Open Map Stack

OpenMapStack Toolkit

Production-grade geospatial workflows with an open-first stack and pragmatic hosted/SaaS choices when global scale, latency, SLA, or data quality makes local processing a poor fit. Cloud-native by default: STAC for discovery, GeoParquet + COG + PMTiles for storage, DuckDB and PostGIS for compute, MapLibre and Martin for delivery.

Reproducible project-first contract

Core principle: reasoning may be exploratory; the delivered analysis must be deterministic, inspectable, and reproducible.

For any material multi-stage GIS analysis, do not optimize for reaching the final map, dashboard, or answer quickly. A one-off polished dashboard is not the deliverable — a reproducible technical GIS project is. First establish a reusable project artifact, then derive the map/dashboard/report from it.

Before treating an analysis as complete, you MUST compile (or maintain) a project like examples/tartu-development: a canonical project.yaml (openmapstack-project/v1), pipeline.py, and README.md; pinned sources with timestamps, selections and licensing; explicit assumptions; every manual addition or correction stored as real geodata; deterministic ordered steps with explicit CRS; machine-readable validation rules plus the report from the run; and output definitions with semantic presentation intent and provenance surfaced in the rendered view. references/project-spec.md is the full schema — read it before compiling a project.

Workflow (agent may retry/experiment internally, but the accepted analysis is recompiled deterministically):

USER QUESTION
    ↓
interpretation / exploration        (internal, may be ad-hoc)
    ↓
COMPILE GIS PROJECT                 project.yaml + pipeline + manifest + overrides + validation
    ↓
EXECUTE PROJECT
    ↓
validated derived datasets
    ↓
QGIS project / standardized web view
    ↓
FINAL ANALYSIS / DASHBOARD / ANSWER

The polished map/dashboard is a view over the project, not the canonical definition of the analysis.

Hard rules for every material analysis — each is expanded in references/project-spec.md:

• Real source data mandatory; never hallucinate coordinates. Fabricating coordinates or synthesizing baseline geometry is forbidden without explicit, informed user consent. Hypothetical or planned features go in data/overrides/ with provenance, rationale, and evidence.
• Never mutate source data. Immutable source + project override layer = effective input. Distinguish external facts, transformations, corrections, assumptions, and hypothetical data.
• Overrides must be executable and verified. Target a real source feature; for attribute changes the asserted prior value must match. Evidence must be non-placeholder. Validation reports each override applied, rejected, or not_testable — listing one is not applying it. Scenario features stay labeled hypothetical and visually distinct from authoritative layers.
• Record, don't memoize on chat. Encode every manual fix as data or pipeline logic. A fresh environment with the documented sources must reproduce the project; the transcript is not part of the dependency graph.
• Prove semantic predicates from data. Ownership, active status, public access, legal designation: the source must expose an authoritative field or documented mapping. Preserve unknown as unknown; never default a missing value to the desired class.
• Bounded APIs must prove completeness. Record numberMatched/equivalent and page until returned == matched. A response filled to the request limit is incomplete until proven otherwise.
• Validation is a pipeline stage, not prose advice, with machine-readable results. Every declared check appears exactly once in the report; warning/not_testable propagate to run and project status; run IDs and hashes resolve to a real runs/*.json record.
• Run the project CLI when available. Use openmapstack validate project.yaml before delivery and openmapstack run project.yaml for the canonical execution path. The CLI audits the manifest, provenance, graph, artifacts, report, and run record; it does not replace domain GIS checks performed by the pipeline.
• The manifest must resolve. Every step input is a source key or an earlier step's output, spelled as the producer declared it; every generated_by names a real step (manifest_graph_resolves).
• One canonical implementation creates every declared output. Convenience/E2E entrypoints may wrap pipeline.py but must not duplicate its processing, QGIS, or report logic.
• Build a layer- and style-perfect QGIS project (project.qgz) mirroring the web view: matching layer-tree groups, identical categorized styles, ./path.gpkg|layername=name datasources, and a regional tiled basemap. Success means valid layers, not exit code 0 — pin the runtime, and when PyQGIS is available require every layer isValid(); otherwise record not_testable, never an implicit pass. Two traps make a project that passes every one of those checks still show the wrong map, so check them explicitly:
  • Every layer declares a complete <srs>, basemaps included, and project reprojection is enabled. A layer without one is assumed to be in the project CRS and never reprojected. An auth-id-only <spatialrefsys> is also broken: QGIS can still report EPSG:3301 while treating the CRS as invalid and silently painting nothing. Emit WKT or PROJ alongside the identifiers and set SpatialRefSys/ProjectionsEnabled to 1. Prefer building layers through the PyQGIS API, which serializes these fields for you; the trap is specific to hand-written .qgs XML.
  • A QGIS layer tree stacks the opposite way to a web map. presentation.map.layers is ordered bottom-to-top, while a layer tree paints its first entry on top, so write the tree in reverse manifest order with the basemap last. Copying the manifest order verbatim puts opaque analysis fills over the point layers that belong above them, and the points vanish.
• Separate analysis semantics from rendering. Declare semantic presentation roles; don't reinvent layout/colors/UX per run.
• Ship a reconfigurable view, and never let it misrepresent the run. Organise the sidebar into tabs of collapsible sections, give every layer group an on/off control, and expose the analysis parameters and scenario overrides as live controls. Each control opens at the value declared in presentation.controls and returning there must reproduce the published numbers; any other position labels itself exploratory and offers a reset. The browser re-applies published rules to values the pipeline measured — it never measures geometry, and a control that changes a shape switches between buffers the pipeline materialised.
• Labels must match the operation. A Euclidean buffer is a "2 km straight-line proxy", not a walking catchment, and column names must say so too. State the measurement basis (nearest edge vs centroid) as an assumption — it changes which features qualify.
• Cheat-sheet: references/project-spec.md defines the full schema; templates/ gives ready scaffolds; examples/tartu-development is a worked reference project matching the acceptance scenario.

Modules — read the relevant reference(s) before starting work

For simple one-shot questions (single CRS conversion, one ogr2ogr invocation), the relevant reference alone is sufficient — a full project artifact is not needed. For multi-stage pipelines, read data-sources.md and processing.md together, and see project-spec.md + templates/ to compile analysis into a rerunnable project. For end-to-end "from raw data to web map" tasks, also read web-delivery.md.

Global defaults — apply unless the user specifies otherwise

• Storage formats: GeoParquet (vector analytics), COG (raster), PMTiles (tile delivery), GeoPackage (desktop interchange). Never produce Shapefile as new output.
• CRS: WGS84 (EPSG:4326) for storage; Web Mercator (EPSG:3857) for web rendering; local projected CRS for any metric computation (distance, area, buffer). For Estonia, EPSG:3301 (L-EST97).
• Compute placement: push spatial joins and aggregations to DuckDB or PostGIS — not Python loops. R-tree / GIST / spatial indexing is mandatory at scale.
• Discovery first: check STAC catalogs (Microsoft Planetary Computer, Earth Search, Overture STAC) before downloading anything. Lazy load with odc-stac or stackstac and only materialize what's needed.
• Cloud-native access: prefer querying remote GeoParquet/COG over downloading. DuckDB with httpfs extension is the default pattern for Overture and similar S3-hosted datasets.
• Scale first: local tools are fine for city/state work; at continental/global scale prefer cloud-native partitioned datasets, precomputed tiles, hosted APIs, or SaaS when they are more reliable than local batch processing.
• License hygiene: preserve license metadata through every transformation. OSM is ODbL (share-alike); Overture varies by source; Sentinel is free-with-attribution; national data varies.
• Runtime hygiene: prefer conda-forge environments or containers for GDAL/PROJ/GEOS/QGIS stacks. Avoid pip-only geospatial environments unless the project already proves they work.

Format decision matrix

Compute decision matrix

Universal anti-patterns — flag and correct

• Hallucinating or fabricating mock coordinates and geometries instead of retrieving real source data (unless the user gave explicit, informed consent for a synthetic mock test)
• Generating a QGIS project that lacks the web dashboard's layers, omits basemaps, or uses broken OGR datasource syntax (path.gpkg|layer without layername=), causing layers to load as non-spatial attribute tables
• Writing .qgs XML by hand with no <srs>, an auth-id-only CRS block, or no ProjectionsEnabled, or copying the manifest's layer order straight into the layer tree — these produce a project where every layer is valid and every datasource resolves, yet the map shows the wrong place or silently hides a layer
• Producing Shapefile as new output (column truncation, 2GB limit, no UTF-8, multi-file)
• Calling .distance(), .buffer(), or .area on geographic CRS (EPSG:4326) — degrees are not meters; unless specific tool explicitly supports wgs84 based geodesic calculations
• Web Mercator (EPSG:3857) for area or distance calculations — it is not equal-area, and the units are not in meters except at the equator
• Spatial joins in Python loops when DuckDB / PostGIS / R-tree-backed sjoin is one line away
• Using bbox containment for area queries when features can cross the boundary — use bbox overlap as the scan gate, then an exact spatial predicate
• Downloading entire datasets when STAC + cloud-native formats allow lazy/range-request access
• Running planet-scale local processing for lookup/search problems when reliable hosted services or precomputed global products already exist
• Treating MBTiles as the default for new web deployments — PMTiles is the modern default
• Using GeoTIFF when COG is one flag away (-of COG)
• Mixing CRS silently — every join must assert matching CRS
• Hand-rolling routing or geocoding when OSRM, Valhalla, or Nominatim are one Docker pull away
• Pinning data to "latest" in a reproducible pipeline — pin Overture release version and STAC item IDs, not just collections. For Overture, verify the pinned release is still available or mirror it.

Quick triage — recognize the request type

Before diving into a task, classify it:

1. Discovery ("what data exists for…?", "is there a dataset of…?", "read this catalog") → start with data-sources.md. STAC search if raster; Overture or OSM if vector basemap; for a Portolan catalog read its AGENTS.md before querying.
2. Conversion / CRS ("convert this to…", "reproject to…", "the projection looks wrong") → formats-and-crs.md. Usually one ogr2ogr or gdalwarp call.
3. Analysis ("what's the average elevation in…", "how many buildings within 500m of…", "where are the hotspots?") → analytics.md and likely processing.md. Push to DuckDB/PostGIS first.
4. Delivery ("publish this as a web map", "generate tiles for…") → web-delivery.md. PMTiles + Martin + MapLibre is the default.
5. Desktop / cartography ("style this in QGIS", "make a print map", "automate this in QGIS") → qgis.md. Consider QGIS MCP for agentic workflows.

Most real tasks span 2–3 of these — read the relevant references in order.

Reproducibility checklist for any pipeline you produce

• Pin dataset versions (Overture release, STAC item IDs, OSM extract dates) — never "latest"
• Document CRS at every stage; never assume
• Use conda-forge envs or pinned container images (ghcr.io/osgeo/gdal:alpine-small-latest for GDAL, qgis/qgis:<tag> for PyQGIS); pip-only geospatial envs break frequently
• Validate outputs: gpq for GeoParquet, rio-cogeo validate for COG, pmtiles show for PMTiles, is_valid for geometries
• Preserve license metadata in column or sidecar JSON and carry required attribution into maps/APIs

Command-level detail for each of these lives in references/validation-and-ops.md.