manage/openspec/changes/archive/2026-07-08-prometheus-direct-charting/design.md

SDD Design: Prometheus Direct Charting (drop Grafana middleman)

Change: prometheus-direct-charting Phase: design Date: 2026-07-08

0. Source findings (read before anything else)

The proposal and spec were written against a stale project map. Reading actual source surfaced deviations the design must account for. Trust source, not the map.

Spec claim	Actual source reality	Design impact
SC-118: "remove the Grafana branch from ObservabilityPage"	ObservabilityPage.tsx no longer exists. It was refactored into a per-service-type service-tabs/ architecture (confirmed by LinksTab.tsx docstring: "Lifts the Grafana deep-link content from the old cross-service ObservabilityPage into an instance-scoped tab"). The stale reference survives only in .pi-map.md files.	Removal targets are service-tabs/LinksTab.tsx + its test, the grafana case in service-tabs/index.ts, the navEntries.ts grafana entry, Dashboard.tsx's OBSERVABILITY_TYPES set, ServicesPage.tsx empty-state text, and the useGrafanaStatus hook. See §5.
"No recharts/d3 in use" (config rule to repeal)	recharts ^3.9.2 is declared, installed, and imported by GrafanaChartWidget.tsx. Confirmed.	Repeal is a doc fix matching reality; recharts is the sanctioned renderer.
Proposal §5.1: "extend PrometheusWidgetSource ... to handle chart"	PrometheusWidgetSource.fetch currently only does instant /api/v1/query and returns {"result": data}. The frontend PrometheusMetricWidget consumes data.result.	The chart/mean paths return a different shape ({series} / {value}); dispatch inside .fetch() by widget_kind. See §2.2.
Map said registry has 7 service types	Source registry has 8: alertmanager, authentik, backups, grafana, jellyfin, nextcloud, prometheus, ssh_tasks. (authentik and backups were added after the map froze.)	No design impact beyond acknowledging grafana is one of eight, not one of seven.
Frontend SERVICE_REGISTRY (registry.ts)	Has a grafana binding with link + chart kinds, and a prometheus binding with only metric. The chart kind must move grafana→prometheus.	Confirmed; §3.2 details the rebinding.

No proposal/spec scope change is required — the intent (remove Grafana, direct Prom charting) still holds. Only the removal targets differ from what SC-118 literally names. This is flagged explicitly so the tasks phase and reviewer aren't surprised.

---

1. Architecture overview

This change cuts Grafana out of the chart data path and makes Prometheus the direct source. The existing recharts renderer is reused unchanged for chart; two new widget kinds (gauge, mean) are added to the prometheus service. Grafana is removed entirely.

 BEFORE                                   AFTER
 ───────                                   ─────
 WidgetData fetch                         WidgetData fetch
   └► GrafanaWidgetSource                   └► PrometheusWidgetSource
        └► POST {grafana}/api/ds/query            ├► kind=metric  → /api/v1/query      {result}     [unchanged]
             (datasource.type=prometheus)         ├► kind=chart   → /api/v1/query_range {series}     [NEW]
        └► normalize frames → {series}            ├► kind=gauge   → /api/v1/query      {value,...}  [NEW]
                                                  └► kind=mean    → /api/v1/query_range {value}      [NEW]

 Frontend GrafanaChartWidget (recharts)   Frontend PrometheusChartWidget (recharts)  [renamed, reused]
                                          + PrometheusGaugeWidget (recharts RadialBarChart)         [NEW]
                                          + PrometheusMeanWidget (MetricCard-style)                 [NEW]

 grafana service type / LinkWidget /      [REMOVED entirely]
 LinksTab / navEntry / status endpoint

Key constraints carried from the spec:

• Reuse the recharts chart renderer unchanged (SC-106) — the rename is structural.
• Shared series normalization (SC-102) — one helper, no duplication.
• Step derived from window presets (SC-104), users never set step.
• gauge/mean are scalar-only (SC-111, SC-114); chart stays multi-series (SC-107).
• Window presets: 1h, 6h, 24h, 7d (SC-108, SC-112).
• All adapter errors return {"error": str}, never raise (SC-103).
• No Grafana references remain (SC-115, SC-116, SC-117, SC-120).

---

2. Backend design

2.1 New module: widgets/prometheus_range.py (shared helpers)

Holds the two pieces the spec's downstream-notes asked for: the step-derivation function (SC-104) and the shared normalization helper (SC-102). Putting them in their own module (rather than inside sources.py) makes them unit-testable in isolation and reusable by the future service-storage-harness change's in-service data path, without sources.py growing unbounded.

Window presets and step derivation (SC-104):

WINDOW_PRESETS: dict[str, int] = {
    "1h": 3_600,
    "6h": 21_600,
    "24h": 86_400,
    "7d": 604_800,
}

# Target ~200 points per window. Step is clamped to >= 15s so Prometheus
# doesn't reject sub-15s resolutions on high-cardinality queries.
def step_for_window(window_seconds: int, target_points: int = 200) -> int:
    return max(15, round(window_seconds / target_points))

Resulting table (verified, all within the 100–300 target band):

Preset	Window (s)	Derived step (s)	Points
1h	3,600	max(15, round(3600/200)) = 18	200
6h	21,600	108	200
24h	86,400	432	200
7d	604,800	3,024	200

(At apply time the implementer may round steps to "nicer" values like 15/60/300/1800 for cache-friendliness; the spec only requires the 100–300 band, which the formula satisfies. The formula is the source of truth; the table is illustrative.)

Shared normalization (SC-102) — normalize_prometheus_matrix:

def normalize_prometheus_matrix(
    result: list[dict[str, Any]],
) -> list[dict[str, Any]]:
    """Turn a Prometheus /api/v1/query_range `data.result` matrix into
    `{label, points:[{t:int, v:float|None}]}` series — the exact shape the
    frontend chart renderer consumes.

    Label derivation reuses the rule from the removed Grafana path:
      1. Drop __name__ from metric labels.
      2. If labels remain, join as `k=v k=v`.
      3. Else fall back to "value".
      4. Dedup collisions with a ` (n)` suffix.
    """
    series: list[dict[str, Any]] = []
    seen: dict[str, int] = {}
    for entry in result:
        metric = entry.get("metric") or {}
        values = entry.get("values") or []
        parts = [f"{k}={v}" for k, v in sorted(metric.items()) if not k.startswith("__")]
        label = " ".join(parts) if parts else "value"
        if label in seen:
            seen[label] += 1
            label = f"{label} ({seen[label]})"
        else:
            seen[label] = 0
        points = []
        for ts, raw in values:
            v = float(raw) if raw not in (None, "NaN", "+Inf", "-Inf") else None
            points.append({"t": int(ts), "v": v})
        series.append({"label": label, "points": points})
    return series

This is a direct extraction of the label/dedup logic currently inside GrafanaWidgetSource._fetch_chart, retargeted at the Prometheus matrix shape ({metric, values:[[ts,"str"],...]}) instead of Grafana frames. The dedup rule is identical so users moving a grafana/chart widget to prometheus/chart see the same labels.

2.2 PrometheusWidgetSource — extend .fetch() by widget_kind

Today .fetch() does only the instant-query → {"result": ...} path. Extend it to dispatch by widget_kind while preserving the existing metric behavior byte-for-byte:

class PrometheusWidgetSource:
    async def fetch(self, service, widget_kind, config) -> dict[str, Any]:
        if service is None:
            return {"error": "Prometheus widget is missing its service"}
        base_url = str(service.config.get("base_url") or "").rstrip("/")
        timeout = int(service.config.get("timeout_seconds") or 10)

        if widget_kind == "metric":
            return await self._fetch_instant(base_url, timeout, config)      # unchanged
        if widget_kind == "chart":
            return await self._fetch_chart(base_url, timeout, config)        # NEW
        if widget_kind == "gauge":
            return await self._fetch_gauge(base_url, timeout, config)        # NEW
        if widget_kind == "mean":
            return await self._fetch_mean(base_url, timeout, config)         # NEW
        return {"error": f"Unknown widget kind: {widget_kind}"}

_fetch_chart (SC-101, SC-103, SC-104):

async def _fetch_chart(self, base_url, timeout, config) -> dict[str, Any]:
    promql = config.get("promql")
    if not promql:
        return {"error": "promql is required"}
    window = WINDOW_PRESETS.get(config.get("window", "1h"), WINDOW_PRESETS["1h"])
    step = step_for_window(window)
    end = int(time.time())
    start = end - window
    try:
        resp = await asyncio.wait_for(asyncio.to_thread(
            requests.get, f"{base_url}/api/v1/query_range",
            params={"query": promql, "start": start, "end": end, "step": step},
            timeout=timeout,
        ), timeout=timeout)
        resp.raise_for_status()
        payload = resp.json()
    except asyncio.TimeoutError:
        return {"error": "Prometheus query timed out"}
    except requests.RequestException as exc:
        return {"error": f"Prometheus query failed: {exc}"}
    result = payload.get("data", {}).get("result", [])
    return {"series": normalize_prometheus_matrix(result)}

_fetch_gauge (SC-109, SC-110, SC-111) — instant query, scalar-only:

async def _fetch_gauge(self, base_url, timeout, config) -> dict[str, Any]:
    promql = config.get("promql")
    if not promql:
        return {"error": "promql is required"}
    raw = await self._instant_query(base_url, timeout, promql)   # shared helper
    if "error" in raw:
        return raw
    result = raw["result"]
    if len(result) != 1:
        return {"error": "Gauge requires a single-series query; refine your PromQL"}
    # vector entry: {metric, value:[ts, "str"]}
    try:
        value = float(result[0]["value"][1])
    except (KeyError, IndexError, ValueError, TypeError):
        return {"error": "Gauge query returned no scalar value"}
    return {
        "value": value,
        "warn_at": config.get("warn_at"),
        "crit_at": config.get("crit_at"),
        "min": config.get("min"),
        "max": config.get("max"),
        "unit": config.get("unit"),
    }

_instant_query is extracted from the current metric path so metric/gauge share it; it returns {"result": [...]} or {"error": ...}.

_fetch_mean (SC-112, SC-113, SC-114) — range query, client-side mean, scalar-only:

async def _fetch_mean(self, base_url, timeout, config) -> dict[str, Any]:
    promql = config.get("promql")
    if not promql:
        return {"error": "promql is required"}
    window = WINDOW_PRESETS.get(config.get("window", "1h"), WINDOW_PRESETS["1h"])
    step = step_for_window(window)
    end = int(time.time()); start = end - window
    # ... run query_range, handle errors identically to _fetch_chart ...
    result = payload.get("data", {}).get("result", [])
    if len(result) != 1:
        return {"error": "Mean requires a single-series query; refine your PromQL"}
    points = result[0].get("values") or []
    nums = [float(v) for _, v in points if v not in (None, "NaN", "+Inf", "-Inf")]
    if not nums:
        return {"error": "Mean query returned no numeric samples in the window"}
    mean = sum(nums) / len(nums)
    return {"value": mean, "unit": config.get("unit")}

The range-query HTTP call + error handling is shared between _fetch_chart and _fetch_mean via a private _range_query(base_url, timeout, promql, window) -> dict returning {"matrix": result} or {"error": ...}. This keeps the chart and mean paths DRY without inventing a generic adapter registry (non-goal enforced).

2.3 integrations/prometheus.py — declare new widget kinds

Add two Pydantic widget-config models and two widget_kind(...) entries; leave PrometheusMetricWidgetConfig and the existing metric kind untouched (non-goal: prometheus_metric stays as-is):

class PrometheusChartWidgetConfig(WidgetConfigBase):
    promql: str
    window: str = "1h"          # one of 1h/6h/24h/7d

class PrometheusGaugeWidgetConfig(WidgetConfigBase):
    promql: str
    warn_at: float | None = None
    crit_at: float | None = None
    min: float | None = None
    max: float | None = None
    unit: str | None = None

class PrometheusMeanWidgetConfig(WidgetConfigBase):
    promql: str
    window: str = "1h"
    unit: str | None = None

DEFINITION.widget_kinds gains chart, gauge, mean (refresh 60s for chart/mean, 30s for gauge). Validation that window ∈ {1h,6h,24h,7d} can be a field_validator on the two models that use it, or simply enforced by WINDOW_PRESETS.get(..., default) server-side — prefer the latter (lenient) so a future preset addition doesn't require a model change.

2.4 integrations/grafana.py — delete; registry drops the entry

• Delete the file.
• In integrations/registry.py, drop the from ...grafana import DEFINITION as GRAFANA import and the GRAFANA.service_type: GRAFANA, line from SERVICE_DEFINITIONS. No other registry change.

2.5 widgets/sources.py — drop Grafana; wire nothing new

• Delete GrafanaWidgetSource and its _fetch_chart.
• Remove "grafana": GrafanaWidgetSource(), from SERVICE_ADAPTERS. (The label/dedup logic has already been extracted into prometheus_range.normalize_prometheus_matrix in §2.1; it is not lost when the Grafana class is deleted.)
• PrometheusWidgetSource gains the three new methods from §2.2.

2.6 routers/monitoring.py — remove get_grafana_status

Delete the @router.get("/grafana-status") endpoint (lines ~175–195). No other monitoring change; get_prometheus_status / get_alertmanager_status stay.

2.7 No SettingsStore or DB schema change

Widget instances are stored generically (service_id, widget_kind, config_json). A prometheus chart instance is just a row with service_id=<prom instance> and widget_kind="chart". Existing grafana/chart rows become orphans resolved by the existing "unknown widget" path (SC-119) — no migration code, no schema change (SC-126).

---

3. Frontend design

3.1 Rename + reuse: GrafanaChartWidget → PrometheusChartWidget

• Rename frontend/src/widgets/GrafanaChartWidget.tsx → PrometheusChartWidget.tsx; rename the exported function/component.
• The recharts body (LineChart, Line, XAxis, YAxis, CartesianGrid, Tooltip, ResponsiveContainer, mergeSeries, CHART_COLORS, formatTime) is preserved unchanged (SC-106).
• One tiny text fix: the empty-state Alert copy "Check your query and datasource_uid" becomes "Check your PromQL query and window." (datasource_uid no longer exists.)
• Rename test file __tests__/GrafanaChartWidget.test.tsx → PrometheusChartWidget.test.tsx; update import + the one error-message assertion ("Grafana api_key is required" → a Prom error string).

3.2 integrations/registry.ts — rebind + add

• Delete the entire grafana key from SERVICE_REGISTRY.
• In the prometheus binding's widgets array, add chart, gauge, mean alongside the existing metric. Each has a configSchema mirroring §2.3 (promql, window for chart/mean; promql, warn_at, crit_at, min, max, unit for gauge). chart/mean refresh 60s; gauge 30s.
• Import the renamed PrometheusChartWidget and the two new components.

3.3 New: PrometheusGaugeWidget.tsx (SC-109, SC-110, SC-111)

Renderer choice: recharts RadialBarChart. Justification: recharts is already a dependency (no new dep), RadialBarChart renders a single-value gauge with domain bands natively, and it shares tooltip/styling conventions with the chart widget — keeping the two visualizations consistent. The alternative (a ~50-line bespoke SVG gauge) was rejected because it would introduce a second rendering dialect for no benefit; the proposal's "fall back to SVG if recharts proves heavy" fallback remains documented but is not the default.

Threshold-band rendering: render three stacked RadialBar cells (green 0→warn, amber warn→crit, red crit→max) as the track, and a fourth cell (the actual value) as the needle/bar. When warn_at/crit_at are absent, render a single neutral-color track. min/max default to 0/max(value, 1) when omitted so the gauge has a sane domain. The component reuses SectionCard + Alert/Skeleton for loading/error states, matching every other widget.

Config fields surfaced to the user (matching §2.3): promql, warn_at, crit_at, min, max, unit.

3.4 New: PrometheusMeanWidget.tsx (SC-112, SC-113, SC-114)

A single-value display reusing the existing MetricCard pattern (already used by BackupDashboardWidget / PrometheusMetricWidget-adjacent tiles): big number, optional unit suffix, optional subtext showing the window ("mean over last 1h"). Loading/error/empty states via Skeleton/Alert as usual. No charting library involvement — it's a number, deliberately.

3.5 Grafana removal on the frontend

Per the source findings (§0), the removal targets are not an ObservabilityPage section. They are:

File	Change
widgets/GrafanaLinkWidget.tsx	delete
widgets/GrafanaChartWidget.tsx	rename to PrometheusChartWidget (§3.1) — not a pure delete
widgets/index.ts	drop GrafanaLinkWidget export; rename the chart export
integrations/registry.ts	drop grafana binding (§3.2)
integrations/navEntries.ts	drop the grafana entry from SERVICE_TYPE_NAV_ENTRIES; drop the now-unused Link2 icon import
hooks/useObservability.ts	drop useGrafanaStatus + its fetchGrafanaStatus import
api/client.ts	drop fetchGrafanaStatus
types/index.ts	drop GrafanaStatus interface
pages/service-tabs/LinksTab.tsx	delete (it is the grafana-specific tab; its GrafanaLinkCard + machine deep-links are Grafana-only)
pages/service-tabs/__tests__/LinksTab.test.tsx	delete
pages/service-tabs/index.ts	drop the LinksTab import and the case "grafana": from serviceContentTabs
pages/Dashboard.tsx	drop "grafana" from OBSERVABILITY_TYPES set (line ~68)
pages/ServicesPage.tsx	update empty-state copy "Add a Grafana, Prometheus, …" → "Add a Prometheus, …"
pages/__tests__/Dashboard.test.tsx	the one "Grafana" label literal there is for a shortcut (a website link), unrelated to the Grafana service — leave it (grep-clean criterion SC-116 still passes; it's not a grafana service reference, just user-typed shortcut text in a test fixture). Flag for reviewer.

Spec-text note (not a design change): SC-118 literally names "the ObservabilityPage Grafana section," which no longer exists. The intent of SC-118 ("Grafana status checks are removed") is satisfied by dropping get_grafana_status (§2.6) + useGrafanaStatus. The tasks phase should note this textual drift so the reviewer doesn't treat it as a missed requirement. If the parent prefers, SC-118 can be reworded in spec.md to name LinksTab/useGrafanaStatus instead; this design does not require that edit to proceed.

3.6 Types

types/index.ts: remove GrafanaStatus. No new widget-payload types — chart uses {series} (existing), gauge uses {value, warn_at, crit_at, min, max, unit} (all optional beyond value), mean uses {value, unit?}. These are read off data?.data untyped-as-before; no WidgetDataResponse generic change is needed (it's already data: dict | None).

---

4. Data flow

1. Chart — useWidgetData(widgetId, 60_000) → GET /api/widgets/instances/{id}/data → PrometheusWidgetSource.fetch(kind="chart") → _fetch_chart → GET {prom}/api/v1/query_range?query=...&start=...&end=...&step=... → normalize_prometheus_matrix → {"series":[...]} → PrometheusChartWidget renders recharts (unchanged).
2. Gauge — useWidgetData(widgetId, 30_000) → fetch(kind="gauge") → _fetch_gauge → instant query → scalar-only check → {"value":..., "warn_at":..., ...} → PrometheusGaugeWidget renders RadialBarChart.
3. Mean — useWidgetData(widgetId, 60_000) → fetch(kind="mean") → _fetch_mean → range query → client-side mean → {"value":..., "unit":...} → PrometheusMeanWidget renders a MetricCard.
4. Metric (unchanged) — existing path preserved byte-for-byte.
5. Orphaned grafana widget — resolveWidget finds no grafana binding → returns undefined → WidgetInstanceCard renders its existing "unknown widget" Alert (SC-119). No crash, no migration.

Errors at any step return {"error": str} (SC-103); the per-widget Alert variant="destructive" renders it and siblings keep polling.

---

5. Testing approach

Backend

Extend backend/tests/test_widgets.py (and/or a focused test_prometheus_range.py):

• normalize_prometheus_matrix: feed a sample Prom /api/v1/query_range data.result (two entries, one with __name__, one colliding label) → assert {series} shape, label dedup, null handling for "NaN".
• step_for_window: assert the 1h/6h/24h/7d → step mapping stays within 100–300 points.
• PrometheusWidgetSource chart path: mock requests.get → assert query_range URL + params (start/end/step present, no from_ts/to_ts) and {"series": ...} return.
• Gauge scalar-only: mock instant query returning 2 series → assert {"error": ...}.
• Mean scalar-only + client-side mean: mock range query returning 1 series with known values → assert the arithmetic mean; mock 2 series → assert error.
• Error paths: timeout / RequestException → {"error": ...} (no raise).

Frontend

• PrometheusChartWidget.test.tsx (renamed from GrafanaChartWidget test): loading, error, rendered-data cases (SC-125).
• PrometheusGaugeWidget.test.tsx: loading, error, rendered-with-bands, rendered-without-bands.
• PrometheusMeanWidget.test.tsx: loading, error, rendered value.
• registry.test.ts: assert no grafana binding; assert prometheus binding has metric, chart, gauge, mean.

Build/lint gates

Each slice: PYTHONPATH=src pytest (from backend/), npm run build, npm run lint must be green.

---

6. File-level plan

Create

File	Rationale
backend/src/media_library_viewer_api/widgets/prometheus_range.py	WINDOW_PRESETS, step_for_window, normalize_prometheus_matrix (SC-102, SC-104).
backend/tests/test_prometheus_range.py	Unit tests for the helpers.
frontend/src/widgets/PrometheusGaugeWidget.tsx	recharts RadialBarChart gauge (SC-109/110/111).
frontend/src/widgets/PrometheusMeanWidget.tsx	MetricCard-style mean (SC-112/113/114).
frontend/src/widgets/__tests__/PrometheusGaugeWidget.test.tsx	SC-125.
frontend/src/widgets/__tests__/PrometheusMeanWidget.test.tsx	SC-125.

Rename (git mv)

From → To	Rationale
widgets/GrafanaChartWidget.tsx → widgets/PrometheusChartWidget.tsx	SC-105/106. Body preserved.
widgets/__tests__/GrafanaChartWidget.test.tsx → widgets/__tests__/PrometheusChartWidget.test.tsx	match rename.

Modify

File	Rationale
backend/.../widgets/sources.py	Drop GrafanaWidgetSource; add chart/gauge/mean to PrometheusWidgetSource.
backend/.../integrations/prometheus.py	Add 3 widget-config models + 3 widget_kind entries.
backend/.../integrations/registry.py	Drop grafana import + entry.
backend/.../routers/monitoring.py	Drop get_grafana_status.
backend/tests/test_widgets.py	Drop grafana adapter tests; add prom chart/gauge/mean tests.
frontend/src/integrations/registry.ts	Drop grafana binding; add chart/gauge/mean to prometheus.
frontend/src/integrations/navEntries.ts	Drop grafana entry.
frontend/src/hooks/useObservability.ts	Drop useGrafanaStatus.
frontend/src/api/client.ts	Drop fetchGrafanaStatus.
frontend/src/types/index.ts	Drop GrafanaStatus.
frontend/src/widgets/index.ts	Drop GrafanaLinkWidget export; rename chart export.
frontend/src/pages/service-tabs/index.ts	Drop LinksTab import + grafana case.
frontend/src/pages/Dashboard.tsx	Drop "grafana" from OBSERVABILITY_TYPES.
frontend/src/pages/ServicesPage.tsx	Update empty-state copy.
openspec/config.yaml	Repeal stale thin-dashboard/no-recharts wording (SC-121).
CHANGELOG.md	Migration note (SC-122).

Delete

File	Rationale
backend/.../integrations/grafana.py	SC-117.
frontend/src/widgets/GrafanaLinkWidget.tsx	SC-117.
frontend/src/pages/service-tabs/LinksTab.tsx	Grafana-only tab; no Prom equivalent needed (chart widget covers viz).
frontend/src/pages/service-tabs/__tests__/LinksTab.test.tsx	matches deletion.

---

7. Slice boundaries (≤400 changed lines each)

Each slice leaves pytest + npm run build + npm run lint green and the app in a working state.

Slice 1 — Prom range path + chart rebrand + shared helper (foundational)

Files: create prometheus_range.py + test; rename GrafanaChartWidget→PrometheusChartWidget (+ test); modify sources.py (_fetch_chart, drop nothing yet), integrations/prometheus.py (add chart kind only), registry.ts (move chart grafana→prometheus), widgets/index.ts, registry.test.ts. State after S1: a prometheus/chart widget renders from direct Prom query_range; the grafana/chart binding is gone but GrafanaWidgetSource/GrafanaLinkWidget/grafana service type still exist (removal is S3). Grafana link widgets still work. ~300–380 lines.

Slice 2 — Gauge + mean widgets (additive)

Files: create PrometheusGaugeWidget.tsx (+test), PrometheusMeanWidget.tsx (+test); modify sources.py (_fetch_gauge, _fetch_mean, shared _instant_query/_range_query), integrations/prometheus.py (add gauge/mean kinds), registry.ts (add gauge/mean bindings), test_widgets.py (add adapter tests). State after S2: gauge + mean widgets selectable and rendering; no grafana change. ~300–380 lines.

Slice 3 — Grafana removal + config + changelog (cleanup)

Files: delete integrations/grafana.py, GrafanaLinkWidget.tsx, LinksTab.tsx (+test); modify sources.py (drop GrafanaWidgetSource + SERVICE_ADAPTERS entry), registry.py, monitoring.py, navEntries.ts, useObservability.ts, client.ts, types/index.ts, widgets/index.ts, service-tabs/index.ts, Dashboard.tsx, ServicesPage.tsx, test_widgets.py (drop grafana tests); rewrite config.yaml; add CHANGELOG.md entry. State after S3: grep-clean (SC-115/116), config accurate (SC-121), migration documented (SC-122). ~250–350 lines (mostly deletions).

Order: S1 → S2 → S3. S1 and S2 are independently shippable; S3 must follow S1 (it removes the grafana chart binding S1 replaces).

---

8. Decisions log (answers to spec downstream-notes)

1. Step derivation (SC-104): step_for_window(window_seconds, target_points=200) = max(15, round(window_seconds/200)), in widgets/prometheus_range.py. Table in §2.1.
2. Shared normalization (SC-102): normalize_prometheus_matrix(result) in widgets/prometheus_range.py; extracts the dedup rule from the to-be-deleted Grafana path; retargeted at Prom matrix shape. Signature in §2.1.
3. Gauge renderer (SC-110): recharts RadialBarChart (no new dep, consistent styling). Threshold bands via stacked track cells; neutral single-color when warn_at/crit_at absent. Config fields: promql, warn_at, crit_at, min, max, unit. §3.3.
4. Mean adapter (SC-112/113): _fetch_mean runs query_range over the window preset, averages non-null samples of the single series client-side, returns {value, unit?}. Scalar-only enforced via len(result) != 1 → error. §2.2.
5. Slice plan: 3 slices (S1 range+rebrand+helper, S2 gauge+mean, S3 grafana removal+config+changelog), each ≤400 lines, order S1→S2→S3. §7.

---

9. Open items for the tasks/apply phases

• Implementer should git mv the chart widget/test so history is preserved (not delete+create).
• Implementer should verify the recharts RadialBarChart gauge renders acceptably; if it proves fiddly, the proposal's SVG fallback (~50 lines) is sanctioned — but try recharts first.
• SC-118 text names a file that no longer exists; tasks phase should record this so verify doesn't flag it as a miss. Intent is satisfied by removing get_grafana_status + useGrafanaStatus + LinksTab.
• Dashboard.test.tsx contains the literal "Grafana" in a shortcut test fixture (not a grafana service reference) — grep for SC-116 should be scoped to service/widget references, or that line whitelisted. Flag for reviewer.