winagent-sandbox-mcp.git/src/Agent.Core/Tools, branch main

feat: harden Python workflow and structured planning

2026-08-11T15:02:23+00:00

Replace the uv-specific skill with a portable Python skill and rewrite the default user system prompt as a concise, structured Markdown execution policy. Normalize a single fenced structured-JSON response, retry invalid skill selection with think disabled, validate selected skill output, and make coding plans keep editing and verification in one task. Document the file_edit patch payload, expose that contract in the tool catalog, and convert malformed patch input into a retryable tool failure instead of terminating the agent task. Update configuration and runtime tests for the Python skill, exact tool catalog, structured-output retry, and implementation-and-verification planning guidance. Verified with 82 Agent.Core tests and an agent-mcp Python scheduling scenario with 49 passing unittest cases.

Require task-specific tool evidence

2026-08-11T13:36:21+00:00

Verify task-specific external effect evidence

2026-08-11T13:20:25+00:00

Complete runtime diagnostics and verification coverage

2026-08-11T12:53:57+00:00

Audit and align v1 runtime design

2026-08-11T12:44:53+00:00

Fix PlanAsync fallback and BeginTask TaskGraph conflict

2026-08-10T20:54:22+00:00

Two runtime bugs discovered during MCP integration testing: 1. PlannerService.BuildTaskGraph fallback: If the LLM-generated plan contains invalid tasks (bad estimates, cycles, etc.), graph.AddTask throws PlannerValidationException. Previously this propagated up and caused "No tasks were executed." or "Planning failed" errors. Fix: wrap BuildTaskGraph in try/catch, fall back to single-task graph. Also handle graph.Count == 0 after filtering. 2. TaskGraphToolHandler.BeginTask conflict: PlanAsync stores the planned TaskGraph in session.TaskGraphs[taskId] before RunAsync calls BeginTaskAsync. BeginTask then tried TryAdd which failed because the key already existed, throwing "TaskGraph for '...' already exists in session". Fix: use TryAdd without throwing on conflict — if PlanAsync already created the graph, keep it; otherwise add a new empty one. Test results: Core 61/61, Planner 10/10, Windows 41/41 (elevated). MCP integration: agent_run/agent_status verified via HTTP transport. ACL boundary: C:\Windows\System32 access correctly denied by sandbox. ToolUseGuard: "Required external effect was not completed" fires correctly when LLM does not produce verifiable external effect.

Implement P2/P3: Plan-Run flow separation, PlannerService, and Windows test fixes

2026-08-10T20:17:32+00:00

## P3: Session/Planning Foundation ### PlanAsync / RunAsync flow separation (Design §8.4) - AgentRuntime.ExecuteAsync now follows a two-phase lifecycle: 1. Planning: generate Initial TaskGraph via Ollama structured output 2. Running: execute ready Tasks (dependencies satisfied) in critical-path then creation order - Sandbox Task Resource (BeginTask/EndTask) is applied at RunAsync time, not PlanAsync time (Design constraint) - QueuedTaskWorker transitions to Planning; ExecuteAsync transitions to Running internally ### PlannerService (new: src/Agent.Core/Runtime/PlannerService.cs) - PlanAsync: uses Ollama ChatStructuredAsync with JSON Schema to generate a TaskGraph with per-Task requires_external_effect - ReplanAsync: rebuilds TaskGraph after policy deny, preserving completed tasks (Design §8.4) - Fallback: single-task plan with conservative estimates when structured output fails or returns empty - Pure analysis tasks get RequiresExternalEffect=false with at least 1 Task (Design constraint) ### PlanResult / PlanSchema (new: src/Agent.Core/Runtime/PlanResult.cs) - JSON model for LLM structured output during Planning - JSON Schema with tasks array: description, dependencies (by index), budget_class, PERT estimates, requires_external_effect ### requires_external_effect integration (P2 + P3) - Removed keyword heuristic RequiresExternalEffect() from AgentRuntime - Removed tool-name-counting RequiredToolCalls() from AgentRuntime - ToolUseGuard now receives PlannerTask.RequiresExternalEffect (set by Planner during Planning, not by keyword matching) - requires_external_effect is NOT always true; it is a per-Task attribute determined by the Planner (Design constraint) ### RunAsync / ready Task execution - SelectNextReadyTask: dependencies satisfied, not completed/failed, critical path first then creation order - Each PlannerTask executed via ExecutePlannerTaskAsync (LLM chat-with-tools iteration loop) - TaskGraph.AddUsage records per-Task token usage for CPM tracking ### Replan implementation - Policy deny (InvalidOperationException with "denied by sandbox policy") triggers ReplanAsync, up to max 3 replan attempts - Failed task is marked Failed in TaskGraph before replan - Registry transitions back to Running for replan iteration ## P3-A: AgentSession + TaskGraph ownership ### AgentSession expansion (src/Agent.Core/Sessions/AgentSession.cs) - Added Summary, ExecutionResults, TaskGraphs (ConcurrentDictionary) - TaskGraph is per-MCP-Task (1 agent_run = 1 TaskGraph), NOT shared per-Session (Design constraint) - Added CompactConversation, CompactExecutionResults, SetSummary methods for context summary compaction ### ExecutionResult (new: src/Agent.Core/Sessions/ExecutionResult.cs) - Records tool execution results in session for context summary - Fields: TaskId, ToolName, Success, Output ### SessionStore (src/Agent.Core/Sessions/SessionStore.cs) - Added Sessions property for TaskGraphToolHandler to resolve graphs by TaskId across sessions ### sessionId threading - IInternalToolBackend.BeginTaskAsync signature changed to include sessionId parameter - InternalToolDispatcher, WindowsToolBackend, TaskGraphToolHandler updated to thread sessionId through - task_update schema includes requires_external_effect property - McpHost and Program.cs pass SessionStore to TaskRegistry ## P3: Context Summary (Design §9.3) - Context utilization = latest prompt_eval_count / num_ctx (NOT cumulative InputTokens) (Design constraint) - TryContextSummary: compacts old Conversation/ExecutionResults when utilization exceeds 80% threshold - AgentRuntimeSettings: added NumCtx, ContextSummaryThreshold, ConversationKeepLastN, ExecutionResultsKeepLastN - OllamaResult: added PromptEvalCount field - OllamaClient: passes PromptEvalCount from OllamaChatResponse ## P3: ExecutionResults Session recording - ExecutePlannerTaskAsync records each tool execution result via session.AppendExecutionResult for context summary and audit ## Windows test fixes ### SandboxPolicyException catch (src/Agent.Windows/Tools/WindowsToolBackend.cs) - ExecuteAsync now catches SandboxPolicyException and returns ToolExecutionResult(false, ..., RetryAllowed: false) instead of propagating the exception - Fixes Dir_mkdir_outside_approved_root_is_denied and File_edit_rename_rejects_outside_approved_root tests ### LoopbackExemption cleanup (tests/Agent.Windows.Tests/SecurityPolicyTests.cs) - Recovery_after_crash: added pre-clean and try/finally to ensure LoopbackExemption is always removed, even on assertion failure - Independent_process_crash: added pre-clean to remove residual loopback exemption from prior test failures - Fixes 7 Independent_process_crash scenarios that were failing due to LoopbackExemption residue from Recovery_after_crash ## Test updates (tests/Agent.Core.Tests/PromptAndRuntimeTests.cs) - FakeOllamaClient implementations updated with ChatStructuredAsync and ChatWithToolsAsync overrides to support Plan/Run flow - RecordingOllamaClient: returns single-task plan JSON for ChatStructuredAsync, tracks all call types - ToolCallingOllamaClient: returns plan with requires_external_effect=true for file edit tasks - Test expectations updated for Plan+Run call counts and token accounting (Plan tokens not added to registry usage) ## Documentation updates - Design.md §7.7: requires_external_effect is set by Planner during Planning, not keyword heuristic - Design.md §8.4: Plan/Run phase separation, replan after policy deny - Design.md §9.3: context utilization formula clarified - ImplementationPlan.md §10.3: Plan/Run flow, replan, Context Summary ## Test results (all passing) - Agent.Core.Tests: 61/61 - Agent.Planner.Tests: 10/10 - Agent.Windows.Tests (elevated): 41/41 - Build: 0 errors, 0 warnings

feat: P1 tool improvements and P5 Agent Skills standard compliance

2026-08-10T15:04:36+00:00

P1: Tool小改善 1. tailバイナリ判定 (TextFileDetector): - file_search tailモードでtext/binary判定を追加 - 先頭4KBのみ読み、BOM確認(UTF-8/UTF-16 LE/BE)→UTF-8 strict byte検証 - NULバイトのみ判定不可(UTF-16でNULが正常出現するため) - binaryファイルはToolValidationExceptionで拒否 - TextFileDetectorクラスへ判定ロジックを分離 2. file_edit rename機能: - file_edit.operation=renameを追加(新規toolではない) - new_path property追加、Write capability要求(source/destination双方) - destination既存時は禁止(overwrite不可) - cross-volume renameは不支持、Tool failureとして拒否 - FileRenameTransaction + RecoveryManager file.rename recovery追加 - crash recovery: destination hash == journal記録の場合のみ戻す P5: Agent Skills open standard準拠 3. Skill形式をagentskills.io標準へ移行: - legacy単一Markdown形式(skills/*.md)を廃止 - /SKILL.md標準directory形式のみ対応 - YAML frontmatter必須(name/description)、YamlDotNetで解析 - validation: name必須(1-64文字,lowercase英数字+hyphen,親dir名一致), description必須(1-1024文字), malformed YAML reject, duplicate fatal 4. Progressive Disclosure: - Stage 1 Discovery: name+descriptionのみをsystem messageへ(name+desc一覧) - Stage 3 Instructions: 選択Skillの本文のみ注入 - PromptComposer: 全文埋め込みを廃止、カタログ形式へ変更 - PromptContext: Skills→AvailableSkills+SelectedSkillsへ分割 - 未選択Skill本文はpromptへ入らない 5. Internal Tool境界維持: - skill_read等のInternal Tool追加なし - Internal Toolは6群(file_search/file_read/file_edit/process_run/task_update/dir)のまま - Skill選択・本文ロードはAgent Runtime/Prompt Context Managementの責務 6. AgentSkill data model (Agent.Core/Skills/AgentSkill.cs): - RootPath, Name, Description, Instructions, License?, Compatibility?, Metadata?, AllowedTools? - SkillLoader(Agent.App)がYamlDotNetでfrontmatter解析→AgentSkill構築設計判断: - renameはfile_edit operation拡張(新規toolではない) - rmdir空dirのみ制約は維持 - Skill load != permission grant維持 - allowed-toolsはSandbox permissionとして解釈しない - YamlDotNetはAgent.App側のみ使用(Agent.Coreへ依存追加なし) 変更対象: - TextFileDetector.cs: 新規、text/binary判定 - WindowsToolBackend.cs: tail binary判定統合、rename case追加 - InternalToolCatalog.cs: file_edit operation enum に rename 追加、new_path property - InternalToolValidator.cs: rename時new_path必須検証 - TransactionalFileEditor.cs: Rename() + FileRenameTransaction - RecoveryManager.cs: file.rename recovery - AgentSkill.cs: 新規data model - SkillLoader.cs: 全面書き換え、YamlDotNet frontmatter解析、validation - PromptComposer.cs: Progressive Disclosure、AvailableSkills+SelectedSkills - AgentRuntimeSettings.cs: Skills型変更 - AgentRuntime.cs: SelectedSkills=[]で初期化 - Program.cs: SkillLoader戻り値型対応 - Specifications.md: FS-07 rename、Skill節Agent Skills standard準拠 - Design.md: §7.1 tail binary判定、§7.3 rename、AgentSkill model - ImplementationPlan.md: Prompt順序更新、Skill標準準拠テスト: - SecurityPolicyTests.cs: tail binary拒否/UTF-8/UTF-16 text許可、rename正常/dest既存拒否/許可外拒否 - InternalToolTests.cs: rename validationテスト追加 - ConfigurationTests.cs: SkillLoader標準形式/missing frontmatter/name mismatch/duplicate tests - PromptAndRuntimeTests.cs: PromptContext型変更対応、skill catalog検証検証結果: - Release build: 警告0、エラー0 - Core tests: 61/61 PASS、skip 0 - Planner tests: 10/10 PASS、skip 0 - Internal Tool 6群固定: 確認済み

feat(tools): add dir internal tool for directory operations

2026-08-10T10:38:55+00:00

新規Internal Tool `dir` を追加し、ディレクトリの作成・削除・一覧取得を可能にした。file_editの拡張ではなく独立toolとして実装し、LLMが「ディレクトリ操作はdir tool」と直感的に判断できるようにした。追加した機能: - dir tool (6番目のInternal Tool、MCP非公開) - mkdir: Write capability要求、Directory.CreateDirectory()で親ディレクトリも含めて作成 - rmdir: Delete capability要求、空ディレクトリのみ削除（非空はToolValidationExceptionで拒否） - list: Read capability要求、直下エントリ一覧を返す設計判断: - file_edit拡張ではなく新規toolとした理由: ツール名から機能が推測しやすく、 LLMが直感的に選択できるため - rmdirを空ディレクトリのみに制限した理由: rollback時の内容復元の複雑さを排除、 MemoryBackupLimit超過リスクなし、安全側 - mkdir/rmdirは既存のPathPolicy/ACL/Journal/Recovery枠組みを利用 - root継承ACE方式のため新規作成ディレクトリは自動的に親のACEを継承、追加ACL操作不要変更対象: - InternalToolCatalog.cs: dir定義追加 (5→6 tool) - InternalToolDispatcher.cs: FakeInternalToolBackendの効果判定にdir追加 - InternalToolValidator.cs: dir用コメント追記 - WindowsToolBackend.cs: DirectoryOp()メソッド実装 - TransactionalFileEditor.cs: Mkdir()/Rmdir() transaction + DirectoryMkdirTransaction/DirectoryRmdirTransaction クラス追加 - RecoveryManager.cs: dir.mkdir/dir.rmdir crash recovery追加 - Specifications.md: FS-06追加 - Design.md: 6群表更新、§7.6 dir追加、5群→6群参照更新テスト: - InternalToolTests.cs: dir単体テスト、catalog 6群検証、malformed args拒否 - SecurityPolicyTests.cs: mkdir/rmdir/list/許可外deny テスト追加 - PromptAndRuntimeTests.cs: tool count 5→6 - OllamaClientTests.cs: tool catalog配列長 5→6 検証結果: - Release build: 警告0、エラー0 - Planner tests: 10/10 PASS、skip 0 - Core tests: 56/56 PASS、skip 0 - MCP経由での動作確認: mkdir/list/rmdir/file_edit すべて正常 - Python実行確認: dir mkdir + file_edit + process_run でfib(1)〜fib(10)が正常出力 - Release Bundle publish済み既知の運用上の知見: - uv venvのpyvenv.cfgのhomeがjunctionを指す場合、CanonicalPath.Resolveが reparse pointとして拒否する。実体パスへ書き換えるか、実体ディレクトリを agent.yamlのread/executeに追加する必要がある - venvのpython.exeはuv trampolineで子プロセスを生成するため、 sandboxのchild-process制限で拒否される。pyvenv.cfgが指す実体python.exeを直接起動する必要がある

fix(agent): fail fast on sandbox policy denial

2026-08-09T16:39:40+00:00

Treat Policy Engine denials and UnauthorizedAccessException results as terminal tool failures by adding RetryAllowed metadata to ToolExecutionResult. Stop the Agent loop immediately after recording the denied tool result, transition the Task to failed, and run normal Task resource cleanup instead of returning the same failure to the LLM for another iteration. Correctable validation and process failures retain the existing bounded retry behavior. Expose RetryAllowed in llm.jsonl and document the fail-fast security invariant in Specifications, Design, and ImplementationPlan. Tests: Agent.Core.Tests 50/50 passed with skip 0; Release build passed with 0 warnings and 0 errors. The full Windows suite remained environment-blocked outside an elevated token: 14 tests failed while creating/deriving the fixed AppContainer or creating symlinks; 11 passed and none were skipped. Release: refreshed the win-x64 self-contained folder binaries, publish manifest, ZIP, and checksum while preserving the local agent.yaml. The clean ZIP contains no runtime logs or Task state.