maplibre/maplibre-agent-skills

Skill name maplibre-leaflet-migration (or similar) Description / Use cases I want to migrate away from Leaflet to MapLibre GL JS. There is already a Leaflet migration guide: https://maplibre.org/maplibre-gl-js/docs/guides/leaflet-migration-guide/

What this is The AI failure report asks a reporter for two pieces of code: what the AI produced, and what was actually correct. Both arrive today as prose or a pasted fragment, and a maintainer has to reconstruct a runnable map before they can verify the failure against the pinned model. maplibregl-recorder (MIT) removes that reconstruction step. It records MapLibre GL JS API calls and writes them back out as a single self-contained HTML page that makes the same calls, style object and all. Not a screen capture: the map's constructor options, addSource/addLayer arguments, and settled camera are all there as literal JavaScript. So this is a report-quality change. Nothing here touches eval infrastructure and nothing lands in package.json; the recorder is something a reporter runs in their own app, not a dependency of this repo. What an export does and doesn't carry It carries the arguments. That makes a load failure the class it serves best, not the class it can't reach: a broken tile UR

The problem maplibre-cartography is our most validated skill: status: verified, a passing eval suite, content reviewed against primary sources. It is also the wrong shape, and those two facts are independent. Validation tells you the content is true. It tells you nothing about whether the content is in the right place. Per the station model adopted in #48's roadmap, skills should be shaped around primitives — a component, an asset pipeline, a format — so that other skills can point at one source of truth instead of re-teaching it. maplibre-cartography is shaped around a role: "things a cartographer thinks about." That bundles at least five unrelated primitives under one roof, which is the "Serve Breakfast" failure the station model warns about: ... a binder with one section called 'Serve Breakfast' containing re-transcribed egg procedures... It works for exactly one meal. The concrete symptom is already measurable. sprite appears in seven of nine shipped skills, and maplibre-cart

What happened maplibre-tile-hosting was withdrawn after a baseline run answered all four of its tests correctly with no skill loaded: the tile.openstreetmap.org production ban, the hosting-approach tradeoffs, the ODbL attribution obligations, and the negative case. That result is real, and the withdrawal was right. But it is worth being precise about what it demonstrated: the skill as written cleared no bar. It did not demonstrate that hosting has no gap. Four passing tests show the four questions we happened to ask were already known. Why the questions were the problem Across this repo's baselines, results have split by the kind of knowledge a question asks for, not by difficulty: Models reliably pass questions of architectural judgment — which approach suits which situation, what the tradeoffs are, when you need infrastructure at all. Models reliably fail on specific names, numeric defaults, version boundaries, and published URLs — the maplibre-gl-leaflet package name, raster ti

[Evals] Gap-intake pipeline — design the triage block and the confirming workflow What this is This repo only accepts skill content that closes a demonstrated gap — evidence, gathered via Promptfoo evals, that the model actually fails without the skill. Right now, turning a reported failure into that evidence is entirely manual: a maintainer re-runs the eval by hand for every report. This issue is about automating the confirmation step — not the judgment call, just the test run. Proposed flow: A contributor reports a failure as a freeform issue (already supported — see Report a failure). A maintainer verifies the report against primary sources (style spec, GL JS docs, actual behavior) and, if it holds up, adds a triage block to the issue: a fenced YAML block naming the target skill (or none if it doesn't fit an existing one), the prompt text, the rubric, and the test type — matching the explicit/implicit/anti-pattern/negative categories already used across the eval suite. Th