From 35719799e62a18850efc52d380fb53bf3baa8022 Mon Sep 17 00:00:00 2001 From: tanaka takashi Date: Tue, 11 Aug 2026 05:17:32 +0900 Subject: Implement P2/P3: Plan-Run flow separation, PlannerService, and Windows test fixes MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## 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 --- doc/Design.md | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) (limited to 'doc/Design.md') diff --git a/doc/Design.md b/doc/Design.md index 31d0247..798a134 100644 --- a/doc/Design.md +++ b/doc/Design.md @@ -330,7 +330,7 @@ PlannerがDAG、不変条件、completed immutable、max_tasksを再検査して - 既存の PathPolicy / CanonicalPath / TaskAclManager (root継承ACE) / TaskResourceJournal / RecoveryManager 枠組みを利用する。新規作成ディレクトリは親の継承ACEを自動取得するため追加ACL操作は不要。 - mkdir後crash → ディレクトリ削除、rmdir後crash → 空ディレクトリ再作成、でRecoveryする。 ### 7.7 ToolUseGuard -Taskには `requires_external_effect` をPlanner内部属性として持たせる。trueのTaskで、当該iteration中のtool callが0かつ完了証拠がない場合、LLMへ「未実行」を示すsystem-generated observationを追加して再試行する。同一失敗fingerprintが5回に達したらTaskをfailedにする。ただし、Policy EngineのdenyおよびSandbox backendの`UnauthorizedAccessException`は`RetryAllowed=false`のterminal Tool resultへ変換する。Runtimeはobserverへ結果を1回記録した直後にTaskをfailedへ遷移させ、LLMへ再投入しない。 +Taskには `requires_external_effect` をPlanner内部属性として持たせる。**Planning フェーズで LLM structured output により Task ごとに判定・設定される**（keyword heuristic ではない）。trueのTaskで、当該iteration中のtool callが0かつ完了証拠がない場合、LLMへ「未実行」を示すsystem-generated observationを追加して再試行する。同一失敗fingerprintが5回に達したらTaskをfailedにする。ただし、Policy EngineのdenyおよびSandbox backendの`UnauthorizedAccessException`は`RetryAllowed=false`のterminal Tool resultへ変換する。Runtimeはobserverへ結果を1回記録した直後にTaskをfailedへ遷移させ、LLMへ再投入しない。Policy deny後は再計画（replan）が可能（§8.4）。 --- ## 8. Token Planner設計 ### 8.1 Task model @@ -357,11 +357,14 @@ validation: `0 < O <= M <= P <= BudgetClass.MaxTokens`. ### 8.3 CPM topological sort後にforward passで各Taskの累積expected token costを計算し、最大経路をCritical Pathとする。時間のearliest/latest startという語をそのままUIへ出さず、`cumulative_token_cost`、`critical_path` と表記する。 ### 8.4 再計画 +- **Plan フェーズ**: LLM structured output により Initial TaskGraph を生成。Planner が Task ごとに `requires_external_effect` を判定。pure analysis task は `requires_external_effect=false` で少なくとも1 Task の TaskGraph を持つ。Sandbox Task Resource は Plan 時には適用しない。 +- **Run フェーズ**: ready な Task (dependencies 解決済み) を critical path → creation order で順次実行。Sandbox Task Resource (BeginTask/EndTask) は RunAsync 時に適用する。 - 未完了Taskの追加/dependency追加後にDAG validationとCPMを再計算。 - completed Taskは変更しない。 - 実績がPを超過しても即失敗にはせず、remaining Taskのestimate見直しをLLMへ要求できる。 - `max_tasks=32`超過提案は拒否し、既存Taskの統合をLLMへ求める。 - Budget Classは意味的Taskを分割する唯一の理由にしない。Taskが複数独立成果物を含む場合に分割する。 +- Policy deny時は replan を最大3回まで試行可能。同一 deny が繰り返される場合は Failed へ遷移。 ### 8.5 Ollama usage集計 1 Ollama callの `prompt_eval_count` をTask inputへ、`eval_count` をTask outputへ加算する。tool実行そのものにはtoken costを加算しない。context summary生成に使ったtokenはSession maintenance Taskとして別集計する。 --- @@ -388,7 +391,7 @@ AgentSession 7. previous Summary 8. recent Conversation / Execution Result ### 9.3 Summary policy -`num_ctx`に対する利用率が設定閾値を超えたとき、古いConversation/Execution Resultを要約する。既定閾値は80%とする。次は削除・曖昧化してはならない。 +`num_ctx`に対する利用率が設定閾値を超えたとき、古いConversation/Execution Resultを要約する。利用率 = latest `prompt_eval_count` / `num_ctx`（累積 InputTokens ではない）。既定閾値は80%とする。次は削除・曖昧化してはならない。 - 利用者が明示した制約 - 未完了Taskとdependency - completed Taskの重要成果 -- cgit v1.2.3