Unreleased
Added / Changed / Added / Changed / Security / Fixed
- Arabic spoken answers on MCP (#9722): the seven tools with response_mode: "spoken" also accept spoken_language: "ar" and answer in Modern Standard Arabic. The Arabic wording comes from the same template data path as the English, so it speaks the same numbers, caveats and freshness, with identical citations, stale and asOf. Countries, chokepoints, risk levels, dimensions and categories are named in Arabic, numbers stay in Western digits, and data titles keep their published language. The envelope gains language. spoken_language is rejected with -32602 without the spoken mode. Tool count unchanged.
- Agents can report problems through MCP (#8721, MCP server 1.48.0, 100 tools): the credential-free tool report_issue stores a bug, missing-capability, data-quality or security report for the maintainers, through the same validation, redaction, retry handling and fail-closed limits as POST /api/feedback/v1/submit-report. A Pro or API account whose entitlement check passes reports under its own budget of 30 per day; every other caller reports anonymously at 5 per hour per IP. Identity comes from the request, never the arguments. Arguments over 24 KB are refused before any feedback limit or storage call, and a security report points to the private advisory form. The published MCP Registry description now reads "on conflict" instead of "across conflict" so the tool count fits its 100-character limit.
- Spoken response mode on seven analytical MCP tools (#9722): get_country_risk, get_world_brief, get_chokepoint_status, get_forecast_predictions, get_prediction_markets, get_media_perspective and get_resilience_indicators accept response_mode: "spoken" and answer with 2-3 plain-language sentences for text-to-speech plus a citations array (source, title, URL, asOf, stale) instead of the full payload. Sentences are templated from the tool's structured fields (api/mcp/spoken-narrators.ts), never model-written, so every number spoken is in the data and the same data always reads the same way; stale snapshots, last-known-good briefs, old risk readings, partial media reads and failed upstream reads change the wording. The mode is advertised in each tool's inputSchema and as a {spoken, citations} branch of its outputSchema; other tools and unknown modes are rejected with -32602, jmespath projects the envelope, and summary cannot be combined with it. Tool count unchanged.
- Webcam discovery and recent media through MCP (#5704, MCP server 1.47.0, 99 tools): find_webcams reads the existing seeded viewport catalog; get_webcam_media resolves one selected camera through WebcamService with at most one Windy request. Results separate catalog freshness, capture time and URL resolution time, retain Windy attribution, and report unavailable media without a generic server error. Windy v3 player URLs now come from player. and detail links from urls.detail. Media is recent visual context without guaranteed live conditions.
- Sentinel-5P NO2 activity by country and cluster (#9722, MCP server 1.46.0, 97 tools): GET /api/economic/v1/get-no2-activity serves the global NO2 activity dataset that seed-no2-activity publishes, validated through the shared contract at the Redis boundary and filtered by country_code (the country plus its clusters), region_id and kind. Each region carries the current 56-day window and its valid days, the baseline years, changePercent, direction and an exclusion from the closed vocabulary; baseline_backfill_pending is a first-class state while prior years are still being fetched. The response states generatedAt, stale (72 h, matching /api/health), the methodVersion that pins the method constants, and the Copernicus credit line. unavailableReason separates NOT_COVERED (a filter that names no registered region) from seed, contract and cache faults. The MCP tool get_no2_activity proxies the route at subscription access, weight 2, and takes country names and alpha-3 codes. The tools/list description budget moves from 64 KiB to 68 KiB.
- Industrial activity (NO2) map layer and country brief line (#9722): an opt-in deck.gl layer colors the 167 countries and 32 industrial clusters of the Sentinel-5P NO2 dataset by direction (strengthening, weakening, unchanged), with baseline-pending and excluded regions as their own states, a legend carrying the Copernicus credit, and a hover card with the change, window, valid days and exclusion. The country brief gains an Industrial Activity (NO2) section with the same states. Both read GET /api/economic/v1/get-no2-activity; the layer reads it only while on, never at bootstrap.
- Per-country subsea cable exposure in MCP (#9722): simulate_infrastructure_cascade takes countries (up to 10) and returns data.exposure with each country's landing cables, distinct landing stations, the chokepoints those cables cross and the share each carries, and cables the cable-health monitor flags FAULT or DEGRADED. The MCP cable adapter now keeps route geometry, and a cable that lands inside the Red Sea and also on the Indian Ocean or Mediterranean side counts as crossing Bab el-Mandeb or Suez. Before this, cut_cables_through for chokepoint:bab_el_mandeb or chokepoint:suez reported that no cable crossed. No new tool.
- Five-factor country scorecard REST routes are free (#9742): GET /api/scorecard/v1/get-five-factor-scorecard, get-bloc-scorecard and list-five-factor-scorecards now return 200 for anonymous and free-account callers — removed from ENDPOINT_ENTITLEMENTS (server/_shared/entitlement-check.ts) and PREMIUM_RPC_PATHS (src/shared/premium-paths.ts), and from the widget agent's Pro-only exclusion list (scripts/_widget-data-policy.cjs), which now advertises the three routes to the anonymous widget session. The Country Deep Dive panel's five-factor scorecard section no longer shows a Pro lock or resubscribes on auth/entitlement changes; it always loads. The MCP tools get_five_factor_scorecard and list_five_factor_scorecards are unchanged and still require a WorldMonitor subscription — MCP tool calls authenticate separately from the REST API, and every other MCP data tool besides get_sources requires credentials by design (tests/mcp-free-tier-subset.test.mjs). Every other derived-analytics dataset in the Pro Intelligence Suite stays Pro-gated.