Server has 4 tools with clear names (audit_project, list_findings, preview_fix, apply_fix) and documented schemas using Zod. Descriptions are present and moderately detailed (100-200 chars). However, parameter descriptions lack specificity about constraints, formats, and error recovery. Output schemas are not formally documented, responses are text-only without structured field definitions. Error handling is minimal: most failures return generic text without actionable recovery guidance. The apply_fix tool correctly requires confirm:true to prevent accidental writes, but lacks dry-run capability. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.

• Document output schemas for all tools. E.g., audit_project should return {pocketbase_url: string, summary: {critical: number, high: number, ...}, findings: [{severity: string, title: string, target: string, ...}], ...}. This enables LLM to extract finding_index for list_findings.
• Add format/constraint descriptions to parameters. E.g., 'url: PocketBase base URL (must be HTTPS, e.g., https://my.pb.io)', 'rule_value: PocketBase rule expression (e.g., @request.auth.id != "" && @request.auth.id = ownerId)', 'finding_index: 0-based index from list_findings output'.
• Implement structured error responses. Return {error: string, type: 'retryable'|'user_fixable'|'fatal', suggestion: string}. E.g., 'error: "Invalid credentials", type: "user_fixable", suggestion: "Check POCKETBASE_ADMIN_EMAIL and POCKETBASE_ADMIN_PASSWORD env vars."'
• Add tool annotations to server.registerTool calls. Mark audit_project with readOnlyHint: true, apply_fix with destructiveHint: true. This signals safety properties to LLM clients.
• Implement dry-run for apply_fix. Add optional 'dry_run: boolean' parameter. When true, return what would change without calling PATCH. This reduces accidental destructive calls.