Strengthen verified task execution prompts

4 files changed, 32 insertions, 3 deletions

diff --git a/doc/Design.md b/doc/Design.md
index f8f5d0b..a637e3b 100644
--- a/doc/Design.md
+++ b/doc/Design.md
@@ -399,7 +399,11 @@ AgentSession
 5. Sandbox capability summary
 6. current TaskGraph / current Task
 7. previous Summary
-8. recent Conversation / Execution Result
+8. recent verified Execution Result
+9. current Taskの`requires_external_effect`に基づく実行要件
+10. recent Conversation
+
+実行要件は、外部効果が必要なTaskでは文章上の宣言、計画、JSON例を実行済みと扱わず、対応するInternal Toolを直ちに呼び出すよう指示する。成功済みのread結果はExecution Resultから再利用し、不要な再読込を避けてwrite/processへ進む。外部効果が不要なTaskにはToolを強制しない。
 ### 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。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
diff --git a/src/Agent.Core/Prompts/PromptComposer.cs b/src/Agent.Core/Prompts/PromptComposer.cs
index 69b18f3..2899088 100644
--- a/src/Agent.Core/Prompts/PromptComposer.cs
+++ b/src/Agent.Core/Prompts/PromptComposer.cs
@@ -13,8 +13,10 @@ public sealed record PromptContext(
     string SandboxCapabilitySummary,
     string TaskGraphSummary,
     string PreviousSummary,
+    IReadOnlyList<ExecutionResult> ExecutionResults,
     IReadOnlyList<ConversationMessage> Conversation,
     string Request,
+    bool RequiresExternalEffect,
     IReadOnlyList<string>? Images = null);
 
 public static class PromptComposer
@@ -29,7 +31,9 @@ public static class PromptComposer
             new("system", ComposeSkillCatalog(context.AvailableSkills)),
             new("system", context.SandboxCapabilitySummary),
             new("system", context.TaskGraphSummary),
-            new("system", context.PreviousSummary)
+            new("system", context.PreviousSummary),
+            new("system", ComposeExecutionResults(context.ExecutionResults)),
+            new("system", ComposeExecutionRequirement(context.RequiresExternalEffect))
         };
         // Inject selected skill instructions (Progressive Disclosure Stage 3).
         foreach (AgentSkill skill in context.SelectedSkills)
@@ -39,6 +43,21 @@ public static class PromptComposer
         return messages;
     }
 
+    private static string ComposeExecutionRequirement(bool requiresExternalEffect) => requiresExternalEffect
+        ? "The current Planner Task requires a verified external effect. Call the appropriate available Internal Tool now. " +
+          "A prose description, promise, plan, or fenced JSON example does not perform the operation and cannot complete this Task. " +
+          "After a successful read, use its returned content and proceed to the required write or process operation without repeating the read unnecessarily."
+        : "The current Planner Task does not require an external effect. Use an Internal Tool only when its verified result is needed; do not claim an unverified result.";
+
+    private static string ComposeExecutionResults(IReadOnlyList<ExecutionResult> results)
+    {
+        if (results.Count == 0) return "No prior Execution Results are available for the current Session.";
+        var lines = new List<string> { "Recent verified Execution Results:" };
+        foreach (ExecutionResult result in results)
+            lines.Add($"- {result.ToolName} ({(result.Success ? "success" : "failed")}): {result.Output}");
+        return string.Join("\n", lines);
+    }
+
     /// <summary>
     /// Stage 1 Discovery: only name + description are included. Full instructions are NOT included.
     /// </summary>
diff --git a/src/Agent.Core/Runtime/AgentRuntime.cs b/src/Agent.Core/Runtime/AgentRuntime.cs
index bc9fc5e..f1c7186 100644
--- a/src/Agent.Core/Runtime/AgentRuntime.cs
+++ b/src/Agent.Core/Runtime/AgentRuntime.cs
@@ -218,8 +218,10 @@ public sealed class AgentRuntime(TaskRegistry registry, IOllamaClient ollama, Ag
             settings.SandboxCapabilitySummary,
             BuildTaskGraphSummary(graph, plannerTask),
             session.Summary ?? string.Empty,
+            session.ExecutionResults,
             session.Conversation,
             plannerTask.Description,
+            plannerTask.RequiresExternalEffect,
             task.Images)).ToList();
 
         var guard = new ToolUseGuard(settings.SameFailureLimit);
diff --git a/tests/Agent.Core.Tests/PromptAndRuntimeTests.cs b/tests/Agent.Core.Tests/PromptAndRuntimeTests.cs
index 836ed5e..a40720f 100644
--- a/tests/Agent.Core.Tests/PromptAndRuntimeTests.cs
+++ b/tests/Agent.Core.Tests/PromptAndRuntimeTests.cs
@@ -20,12 +20,16 @@ public sealed class PromptAndRuntimeTests
         };
         IReadOnlyList<OllamaMessage> messages = PromptComposer.Compose(new PromptContext(
             "fixed", "user system", "ja", availableSkills, [], "capability", "graph", "summary",
+            [new ExecutionResult(Guid.NewGuid(), "file_read", true, "verified content")],
             [new ConversationMessage("user", "old"), new ConversationMessage("assistant", "reply")],
-            "new request", ["image"]));
+            "new request", true, ["image"]));
 
         // Check that the skill catalog (name + description) is present, not the full instructions.
         Assert.Contains(messages, m => m.Content.Contains("test-skill") && m.Content.Contains("A test skill."));
         Assert.DoesNotContain(messages, m => m.Content == "skill instructions");
+        Assert.Contains(messages, m => m.Role == "system" && m.Content.Contains("requires a verified external effect", StringComparison.Ordinal));
+        Assert.Contains(messages, m => m.Role == "system" && m.Content.Contains("fenced JSON example does not perform", StringComparison.Ordinal));
+        Assert.Contains(messages, m => m.Role == "system" && m.Content.Contains("file_read (success): verified content", StringComparison.Ordinal));
         Assert.Equal("system", messages[0].Role);
         Assert.Equal("user", messages[^1].Role);
         Assert.Equal(["image"], messages[^1].Images);