refactor: Remove reasoning field from completion logic and update prompts for agentic execution

Author: yaonyan

packages/core/src/executors/sampling/agentic-sampling-executor.ts

@@ -108,13 +108,12 @@ export class SamplingExecutor extends BaseSamplingExecutor {
     const toolCallData = parsedData;
 
     if (toolCallData.decision === "complete") {
-      const reasoning = (toolCallData.reasoning as string) || "Task completed";
-      return this.createCompletionResult(reasoning);
+      return this.createCompletionResult("Task completed");
     }
 
     try {
-      // Extract tool arguments (everything except action and reasoning)
-      const { action: _action, reasoning: _reasoning } = toolCallData;
+      // Extract tool arguments (everything except action and decision)
+      const { action: _action, decision: _decision, ..._toolArgs } = toolCallData;
 
       const toolResult = await this.agenticExecutor.execute(
         toolCallData,
@@ -178,8 +177,7 @@ export class SamplingExecutor extends BaseSamplingExecutor {
 I will now use agentic sampling to complete the following task: "${userRequest}"
 
 When I need to use a tool, I should specify the tool name in 'action' and provide tool-specific parameters as additional properties.
-When the task is complete, I should use "action": "complete".
-I must always provide reasoning for my actions.`;
+When the task is complete, I should use "action": "complete".`;
 
     // Use JSON instruction injection pattern
     return this.injectJsonInstruction({

packages/core/src/executors/sampling/workflow-sampling-executor.ts

@@ -83,8 +83,7 @@ export class WorkflowSamplingExecutor extends BaseSamplingExecutor {
     const toolCallData = parsedData as Record<string, unknown>;
 
     if (toolCallData.decision === "complete") {
-      const reasoning = (toolCallData.reasoning as string) || "Task completed";
-      return this.createCompletionResult(reasoning);
+      return this.createCompletionResult("Task completed");
     }
 
     try {

packages/core/src/factories/args-def-factory.ts

@@ -150,7 +150,7 @@ Workflow step definitions - provide ONLY on initial call.
 
     forAgentic: function (
       toolNameToDetailList: [string, unknown][],
-      sampling: boolean = false,
+      _sampling: boolean = false,
       ACTION_KEY: string = "action",
       NEXT_ACTION_KEY: string = "nextAction",
     ): JSONSchema {
@@ -187,14 +187,6 @@ Workflow step definitions - provide ONLY on initial call.
         ...depGroups,
       };
 
-      // Add reasoning field for sampling mode
-      if (sampling) {
-        baseProperties.reasoning = {
-          type: "string",
-          description: "Explain your reasoning for this action",
-        };
-      }
-
       const requiredFields = [ACTION_KEY, "decision"];
 
       return {

packages/core/src/prompts/index.ts

@@ -60,7 +60,6 @@ export const SystemPrompts = {
 - **ADVANCE STEP**: Set \`decision: "proceed"\` to move to next step  
 - **RETRY STEP**: Set \`decision: "retry"\`
 - **COMPLETE WORKFLOW**: Set \`decision: "complete"\` when ready to finish
-- Use \`reasoning\` action for thinking/analysis
 
 **🚫 Do NOT include \`steps\` parameter during normal execution**
 **✅ Include \`steps\` parameter ONLY when restarting workflow with \`init: true\`**
@@ -84,15 +83,13 @@ export const SystemPrompts = {
 
 **⚡ AGENTIC EXECUTION RULES:**
 - Each response demonstrates autonomous reasoning and decision-making
-- Use "reasoning" field to show your independent thought process
 - Make self-directed choices about which tools to use and when
 - Adapt your approach based on previous results without external guidance
 - Use "decision: complete" when you autonomously determine the task is finished
 
 **🔄 JSON Response Format (Agentic Decision Output):**
 You MUST respond with a JSON object that demonstrates your autonomous decision:
 - action: Your self-selected tool name
-- reasoning: Your independent reasoning and decision-making process
 - decision: "proceed" (continue execution), "retry" (retry current step), or "complete" (finish task)
 - [tool parameters]: Tool-specific parameters you autonomously determine
 
@@ -102,7 +99,6 @@ You MUST respond with a JSON object that demonstrates your autonomous decision:
 **🎯 AGENTIC CONSTRAINTS:**
 - Response must be pure JSON demonstrating autonomous decision-making
 - Invalid JSON indicates failure in agentic reasoning
-- Missing "reasoning" field shows lack of autonomous thought process
 - Tool parameters must reflect your independent analysis and planning`,
 
   /**
@@ -123,7 +119,6 @@ You MUST respond with a JSON object that demonstrates your autonomous decision:
 
 **⚡ AGENTIC EXECUTION RULES (Subsequent Calls):**
 - Each response demonstrates autonomous reasoning and decision-making within workflow context
-- Use "reasoning" field to show your independent thought process for current workflow step
 - Make self-directed choices about step execution, retry, or advancement
 - Adapt your approach based on previous step results without external guidance
 - Balance workflow structure with autonomous flexibility
@@ -133,21 +128,18 @@ You MUST respond with a JSON object for workflow execution:
 
 **For Workflow Initialization (First Call):**
 - action: "{toolName}"
-- reasoning: Your autonomous workflow planning process
 - init: true
 - steps: Autonomously designed workflow steps array
 - [other workflow parameters]: As you autonomously determine
 
 **For Step Execution (Subsequent Calls):**
 - action: "{toolName}"
-- reasoning: Your independent analysis of current step and next decision
 - decision: "proceed" (advance), "retry" (retry), or "complete" (finish - sampling mode only)
 - [step parameters]: Tool-specific parameters you autonomously determine for current step
 
 **🎯 AGENTIC WORKFLOW CONSTRAINTS:**
 - Response must be pure JSON demonstrating autonomous decision-making within workflow structure
 - Invalid JSON indicates failure in agentic workflow reasoning
-- Missing "reasoning" field shows lack of autonomous thought process
 - Tool parameters must reflect your independent analysis and workflow planning
 - Balance autonomous decision-making with structured workflow progression