FieldCure.AssistStudio.Core 0.19.2

FieldCure.AssistStudio.Core

App-level helpers and models for AssistStudio — tool orchestration, specialist agents, workspace context, and MCP server management. AI providers live in FieldCure.Ai.Providers.

Features

• Tool Orchestration — ToolCallExecutor with confirmation handler and parallel execution, ToolResolver for built-in/MCP tool merge
• Specialist Agents — ISpecialist interface for domain-specific sub-agent routing (e.g., Web Search Specialist)
• Knowledge Base — KnowledgeBase model for multi-KB metadata persistence
• Workspace Context — IWorkspaceContext for dynamic system prompt injection
• MCP Server Management — Built-in server lifecycle (install, update, connect) via BuiltInServerHelper
• RAG Support — IContextProvider retrieves ContextChunks for queries
• Structured Logging — DiagnosticLogger with pluggable callbacks
• Hardware Probing — HardwareProbe for GPU/CPU detection (Ollama compatibility)
• Platform-Agnostic — Targets net8.0

Note: AI providers (Claude, OpenAI, Gemini, Ollama, Groq), streaming, and shared models (ChatMessage, AiRequest, IAssistTool, etc.) are in FieldCure.Ai.Providers. See Supported Providers.

Install

dotnet add package FieldCure.AssistStudio.Core

Tool Execution

ToolCallExecutor runs tool calls with optional user confirmation and parallel execution:

using FieldCure.AssistStudio.Core.Helpers;

var executor = new ToolCallExecutor([weatherTool, fileTool, searchTool]);

// User confirmation — returns (Approved, UserNote)
executor.ConfirmationHandler = async (toolName, argsJson) =>
{
    var approved = await ShowConfirmationDialog(toolName, argsJson);
    return (approved, userNote: null);
};

// Execute a single tool call
var result = await executor.ExecuteAsync(toolCall);
Console.WriteLine(result.Text);

// Multimedia results (IMultiContentTool)
foreach (var media in result.MediaItems)
    Console.WriteLine($"  {media.MimeType}: {media.FileName}");

Tool Resolution

ToolResolver merges built-in tools with MCP tools, prefixing names on conflict:

using FieldCure.AssistStudio.Core.Helpers;

// Built-in "read_file" + MCP "read_file" → MCP tool becomes "filesystem_read_file"
var tools = ToolResolver.Resolve(builtInTools, mcpTools, conversationState);

Specialist Agents

Define domain-specific specialists for sub-agent routing via ISpecialist:

using FieldCure.AssistStudio.Core;

public class WebSearchSpecialist : ISpecialist
{
    public string Name => "web_search_specialist";
    public string DisplayName => "Web Search Specialist";
    public string? Icon => null;

    public IReadOnlyList<string> AllowedTools { get; } =
        ["web_search", "web_fetch", "run_javascript"];

    public IReadOnlyList<string> FallbackServers { get; } = ["builtin_essentials"];
    public int MaxRounds => 15;
    public TimeSpan Timeout => TimeSpan.FromMinutes(2);

    public string BuildSystemPrompt(
        string userQuery, IReadOnlyDictionary<string, string>? contextHints = null)
    {
        return $"""
            You are a web research specialist.
            Search the web, read relevant pages, and produce a concise report.

            ## Task
            {userQuery}
            """;
    }
}

Routing hints for the parent

ISpecialist itself only defines the sub-agent's own contract. Implementations can additionally expose a plain const string (by convention named RoutingGuideline) that the host appends to the parent conversation's system prompt. This is how the host steers the parent on when to delegate and how to handle the specialist's result — e.g. forward the returned report verbatim, re-invoke on status: "truncated", or preserve the specialist parameter on retry instead of falling back to generic delegate_task arguments.

WebSearchSpecialist.RoutingGuideline and JudgmentSpecialist.RoutingGuideline in AssistStudio are reference patterns. The constant is not part of the ISpecialist interface — the host that knows about a specialist appends its guideline explicitly when building the parent system prompt.

Tip: model choice per specialist

Different LLM families handle long structured protocols differently. Specialists that spell out an elaborate multi-phase dialectic (e.g. the Judgment protocol's ASSERT → CHALLENGE → SYNTHESIZE loop) behave noticeably differently depending on the underlying model:

• Compression-oriented models tend to execute the protocol with purpose-first summarization — they produce the final report inside a single max_tokens budget and return status: "completed".
• Verbatim-oriented models tend to follow the protocol section by section and more easily hit max_tokens mid-synthesis, returning status: "truncated". The routing guideline's re-invocation rule still recovers a usable report, but at higher token cost.

This is not a correctness issue — the pipeline handles both paths — but it's worth considering when picking the provider preset a specialist runs under. The host already exposes this knob via the specialistPresetProvider delegate passed to SubAgentTool; a future revision of ISpecialist may also carry a PreferredProvider hint.

Workspace Context

Inject dynamic context into the system prompt based on app state:

public class MyWorkspace : IWorkspaceContext
{
    public string? ActiveLabel => "Project: MyApp";
    public Task<string?> GetContextAsync(CancellationToken ct = default)
        => Task.FromResult<string?>($"Current file: {_currentFile}");
}

Diagnostic Logging

Wire up DiagnosticLogger to capture internal events from providers and helpers:

DiagnosticLogger.OnException = ex => logger.LogError(ex, "AssistStudio error");
DiagnosticLogger.OnWarning = msg => logger.LogWarning(msg);
DiagnosticLogger.OnInfo = msg => logger.LogInformation(msg);

Key Types

Core Types

From FieldCure.Ai.Providers (transitive dependency)

Related Packages

• FieldCure.Ai.Providers — AI provider clients (Claude, OpenAI, Gemini, Ollama, Groq) and shared models. Core depends on this.
• FieldCure.Ai.Execution — Agent loop and sub-agent execution engine built on Providers.
• FieldCure.AssistStudio.Controls.WinUI — WinUI 3 chat UI controls built on this library.

License

MIT — Copyright (c) 2026 FieldCure Co., Ltd.