pygpt-net 2.6.59__py3-none-any.whl → 2.6.60__py3-none-any.whl

pygpt_net/core/agents/custom/runner.py

@@ -0,0 +1,454 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# ================================================== #
+# This file is a part of PYGPT package               #
+# Website: https://pygpt.net                         #
+# GitHub:  https://github.com/szczyglis-dev/py-gpt   #
+# MIT License                                        #
+# Created By  : Marcin Szczygliński                  #
+# Updated Date: 2025.09.24 23:00:00                  #
+# ================================================== #
+
+from __future__ import annotations
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+from time import perf_counter
+
+from agents import Runner, TResponseInputItem
+from pygpt_net.core.agents.bridge import ConnectionContext
+from pygpt_net.item.ctx import CtxItem
+from pygpt_net.item.model import ModelItem
+from pygpt_net.item.preset import PresetItem
+from pygpt_net.provider.api.openai.agents.response import StreamHandler
+
+from .logging import Logger, NullLogger
+from .schema import FlowSchema, AgentNode, parse_schema
+from .graph import FlowGraph, build_graph
+from .memory import MemoryManager
+from .factory import AgentFactory
+from .router import parse_route_output
+from .utils import (
+    sanitize_input_items,
+    extract_text_output,
+    patch_last_assistant_output,
+    OptionGetter,
+    resolve_node_runtime,
+)
+from .router_streamer import DelayedRouterStreamer, RealtimeRouterStreamer
+from .debug import items_preview, ellipsize
+
+
+@dataclass
+class FlowResult:
+    ctx: CtxItem
+    final_output: str
+    last_response_id: Optional[str]
+
+
+@dataclass
+class DebugConfig:
+    log_runtime: bool = True
+    log_routes: bool = True
+    log_inputs: bool = False
+    log_outputs: bool = False
+    preview_chars: int = 280
+
+
+class FlowOrchestrator:
+    """
+    Orchestrates dynamic multi-agent flow based on NodeEditor schema.
+    UI semantics follow "bot-to-bot" and supports router stream modes: off/delayed/realtime.
+    Memory/no-memory input policy:
+      - First agent (in the whole flow) gets full initial messages from the app.
+      - Next agent WITHOUT memory gets only last step's displayed content as a single 'user' message.
+      - Agent WITH memory:
+          * if memory has items -> use them;
+          * if memory empty -> seed input from last displayed content (or initial messages as fallback).
+    """
+    def __init__(self, window, logger: Optional[Logger] = None) -> None:
+        self.window = window
+        self.logger = logger or NullLogger()
+
+    async def run_flow(
+        self,
+        schema: List[Dict[str, Any]],
+        messages: List[TResponseInputItem],
+        ctx: CtxItem,
+        bridge: ConnectionContext,
+        agent_kwargs: Dict[str, Any],
+        preset: Optional[PresetItem],
+        model: ModelItem,
+        stream: bool,
+        use_partial_ctx: bool,
+        base_prompt: Optional[str],
+        allow_local_tools_default: bool,
+        allow_remote_tools_default: bool,
+        function_tools: List[dict],
+        trace_id: Optional[str],
+        max_iterations: int = 20,
+        router_stream_mode: str = "off",  # "off" | "delayed" | "realtime"
+        option_get: Optional[OptionGetter] = None,
+    ) -> FlowResult:
+        fs: FlowSchema = parse_schema(schema)
+        g: FlowGraph = build_graph(fs)
+        mem = MemoryManager()
+        factory = AgentFactory(self.window, self.logger)
+        option_get = option_get or (lambda s, k, d=None: d)
+
+        # Debug config
+        dbg = DebugConfig(
+            log_runtime=bool(option_get("debug", "log_runtime", True)),
+            log_routes=bool(option_get("debug", "log_routes", True)),
+            log_inputs=bool(option_get("debug", "log_inputs", True)),
+            log_outputs=bool(option_get("debug", "log_outputs", True)),
+            preview_chars=int(option_get("debug", "preview_chars", 280)),
+        )
+
+        # Entry
+        if g.start_targets:
+            current_ids: List[str] = [g.start_targets[0]]
+            self.logger.info(f"Using explicit START -> {current_ids[0]}")
+        else:
+            default_agent = g.pick_default_start_agent()
+            if default_agent is None:
+                self.logger.error("No START and no agents in schema.")
+                return FlowResult(ctx=ctx, final_output="", last_response_id=None)
+            current_ids = [default_agent]
+            self.logger.info(f"No START found, using lowest-id agent: {default_agent}")
+
+        # State
+        final_output = ""
+        last_response_id: Optional[str] = None
+
+        # Initial messages for the very first agent in flow
+        initial_messages: List[TResponseInputItem] = sanitize_input_items(list(messages or []))
+        first_dispatch_done: bool = False
+        last_plain_output: str = ""  # what was displayed to UI in the previous step
+
+        steps = 0
+
+        # Shared stream handler (bot-to-bot style)
+        handler = StreamHandler(self.window, bridge)
+        begin = True
+
+        while current_ids and (steps < max_iterations or max_iterations == 0) and not bridge.stopped():
+            step_start = perf_counter()
+            current_id = current_ids[0]
+            steps += 1
+
+            # END node
+            if current_id in fs.ends:
+                self.logger.info(f"Reached END node: {current_id}")
+                break
+
+            # Validate agent
+            if current_id not in fs.agents:
+                self.logger.warning(f"Next id {current_id} is not an agent; stopping or jumping to END.")
+                if g.end_nodes:
+                    current_ids = [g.end_nodes[0]]
+                    continue
+                break
+
+            node: AgentNode = fs.agents[current_id]
+            self.logger.debug(f"[step {steps}] agent_id={node.id} name={node.name} outs={node.outputs}")
+
+            # Resolve per-node runtime
+            node_rt = resolve_node_runtime(
+                window=self.window,
+                node=node,
+                option_get=option_get,
+                default_model=model,
+                base_prompt=base_prompt,
+                schema_allow_local=node.allow_local_tools,
+                schema_allow_remote=node.allow_remote_tools,
+                default_allow_local=allow_local_tools_default,
+                default_allow_remote=allow_remote_tools_default,
+            )
+
+            if dbg.log_runtime:
+                instr_preview = ellipsize(node_rt.instructions, dbg.preview_chars)
+                self.logger.debug(
+                    f"[runtime] model={getattr(node_rt.model,'name',str(node_rt.model))} "
+                    f"allow_local={node_rt.allow_local_tools} allow_remote={node_rt.allow_remote_tools} "
+                    f"instructions='{instr_preview}'"
+                )
+
+            # Memory selection and INPUT BUILD (memory-first policy)
+            mem_id = g.agent_to_memory.get(current_id)
+            mem_state = mem.get(mem_id) if mem_id else None
+            if dbg.log_runtime:
+                mem_info = f"{mem_id} (len={len(mem_state.items)})" if mem_state and mem_state.items else (mem_id or "-")
+                self.logger.debug(f"[memory] attached={bool(mem_id)} mem_id={mem_info}")
+
+            input_items: List[TResponseInputItem]
+            input_source = ""
+
+            if mem_state and not mem_state.is_empty():
+                # memory present and already has history
+                input_items = list(mem_state.items)
+                input_source = "memory:existing"
+            elif mem_state:
+                # memory present but empty -> seed from last output or initial messages
+                if last_plain_output and last_plain_output.strip():
+                    input_items = [{"role": "user", "content": last_plain_output}]
+                    input_source = "memory:seeded_from_last"
+                else:
+                    input_items = list(initial_messages)
+                    input_source = "memory:seeded_from_initial"
+            else:
+                # no memory -> first agent gets initial messages; others get only last output
+                if not first_dispatch_done:
+                    input_items = list(initial_messages)
+                    input_source = "no-mem:initial"
+                else:
+                    if last_plain_output and last_plain_output.strip():
+                        input_items = [{"role": "user", "content": last_plain_output}]
+                        input_source = "no-mem:last_output"
+                    else:
+                        input_items = list(initial_messages)
+                        input_source = "no-mem:fallback_initial"
+
+            prepared_items = sanitize_input_items(input_items)
+
+            if dbg.log_inputs:
+                self.logger.debug(f"[input] source={input_source} items={len(prepared_items)} "
+                                  f"preview={items_preview(prepared_items, dbg.preview_chars)}")
+
+            # Build agent with per-node runtime
+            built = factory.build(
+                node=node,
+                node_runtime=node_rt,
+                preset=preset,
+                function_tools=function_tools,
+                force_router=False,  # auto on multi-output
+                friendly_map={aid: a.name or aid for aid, a in fs.agents.items()},
+                handoffs_enabled=True,
+                context=agent_kwargs.get("context"),
+            )
+            agent = built.instance
+            multi_output = built.multi_output
+            allowed_routes = built.allowed_routes
+
+            if dbg.log_runtime and multi_output:
+                self.logger.debug(f"[routing] multi_output=True routes={allowed_routes} mode={router_stream_mode}")
+
+            # Prepare run kwargs
+            run_kwargs: Dict[str, Any] = {
+                "input": prepared_items,
+                "max_turns": int(agent_kwargs.get("max_iterations", max_iterations)),
+            }
+            if trace_id:
+                run_kwargs["trace_id"] = trace_id
+
+            # Header for UI
+            title = f"\n\n**{built.name}**\n\n"
+            ctx.stream = title
+            bridge.on_step(ctx, begin)
+            begin = False
+            handler.begin = begin
+            if not use_partial_ctx:
+                handler.to_buffer(title)
+
+            display_text = ""  # what we show to UI for this step
+            next_id: Optional[str] = None
+
+            # --- EXECUTION ---
+            if stream and not multi_output:
+                # Full token streaming (single-output agent)
+                result = Runner.run_streamed(agent, **run_kwargs)
+                handler.reset()
+
+                async for event in result.stream_events():
+                    if bridge.stopped():
+                        result.cancel()
+                        bridge.on_stop(ctx)
+                        break
+                    display_text, last_response_id = handler.handle(event, ctx)
+
+                # Prepare next inputs from result for memory update (if any)
+                input_items_next = result.to_input_list()
+                input_items_next = sanitize_input_items(input_items_next)
+
+                # Update memory only if attached
+                if mem_state:
+                    mem_state.update_from_result(input_items_next, last_response_id)
+
+                # Route: first edge or END
+                outs = g.get_next(current_id)
+                next_id = outs[0] if outs else g.first_connected_end(current_id)
+
+            else:
+                # Multi-output or non-stream
+                if multi_output:
+                    mode = (router_stream_mode or "off").lower()
+                    if stream and mode == "realtime":
+                        # Realtime router streaming: stream content as tokens arrive
+                        result = Runner.run_streamed(agent, **run_kwargs)
+                        rts = RealtimeRouterStreamer(
+                            window=self.window,
+                            bridge=bridge,
+                            handler=handler if not use_partial_ctx else None,
+                            buffer_to_handler=(not use_partial_ctx),
+                            logger=self.logger,
+                        )
+                        rts.reset()
+
+                        async for event in result.stream_events():
+                            if bridge.stopped():
+                                result.cancel()
+                                bridge.on_stop(ctx)
+                                break
+                            rts.handle_event(event, ctx)
+                            if rts.last_response_id:
+                                last_response_id = rts.last_response_id
+
+                        raw_text = rts.buffer or ""
+                        decision = parse_route_output(raw_text, allowed_routes)
+                        display_text = decision.content or ""
+
+                        # Prepare next inputs from streamed result, patch assistant content -> content
+                        input_items_next = result.to_input_list()
+                        input_items_next = patch_last_assistant_output(input_items_next, decision.content or "")
+
+                        # Update memory if attached
+                        if mem_state:
+                            mem_state.update_from_result(input_items_next, last_response_id)
+
+                        # Route decision
+                        if decision.valid:
+                            next_id = decision.route
+                        else:
+                            if dbg.log_routes:
+                                self.logger.warning("[router-realtime] Invalid JSON; fallback to first route.")
+                            next_id = allowed_routes[0] if allowed_routes else None
+
+                    elif stream and mode == "delayed":
+                        # Delayed router streaming: collect tokens silently, reveal once
+                        result = Runner.run_streamed(agent, **run_kwargs)
+                        delayed = DelayedRouterStreamer(self.window, bridge)
+                        delayed.reset()
+
+                        async for event in result.stream_events():
+                            if bridge.stopped():
+                                result.cancel()
+                                bridge.on_stop(ctx)
+                                break
+                            _, rid = delayed.handle_event(event, ctx)
+                            if rid:
+                                last_response_id = rid
+
+                        raw_text = delayed.buffer or ""
+                        decision = parse_route_output(raw_text, allowed_routes)
+                        display_text = decision.content or ""
+                        if display_text:
+                            ctx.stream = display_text
+                            bridge.on_step(ctx, False)
+                            if not use_partial_ctx:
+                                handler.to_buffer(display_text)
+
+                        input_items_next = result.to_input_list()
+                        input_items_next = patch_last_assistant_output(input_items_next, display_text)
+
+                        if mem_state:
+                            mem_state.update_from_result(input_items_next, last_response_id)
+
+                        if decision.valid:
+                            next_id = decision.route
+                        else:
+                            if dbg.log_routes:
+                                self.logger.warning(f"[router-delayed] Invalid JSON: {decision.error}; fallback first route.")
+                            next_id = allowed_routes[0] if allowed_routes else None
+                    else:
+                        # No streaming for router: run to completion, then emit only content
+                        result = await Runner.run(agent, **run_kwargs)
+                        last_response_id = getattr(result, "last_response_id", None)
+                        raw_text = extract_text_output(result)
+                        decision = parse_route_output(raw_text, allowed_routes)
+                        display_text = decision.content or ""
+                        if display_text:
+                            ctx.stream = display_text
+                            bridge.on_step(ctx, False)
+                            if not use_partial_ctx:
+                                handler.to_buffer(display_text)
+
+                        input_items_next = result.to_input_list()
+                        input_items_next = patch_last_assistant_output(input_items_next, display_text)
+
+                        if mem_state:
+                            mem_state.update_from_result(input_items_next, last_response_id)
+
+                        if decision.valid:
+                            next_id = decision.route
+                        else:
+                            if dbg.log_routes:
+                                self.logger.warning(f"[router-off] Invalid JSON: {decision.error}; fallback first route.")
+                            next_id = allowed_routes[0] if allowed_routes else None
+                else:
+                    # Single-output, non-stream path
+                    result = await Runner.run(agent, **run_kwargs)
+                    last_response_id = getattr(result, "last_response_id", None)
+                    raw_text = extract_text_output(result)
+                    display_text = raw_text or ""
+                    if display_text:
+                        ctx.stream = display_text
+                        bridge.on_step(ctx, False)
+                        if not use_partial_ctx:
+                            handler.to_buffer(display_text)
+
+                    input_items_next = result.to_input_list()
+                    input_items_next = sanitize_input_items(input_items_next)
+
+                    if mem_state:
+                        mem_state.update_from_result(input_items_next, last_response_id)
+
+                    outs = g.get_next(current_id)
+                    next_id = outs[0] if outs else g.first_connected_end(current_id)
+
+            # DEBUG: output + route
+            if dbg.log_outputs:
+                self.logger.debug(f"[output] preview='{ellipsize(display_text, dbg.preview_chars)}' "
+                                  f"last_response_id={last_response_id}")
+
+            if dbg.log_routes:
+                self.logger.debug(f"[route] current={current_id} -> next={next_id} "
+                                  f"(end_connected={g.first_connected_end(current_id)})")
+
+            # Mark dispatch and remember last plain output (for next no-memory agent)
+            first_dispatch_done = True
+            last_plain_output = display_text
+            final_output = display_text
+
+            # Resolve next id / END
+            if isinstance(next_id, str) and next_id.lower() == "end":
+                end_id = g.first_connected_end(current_id) or (g.end_nodes[0] if g.end_nodes else None)
+                current_ids = [end_id] if end_id else []
+            elif next_id:
+                current_ids = [next_id]
+            else:
+                end_id = g.first_connected_end(current_id)
+                current_ids = [end_id] if end_id else []
+
+            # UI separation after each agent step
+            is_end = True if not current_ids else (current_ids[0] in fs.ends)
+            if use_partial_ctx:
+                ctx = bridge.on_next_ctx(
+                    ctx=ctx,
+                    input="",
+                    output=display_text,
+                    response_id=last_response_id or "",
+                    finish=is_end,
+                    stream=True,
+                )
+                handler.new()
+            else:
+                bridge.on_next(ctx)
+
+            # Step duration
+            dur = perf_counter() - step_start
+            self.logger.debug(f"[step {steps}] duration={dur:.3f}s")
+
+        if bridge.stopped():
+            bridge.on_stop(ctx)
+
+        self.logger.info(f"Flow finished. steps={steps} final_len={len(final_output)}")
+        return FlowResult(ctx=ctx, final_output=final_output, last_response_id=last_response_id)

pygpt_net/core/agents/custom/schema.py

@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# ================================================== #
+# This file is a part of PYGPT package               #
+# Website: https://pygpt.net                         #
+# GitHub:  https://github.com/szczyglis-dev/py-gpt   #
+# MIT License                                        #
+# Created By  : Marcin Szczygliński                  #
+# Updated Date: 2025.09.24 23:00:00                  #
+# ================================================== #
+
+from __future__ import annotations
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Any
+
+
+@dataclass
+class BaseNode:
+    id: str
+    type: str
+    slots: Dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class AgentNode(BaseNode):
+    name: str = ""
+    instruction: str = ""
+    allow_remote_tools: bool = True
+    allow_local_tools: bool = True
+    outputs: List[str] = field(default_factory=list)
+    inputs: List[str] = field(default_factory=list)
+    memory_out: Optional[str] = None  # single mem by spec
+    memory_in: List[str] = field(default_factory=list)  # not used, but kept for completeness
+
+
+@dataclass
+class StartNode(BaseNode):
+    outputs: List[str] = field(default_factory=list)
+
+
+@dataclass
+class EndNode(BaseNode):
+    inputs: List[str] = field(default_factory=list)
+
+
+@dataclass
+class MemoryNode(BaseNode):
+    name: str = ""
+    agents: List[str] = field(default_factory=list)  # agents connected to this memory
+
+
+@dataclass
+class FlowSchema:
+    agents: Dict[str, AgentNode] = field(default_factory=dict)
+    memories: Dict[str, MemoryNode] = field(default_factory=dict)
+    starts: Dict[str, StartNode] = field(default_factory=dict)
+    ends: Dict[str, EndNode] = field(default_factory=dict)
+
+
+def _safe_get(d: Dict[str, Any], *keys, default=None):
+    cur = d
+    for k in keys:
+        if not isinstance(cur, dict):
+            return default
+        if k not in cur:
+            return default
+        cur = cur[k]
+    return cur
+
+
+def parse_schema(schema: List[Dict[str, Any]]) -> FlowSchema:
+    """
+    Parse NodeEditor-exported schema list into FlowSchema.
+    """
+    fs = FlowSchema()
+    for raw in schema:
+        ntype = raw.get("type")
+        nid = raw.get("id")
+        slots = raw.get("slots", {}) or {}
+
+        if ntype == "agent":
+            node = AgentNode(
+                id=nid,
+                type=ntype,
+                slots=slots,
+                name=_safe_get(slots, "name", default=""),
+                instruction=_safe_get(slots, "instruction", default=""),
+                allow_remote_tools=bool(_safe_get(slots, "remote_tools", default=True)),
+                allow_local_tools=bool(_safe_get(slots, "local_tools", default=True)),
+                outputs=list(_safe_get(slots, "output", "out", default=[])) or [],
+                inputs=list(_safe_get(slots, "input", "in", default=[])) or [],
+                memory_out=(_safe_get(slots, "memory", "out", default=[None]) or [None])[0],
+                memory_in=list(_safe_get(slots, "memory", "in", default=[])) or [],
+            )
+            fs.agents[nid] = node
+
+        elif ntype == "start":
+            node = StartNode(
+                id=nid,
+                type=ntype,
+                slots=slots,
+                outputs=list(_safe_get(slots, "output", "out", default=[])) or [],
+            )
+            fs.starts[nid] = node
+
+        elif ntype == "end":
+            node = EndNode(
+                id=nid,
+                type=ntype,
+                slots=slots,
+                inputs=list(_safe_get(slots, "input", "in", default=[])) or [],
+            )
+            fs.ends[nid] = node
+
+        elif ntype == "memory":
+            node = MemoryNode(
+                id=nid,
+                type=ntype,
+                slots=slots,
+                name=_safe_get(slots, "name", default=""),
+                agents=list(_safe_get(slots, "input", "in", default=[])) or [],
+            )
+            fs.memories[nid] = node
+
+    return fs