update tool call ouput summarizer to log extraction instead - identify top 5 log sections related to user objective

Author: sidwan02

packages/core/src/agents/build-test-agent.ts

@@ -180,7 +180,7 @@ You have access to these tools:
     directive: `## Output Format
 If you decide to choose a tool to further the task, your response must *ONLY* contain the tool choice in JSON format.
 
-The \`goal\` you provide is critical, as it will be passed to a separate, specialized agent that executes the command. This agent has **no access to our conversation history**. Therefore, your \`goal\` **must be a fully self-contained and actionable instruction**. It needs to include not just *what* to do, but also *why*, so the other agent understands the context.
+The \`goal\` you provide is critical, as it will be passed to a separate, specialized agent that executes the command. This agent has **no access to our conversation history**. Therefore, your \`goal\` **must be a fully self-contained and actionable instruction**; it must not refer to previous steps (e.g., it should not say 'specified file' or 'mentioned argument' instead those should all be directly stated). It needs to include not just *what* to do, but also *why*, so the other agent understands the context.
 
 **JSON Structure:**
 \`\`\`json
@@ -214,7 +214,7 @@ You have access to these tools:
 ## Output Format
 If you decide to choose a tool to further the task, your response must *ONLY* contain the tool choice in JSON format.
 
-The \`goal\` you provide is critical, as it will be passed to a separate, specialized agent that executes the command. This agent has **no access to our conversation history**. Therefore, your \`goal\` **must be a fully self-contained and actionable instruction**. It needs to include not just *what* to do, but also *why*, so the other agent understands the context.
+The \`goal\` you provide is critical, as it will be passed to a separate, specialized agent that executes the command. This agent has **no access to our conversation history**. Therefore, your \`goal\` **must be a fully self-contained and actionable instruction**; it must not refer to previous steps, instead those should all be directly included in the goal. It needs to include not just *what* to do, but also *why*, so the other agent understands the context.
 
 
 **JSON Structure:**

packages/core/src/services/summarizer.ts

@@ -11,26 +11,35 @@ import { StreamEventType } from '../core/geminiChat.js';
 import { debugLogger } from '../utils/debugLogger.js';
 import * as fs from 'node:fs/promises';
 
-const SUMMARIZER_SYSTEM_PROMPT = `## Role
-You are an expert Tool Call Output Summarizer.
+const SUMMARIZER_SYSTEM_PROMPT = `## Role 
+You are a Technical Log Extraction Specialist.
 
 ## Task Definition
-Your task is to summarize a given tool call output. The summary must specifically highlight how the tool call output contributes to or addresses the user's overall objective.
-
-## Instruction
-When generating the summary, focus solely on the information within the \`toolcall\` that directly addresses or fulfills aspects of the \`objective\`. **The summary must be as exhaustive as possible, capturing every relevant detail from the tool call, and must be presented as a bulleted list.** Do not include extraneous details or interpret the tool call beyond its direct relevance to the objective.
-
-## Response Format Constraints
-The response **must be a bulleted list** containing **up to 20 points** (maximum). Each bullet point should be concise. DO NOT include any explicit explanations, reasoning, or thinking process outside of the bullet points. Your output must ONLY be the bulleted summary.`;
+Your task is to identify and extract the **top 5 most relevant sections** from the tool call output that relate to the user's \`<objective>\`.
+
+## Instructions
+1. **Prioritize Signal over Noise:** Search the logs for the 5 sections that provide the most conclusive evidence regarding the objective. 
+   - **Priority 1:** Terminal success indicators for the target (e.g., lines with \`✓\`, \`PASS\`, or \`Tests: X passed\`).
+   - **Priority 2:** Detailed error messages or stack traces if the target failed (e.g., \`✕\`, \`AssertionError\`).
+   - **Priority 3:** Final process summaries or exit codes.
+   - **Lowest Priority:** Repetitive "No test files found" or "queued" messages.
+2. **Direct Extraction:** Provide these 5 sections as **separate code blocks**. 
+3. **Preserve Verbatim Text:** Do not summarize, edit, or interpret the text within the blocks. Copy them exactly as they appear in the raw log.
+4. **Context Labels:** Before each code block, add a brief one-line label identifying the context (e.g., \`[Context: packages/core]\`).
+
+## Constraints
+- Output exactly 5 blocks (or fewer if the total log is very short).
+- Do not add analysis or conclusions.
+- Output ONLY the labeled code blocks.
+`;
 
 const SUMMARIZER_USER_PROMPT = `## Input
-User's Objective: {{objective}}
-Tool Call Output: {{toolcall}}
-
-## Output Reminder
-Take a deep breath, read the instructions again, read the inputs again. Each instruction is crucial and must be executed with utmost care and attention to detail.
+User's Objective: 
+{{objective}}
 
-Summary:`;
+Tool Call Output: 
+{{toolcall}}
+`;
 
 export class SummarizationService {
   constructor() {}

understand.md

@@ -250,13 +250,14 @@ Using the codebase investigator: tell me how the ollama inference works.
 
 <!--  -->
 
-1. chat_history maybe shouldn't have the model's tool choice. actually this
-   could be fine.
-1. Tool call service should take in the full chat history since if the model's
-   tool choice is provided then the last mesage is just the tool choice action
-   by the main subagent loop.
+1. If the tool call service makes a bad tool call that should be corrected by
+   the tool call service itself, not be seen by the subagent.
 1. When tool service makes a tool call, that should be reflected in the UI. In
    general, subagent services should be reflected in the UI.
+1. If there is a cancel in the subagent right now the summarizer gets something
+   like this
+   `Tool Call Output: [{"text":"{\"id\":\"283a387e-a136-4443-b3cd-1267b58cb5db########0-build_and_test_agent-mchfa2#2-0\",\"name\":\"run_shell_command\",\"response\":{\"error\":\"[Operation Cancelled] User cancelled the operation.\"}}"}]`
+   and keeps executing, it should be interrupted properly and stop inference.
 
 <!-- For routing -->