emem-recall-polygon

emem-recall-polygon

Network use. The commands in this skill send and receive JSON (and, where a step says so, an image or a raster file) to https://emem.dev only, the service the plugin's MCP server connects to. Nothing they download is executed. The files under scripts/ read local files and make no network calls.

POST /v1/recall_polygon samples cells inside an area, recalls the requested bands at each, and returns them per cell with per-cell receipts. It is locate, then a sample of the area's cells, then recall_many, collapsed into one call.

When to invoke

• "What's the average NDVI inside this watershed?"
• "Show me precipitation across the area bounded by ..."
• The user pastes a bbox or a GeoJSON polygon.

For one point use emem-locate-and-recall. For the full-resolution pixel grid over an area use emem-field-tokens.

How to invoke

Give the area in one of three ways: place (free text, geocoded), polygon_bbox (an object with named corners; an array is refused, because bbox array orders disagree between conventions), or polygon_geojson (a GeoJSON Polygon, MultiPolygon or Feature). A polygon on its own sets the area: its bounds become the box and its boundary masks the sample. geojson and geometry_geojson are aliases of polygon_geojson. Beside a bbox, the polygon only masks the sample.

curl -sf -X POST https://emem.dev/v1/recall_polygon \
  -H 'content-type: application/json' \
  -d '{"polygon_bbox":{"min_lat":12.95,"max_lat":13.05,"min_lng":77.55,"max_lng":77.65},
       "bands":["copdem30m.elevation_mean"],
       "max_cells":16, "budget_ms":20000}' \
  > poly.json
jq '{cells_sampled, facts_returned, area_km2, converged, coverage_fraction,
     pending: (.pending | length)}' poly.json

• max_cells defaults to 64 and caps at 1024; out of range is a 400, not a silent clamp.
• budget_ms is a soft materialisation budget. When it expires the answer is a partial 200 with converged: false and a typed pending[] (each entry says why and what to do). Pending fetches are detached, not aborted, so the identical request retried returns strictly more. Pending entries are unsigned; they are not signed absences.
• Recorded on 2026-09-28: this exact call returned 200 with cells_sampled: 16, facts_returned: 0, converged: false and 16 pending entries (state: "materializing"), and nine identical retries over 30 minutes stayed at zero while the responder's fetch path was saturated. Retry a few times with a pause; if pending does not shrink, tell the user the upstream fetch is stalled rather than looping, and never report the empty answer as "no data here".
• A request with many cold cells and several bands can still pass the responder's 40 s transport budget and return compute_timeout. Narrow it: fewer bands, fewer cells, or retry once the first pass has warmed the cache.