Server has 6 tools with explicit schemas and descriptions. Naming follows verb_noun pattern (searchPlaces, searchJourneys, getDepartures, listDemands, createDemand, updateDemand). Descriptions are present and moderately detailed (100-250 chars), exceeding the 20-char minimum. However, parameter descriptions lack specificity on formats, constraints, and error recovery. Output schemas are not documented, responses return structuredContent but LLMs cannot predict field structure. Error handling is minimal (try/catch with generic messages). No tool annotations (readOnlyHint/destructiveHint) despite clear WRITE vs READ_ONLY risk levels. Composition is reasonable: searchPlaces feeds into searchJourneys/getDepartures; listDemands/createDemand/updateDemand form a coherent demand workflow. Missing: pagination for listDemands, idempotency guidance, and per-parameter type constraints.

• Add explicit output schema documentation to each tool. Define return type as a TypeScript interface or JSON Schema and include it in the tool definition comment or a separate schema field. Example: searchJourneys returns {journeys: Array<{departure: string, arrival: string, duration_min: number, train: string, mode: string, from: string, to: string}>}.
• Add tool annotations: mark searchPlaces/searchJourneys/getDepartures with readOnlyHint: true; mark createDemand/updateDemand with destructiveHint: true. This enables agents to reason about side effects.
• Enhance parameter descriptions with format constraints. For datetime, add '(ISO 8601 format: YYYYMMDDTHHmmss, e.g. 20260501T090000)'. For count, add '(1 - 50, default 5)'. For email, add '(valid email format or omit for all demands)'.
• Implement structured error responses. Instead of 'ERREUR searchPlaces: {message}', return {error: 'not_found', message: 'City not found', suggestion: 'Try a partial name or nearby city'}. Categorize as retryable|user_fixable|fatal.
• Add pagination to listDemands: accept limit (1 - 100, default 20) and offset (default 0) parameters. Return {demands: [...], total: N, hasMore: boolean}. Cap results at 50 to prevent context explosion.

Inferred effective spec: 2026-07-28+.

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained