summaryrefslogtreecommitdiff
path: root/src/Agent.Core/Tools
diff options
context:
space:
mode:
authortanaka takashi <isazaka@gmail.com>2026-08-09 21:08:20 +0900
committertanaka takashi <isazaka@gmail.com>2026-08-09 21:08:20 +0900
commitb07b41cd435991cd254957af79055080c85b346e (patch)
tree27720d3ce14e4d0b84670a55ac26a2f877ce3a20 /src/Agent.Core/Tools
parent2fefd29e86c3b926d0cf4a362251d3f119688612 (diff)
feat: AI Agent Sandbox MCP Server v1を実装
Specifications、Design、ImplementationPlanの優先順位を整理し、単一Task実行モデル、typed option優先、設定可能なresource上限、local Ollamaを必須とするRelease Gateを文書化。 .NET 10/C# 14の5 production project、3 test project、E2E Clientを構築。typed YAML、en/ja locale、fixed/user prompt、stdioおよびloopback限定Streamable HTTP、agent_run/agent_status、Session、Ollama text/tool/Vision、TaskGraphとtoken accountingを実装。 Windows sandboxとしてfixed AppContainer、canonical/reparse検査、能力別ACL、path+SHA-256 executable approval、child-process restriction、Job Object、task journal、atomic file transaction、crash recoveryを実装。WFPのIPv4/IPv6 default denyとexact allow、dynamic session、temporary loopback exemption、管理者/BFE診断を追加。 Security/Recovery test、MCP/Ollama integration test、stdio/HTTP Operational E2E、Release Bundle生成・checksum・manifest・fresh extraction検査を追加。Requirement TraceabilityとPhase 0〜11の進捗、初回セットアップ可能なREADMEを整備。 Debug/Release/publish/objをrepository rootのbuild/へ集約し、旧project-local bin/objへの依存を除去。artifacts/とbuild/をGit管理外に設定し、clean checkoutでもsolution buildからE2E Clientまで生成可能にした。
Diffstat (limited to 'src/Agent.Core/Tools')
-rw-r--r--src/Agent.Core/Tools/InternalToolCatalog.cs32
-rw-r--r--src/Agent.Core/Tools/InternalToolDefinition.cs5
-rw-r--r--src/Agent.Core/Tools/InternalToolDispatcher.cs29
-rw-r--r--src/Agent.Core/Tools/InternalToolValidator.cs86
-rw-r--r--src/Agent.Core/Tools/TaskGraphToolHandler.cs123
-rw-r--r--src/Agent.Core/Tools/ToolPolicy.cs31
-rw-r--r--src/Agent.Core/Tools/ToolUseGuard.cs22
-rw-r--r--src/Agent.Core/Tools/ToolValidationException.cs3
8 files changed, 331 insertions, 0 deletions
diff --git a/src/Agent.Core/Tools/InternalToolCatalog.cs b/src/Agent.Core/Tools/InternalToolCatalog.cs
new file mode 100644
index 0000000..1563c16
--- /dev/null
+++ b/src/Agent.Core/Tools/InternalToolCatalog.cs
@@ -0,0 +1,32 @@
+using System.Text.Json;
+
+namespace Agent.Core.Tools;
+
+public static class InternalToolCatalog
+{
+ private static readonly IReadOnlyDictionary<string, InternalToolDefinition> Definitions = CreateDefinitions();
+
+ public static IReadOnlyList<InternalToolDefinition> All { get; } = Definitions.Values.ToArray();
+ public static InternalToolDefinition Get(string name) => Definitions.TryGetValue(name, out InternalToolDefinition? definition)
+ ? definition
+ : throw new ToolValidationException($"Unknown internal tool '{name}'.");
+
+ private static IReadOnlyDictionary<string, InternalToolDefinition> CreateDefinitions()
+ {
+ InternalToolDefinition[] definitions =
+ [
+ Define("file_search", "Search files under an approved root.", """{"type":"object","additionalProperties":false,"required":["root","mode","query"],"properties":{"root":{"type":"string"},"mode":{"type":"string","enum":["glob","name","text","regex","tail"]},"query":{"type":"string"},"max_results":{"type":"integer","minimum":1,"maximum":1000},"tail_lines":{"type":"integer","minimum":1,"maximum":10000}}}"""),
+ Define("file_read", "Read an approved file.", """{"type":"object","additionalProperties":false,"required":["path","mode"],"properties":{"path":{"type":"string"},"mode":{"type":"string","enum":["whole","range","metadata"]},"start":{"type":"integer","minimum":0},"length":{"type":"integer","minimum":0}}}"""),
+ Define("file_edit", "Edit an approved file transactionally.", """{"type":"object","additionalProperties":false,"required":["path","operation"],"properties":{"path":{"type":"string"},"operation":{"type":"string","enum":["overwrite","patch","atomic_replace","delete"]},"content":{"type":["string","null"]},"patch":{"type":["string","null"]},"expected_sha256":{"type":["string","null"]}}}"""),
+ Define("process_run", "Run one approved executable with separated arguments.", """{"type":"object","additionalProperties":false,"required":["executable","args","cwd"],"properties":{"executable":{"type":"string"},"args":{"type":"array","items":{"type":"string"}},"cwd":{"type":"string"},"timeout_seconds":{"type":"integer","minimum":1},"stdin":{"type":["string","null"]}}}"""),
+ Define("task_update", "Update the deterministic task graph.", """{"type":"object","additionalProperties":false,"required":["operation"],"properties":{"operation":{"type":"string","enum":["add","add_dependency","set_state","set_estimate"]},"task_id":{"type":["string","null"]},"description":{"type":["string","null"]},"depends_on":{"type":"array","items":{"type":"string"}},"budget_class":{"type":["string","null"],"enum":["1K","2K","4K","8K","16K","32K","64K","128K",null]},"optimistic_tokens":{"type":"integer","minimum":0},"most_likely_tokens":{"type":"integer","minimum":0},"pessimistic_tokens":{"type":"integer","minimum":0},"state":{"type":["string","null"],"enum":["queued","planning","running","completed","failed","cancelled",null]}}}""")
+ ];
+ return definitions.ToDictionary(static definition => definition.Name, StringComparer.Ordinal);
+ }
+
+ private static InternalToolDefinition Define(string name, string description, string schema)
+ {
+ using JsonDocument document = JsonDocument.Parse(schema);
+ return new(name, description, document.RootElement.Clone());
+ }
+}
diff --git a/src/Agent.Core/Tools/InternalToolDefinition.cs b/src/Agent.Core/Tools/InternalToolDefinition.cs
new file mode 100644
index 0000000..8f06764
--- /dev/null
+++ b/src/Agent.Core/Tools/InternalToolDefinition.cs
@@ -0,0 +1,5 @@
+using System.Text.Json;
+
+namespace Agent.Core.Tools;
+
+public sealed record InternalToolDefinition(string Name, string Description, JsonElement Parameters);
diff --git a/src/Agent.Core/Tools/InternalToolDispatcher.cs b/src/Agent.Core/Tools/InternalToolDispatcher.cs
new file mode 100644
index 0000000..7302c1f
--- /dev/null
+++ b/src/Agent.Core/Tools/InternalToolDispatcher.cs
@@ -0,0 +1,29 @@
+using System.Text.Json;
+
+namespace Agent.Core.Tools;
+
+public sealed class InternalToolDispatcher(IToolPolicyEngine policy, IInternalToolBackend backend)
+{
+ public Task BeginTaskAsync(Guid taskId, CancellationToken cancellationToken) => backend.BeginTaskAsync(taskId, cancellationToken);
+ public Task EndTaskAsync(Guid taskId, bool success, CancellationToken cancellationToken) => backend.EndTaskAsync(taskId, success, cancellationToken);
+
+ public async Task<ToolExecutionResult> DispatchAsync(Guid taskId, string toolName, JsonElement arguments, CancellationToken cancellationToken)
+ {
+ InternalToolValidator.Validate(toolName, arguments);
+ ToolPolicyDecision decision = policy.Evaluate(toolName, arguments);
+ if (!decision.Allowed) return new(false, decision.Reason, false, $"policy:{toolName}:{decision.Reason}");
+ return await backend.ExecuteAsync(taskId, toolName, arguments, cancellationToken).ConfigureAwait(false);
+ }
+}
+
+public sealed class FakeInternalToolBackend : IInternalToolBackend
+{
+ private readonly List<(string Name, JsonElement Arguments)> _calls = [];
+ public IReadOnlyList<(string Name, JsonElement Arguments)> Calls => _calls;
+ public Task<ToolExecutionResult> ExecuteAsync(Guid taskId, string toolName, JsonElement arguments, CancellationToken cancellationToken)
+ {
+ _calls.Add((toolName, arguments.Clone()));
+ bool effect = toolName is "file_edit" or "process_run";
+ return Task.FromResult(new ToolExecutionResult(true, "fake-ok", effect));
+ }
+}
diff --git a/src/Agent.Core/Tools/InternalToolValidator.cs b/src/Agent.Core/Tools/InternalToolValidator.cs
new file mode 100644
index 0000000..c48719e
--- /dev/null
+++ b/src/Agent.Core/Tools/InternalToolValidator.cs
@@ -0,0 +1,86 @@
+using System.Text.Json;
+
+namespace Agent.Core.Tools;
+
+public static class InternalToolValidator
+{
+ public static void Validate(string name, JsonElement arguments)
+ {
+ InternalToolDefinition definition = InternalToolCatalog.Get(name);
+ if (arguments.ValueKind != JsonValueKind.Object) throw new ToolValidationException("Tool arguments must be an object.");
+ JsonElement properties = definition.Parameters.GetProperty("properties");
+ var known = properties.EnumerateObject().Select(static item => item.Name).ToHashSet(StringComparer.Ordinal);
+ foreach (JsonProperty argument in arguments.EnumerateObject())
+ {
+ if (!known.Contains(argument.Name)) throw new ToolValidationException($"Unknown argument '{argument.Name}'.");
+ ValidateValue(argument.Name, argument.Value, properties.GetProperty(argument.Name));
+ }
+ foreach (JsonElement required in definition.Parameters.GetProperty("required").EnumerateArray())
+ {
+ string requiredName = required.GetString()!;
+ if (!arguments.TryGetProperty(requiredName, out _)) throw new ToolValidationException($"Missing required argument '{requiredName}'.");
+ }
+ ValidateOperationSpecific(name, arguments);
+ }
+
+ private static void ValidateValue(string name, JsonElement value, JsonElement schema)
+ {
+ if (value.ValueKind == JsonValueKind.Null && AllowsNull(schema)) return;
+ bool validType = GetTypes(schema).Any(type => type switch
+ {
+ "string" => value.ValueKind == JsonValueKind.String,
+ "integer" => value.ValueKind == JsonValueKind.Number && value.TryGetInt64(out _),
+ "array" => value.ValueKind == JsonValueKind.Array,
+ _ => false
+ });
+ if (!validType) throw new ToolValidationException($"Argument '{name}' has an invalid type.");
+ if (schema.TryGetProperty("enum", out JsonElement enumValues) && !enumValues.EnumerateArray().Any(item => JsonElement.DeepEquals(item, value)))
+ throw new ToolValidationException($"Argument '{name}' has an unsupported value.");
+ if (value.ValueKind == JsonValueKind.Number)
+ {
+ long number = value.GetInt64();
+ if (schema.TryGetProperty("minimum", out JsonElement minimum) && number < minimum.GetInt64()) throw new ToolValidationException($"Argument '{name}' is below its minimum.");
+ if (schema.TryGetProperty("maximum", out JsonElement maximum) && number > maximum.GetInt64()) throw new ToolValidationException($"Argument '{name}' exceeds its maximum.");
+ }
+ if (value.ValueKind == JsonValueKind.Array && schema.TryGetProperty("items", out JsonElement items))
+ foreach (JsonElement item in value.EnumerateArray()) ValidateValue(name, item, items);
+ }
+
+ private static IEnumerable<string> GetTypes(JsonElement schema)
+ {
+ JsonElement type = schema.GetProperty("type");
+ return type.ValueKind == JsonValueKind.Array
+ ? type.EnumerateArray().Select(static item => item.GetString()!).ToArray()
+ : [type.GetString()!];
+ }
+ private static bool AllowsNull(JsonElement schema) => GetTypes(schema).Contains("null", StringComparer.Ordinal);
+
+ private static void ValidateOperationSpecific(string name, JsonElement arguments)
+ {
+ if (name == "file_read" && arguments.GetProperty("mode").GetString() == "range" && !arguments.TryGetProperty("length", out _))
+ throw new ToolValidationException("Range reads require length.");
+ if (name == "file_edit")
+ {
+ string operation = arguments.GetProperty("operation").GetString()!;
+ if (operation is "overwrite" or "atomic_replace" && (!arguments.TryGetProperty("content", out JsonElement content) || content.ValueKind != JsonValueKind.String))
+ throw new ToolValidationException($"{operation} requires content.");
+ if (operation == "patch" && (!arguments.TryGetProperty("patch", out JsonElement patch) || patch.ValueKind != JsonValueKind.String))
+ throw new ToolValidationException("patch requires patch text.");
+ }
+ if (name == "task_update")
+ {
+ string operation = arguments.GetProperty("operation").GetString()!;
+ if (operation == "add") Require(arguments, "description", "budget_class", "optimistic_tokens", "most_likely_tokens", "pessimistic_tokens");
+ if (operation == "add_dependency") Require(arguments, "task_id", "depends_on");
+ if (operation == "set_state") Require(arguments, "task_id", "state");
+ if (operation == "set_estimate") Require(arguments, "task_id", "budget_class", "optimistic_tokens", "most_likely_tokens", "pessimistic_tokens");
+ }
+
+ static void Require(JsonElement value, params string[] properties)
+ {
+ foreach (string property in properties)
+ if (!value.TryGetProperty(property, out JsonElement item) || item.ValueKind == JsonValueKind.Null)
+ throw new ToolValidationException($"{property} is required for this operation.");
+ }
+ }
+}
diff --git a/src/Agent.Core/Tools/TaskGraphToolHandler.cs b/src/Agent.Core/Tools/TaskGraphToolHandler.cs
new file mode 100644
index 0000000..03daa56
--- /dev/null
+++ b/src/Agent.Core/Tools/TaskGraphToolHandler.cs
@@ -0,0 +1,123 @@
+using System.Collections.Concurrent;
+using System.Text.Json;
+using Agent.Planner;
+
+namespace Agent.Core.Tools;
+
+public interface ITaskUpdateHandler
+{
+ void BeginTask(Guid taskId);
+ ToolExecutionResult Execute(Guid taskId, JsonElement arguments);
+ void EndTask(Guid taskId);
+}
+
+public sealed class TaskGraphToolHandler(int maxTasks = 32) : ITaskUpdateHandler
+{
+ private readonly ConcurrentDictionary<Guid, TaskGraph> _graphs = new();
+
+ public void BeginTask(Guid taskId)
+ {
+ if (!_graphs.TryAdd(taskId, new TaskGraph(maxTasks)))
+ throw new InvalidOperationException($"TaskGraph for '{taskId}' already exists.");
+ }
+
+ public ToolExecutionResult Execute(Guid taskId, JsonElement arguments)
+ {
+ if (!_graphs.TryGetValue(taskId, out TaskGraph? graph))
+ throw new InvalidOperationException("TaskGraph scope is not active.");
+
+ string operation = arguments.GetProperty("operation").GetString()!;
+ switch (operation)
+ {
+ case "add":
+ graph.AddTask(
+ RequiredString(arguments, "description"),
+ Budget(arguments),
+ Estimate(arguments),
+ Ids(arguments, "depends_on"),
+ id: OptionalId(arguments, "task_id"));
+ break;
+ case "add_dependency":
+ Guid taskIdValue = RequiredId(arguments, "task_id");
+ Guid[] dependencies = Ids(arguments, "depends_on");
+ if (dependencies.Length == 0) throw new ToolValidationException("depends_on must contain at least one task id.");
+ foreach (Guid dependency in dependencies) graph.AddDependency(taskIdValue, dependency);
+ break;
+ case "set_state":
+ graph.SetState(RequiredId(arguments, "task_id"), ParseState(RequiredString(arguments, "state")));
+ break;
+ case "set_estimate":
+ graph.SetEstimate(RequiredId(arguments, "task_id"), Budget(arguments), Estimate(arguments));
+ break;
+ default:
+ throw new ToolValidationException($"Unsupported task_update operation '{operation}'.");
+ }
+
+ string output = JsonSerializer.Serialize(new
+ {
+ tasks = graph.Tasks.Select(static task => new
+ {
+ task_id = task.Id,
+ description = task.Description,
+ depends_on = task.Dependencies,
+ state = task.State.ToString().ToLowerInvariant(),
+ expected_tokens = task.ExpectedTokens
+ }),
+ critical_path = graph.CriticalPath.TaskIds,
+ critical_path_tokens = graph.CriticalPath.CumulativeTokenCost,
+ total_token_budget = graph.TotalTokenBudget
+ });
+ return new ToolExecutionResult(true, output, false);
+ }
+
+ public void EndTask(Guid taskId) => _graphs.TryRemove(taskId, out _);
+
+ private static TokenBudgetClass Budget(JsonElement arguments) =>
+ TokenBudgetClassExtensions.Parse(RequiredString(arguments, "budget_class"));
+
+ private static TokenEstimate Estimate(JsonElement arguments) => new(
+ RequiredInt64(arguments, "optimistic_tokens"),
+ RequiredInt64(arguments, "most_likely_tokens"),
+ RequiredInt64(arguments, "pessimistic_tokens"));
+
+ private static string RequiredString(JsonElement arguments, string property) =>
+ arguments.TryGetProperty(property, out JsonElement value) && value.ValueKind == JsonValueKind.String && !string.IsNullOrWhiteSpace(value.GetString())
+ ? value.GetString()!
+ : throw new ToolValidationException($"{property} is required for this operation.");
+
+ private static long RequiredInt64(JsonElement arguments, string property) =>
+ arguments.TryGetProperty(property, out JsonElement value) && value.ValueKind == JsonValueKind.Number
+ ? value.GetInt64()
+ : throw new ToolValidationException($"{property} is required for this operation.");
+
+ private static Guid RequiredId(JsonElement arguments, string property) =>
+ OptionalId(arguments, property) ?? throw new ToolValidationException($"{property} must be a UUID for this operation.");
+
+ private static Guid? OptionalId(JsonElement arguments, string property)
+ {
+ if (!arguments.TryGetProperty(property, out JsonElement value) || value.ValueKind == JsonValueKind.Null) return null;
+ return value.ValueKind == JsonValueKind.String && Guid.TryParse(value.GetString(), out Guid id)
+ ? id
+ : throw new ToolValidationException($"{property} must be a UUID.");
+ }
+
+ private static Guid[] Ids(JsonElement arguments, string property)
+ {
+ if (!arguments.TryGetProperty(property, out JsonElement value)) return [];
+ var ids = new List<Guid>();
+ foreach (JsonElement item in value.EnumerateArray())
+ ids.Add(Guid.TryParse(item.GetString(), out Guid id) ? id : throw new ToolValidationException($"{property} must contain only UUIDs."));
+ return [.. ids];
+ }
+
+ private static PlannerTaskState ParseState(string state) => state switch
+ {
+ "queued" => PlannerTaskState.Queued,
+ "planning" => PlannerTaskState.Planning,
+ "running" => PlannerTaskState.Running,
+ "completed" => PlannerTaskState.Completed,
+ "failed" => PlannerTaskState.Failed,
+ "cancelled" => PlannerTaskState.Cancelled,
+ _ => throw new ToolValidationException($"Unknown task state '{state}'.")
+ };
+}
diff --git a/src/Agent.Core/Tools/ToolPolicy.cs b/src/Agent.Core/Tools/ToolPolicy.cs
new file mode 100644
index 0000000..c9f216a
--- /dev/null
+++ b/src/Agent.Core/Tools/ToolPolicy.cs
@@ -0,0 +1,31 @@
+using System.Text.Json;
+
+namespace Agent.Core.Tools;
+
+public sealed record ToolPolicyDecision(bool Allowed, string Reason)
+{
+ public static ToolPolicyDecision Allow() => new(true, "allowed");
+ public static ToolPolicyDecision Deny(string reason) => new(false, reason);
+}
+
+public interface IToolPolicyEngine
+{
+ ToolPolicyDecision Evaluate(string toolName, JsonElement arguments);
+}
+
+public interface IInternalToolBackend
+{
+ Task BeginTaskAsync(Guid taskId, CancellationToken cancellationToken) => Task.CompletedTask;
+ Task<ToolExecutionResult> ExecuteAsync(Guid taskId, string toolName, JsonElement arguments, CancellationToken cancellationToken);
+ Task EndTaskAsync(Guid taskId, bool success, CancellationToken cancellationToken) => Task.CompletedTask;
+}
+
+public sealed record ToolExecutionResult(bool Success, string Output, bool ExternalEffect, string? FailureFingerprint = null);
+
+public sealed class DenyByDefaultPolicy(IReadOnlySet<string>? allowedTools = null) : IToolPolicyEngine
+{
+ private readonly IReadOnlySet<string> _allowedTools = allowedTools ?? new HashSet<string>(StringComparer.Ordinal);
+ public ToolPolicyDecision Evaluate(string toolName, JsonElement arguments) => _allowedTools.Contains(toolName)
+ ? ToolPolicyDecision.Allow()
+ : ToolPolicyDecision.Deny("Capability is not allowed by policy.");
+}
diff --git a/src/Agent.Core/Tools/ToolUseGuard.cs b/src/Agent.Core/Tools/ToolUseGuard.cs
new file mode 100644
index 0000000..633d34c
--- /dev/null
+++ b/src/Agent.Core/Tools/ToolUseGuard.cs
@@ -0,0 +1,22 @@
+namespace Agent.Core.Tools;
+
+public enum ToolUseGuardDecision { Complete, Retry, Fail }
+
+public sealed record ToolUseGuardResult(ToolUseGuardDecision Decision, string Observation, int IdenticalFailureCount);
+
+public sealed class ToolUseGuard(int sameFailureLimit = 5)
+{
+ private readonly Dictionary<string, int> _failures = new(StringComparer.Ordinal);
+ public int SameFailureLimit { get; } = sameFailureLimit > 0 ? sameFailureLimit : throw new ArgumentOutOfRangeException(nameof(sameFailureLimit));
+
+ public ToolUseGuardResult Evaluate(bool requiresExternalEffect, bool hadToolCall, bool hasExternalEffectEvidence, string? failureFingerprint = null)
+ {
+ if (!requiresExternalEffect) return new(ToolUseGuardDecision.Complete, string.Empty, 0);
+ if (hasExternalEffectEvidence) return new(ToolUseGuardDecision.Complete, string.Empty, 0);
+ string fingerprint = string.IsNullOrWhiteSpace(failureFingerprint) ? "missing-external-effect" : failureFingerprint;
+ int count = _failures.TryGetValue(fingerprint, out int prior) ? prior + 1 : 1;
+ _failures[fingerprint] = count;
+ if (count >= SameFailureLimit) return new(ToolUseGuardDecision.Fail, "Required external effect was not completed after repeated identical failures.", count);
+ return new(ToolUseGuardDecision.Retry, "No verified external effect occurred; perform the required operation using an internal tool.", count);
+ }
+}
diff --git a/src/Agent.Core/Tools/ToolValidationException.cs b/src/Agent.Core/Tools/ToolValidationException.cs
new file mode 100644
index 0000000..a99472c
--- /dev/null
+++ b/src/Agent.Core/Tools/ToolValidationException.cs
@@ -0,0 +1,3 @@
+namespace Agent.Core.Tools;
+
+public sealed class ToolValidationException(string message) : InvalidOperationException(message);