Platform-Inventory

Canonical, version-controlled home for the SERVIR Platform Ecosystem engineering plan — so the plan, its decisions, and their rationale are readable by people and by other LLM sessions without depending on a specific claude.ai conversation.

Live site: https://servir-ai.github.io/Platform-Inventory/

Contents

The plan, as a site

Page	Section
index.html	Overview and contents
01-architecture.html	How to build this — the shared build order
02-gateway.html	Gateway & integration architecture · 259-repo triage
03-geoai.html	GeoAI strategy alignment
04-global-risk.html	Global Risk (RiskMap) — 12 perils × 3 clocks
05-food-security.html	Food Security (AgriNexus)
06-nrm.html	Natural Resource Management (LCEM)
07-shared-infra.html	Cross-platform shared infrastructure
08-decisions.html	22 tracked decisions · phased roadmap
09-services.html	123 hub services, with status and user institution
10-nasa.html	34 NASA programme assets beyond SERVIR
11-appendix.html	Terminology and the source record
full-plan.html	The entire plan on one page — for Ctrl-F and printing

Shared styling lives in assets/style.css. .nojekyll stops GitHub Pages running the files through Jekyll.

Supporting records

• servir-platform-inventory-repo.md — decision/context record: why this repo is the canonical source, what was and wasn't carried over, and constraints future sessions (Claude or otherwise) should know about before changing how this is maintained.
• RED_TEAM_REVIEW.md — adversarial review of the engineering plan: architectural risk, feasibility gaps, governance/equity issues, internal inconsistencies, and a prioritized list of the top 5 fixes needed before the plan should be treated as fundable/buildable.
• OPEN_DATA_SOURCES.md — verified reference of open datasets and microservices the Gateway could integrate with: world population, building/settlement footprints, crop data, field boundaries, and adjacent open EO/hazard APIs — each with access method, license, and integration notes.
• planscape-review.md — 10 Sept 2026 check of Planscape, Collect Earth Online and Forest Data Partnership against the plan; Planscape was the gap, now §06 N7 and Decision #23.
• fdp-review.md — 10 Sept 2026 correction: FDP is a code/model ecosystem, not a map feed; palm oil moves from New build to Partner (palm_model_2026a), Whisp attribution fixed (§05 F7).
• brand/README.md — SERVIR design-system rules applied to the site and where the logo (assets/servir-logo.png) came from.

Other formats

• SERVIR-Engineering-Plan.pptx — the plan as a 48-slide deck, generated from the same corrected content as the site (v5.0). Headings use Roboto Condensed where available, substituting on machines without it.
• SERVIR-Engineering-Plan-doc.html — a Google Docs–importable export of the plan, with the SERVIR masthead composited onto navy for light-background use.

Maintenance

Keeping this register current is a repo job, not a person's job. .github/workflows/inventory-sweep.yml runs scripts/sweep.py on a schedule (Mondays 14:00 UTC) and on demand via Actions → Inventory sweep → Run workflow.

The sweep refreshes the two mechanical parts of the register:

1. the SERVIR GitHub organisations' repository inventory, and
2. the SERVIR App Center service status badges.

It writes data/servir-repos.json, data/appcenter.json and data/CHANGES.md, and opens a pull request only when something actually changed — a quiet week produces no PR. It runs on the built-in GITHUB_TOKEN; no secret needs configuring.

Three deliberate safeguards, because a silent failure here is worse than no sweep at all:

• Empty fetch. If a fetch returns nothing for every org, the sweep keeps the previous snapshot rather than overwriting the baseline with an empty one — otherwise one blocked run would make the next run report every service as new.
• Unparseable badges. If App Center pages fetch fine but 80%+ carry no recognisable status badge, the sweep treats that as a markup change, keeps the previous snapshot and says so. Without this, a changed page template would give every service an empty status, the diff would go quiet, and the sweep would look healthy while reporting nothing at all.
• Per-subsystem, not global. A stale subsystem has its snapshot substituted back in, so its own diff is empty by construction. The PR gate is changed = bool(changes) and nothing more — an earlier version also gated on "did anything go stale", which meant a blocked GitHub fetch would suppress a real App Center status change in the same run. It no longer does.

Any run that fell back says so under Attention in data/CHANGES.md.

Verified against the live site, 9 September 2026. The crawler parses the listing page, not detail pages: status is a CSS class on each card (inactive, or absent), which is the only status this source publishes — a binary, not a graded badge. Detail pages carry no status at all and their <title> is boilerplate identical for every app. An earlier version of this script walked 95 detail pages looking for status words in the page text; it would have found none and reported silence forever. The shipped pattern was checked against the live DOM: 79 of 79 cards parsed, all 5 inactive entries correct. One request now, not 95.

Note for anyone re-testing from a Claude sandbox: appcenter.servirglobal.net is not in the default egress allowlist there, so the crawl returns nothing and the first safeguard fires. That is the sandbox, not the script. GitHub Actions runners reach it fine.