config.alerting[0].direction or "any").lower()
+    return raw if raw in ("up", "down", "any") else "any"
+
+
 def build_tune_payload(
     *,
     metric_config: MetricConfig,
@@ -141,6 +175,7 @@ def build_tune_payload(
     consecutive = 3
     if metric_config.alerting:
         consecutive = metric_config.alerting[0].consecutive_anomalies
+    direction = _seed_direction(metric_config)
 
     empty: dict[str, Any] = {
         "metric": name,
@@ -153,6 +188,7 @@ def build_tune_payload(
         "seasonality_columns": [],
         "detector": seed,
         "consecutive_anomalies": consecutive,
+        "direction": direction,
         "save_url": save_url,
     }
     if start is None or end is None:
@@ -190,5 +226,6 @@ def build_tune_payload(
         "seasonality_columns": season_cols,
         "detector": seed,
         "consecutive_anomalies": consecutive,
+        "direction": direction,
         "save_url": save_url,
     }