fix (tasks): planner fixes

Author: Kabeer2004

src/client/components/tasks/TaskDetailsContent.js

@@ -396,6 +396,20 @@ const TaskDetailsContent = ({
 				<SwarmDetailsSection swarmDetails={displayTask.swarm_details} />
 			)}
 
+			{/* --- Researched Context (if planning) --- */}
+			{displayTask.status === "planning" && displayTask.found_context && (
+				<div>
+					<label className="text-sm font-medium text-neutral-400 block mb-2">
+						Researched Context
+					</label>
+					<div className="bg-neutral-800/50 p-3 rounded-lg text-sm text-neutral-300 whitespace-pre-wrap border border-neutral-700/50">
+						<ReactMarkdown className="prose prose-sm prose-invert">
+							{displayTask.found_context}
+						</ReactMarkdown>
+					</div>
+				</div>
+			)}
+
 			{/* --- META INFO & ASSIGNEE --- */}
 			<div className="grid grid-cols-2 gap-6">
 				<div>

src/server/workers/planner/prompts.py

@@ -10,14 +10,16 @@
 {retrieved_context}
 
 **Core Directives:**
-1.  **CRITICAL: USE THE RETRIEVED CONTEXT:** The "Retrieved Context" section contains vital information gathered by a research agent. You **MUST** use this information (e.g., email addresses, document contents, specific details) to create your plan. Do not ask for information that is already provided in the context.
-2.  **CRITICAL: HANDLE CHANGE REQUESTS:** If the context includes `chat_history` and `previous_result`, you are modifying a previous task. Your new plan **MUST** use information from `previous_result` (like a `document_id` or `url`) to **MODIFY** the existing entity. **DO NOT** create a new one unless explicitly asked. The user's latest message in `chat_history` is your primary instruction.
-3.  **Use Memory for Personalization:** If the user's request is personal or lacks details not found in the retrieved context or previous results (e.g., "email my manager"), your plan's **FIRST STEP** MUST be to call the `memory` tool to retrieve the necessary context.
-4.  **Analyze the Goal:** After checking context and memory, deeply understand the user's objective.
-5.  **Think Step-by-Step:** Deconstruct the goal into a logical sequence of steps.
-6.  **Be Resourceful:** Use the provided list of tools creatively. A single action item might require multiple tool calls.
-7.  **Anticipate Information Gaps:** If crucial information is still missing after checking all context, the first step should be to use a tool to find it (e.g., `internet_search` for public information, `gpeople` for contacts).
-8.  **Output a Clear Plan:** Your final output must be a single, valid JSON object containing a concise description of the overall goal and a list of specific, actionable steps for the executor.
+1.  **CRITICAL: USE THE RETRIEVED CONTEXT:** The "Retrieved Context" section contains vital information gathered by a research agent. You **MUST** use this information to inform your plan, but **DO NOT** copy the context directly into your plan steps.
+2.  **Executor Has Context:** The executor agent who will follow your plan will also receive the full 'Retrieved Context'. Therefore, your plan steps should be high-level instructions (e.g., 'Summarize the key findings from the context') rather than copy-pasting the context itself into the plan.
+3.  **Decompose the Goal:** Break down complex goals into smaller, sequential steps. For example, instead of one step 'Create a document with sections A, B, and C', create separate steps: 'Create the document', 'Add section A to the document', 'Add section B to the document', and 'Add section C to the document'.
+4.  **CRITICAL: HANDLE CHANGE REQUESTS:** If the context includes `chat_history` and `previous_result`, you are modifying a previous task. Your new plan **MUST** use information from `previous_result` (like a `document_id` or `url`) to **MODIFY** the existing entity. **DO NOT** create a new one unless explicitly asked. The user's latest message in `chat_history` is your primary instruction.
+5.  **Use Memory for Personalization:** If the user's request is personal or lacks details not found in the retrieved context or previous results (e.g., "email my manager"), your plan's **FIRST STEP** MUST be to call the `memory` tool to retrieve the necessary context.
+6.  **Analyze the Goal:** After checking context and memory, deeply understand the user's objective.
+7.  **Think Step-by-Step:** Deconstruct the goal into a logical sequence of steps.
+8.  **Be Resourceful:** Use the provided list of tools creatively. A single action item might require multiple tool calls.
+9.  **Anticipate Information Gaps:** If crucial information is still missing after checking all context, the first step should be to use a tool to find it (e.g., `internet_search` for public information, `gpeople` for contacts).
+10. **Output a Clear Plan:** Your final output must be a single, valid JSON object containing a concise description of the overall goal and a list of specific, actionable steps for the executor.
 
 **Tool Usage Guidelines:**
 -   `internet_search`: Use ONLY for searching for public, factual information on the web. DO NOT use it to find personal information like calendars or contacts.