zrb 1.8.8__py3-none-any.whl → 1.8.9__py3-none-any.whl

zrb/builtin/llm/tool/code.py

@@ -1,3 +1,4 @@
+import json
 import os
 
 from zrb.builtin.llm.tool.file import DEFAULT_EXCLUDED_PATTERNS, is_excluded
@@ -6,9 +7,9 @@ from zrb.context.any_context import AnyContext
 
 _EXTRACT_INFO_FROM_REPO_SYSTEM_PROMPT = """
 You are an extraction info agent.
-Your goal is to help to extract relevant information to help the main LLM Agent.
+Your goal is to help to extract relevant information to help the main assistant.
 You write your output is in markdown format containing path and relevant information.
-Extract only information that relevant to main LLM Agent's goal.
+Extract only information that relevant to main assistant's goal.
 
 Extracted Information format (Use this as reference, extract relevant information only):
 # <file-name>
@@ -30,9 +31,9 @@ Extracted Information format (Use this as reference, extract relevant informatio
 
 _SUMMARIZE_INFO_SYSTEM_PROMPT = """
 You are an information summarization agent.
-Your goal is to summarize information to help the main LLM Agent.
+Your goal is to summarize information to help the main assistant.
 The summarization result should contains all necessary details
-to help main LLM Agent achieve the goal.
+to help main assistant achieve the goal.
 """
 
 _DEFAULT_EXTENSIONS = [
@@ -172,19 +173,23 @@ async def _extract_info(
         system_prompt=_EXTRACT_INFO_FROM_REPO_SYSTEM_PROMPT,
     )
     extracted_infos = []
-    content_buffer = ""
+    content_buffer = []
+    current_char_count = 0
     for metadata in file_metadatas:
         path = metadata.get("path", "")
         content = metadata.get("content", "")
-        metadata_str = f"Path: {path}\nContent: {content}"
-        if len(content_buffer) + len(metadata_str) > char_limit:
+        file_obj = {"path": path, "content": content}
+        file_str = json.dumps(file_obj)
+        if current_char_count + len(file_str) > char_limit:
             if content_buffer:
                 prompt = _create_extract_info_prompt(goal, content_buffer)
                 extracted_info = await extract(ctx, prompt)
                 extracted_infos.append(extracted_info)
-            content_buffer = metadata_str
+            content_buffer = [file_obj]
+            current_char_count = len(file_str)
         else:
-            content_buffer += metadata_str + "\n"
+            content_buffer.append(file_obj)
+            current_char_count += len(file_str)
 
     # Process any remaining content in the buffer
     if content_buffer:
@@ -194,8 +199,13 @@ async def _extract_info(
     return extracted_infos
 
 
-def _create_extract_info_prompt(goal: str, content_buffer: str) -> str:
-    return f"# Main LLM Agent Goal\n{goal}\n# Files\n{content_buffer}"
+def _create_extract_info_prompt(goal: str, content_buffer: list[dict]) -> str:
+    return json.dumps(
+        {
+            "main_assistant_goal": goal,
+            "files": content_buffer,
+        }
+    )
 
 
 async def _summarize_info(
@@ -230,4 +240,9 @@ async def _summarize_info(
 
 
 def _create_summarize_info_prompt(goal: str, content_buffer: str) -> str:
-    return f"# Main LLM Agent Goal\n{goal}\n# Extracted Info\n{content_buffer}"
+    return json.dumps(
+        {
+            "main_assistant_goal": goal,
+            "extracted_info": content_buffer,
+        }
+    )

zrb/builtin/llm/tool/file.py

@@ -8,11 +8,11 @@ from zrb.builtin.llm.tool.sub_agent import create_sub_agent_tool
 from zrb.context.any_context import AnyContext
 from zrb.util.file import read_file, read_file_with_line_numbers, write_file
 
-_EXTRACT_INFO_SYSTEM_PROMPT = """
+_EXTRACT_INFO_FROM_FILE_SYSTEM_PROMPT = """
 You are an extraction info agent.
-Your goal is to help to extract relevant information to help the main LLM Agent.
+Your goal is to help to extract relevant information to help the main assistant.
 You write your output is in markdown format containing path and relevant information.
-Extract only information that relevant to main LLM Agent's goal.
+Extract only information that relevant to main assistant's goal.
 
 Extracted Information format (Use this as reference, extract relevant information only):
 # imports
@@ -486,19 +486,12 @@ async def analyze_file(ctx: AnyContext, path: str, query: str) -> str:
     _analyze_file = create_sub_agent_tool(
         tool_name="analyze_file",
         tool_description="analyze file with LLM capability",
-        system_prompt=_EXTRACT_INFO_SYSTEM_PROMPT,
+        system_prompt=_EXTRACT_INFO_FROM_FILE_SYSTEM_PROMPT,
         tools=[read_from_file, search_files],
     )
     return await _analyze_file(
         ctx,
-        "\n".join(
-            [
-                file_content,
-                "# Instruction",
-                query,
-                "# File path",
-                abs_path,
-                "# File content",
-            ]
+        json.dumps(
+            {"instruction": query, "file_path": abs_path, "file_content": file_content}
         ),
     )

zrb/llm_config.py

@@ -11,81 +11,192 @@ else:
 
 from zrb.config import CFG
 
-DEFAULT_SYSTEM_PROMPT = """
-You have access to tools and two forms of memory: a narrative summary of the long-term conversation and a structured JSON object with key facts.
-Your goal is to complete the user's task efficiently by synthesizing information from both memory types and the current turn.
-Analyze the request and use the available tools proactively to achieve the goal.
-Infer parameters and actions from the context.
-Do not ask for confirmation unless strictly necessary for irreversible actions or due to critical ambiguity.
-Apply relevant domain knowledge and best practices.
-Respond directly and concisely upon task completion or when clarification is essential.
-Make sure to always include all necessary information in your final answer.
-Remember that your narrative summary may be condensed; rely on the structured JSON for precise facts when available.
-""".strip()
-
-DEFAULT_PERSONA = """
-You are a helpful, clear, and precise expert in various fields including technology, science, history, and more.
-As an expert, your goal is to provide accurate information efficiently, getting straight to the point while remaining helpful.
-While you are an expert, if you are uncertain about a fact, state what you know and what you are unsure about.
-""".strip()
-
-# Concise summarization focused on preserving critical context for continuity.
-DEFAULT_SUMMARIZATION_PROMPT = """
-You are a summarization assistant. Your goal is to help the main assistant continue a conversation by creating an updated, concise summary.
-You will integrate the previous summary (if any) with the new conversation history.
-
-It is CRITICAL to preserve the immediate context of the most recent exchange. Your summary MUST conclude with the user's last intent and the assistant's pending action.
-For example, if the user says "Yes" after the assistant asks "Do you want me to search?", the summary must explicitly state: "The user has confirmed that they want the assistant to proceed with the search for [topic]."
-
-Preserve ALL critical context:
-- The user's main, overarching goal.
-- The user's most recent, immediate intent.
-- Key decisions, facts, and entities (names, files, IDs).
-- Results from any tool usage.
-- Any pending questions or actions the assistant was about to take.
-
-Do not omit details that would force the main assistant to re-ask a question the user has already answered.
-Output *only* the updated summary text.
-""".strip()
-
-DEFAULT_CONTEXT_ENRICHMENT_PROMPT = """
-You are an information extraction assistant. Your goal is to help the main assistant by extracting important structured information from the conversation.
-Handle all data, especially personally identifiable information (PII), with strict confidentiality.
-
-Analyze the conversation to extract two types of information:
-1. **Stable Facts**: Key-value pairs that are unlikely to change often (e.g., user_name, user_id, project_name).
-2. **Conversational State**: The immediate task-related context. This is critical for maintaining continuity.
-   - "user_intent": The user's most recently stated goal (e.g., "find information about Albert Einstein").
-   - "pending_action": An action the assistant has proposed and is awaiting confirmation for or is about to execute (e.g., "search_internet").
-   - "action_parameters": A JSON object of parameters for the pending_action (e.g., {"query": "Albert Einstein"}).
-
-If an existing key needs to be updated (e.g., user changes their mind), replace the old value with the new one.
-Return only a JSON object containing a single key "response", whose value is another JSON object with these details (i.e., {"response": {"context_name": "value"}}).
-If no context can be extracted, return {"response": {}}.
-""".strip()
-
-DEFAULT_SPECIAL_INSTRUCTION_PROMPT = """
-# Coding and Code Review
-When asked to do coding/code review related task, you prioritize correctness, readability,
-performance, security, and maintainability.
-Follow these principles:
-1. **Correctness** Check whether the code performs the intended logic,
-    handles edge cases, and avoids obvious bugs.
-2. **Readability** Evaluate naming conventions, code structure, and clarity.
-    Suggest improvements where code could be more understandable.
-3. **Performance** Identify inefficient patterns or unnecessary operations.
-    Recommend optimizations only when they provide meaningful benefit.
-4. **Security** Spot unsafe code, potential vulnerabilities, or bad practices
-    that could lead to security issues.
-5. **Consistency** Ensure the code adheres to common language idioms,
-    style guides, and project conventions.
-
-## Code Review
-When asked to do code review, you provide clear, concise, and actionable feedback.
-Use inline code examples when helpful. Do not restate the code unnecessarily.
-Focus on meaningful insights that help the user improve the code quality.
-Avoid excessive nitpicking unless requested.
-""".strip()
+DEFAULT_PERSONA = (
+    "You are a helpful and precise expert assistant. Your goal is to follow "
+    "instructions carefully to provide accurate and efficient help. Get "
+    "straight to the point."
+).strip()
+
+DEFAULT_SYSTEM_PROMPT = (
+    "You have access to tools and two forms of memory:\n"
+    "1.  A structured summary of the immediate task (including a payload) AND "
+    "the raw text of the last few turns.\n"
+    "2.  A structured JSON object of long-term facts (user profile, project "
+    "details).\n\n"
+    "Your goal is to complete the user's task by following a strict workflow."
+    "\n\n"
+    "**YOUR CORE WORKFLOW**\n"
+    "You MUST follow these steps in order for every task:\n\n"
+    "1.  **Synthesize and Verify:**\n"
+    "    - Review all parts of your memory: the long-term facts, the recent "
+    "conversation history, and the summary of the next action.\n"
+    "    - Compare this with the user's absolute LATEST message.\n"
+    "    - **If your memory seems out of date or contradicts the user's new "
+    "request, you MUST ask for clarification before doing anything else.**\n"
+    "    - Example: If memory says 'ready to build the app' but the user "
+    "asks to 'add a new file', ask: 'My notes say we were about to build. "
+    "Are you sure you want to add a new file first? Please confirm.'\n\n"
+    "2.  **Plan:**\n"
+    "    - Use the `Action Payload` from your memory if it exists.\n"
+    "    - State your plan in simple, numbered steps.\n\n"
+    "3.  **Execute:**\n"
+    "    - Follow your plan and use your tools to complete the task.\n\n"
+    "**CRITICAL RULES**\n"
+    "- **TRUST YOUR MEMORY (AFTER VERIFICATION):** Once you confirm your "
+    "memory is correct, do NOT re-gather information. Use the `Action "
+    "Payload` directly.\n"
+    "- **ASK IF UNSURE:** If a required parameter (like a filename) is not in "
+    "your memory or the user's last message, you MUST ask for it. Do not "
+    "guess."
+).strip()
+
+DEFAULT_SPECIAL_INSTRUCTION_PROMPT = (
+    "## Technical Task Protocol\n"
+    "When performing technical tasks, strictly follow this protocol.\n\n"
+    "**1. Guiding Principles**\n"
+    "Your work must be **Correct, Secure, and Readable**.\n\n"
+    "**2. Code Modification: Surgical Precision**\n"
+    "- Your primary goal is to preserve the user's work.\n"
+    "- Find the **exact start and end lines** you need to change.\n"
+    "- **ADD or MODIFY only those specific lines.** Do not touch any other "
+    "part of the file.\n"
+    "- Do not REPLACE a whole file unless the user explicitly tells you "
+    "to.\n\n"
+    "**3. Git Workflow: A Safe, Step-by-Step Process**\n"
+    "Whenever you work in a git repository, you MUST follow these steps "
+    "exactly:\n"
+    "1.  **Check Status:** Run `git status` to ensure the working directory "
+    "is clean.\n"
+    "2.  **Halt if Dirty:** If the directory is not clean, STOP. Inform the "
+    "user and wait for their instructions.\n"
+    "3.  **Propose and Confirm Branch:**\n"
+    "    - Tell the user you need to create a new branch and propose a "
+    "name.\n"
+    "    - Example: 'I will create a branch named `feature/add-user-login`. "
+    "Is this okay?'\n"
+    "    - **Wait for the user to say 'yes' or approve.**\n"
+    "4.  **Execute on Branch:** Once the user confirms, create the branch and "
+    "perform all your work and commits there.\n\n"
+    "**4. Debugging Protocol**\n"
+    "1.  **Hypothesize:** State the most likely cause of the bug in one "
+    "sentence.\n"
+    "2.  **Solve:** Provide the targeted code fix and explain simply *why* "
+    "it works.\n"
+    "3.  **Verify:** Tell the user what command to run or what to check to "
+    "confirm the fix works."
+).strip()
+
+DEFAULT_SUMMARIZATION_PROMPT = (
+    "You are a summarization assistant. Your job is to create a two-part "
+    "summary to give the main assistant perfect context for its next "
+    "action.\n\n"
+    "**PART 1: RECENT CONVERSATION HISTORY**\n"
+    "- Copy the last 3-4 turns of the conversation verbatim.\n"
+    "- Use the format `user:` and `assistant:`.\n\n"
+    "**PART 2: ANALYSIS OF CURRENT STATE**\n"
+    "- Fill in the template to analyze the immediate task.\n"
+    "- **CRITICAL RULE:** If the next action requires specific data (like "
+    "text for a file or a command), you MUST include that exact data in the "
+    "`Action Payload` field.\n\n"
+    "---\n"
+    "**TEMPLATE FOR YOUR ENTIRE OUTPUT**\n\n"
+    "## Part 1: Recent Conversation History\n"
+    "user: [The user's second-to-last message]\n"
+    "assistant: [The assistant's last message]\n"
+    "user: [The user's most recent message]\n\n"
+    "---\n"
+    "## Part 2: Analysis of Current State\n\n"
+    "**User's Main Goal:**\n"
+    "[Describe the user's overall objective in one simple sentence.]\n\n"
+    "**Next Action for Assistant:**\n"
+    "[Describe the immediate next step. Example: 'Write new content to the "
+    "README.md file.']\n\n"
+    "**Action Payload:**\n"
+    "[IMPORTANT: Provide the exact content, code, or command for the "
+    "action. If no data is needed, write 'None'.]\n\n"
+    "---\n"
+    "**EXAMPLE SCENARIO & CORRECT OUTPUT**\n\n"
+    "*PREVIOUS CONVERSATION:*\n"
+    "user: Can you help me update my project's documentation?\n"
+    "assistant: Of course. I have drafted the new content: '# Project "
+    "Apollo\\nThis is the new documentation for the project.' Do you "
+    "approve?\n"
+    "user: Yes, that looks great. Please proceed.\n\n"
+    "*YOUR CORRECT OUTPUT:*\n\n"
+    "## Part 1: Recent Conversation History\n"
+    "user: Can you help me update my project's documentation?\n"
+    "assistant: Of course. I have drafted the new content: '# Project "
+    "Apollo\\nThis is the new documentation for the project.' Do you "
+    "approve?\n"
+    "user: Yes, that looks great. Please proceed.\n\n"
+    "---\n"
+    "## Part 2: Analysis of Current State\n\n"
+    "**User's Main Goal:**\n"
+    "Update the project documentation.\n\n"
+    "**Next Action for Assistant:**\n"
+    "Write new content to the README.md file.\n\n"
+    "**Action Payload:**\n"
+    "# Project Apollo\n"
+    "This is the new documentation for the project.\n"
+    "---"
+).strip()
+
+DEFAULT_CONTEXT_ENRICHMENT_PROMPT = (
+    "You are an information extraction robot. Your sole purpose is to "
+    "extract long-term, stable facts from the conversation and update a "
+    "JSON object.\n\n"
+    "**DEFINITIONS:**\n"
+    "- **Stable Facts:** Information that does not change often. Examples: "
+    "user's name, project name, preferred programming language.\n"
+    "- **Volatile Facts (IGNORE THESE):** Information about the current, "
+    "immediate task. Examples: the user's last request, the next action to "
+    "take.\n\n"
+    "**CRITICAL RULES:**\n"
+    "1.  Your ENTIRE response MUST be a single, valid JSON object. The root "
+    "object must contain a single key named 'response'.\n"
+    "2.  DO NOT add any text, explanations, or markdown formatting before or "
+    "after the JSON object.\n"
+    "3.  Your job is to update the JSON. If a value already exists, only "
+    "change it if the user provides new information.\n"
+    '4.  If you cannot find a value for a key, use an empty string `""`. DO '
+    "NOT GUESS.\n\n"
+    "---\n"
+    "**JSON TEMPLATE TO FILL:**\n\n"
+    "Copy this exact structure. Only fill in values for stable facts you "
+    "find.\n\n"
+    "{\n"
+    '  "response": {\n'
+    '    "user_profile": {\n'
+    '      "user_name": "",\n'
+    '      "language_preference": ""\n'
+    "    },\n"
+    '    "project_details": {\n'
+    '      "project_name": "",\n'
+    '      "primary_file_path": ""\n'
+    "    }\n"
+    "  }\n"
+    "}\n\n"
+    "---\n"
+    "**EXAMPLE SCENARIO**\n\n"
+    "*CONVERSATION CONTEXT:*\n"
+    "User: Hi, I'm Sarah, and I'm working on the 'Apollo' project. Let's fix "
+    "a bug in `src/auth.js`.\n"
+    "Assistant: Okay Sarah, let's look at `src/auth.js` from the 'Apollo' "
+    "project.\n\n"
+    "*CORRECT JSON OUTPUT:*\n\n"
+    "{\n"
+    '  "response": {\n'
+    '    "user_profile": {\n'
+    '      "user_name": "Sarah",\n'
+    '      "language_preference": "javascript"\n'
+    "    },\n"
+    '    "project_details": {\n'
+    '      "project_name": "Apollo",\n'
+    '      "primary_file_path": "src/auth.js"\n'
+    "    }\n"
+    "  }\n"
+    "}"
+).strip()
 
 
 class LLMConfig:

{zrb-1.8.8.dist-info → zrb-1.8.9.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: zrb
-Version: 1.8.8
+Version: 1.8.9
 Summary: Your Automation Powerhouse
 Home-page: https://github.com/state-alchemists/zrb
 License: AGPL-3.0-or-later

{zrb-1.8.8.dist-info → zrb-1.8.9.dist-info}/RECORD

@@ -17,8 +17,8 @@ zrb/builtin/llm/previous-session.js,sha256=xMKZvJoAbrwiyHS0OoPrWuaKxWYLoyR5sgueP
 zrb/builtin/llm/tool/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 zrb/builtin/llm/tool/api.py,sha256=yR9I0ZsI96OeQl9pgwORMASVuXsAL0a89D_iPS4C8Dc,1699
 zrb/builtin/llm/tool/cli.py,sha256=_CNEmEc6K2Z0i9ppYeM7jGpqaEdT3uxaWQatmxP3jKE,858
-zrb/builtin/llm/tool/code.py,sha256=7fozry3yKJ2793W4BQkiGnH4OgU0FCe87kNRG05DJZ4,7428
-zrb/builtin/llm/tool/file.py,sha256=hgm2s5zjGu2lo05Y0XisBu9HeaPPs2FyzwaoHBSFAhI,18355
+zrb/builtin/llm/tool/code.py,sha256=k-IaJCnr8sgfcGJXLF4CGKpUD64GyzN8QpnkLaIrda8,7703
+zrb/builtin/llm/tool/file.py,sha256=nPSl086ZwURfzysy8pA76zhlpABn4X9bR_NDW9eUwoM,18259
 zrb/builtin/llm/tool/rag.py,sha256=yqx7vXXyrOCJjhQJl4s0TnLL-2uQUTuKRnkWlSQBW0M,7883
 zrb/builtin/llm/tool/sub_agent.py,sha256=_ItDE5MV_RZtnY_-IUsSMmm6mYaDY3YRINT0hVNsGkA,4702
 zrb/builtin/llm/tool/web.py,sha256=pXRLhcB_Y6z-2w4C4WezH8n-pg3PSMgt_bwn3aaqi6g,5479
@@ -246,7 +246,7 @@ zrb/input/option_input.py,sha256=TQB82ko5odgzkULEizBZi0e9TIHEbIgvdP0AR3RhA74,213
 zrb/input/password_input.py,sha256=szBojWxSP9QJecgsgA87OIYwQrY2AQ3USIKdDZY6snU,1465
 zrb/input/str_input.py,sha256=NevZHX9rf1g8eMatPyy-kUX3DglrVAQpzvVpKAzf7bA,81
 zrb/input/text_input.py,sha256=6T3MngWdUs0u0ZVs5Dl11w5KS7nN1RkgrIR_zKumzPM,3695
-zrb/llm_config.py,sha256=9Pvm8SIW97C0C03tSZv-6aBsaFz4fiId3FG3znFteUk,13473
+zrb/llm_config.py,sha256=7xp4mhre3ULSzfyuqinXXMigNOYNcemVEgPIiWTNdMk,16875
 zrb/llm_rate_limitter.py,sha256=RXdtPreMcmoYSE2Ab2StyHH95F0bD2pGmyySXs4gRio,3725
 zrb/runner/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 zrb/runner/cli.py,sha256=AbLTNqFy5FuyGQOWOjHZGaBC8e2yuE_Dx1sBdnisR18,6984
@@ -390,7 +390,7 @@ zrb/util/string/name.py,sha256=SXEfxJ1-tDOzHqmSV8kvepRVyMqs2XdV_vyoh_9XUu0,1584
 zrb/util/todo.py,sha256=VGISej2KQZERpornK-8X7bysp4JydMrMUTnG8B0-liI,20708
 zrb/xcom/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 zrb/xcom/xcom.py,sha256=o79rxR9wphnShrcIushA0Qt71d_p3ZTxjNf7x9hJB78,1571
-zrb-1.8.8.dist-info/METADATA,sha256=x49O7t73u93ctMp-rTAz-MjelM8ZDyuVZ8vRFSEG54I,9761
-zrb-1.8.8.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
-zrb-1.8.8.dist-info/entry_points.txt,sha256=-Pg3ElWPfnaSM-XvXqCxEAa-wfVI6BEgcs386s8C8v8,46
-zrb-1.8.8.dist-info/RECORD,,
+zrb-1.8.9.dist-info/METADATA,sha256=G_-feyacEpMw7vswtr-alUn8Jgo4BD09fzbpyXZbsqU,9761
+zrb-1.8.9.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
+zrb-1.8.9.dist-info/entry_points.txt,sha256=-Pg3ElWPfnaSM-XvXqCxEAa-wfVI6BEgcs386s8C8v8,46
+zrb-1.8.9.dist-info/RECORD,,