zrb 1.21.17__py3-none-any.whl → 1.21.33__py3-none-any.whl

zrb/task/llm/history_processor.py

@@ -0,0 +1,206 @@
+import json
+import sys
+import traceback
+from typing import TYPE_CHECKING, Any, Callable, Coroutine
+
+from zrb.config.llm_config import llm_config
+from zrb.config.llm_rate_limitter import LLMRateLimitter
+from zrb.config.llm_rate_limitter import llm_rate_limitter as default_llm_rate_limitter
+from zrb.context.any_context import AnyContext
+from zrb.task.llm.agent_runner import run_agent_iteration
+from zrb.util.cli.style import stylize_faint
+from zrb.util.markdown import make_markdown_section
+
+if sys.version_info >= (3, 12):
+    from typing import TypedDict
+else:
+    from typing_extensions import TypedDict
+
+
+if TYPE_CHECKING:
+    from pydantic_ai import ModelMessage
+    from pydantic_ai.models import Model
+    from pydantic_ai.settings import ModelSettings
+
+
+class SingleMessage(TypedDict):
+    """
+    SingleConversation
+
+    Attributes:
+        role: Either AI, User, Tool Call, or Tool Result
+        time: yyyy-mm-ddTHH:MM:SSZ:
+        content: The content of the message (summarize if too long)
+    """
+
+    role: str
+    time: str
+    content: str
+
+
+class ConversationSummary(TypedDict):
+    """
+    Conversation history
+
+    Attributes:
+        transcript: Several last transcript of the conversation
+        summary: Descriptive conversation summary
+    """
+
+    transcript: list[SingleMessage]
+    summary: str
+
+
+def save_conversation_summary(conversation_summary: ConversationSummary):
+    """
+    Write conversation summary for main assistant to continue conversation.
+    """
+    return conversation_summary
+
+
+def create_summarize_history_processor(
+    ctx: AnyContext,
+    system_prompt: str,
+    rate_limitter: LLMRateLimitter | None = None,
+    summarization_model: "Model | str | None" = None,
+    summarization_model_settings: "ModelSettings | None" = None,
+    summarization_system_prompt: str | None = None,
+    summarization_token_threshold: int | None = None,
+    summarization_retries: int = 2,
+) -> Callable[[list["ModelMessage"]], Coroutine[None, None, list["ModelMessage"]]]:
+    from pydantic_ai import Agent, ModelMessage, ModelRequest
+    from pydantic_ai.messages import ModelMessagesTypeAdapter, UserPromptPart
+
+    if rate_limitter is None:
+        rate_limitter = default_llm_rate_limitter
+    if summarization_model is None:
+        summarization_model = llm_config.default_small_model
+    if summarization_model_settings is None:
+        summarization_model_settings = llm_config.default_small_model_settings
+    if summarization_system_prompt is None:
+        summarization_system_prompt = llm_config.default_summarization_prompt
+    if summarization_token_threshold is None:
+        summarization_token_threshold = (
+            llm_config.default_history_summarization_token_threshold
+        )
+
+    async def maybe_summarize_history(
+        messages: list[ModelMessage],
+    ) -> list[ModelMessage]:
+        history_list = json.loads(ModelMessagesTypeAdapter.dump_json(messages))
+        history_json_str = json.dumps(history_list)
+        # Estimate token usage
+        # Note: Pydantic ai has run context parameter
+        # (https://ai.pydantic.dev/message-history/#runcontext-parameter)
+        # But we cannot use run_ctx.usage.total_tokens because total token keep increasing
+        # even after summariztion.
+        estimated_token_usage = rate_limitter.count_token(history_json_str)
+        _print_request_info(
+            ctx, estimated_token_usage, summarization_token_threshold, messages
+        )
+        if estimated_token_usage < summarization_token_threshold or len(messages) == 1:
+            return messages
+        history_list_without_instruction = [
+            {
+                key: obj[key]
+                for key in obj
+                if index == len(history_list) - 1 or key != "instructions"
+            }
+            for index, obj in enumerate(history_list)
+        ]
+        history_json_str_without_instruction = json.dumps(
+            history_list_without_instruction
+        )
+        summarization_message = f"Summarize the following conversation: {history_json_str_without_instruction}"
+        summarization_agent = Agent[None, ConversationSummary](
+            model=summarization_model,
+            output_type=save_conversation_summary,
+            instructions=summarization_system_prompt,
+            model_settings=summarization_model_settings,
+            retries=summarization_retries,
+        )
+        try:
+            _print_info(ctx, "📝 Rollup Conversation", 2)
+            summary_run = await run_agent_iteration(
+                ctx=ctx,
+                agent=summarization_agent,
+                user_prompt=summarization_message,
+                attachments=[],
+                history_list=[],
+                rate_limitter=rate_limitter,
+                log_indent_level=2,
+            )
+            if summary_run and summary_run.result and summary_run.result.output:
+                usage = summary_run.result.usage()
+                _print_info(ctx, f"📝 Rollup Conversation Token: {usage}", 2)
+                ctx.print(plain=True)
+                ctx.log_info("History summarized and updated.")
+                condensed_message = make_markdown_section(
+                    header="Past Conversation",
+                    content="\n".join(
+                        [
+                            make_markdown_section(
+                                "Summary", _extract_summary(summary_run.result.output)
+                            ),
+                            make_markdown_section(
+                                "Past Trancript",
+                                _extract_transcript(summary_run.result.output),
+                            ),
+                        ]
+                    ),
+                )
+                return [
+                    ModelRequest(
+                        instructions=system_prompt,
+                        parts=[UserPromptPart(condensed_message)],
+                    )
+                ]
+            ctx.log_warning("History summarization failed or returned no data.")
+        except BaseException as e:
+            ctx.log_warning(f"Error during history summarization: {e}")
+            traceback.print_exc()
+        return messages
+
+    return maybe_summarize_history
+
+
+def _print_request_info(
+    ctx: AnyContext,
+    estimated_token_usage: int,
+    summarization_token_threshold: int,
+    messages: list["ModelMessage"],
+):
+    _print_info(ctx, f"Current request token (estimated): {estimated_token_usage}")
+    _print_info(ctx, f"Summarization token threshold: {summarization_token_threshold}")
+    _print_info(ctx, f"History length: {len(messages)}")
+
+
+def _print_info(ctx: AnyContext, text: str, log_indent_level: int = 0):
+    log_prefix = (2 * (log_indent_level + 1)) * " "
+    ctx.print(stylize_faint(f"{log_prefix}{text}"), plain=True)
+
+
+def _extract_summary(summary_result_output: dict[str, Any] | str) -> str:
+    summary = (
+        summary_result_output.get("summary", "")
+        if isinstance(summary_result_output, dict)
+        else ""
+    )
+    return summary
+
+
+def _extract_transcript(summary_result_output: dict[str, Any] | str) -> str:
+    transcript_list = (
+        summary_result_output.get("transcript", [])
+        if isinstance(summary_result_output, dict)
+        else []
+    )
+    transcript_list = [] if not isinstance(transcript_list, list) else transcript_list
+    return "\n".join(_format_transcript_message(message) for message in transcript_list)
+
+
+def _format_transcript_message(message: dict[str, str]) -> str:
+    role = message.get("role", "Message")
+    time = message.get("time", "<unknown>")
+    content = message.get("content", "<empty>")
+    return f"{role} ({time}): {content}"

zrb/task/llm/history_summarization.py

@@ -1,36 +1,7 @@
-import json
-import traceback
-from typing import TYPE_CHECKING
-
-from zrb.attr.type import BoolAttr, IntAttr
+from zrb.attr.type import IntAttr
 from zrb.config.llm_config import llm_config
-from zrb.config.llm_rate_limitter import LLMRateLimiter, llm_rate_limitter
 from zrb.context.any_context import AnyContext
-from zrb.task.llm.agent import run_agent_iteration
-from zrb.task.llm.conversation_history import (
-    count_part_in_history_list,
-    inject_conversation_history_notes,
-    replace_system_prompt_in_history,
-)
-from zrb.task.llm.conversation_history_model import ConversationHistory
-from zrb.task.llm.history_summarization_tool import (
-    create_history_summarization_tool,
-)
-from zrb.task.llm.typing import ListOfDict
-from zrb.util.attr import get_bool_attr, get_int_attr
-from zrb.util.cli.style import stylize_faint
-from zrb.util.markdown import make_markdown_section
-from zrb.util.truncate import truncate_str
-
-if TYPE_CHECKING:
-    from pydantic_ai.models import Model
-    from pydantic_ai.settings import ModelSettings
-
-
-def _count_token_in_history(history_list: ListOfDict) -> int:
-    """Counts the total number of tokens in a conversation history list."""
-    text_to_count = json.dumps(history_list)
-    return llm_rate_limitter.count_token(text_to_count)
+from zrb.util.attr import get_int_attr
 
 
 def get_history_summarization_token_threshold(
@@ -52,151 +23,3 @@ def get_history_summarization_token_threshold(
             "Defaulting to -1 (no threshold)."
         )
         return -1
-
-
-def should_summarize_history(
-    ctx: AnyContext,
-    history_list: ListOfDict,
-    should_summarize_history_attr: BoolAttr | None,
-    render_summarize_history: bool,
-    history_summarization_token_threshold_attr: IntAttr | None,
-    render_history_summarization_token_threshold: bool,
-) -> bool:
-    """Determines if history summarization should occur based on token length and config."""
-    history_part_count = count_part_in_history_list(history_list)
-    if history_part_count == 0:
-        return False
-    summarization_token_threshold = get_history_summarization_token_threshold(
-        ctx,
-        history_summarization_token_threshold_attr,
-        render_history_summarization_token_threshold,
-    )
-    history_token_count = _count_token_in_history(history_list)
-    if (
-        summarization_token_threshold == -1
-        or summarization_token_threshold > history_token_count
-    ):
-        return False
-    return get_bool_attr(
-        ctx,
-        should_summarize_history_attr,
-        llm_config.default_summarize_history,
-        auto_render=render_summarize_history,
-    )
-
-
-async def summarize_history(
-    ctx: AnyContext,
-    model: "Model | str | None",
-    settings: "ModelSettings | None",
-    system_prompt: str,
-    conversation_history: ConversationHistory,
-    rate_limitter: LLMRateLimiter | None = None,
-    retries: int = 3,
-) -> ConversationHistory:
-    """Runs an LLM call to update the conversation summary."""
-    from pydantic_ai import Agent
-
-    inject_conversation_history_notes(conversation_history)
-    ctx.log_info("Attempting to summarize conversation history...")
-    # Construct the user prompt for the summarization agent
-    user_prompt = "\n".join(
-        [
-            make_markdown_section(
-                "Past Conversation",
-                "\n".join(
-                    [
-                        make_markdown_section(
-                            "Summary",
-                            conversation_history.past_conversation_summary,
-                            as_code=True,
-                        ),
-                        make_markdown_section(
-                            "Last Transcript",
-                            conversation_history.past_conversation_transcript,
-                            as_code=True,
-                        ),
-                    ]
-                ),
-            ),
-            make_markdown_section(
-                "Recent Conversation (JSON)",
-                json.dumps(truncate_str(conversation_history.history, 1000)),
-                as_code=True,
-            ),
-        ]
-    )
-    summarize = create_history_summarization_tool(conversation_history)
-    summarization_agent = Agent[None, str](
-        model=model,
-        output_type=summarize,
-        system_prompt=system_prompt,
-        model_settings=settings,
-        retries=retries,
-    )
-    try:
-        ctx.print(stylize_faint("  📝 Rollup Conversation"), plain=True)
-        summary_run = await run_agent_iteration(
-            ctx=ctx,
-            agent=summarization_agent,
-            user_prompt=user_prompt,
-            attachments=[],
-            history_list=[],
-            rate_limitter=rate_limitter,
-            log_indent_level=2,
-        )
-        if summary_run and summary_run.result and summary_run.result.output:
-            usage = summary_run.result.usage()
-            ctx.print(
-                stylize_faint(f"  📝 Rollup Conversation Token: {usage}"), plain=True
-            )
-            ctx.print(plain=True)
-            ctx.log_info("History summarized and updated.")
-        else:
-            ctx.log_warning("History summarization failed or returned no data.")
-    except BaseException as e:
-        ctx.log_warning(f"Error during history summarization: {e}")
-        traceback.print_exc()
-    # Return the original summary if summarization fails
-    return conversation_history
-
-
-async def maybe_summarize_history(
-    ctx: AnyContext,
-    conversation_history: ConversationHistory,
-    should_summarize_history_attr: BoolAttr | None,
-    render_summarize_history: bool,
-    history_summarization_token_threshold_attr: IntAttr | None,
-    render_history_summarization_token_threshold: bool,
-    model: "str | Model | None",
-    model_settings: "ModelSettings | None",
-    summarization_prompt: str,
-    rate_limitter: LLMRateLimiter | None = None,
-) -> ConversationHistory:
-    """Summarizes history and updates context if enabled and threshold met."""
-    shorten_history = replace_system_prompt_in_history(conversation_history.history)
-    if should_summarize_history(
-        ctx,
-        shorten_history,
-        should_summarize_history_attr,
-        render_summarize_history,
-        history_summarization_token_threshold_attr,
-        render_history_summarization_token_threshold,
-    ):
-        original_history = conversation_history.history
-        conversation_history.history = shorten_history
-        conversation_history = await summarize_history(
-            ctx=ctx,
-            model=model,
-            settings=model_settings,
-            system_prompt=summarization_prompt,
-            conversation_history=conversation_history,
-            rate_limitter=rate_limitter,
-        )
-        conversation_history.history = original_history
-        if (
-            conversation_history.past_conversation_summary != ""
-            and conversation_history.past_conversation_transcript != ""
-        ):
-            conversation_history.history = []
-    return conversation_history

zrb/task/llm/print_node.py

@@ -2,6 +2,7 @@ import json
 from collections.abc import Callable
 from typing import Any
 
+from zrb.config.config import CFG
 from zrb.util.cli.style import stylize_faint
 
 
@@ -104,12 +105,20 @@ async def print_node(
                         and event.tool_call_id
                     ):
                         call_id = event.tool_call_id
-                        result_content = event.result.content
-                        print_func(
-                            _format_content(
-                                f"{call_id} | {result_content}", log_indent_level
+                        if CFG.LLM_SHOW_TOOL_CALL_RESULT:
+                            result_content = event.result.content
+                            print_func(
+                                _format_content(
+                                    f"{call_id} | Return {result_content}",
+                                    log_indent_level,
+                                )
+                            )
+                        else:
+                            print_func(
+                                _format_content(
+                                    f"{call_id} | Executed", log_indent_level
+                                )
                             )
-                        )
         except UnexpectedModelBehavior as e:
             print_func("")  # ensure newline consistency
             print_func(

zrb/task/llm/prompt.py

@@ -115,11 +115,11 @@ def _construct_system_prompt(
                             ),
                         ),
                         make_markdown_section(
-                            "🧠 Long Term Note",
+                            "🧠 Long Term Note Content",
                             conversation_history.long_term_note,
                         ),
                         make_markdown_section(
-                            "📝 Contextual Note",
+                            "📝 Contextual Note Content",
                             conversation_history.contextual_note,
                         ),
                         make_markdown_section(
@@ -129,21 +129,6 @@ def _construct_system_prompt(
                     ]
                 ),
             ),
-            make_markdown_section(
-                "💬 PAST CONVERSATION",
-                "\n".join(
-                    [
-                        make_markdown_section(
-                            "Narrative Summary",
-                            conversation_history.past_conversation_summary,
-                        ),
-                        make_markdown_section(
-                            "Past Transcript",
-                            conversation_history.past_conversation_transcript,
-                        ),
-                    ]
-                ),
-            ),
         ]
     )
 

zrb/task/llm/subagent_conversation_history.py

@@ -0,0 +1,41 @@
+from zrb.context.any_context import AnyContext
+from zrb.task.llm.conversation_history_model import ConversationHistory
+from zrb.task.llm.typing import ListOfDict
+from zrb.xcom.xcom import Xcom
+
+
+def inject_subagent_conversation_history_into_ctx(
+    ctx: AnyContext, conversation_history: ConversationHistory
+):
+    subagent_messages_xcom = _get_global_subagent_history_xcom(ctx)
+    existing_subagent_history = subagent_messages_xcom.get({})
+    subagent_messages_xcom.set(
+        {**existing_subagent_history, **conversation_history.subagent_history}
+    )
+
+
+def extract_subagent_conversation_history_from_ctx(
+    ctx: AnyContext,
+) -> dict[str, ListOfDict]:
+    subagent_messsages_xcom = _get_global_subagent_history_xcom(ctx)
+    return subagent_messsages_xcom.get({})
+
+
+def get_ctx_subagent_history(ctx: AnyContext, subagent_name: str) -> ListOfDict:
+    subagent_history = extract_subagent_conversation_history_from_ctx(ctx)
+    return subagent_history.get(subagent_name, [])
+
+
+def set_ctx_subagent_history(ctx: AnyContext, subagent_name: str, messages: ListOfDict):
+    subagent_history = extract_subagent_conversation_history_from_ctx(ctx)
+    subagent_history[subagent_name] = messages
+    subagent_messages_xcom = _get_global_subagent_history_xcom(ctx)
+    subagent_messages_xcom.set(subagent_history)
+
+
+def _get_global_subagent_history_xcom(ctx: AnyContext) -> Xcom:
+    if "_global_subagents" not in ctx.xcom:
+        ctx.xcom["_global_subagents"] = Xcom([{}])
+    if not isinstance(ctx.xcom["_global_subagents"], Xcom):
+        raise ValueError("ctx.xcom._global_subagents must be an Xcom")
+    return ctx.xcom["_global_subagents"]

zrb/task/llm/tool_confirmation_completer.py

@@ -0,0 +1,41 @@
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from prompt_toolkit.completion import Completer
+
+
+def get_tool_confirmation_completer(
+    options: list[str], meta_dict: dict[str, str]
+) -> "Completer":
+    from prompt_toolkit.completion import Completer, Completion
+
+    class ToolConfirmationCompleter(Completer):
+        """Custom completer for tool confirmation that doesn't auto-complete partial words."""
+
+        def __init__(self, options, meta_dict):
+            self.options = options
+            self.meta_dict = meta_dict
+
+        def get_completions(self, document, complete_event):
+            text = document.text.strip()
+            # 1. Input is empty, OR
+            # 2. Input exactly matches the beginning of an option
+            if text == "":
+                # Show all options when nothing is typed
+                for option in self.options:
+                    yield Completion(
+                        option,
+                        start_position=0,
+                        display_meta=self.meta_dict.get(option, ""),
+                    )
+                return
+            # Only complete if text exactly matches the beginning of an option
+            for option in self.options:
+                if option.startswith(text):
+                    yield Completion(
+                        option,
+                        start_position=-len(text),
+                        display_meta=self.meta_dict.get(option, ""),
+                    )
+
+    return ToolConfirmationCompleter(options, meta_dict)

zrb/task/llm/tool_wrapper.py

@@ -11,11 +11,11 @@ from zrb.config.llm_rate_limitter import llm_rate_limitter
 from zrb.context.any_context import AnyContext
 from zrb.task.llm.error import ToolExecutionError
 from zrb.task.llm.file_replacement import edit_replacement, is_single_path_replacement
+from zrb.task.llm.tool_confirmation_completer import get_tool_confirmation_completer
 from zrb.util.callable import get_callable_name
 from zrb.util.cli.markdown import render_markdown
 from zrb.util.cli.style import (
     stylize_blue,
-    stylize_error,
     stylize_faint,
     stylize_green,
     stylize_yellow,
@@ -185,7 +185,7 @@ async def _handle_user_response(
             ]
         )
         ctx.print(complete_confirmation_message, plain=True)
-        user_response = await _read_line()
+        user_response = await _read_line(args, kwargs)
         ctx.print("", plain=True)
         new_kwargs, is_edited = _get_edited_kwargs(ctx, user_response, kwargs)
         if is_edited:
@@ -250,13 +250,7 @@ def _get_user_approval_and_reason(
     try:
         approved = True if approval_str.strip() == "" else to_boolean(approval_str)
         if not approved and reason == "":
-            ctx.print(
-                stylize_error(
-                    f"You must specify rejection reason (i.e., No, <why>) for {func_call_str}"  # noqa
-                ),
-                plain=True,
-            )
-            return None
+            reason = "User disapproving the tool execution"
         return approved, reason
     except Exception:
         return False, user_response
@@ -300,11 +294,21 @@ def _truncate_arg(arg: str, length: int = 19) -> str:
     return normalized_arg
 
 
-async def _read_line():
+async def _read_line(args: list[Any] | tuple[Any], kwargs: dict[str, Any]):
     from prompt_toolkit import PromptSession
 
+    options = ["yes", "no", "edit"]
+    meta_dict = {
+        "yes": "Approve the execution",
+        "no": "Disapprove the execution",
+        "edit": "Edit tool execution parameters",
+    }
+    for key in kwargs:
+        options.append(f"edit {key}")
+        meta_dict[f"edit {key}"] = f"Edit tool execution parameter: {key}"
+    completer = get_tool_confirmation_completer(options, meta_dict)
     reader = PromptSession()
-    return await reader.prompt_async()
+    return await reader.prompt_async(completer=completer)
 
 
 def _adjust_signature(wrapper: Callable, original_sig: inspect.Signature):