[damienesp]

Summary

Four changes to the MCP Apps boards, on top of #290 / #291:

1. The Route Planner is the second MCP Apps view, for leadbay_followups_map, on /chatgpt/mcp and /apps/mcp.
2. Boards load once: they render the tool call that opened them instead of calling it again.
3. Boards in French: both boards and the kit's status picker follow the host's or the browser's language.
4. Trip plans open the planner in ChatGPT: tour_plan's "Route planner" step now works on MCP Apps hosts.

Claude is unchanged: /mcp and /fr/mcp carry no board pointer, the followups_map description there is byte-for-byte the same, and tour_plan's steps on Claude are identical.

Changes

Route Planner as an MCP App

• apps.ts: ui://leadbay/route-planner for leadbay_followups_map. It declares https://cdnjs.cloudflare.com (Leaflet) as its only resourceDomains; the triage board still declares no CSP.
• The planner hard-wired window.claude.use("mcp"), so it was dead in cowork and in MCP Apps hosts. It now uses claude.ai's bridge only on claude.ai, and the kit's transport everywhere else.
• Sizing: its 100% / vh layout collapses in a frame sized from the page's own height. Under data-lb-surface="mcp-app", set by the bridge after the handshake, it gets:
  • 680px, or the host's containerDimensions (--lb-frame-height / --lb-frame-max-height);
  • a 360px map on a phone.
• It opens on the agent's city / city_id.
• followups-map.md.tmpl: an {{apps}} sentence says the board is the rendering, so ChatGPT doesn't also draw place cards. That description is 16,725 chars on the apps surfaces.

Load once

• The bridge keeps the first ui/notifications/tool-input and tool-result.
• lb.openingResult() hands that result to the board. It's null on cowork and claude.ai, and on an error result.
• Both boards render it, then continue with the agent's count and lensId (the triage board pins result.lens.id). That means one call and one LEAD_SEEN receipt instead of two.

French

• lb.locale() returns "fr" | "en": the MCP Apps host's locale once it's reported, otherwise navigator.languages[0].
• lb.i18n({ en, fr }) returns t(key, vars), falling back to English per key.
• Both boards render every label through it, and the fixed markup carries data-i18n / data-i18n-aria / data-i18n-placeholder.
• lb.leadStatus options are localized.
• Wording follows the team's choices: "Wanted" → En cours, "Still chasing" → Toujours en cours, "Meeting planned" → Rendez-vous planifié, "Monitor" → My leads / Mes leads.

Trip plans on MCP Apps hosts

• tour_plan's "Route planner" step told the agent to publish the route_planner artifact, which these hosts can't do.
• client.apps is now set by buildServer from includeApps, following the same pattern as client.commerce.
• On apps hosts the step calls leadbay_followups_map with the tour's city (kind: "open_board"), and the artifact-only lead-desk step is dropped.

Docs and release

• Kit guide: a new "One page, three surfaces" section (never hard-wire a host runtime, load once, size for the frame, label in the viewer's language), the three surface helpers, and both recipes updated.
• CLAUDE.md lists both views.
• Version 0.45.0: CHANGELOG, package.json, server.json, plugin.json.

Tests

New files:

• components/test/mcp-app-opening.test.ts (12): opening result, first result kept, cancel/timeout, frame variables, locale, i18n, French status picker.
• components/test/boards-as-mcp-app.test.ts (9): both real boards booted in jsdom, as a fake MCP Apps host and as Claude artifacts.
  • Triage board: no second pull_leads, and next page continues with the agent's count and lensId.
  • French via the host locale and via the browser.
  • Route Planner: opens on the agent's city with no second call; loads on cowork; keeps its claude.ai bridge.
• components/test/guide-one-page-three-surfaces.test.ts (3).
• mcp/test/unit/mcp-apps-route-planner.test.ts (6): Claude side clean, pointer and sentence on both apps surfaces, CSP exactly cdnjs.
• core/test/unit/composite/tour-plan-board-on-apps-hosts.test.ts (6) and mcp/test/unit/mcp-apps-client-flag.test.ts (2).

Existing tests touched: five lines in mcp-apps-gate, mcp-apps-path and mcp-apps-description (from #290). They pinned a single board (toEqual(["leadbay_pull_leads"]), toEqual([TRIAGE_URI]), skipping only pull_leads) and are widened to admit the second one. No Claude-side assertion is loosened.

Results:

• Components: 632/632. Core: 1942/1942.
• MCP: all green except the 8 known Windows-only failures (release scripts, file mode, symlink). The promptforge snippet-path test is also Windows-only. All of these fail identically on main.
• Generated files are up to date.

How to verify

1. Deploy, then in ChatGPT developer mode on /chatgpt/mcp:
   • "Show me my follow-ups in Lyon on a map": the planner opens in the chat on Lyon, fills its frame (no collapsed map), and doesn't load the leads twice.
   • "I'm in Lyon Thursday, who should I meet?": then pick "Route planner"; the planner opens on Lyon.
   • "Show me today's leads": the triage board opens; "Next" continues the same batch.
   • With ChatGPT or the browser in French: labels are in French.
2. In Claude: the trip-plan "Route planner" step still publishes the artifact, which loads as before (now also in cowork) and is labeled in French for a French browser.

🤖 Generated with Claude Code