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

pygpt_net/core/agents/custom/llama_index/router_streamer.py

@@ -0,0 +1,106 @@
+#!/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
+import json
+import re
+from typing import Optional
+
+
+class DelayedRouterStreamerLI:
+    """Collect raw JSON stream (no UI)."""
+    def __init__(self):
+        self._raw = ""
+    def reset(self):
+        self._raw = ""
+    def handle_delta(self, delta: str):
+        self._raw += delta or ""
+    @property
+    def buffer(self) -> str:
+        return self._raw
+
+
+class RealtimeRouterStreamerLI:
+    """
+    Stream only JSON 'content' string incrementally.
+    handle_delta(delta) -> returns decoded content suffix to emit (may be '').
+    """
+    CONTENT_PATTERN = re.compile(r'"content"\s*:\s*"')
+
+    def __init__(self):
+        self._raw = ""
+        self._content_started = False
+        self._content_closed = False
+        self._content_start_idx = -1
+        self._content_raw = ""
+        self._content_decoded = ""
+
+    def reset(self):
+        self._raw = ""
+        self._content_started = False
+        self._content_closed = False
+        self._content_start_idx = -1
+        self._content_raw = ""
+        self._content_decoded = ""
+
+    def handle_delta(self, delta: str) -> str:
+        if not delta:
+            return ""
+        self._raw += delta
+        if not self._content_started:
+            m = self.CONTENT_PATTERN.search(self._raw)
+            if m:
+                self._content_started = True
+                self._content_start_idx = m.end()
+            else:
+                return ""
+        if self._content_started and not self._content_closed:
+            return self._process()
+        return ""
+
+    def _process(self) -> str:
+        sub = self._raw[self._content_start_idx:]
+        close_idx = self._find_unescaped_quote(sub)
+        if close_idx is not None:
+            portion = sub[:close_idx]
+            self._content_closed = True
+        else:
+            portion = sub
+        new_raw_piece = portion[len(self._content_raw):]
+        if not new_raw_piece:
+            return ""
+        self._content_raw += new_raw_piece
+        try:
+            decoded_full = json.loads(f'"{self._content_raw}"')
+            new_suffix = decoded_full[len(self._content_decoded):]
+            self._content_decoded = decoded_full
+            return new_suffix
+        except Exception:
+            return ""
+
+    @staticmethod
+    def _find_unescaped_quote(s: str) -> Optional[int]:
+        i = 0
+        while i < len(s):
+            if s[i] == '"':
+                j = i - 1
+                bs = 0
+                while j >= 0 and s[j] == '\\':
+                    bs += 1
+                    j -= 1
+                if bs % 2 == 0:
+                    return i
+            i += 1
+        return None
+
+    @property
+    def buffer(self) -> str:
+        return self._raw

pygpt_net/core/agents/custom/llama_index/runner.py

@@ -0,0 +1,529 @@
+#!/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
+import asyncio
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+from time import perf_counter
+from pydantic import ValidationError
+
+from agents import TResponseInputItem
+from pygpt_net.item.model import ModelItem
+from pygpt_net.item.preset import PresetItem
+
+# Shared (reused) components from OpenAI flow core
+from ..logging import Logger, NullLogger
+from ..schema import FlowSchema, AgentNode, parse_schema
+from ..graph import FlowGraph, build_graph
+from ..memory import MemoryManager
+from ..router import parse_route_output
+from ..debug import ellipsize
+
+# LI-specific utils/factory
+from .utils import (
+    sanitize_input_items,
+    NodeRuntime,
+    OptionGetter,
+    resolve_node_runtime,
+    to_li_chat_messages,
+    resolve_llm,
+    extract_agent_text,
+    strip_role_prefixes,
+)
+from .factory import AgentFactoryLI
+
+# LlamaIndex Workflow primitives + events
+from llama_index.core.workflow import Workflow, Context, StartEvent, StopEvent, Event, step
+from llama_index.core.agent.workflow import AgentStream
+
+# App-specific event used as visual separator between agents
+from pygpt_net.provider.agents.llama_index.workflow.events import StepEvent
+
+# Cooperative cancellation error used by main runner
+from workflows.errors import WorkflowCancelledByUser
+
+
+@dataclass
+class DebugConfig:
+    """Config values controlled via get_option(..., 'debug', ...)."""
+    log_runtime: bool = True
+    log_routes: bool = True
+    log_inputs: bool = True
+    log_outputs: bool = True
+    log_tools: bool = True
+    log_llm: bool = True
+    log_schema: bool = True
+    log_memory_dump: bool = True
+    preview_chars: int = 280
+    step_timeout_sec: int = 0
+    timeit_agent_run: bool = True
+    event_echo: bool = True
+
+
+class FlowStartEvent(StartEvent):
+    query: str
+
+
+class FlowTickEvent(Event):
+    """Tick event drives one step of the flow loop."""
+    pass
+
+
+class FlowStopEvent(StopEvent):
+    final_answer: str = ""
+
+
+class DynamicFlowWorkflowLI(Workflow):
+    """
+    LlamaIndex Workflow mirroring OpenAI dynamic flow:
+    - Emits AgentStream header + AgentStream content per step (no token-by-token from single LI agents).
+    - Emits StepEvent between agents (your runner uses it to separate UI blocks).
+    - Routing via JSON {route, content} (UI sees only content; JSON never leaks).
+    - Memory/no-memory with baton policy:
+        * first agent w/o memory: initial messages (or query),
+        * next w/o memory: only last displayed content as single user message,
+        * agent with memory: if empty -> seed from last output (fallback initial); else use its history.
+      After step, memory is updated with [user baton, assistant content].
+    - Per-agent model/prompt/allow_* via get_option(node_id, ...).
+    """
+
+    def __init__(
+        self,
+        *,
+        window,
+        logger: Optional[Logger],
+        schema: List[Dict[str, Any]],
+        initial_messages: Optional[List[TResponseInputItem]],
+        preset: Optional[PresetItem],
+        default_model: ModelItem,
+        option_get: OptionGetter,
+        router_stream_mode: str,
+        allow_local_tools_default: bool,
+        allow_remote_tools_default: bool,
+        max_iterations: int,
+        llm: Any,
+        tools: List[Any],
+        stream: bool,
+        base_prompt: Optional[str],
+        timeout: int = 120,
+        verbose: bool = True,
+    ):
+        super().__init__(timeout=timeout, verbose=verbose)
+        self.window = window
+        self.logger = logger or NullLogger()
+
+        # Graph/schema
+        self.fs: FlowSchema = parse_schema(schema or [])
+        self.g: FlowGraph = build_graph(self.fs)
+        self.mem = MemoryManager()
+        self.factory = AgentFactoryLI(window, self.logger)
+
+        # Options
+        self.preset = preset
+        self.default_model = default_model
+        self.option_get = option_get
+        self.router_stream_mode = (router_stream_mode or "off").lower()  # LI: agents don't token-stream
+        self.allow_local_tools_default = allow_local_tools_default
+        self.allow_remote_tools_default = allow_remote_tools_default
+        self.max_iterations = int(max_iterations or 20)
+
+        # Base LLM/tools from app (per-node model override via resolve_llm)
+        self.llm_base = llm
+        self.tools_base = tools or []
+        self.stream = bool(stream)  # kept for symmetry with OpenAI; LI agents don't stream tokens
+        self.base_prompt = base_prompt or ""
+
+        # Runtime
+        self._on_stop = None
+        self._cancelled = False
+        self._current_ids: List[str] = []
+        self._steps = 0
+        self._first_dispatch_done = False
+        self._last_plain_output = ""
+        self._initial_items = sanitize_input_items(initial_messages or [])
+        self._initial_chat = []  # List[ChatMessage]
+
+        # Debug config (read once)
+        self.dbg = DebugConfig(
+            log_runtime=bool(self.option_get("debug", "log_runtime", True)),
+            log_routes=bool(self.option_get("debug", "log_routes", True)),
+            log_inputs=bool(self.option_get("debug", "log_inputs", True)),
+            log_outputs=bool(self.option_get("debug", "log_outputs", True)),
+            log_tools=bool(self.option_get("debug", "log_tools", True)),
+            log_llm=bool(self.option_get("debug", "log_llm", True)),
+            log_schema=bool(self.option_get("debug", "log_schema", True)),
+            log_memory_dump=bool(self.option_get("debug", "log_memory_dump", True)),
+            preview_chars=int(self.option_get("debug", "preview_chars", 280)),
+            step_timeout_sec=int(self.option_get("debug", "step_timeout_sec", 0)),
+            timeit_agent_run=bool(self.option_get("debug", "timeit_agent_run", True)),
+            event_echo=bool(self.option_get("debug", "event_echo", True)),
+        )
+
+        # One-shot schema/graph dump
+        if self.dbg.log_schema:
+            try:
+                self._dump_schema_debug()
+            except Exception as e:
+                self.logger.error(f"[debug] schema dump failed: {e}")
+
+    # ============== Debug helpers ==============
+
+    def _dump_schema_debug(self):
+        self.logger.info("[debug] ===== Schema/Graph =====")
+        self.logger.info(f"[debug] agents={len(self.fs.agents)} memories={len(self.fs.memories)} "
+                         f"starts={len(self.fs.starts)} ends={len(self.fs.ends)}")
+        self.logger.info(f"[debug] start_targets={self.g.start_targets} end_nodes={self.g.end_nodes}")
+        for nid, outs in self.g.adjacency.items():
+            self.logger.info(f"[debug] edge {nid} -> {outs}")
+        for aid, mem_id in self.g.agent_to_memory.items():
+            self.logger.info(f"[debug] agent_to_memory {aid} -> {mem_id}")
+        self.logger.info("[debug] ========================")
+
+    def _tool_names(self, tools: List[Any]) -> List[str]:
+        names = []
+        for t in tools or []:
+            n = getattr(getattr(t, "metadata", None), "name", None) or getattr(t, "name", None) or t.__class__.__name__
+            names.append(str(n))
+        return names
+
+    # ============== Internal helpers ==============
+
+    def _is_stopped(self) -> bool:
+        if self._cancelled:
+            return True
+        if callable(self._on_stop):
+            try:
+                return bool(self._on_stop())
+            except Exception:
+                return False
+        return False
+
+    def _friendly_map(self) -> Dict[str, str]:
+        return {aid: a.name or aid for aid, a in self.fs.agents.items()}
+
+    async def _emit(self, ctx: Context, ev: Any):
+        if self.dbg.event_echo:
+            self.logger.debug(f"[event] emit {ev.__class__.__name__}")
+        ctx.write_event_to_stream(ev)
+
+    async def _emit_agent_text(self, ctx: Context, text: str, agent_name: str = "Agent"):
+        """
+        Emit AgentStream(delta=text) robustly. If your env requires extra fields,
+        fall back to extended AgentStream like in your SupervisorWorkflow.
+        """
+        try:
+            if self.dbg.event_echo:
+                self.logger.debug(f"[event] AgentStream delta len={len(text or '')}")
+            ctx.write_event_to_stream(AgentStream(delta=text or ""))
+        except ValidationError:
+            if self.dbg.event_echo:
+                self.logger.debug("[event] AgentStream ValidationError -> using extended fields")
+            ctx.write_event_to_stream(
+                AgentStream(
+                    delta=text or "",
+                    response=text or "",
+                    current_agent_name=agent_name or "Agent",
+                    tool_calls=[],
+                    raw={},
+                )
+            )
+
+    async def _emit_header(self, ctx: Context, name: str):
+        if self.dbg.event_echo:
+            self.logger.debug(f"[event] header emit begin name='{name}'")
+        await self._emit_agent_text(ctx, f"\n\n**{name}**\n\n", agent_name=name)
+        if self.dbg.event_echo:
+            self.logger.debug("[event] header emit done")
+
+    async def _emit_step_sep(self, ctx: Context, node_id: str):
+        try:
+            await self._emit(ctx, StepEvent(name="next", index=self._steps, total=self.max_iterations, meta={"node": node_id}))
+        except Exception as e:
+            self.logger.error(f"[event] StepEvent emit failed: {e}")
+
+    def _resolve_node_runtime(self, node: AgentNode) -> NodeRuntime:
+        return resolve_node_runtime(
+            window=self.window,
+            node=node,
+            option_get=self.option_get,
+            default_model=self.default_model,
+            base_prompt=self.base_prompt,
+            schema_allow_local=node.allow_local_tools,
+            schema_allow_remote=node.allow_remote_tools,
+            default_allow_local=self.allow_local_tools_default,
+            default_allow_remote=self.allow_remote_tools_default,
+        )
+
+    def _build_input_for_node(self, node_id: str) -> tuple[str, list, str]:
+        """
+        Return (user_msg_text, chat_history_msgs, source_tag) using baton/memory policy.
+        """
+        mem_id = self.g.agent_to_memory.get(node_id)
+        mem_state = self.mem.get(mem_id) if mem_id else None
+
+        # memory with history
+        if mem_state and mem_state.items:
+            base_items = list(mem_state.items[:-1]) if len(mem_state.items) >= 1 else []
+            if self._last_plain_output.strip():
+                user_msg_text = self._last_plain_output
+            else:
+                last_ass = mem_state.items[-1] if isinstance(mem_state.items[-1], dict) else {}
+                if isinstance(last_ass.get("content"), str):
+                    user_msg_text = last_ass.get("content", "")
+                elif isinstance(last_ass.get("content"), list) and last_ass["content"]:
+                    user_msg_text = last_ass["content"][0].get("text", "")
+                else:
+                    user_msg_text = ""
+            chat_history_msgs = to_li_chat_messages(base_items)
+            return user_msg_text, chat_history_msgs, "memory:existing_to_user_baton"
+
+        # memory empty
+        if mem_state:
+            if self._last_plain_output.strip():
+                return self._last_plain_output, [], "memory:seed_from_last_output"
+            else:
+                chat_hist = self._initial_chat[:-1] if self._initial_chat else []
+                user_msg = self._initial_chat[-1].content if self._initial_chat else ""
+                return user_msg, chat_hist, "memory:seed_from_initial"
+
+        # no memory
+        if not self._first_dispatch_done:
+            chat_hist = self._initial_chat[:-1] if self._initial_chat else []
+            user_msg = self._initial_chat[-1].content if self._initial_chat else ""
+            return user_msg, chat_hist, "no-mem:first_initial"
+        else:
+            user_msg = self._last_plain_output if self._last_plain_output.strip() else (
+                self._initial_chat[-1].content if self._initial_chat else ""
+            )
+            return user_msg, [], "no-mem:last_output"
+
+    async def _update_memory_after_step(self, node_id: str, user_msg_text: str, display_text: str):
+        mem_id = self.g.agent_to_memory.get(node_id)
+        mem_state = self.mem.get(mem_id) if mem_id else None
+        if not mem_state:
+            if self.dbg.log_memory_dump:
+                self.logger.debug(f"[memory] no memory for {node_id}; skip update.")
+            return
+        before_len = len(mem_state.items)
+        base_items = list(mem_state.items[:-1]) if mem_state.items else []
+        new_mem = (base_items or []) + [
+            {"role": "user", "content": user_msg_text},
+            {"role": "assistant", "content": [{"type": "output_text", "text": display_text}]},
+        ]
+        mem_state.set_from(new_mem, None)
+        after_len = len(mem_state.items)
+        if self.dbg.log_memory_dump:
+            self.logger.debug(
+                f"[memory] {node_id} updated mem_id={mem_id} len {before_len} -> {after_len} "
+                f"user='{ellipsize(user_msg_text, self.dbg.preview_chars)}' "
+                f"assist='{ellipsize(display_text, self.dbg.preview_chars)}'"
+            )
+
+    # ============== Workflow steps ==============
+
+    def run(self, query: str, ctx: Optional[Context] = None, memory: Any = None, verbose: bool = False, on_stop=None):
+        """Entry point used by your LlamaWorkflow runner."""
+        self._on_stop = on_stop
+
+        # Build initial chat once
+        if self._initial_items:
+            self._initial_chat = to_li_chat_messages(self._initial_items)
+            if self.dbg.log_inputs:
+                prev = ellipsize(str(self._initial_items), self.dbg.preview_chars)
+                self.logger.debug(f"[debug] initial_items count={len(self._initial_items)} preview={prev}")
+        else:
+            from llama_index.core.llms import ChatMessage, MessageRole
+            self._initial_chat = [ChatMessage(role=MessageRole.USER, content=query or "")]
+            if self.dbg.log_inputs:
+                self.logger.debug(f"[debug] initial from query='{ellipsize(query, self.dbg.preview_chars)}'")
+
+        # Pick START
+        if self.g.start_targets:
+            self._current_ids = [self.g.start_targets[0]]
+            self.logger.info(f"[step] START -> {self._current_ids[0]}")
+        else:
+            default_agent = self.g.pick_default_start_agent()
+            if default_agent is None:
+                self.logger.error("[step] No START and no agents in schema.")
+                return super().run(ctx=ctx, start_event=FlowStartEvent(query=query or ""))
+            self._current_ids = [default_agent]
+            self.logger.info(f"[step] START (auto lowest-id) -> {default_agent}")
+
+        self._steps = 0
+        self._first_dispatch_done = False
+        self._last_plain_output = ""
+
+        return super().run(ctx=ctx, start_event=FlowStartEvent(query=query or ""))
+
+    @step
+    async def start_step(self, ctx: Context, ev: FlowStartEvent) -> FlowTickEvent | FlowStopEvent:
+        self.logger.debug("[step] start_step")
+        if not self._current_ids:
+            await self._emit_agent_text(ctx, "Flow has no start and no agents.\n", agent_name="Flow")
+            return FlowStopEvent(final_answer="")
+        return FlowTickEvent()
+
+    @step
+    async def loop_step(self, ctx: Context, ev: FlowTickEvent) -> FlowTickEvent | FlowStopEvent:
+        # Cooperative stop
+        if self._is_stopped():
+            self.logger.info("[step] loop_step: stopped() -> cancelling")
+            raise WorkflowCancelledByUser()
+
+        # Termination conditions
+        if not self._current_ids or (self._steps >= self.max_iterations > 0):
+            self.logger.info(f"[step] loop_step: done (ids={self._current_ids}, steps={self._steps})")
+            return FlowStopEvent(final_answer=self._last_plain_output or "")
+
+        current_id = self._current_ids[0]
+        self._steps += 1
+        self.logger.debug(f"[step] loop_step#{self._steps} current_id={current_id}")
+
+        # Reached END?
+        if current_id in self.fs.ends:
+            self.logger.info(f"[step] loop_step: reached END {current_id}")
+            return FlowStopEvent(final_answer=self._last_plain_output or "")
+
+        # If unknown id: jump to END if any
+        if current_id not in self.fs.agents:
+            self.logger.warning(f"[step] loop_step: {current_id} not an agent; jumping to END if any.")
+            end_id = self.g.end_nodes[0] if self.g.end_nodes else None
+            self._current_ids = [end_id] if end_id else []
+            return FlowTickEvent() if self._current_ids else FlowStopEvent(final_answer=self._last_plain_output or "")
+
+        node: AgentNode = self.fs.agents[current_id]
+        if self._steps > 1:
+            await self._emit_step_sep(ctx, current_id)
+        await self._emit_header(ctx, node.name or current_id)
+
+        # Resolve runtime + per-node LLM/tools
+        node_rt = self._resolve_node_runtime(node)
+        if self.dbg.log_runtime:
+            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='{ellipsize(node_rt.instructions, self.dbg.preview_chars)}'"
+            )
+
+        llm_node = resolve_llm(self.window, node_rt.model, self.llm_base, self.stream)
+        if self.dbg.log_llm:
+            self.logger.debug(f"[llm] using={llm_node.__class__.__name__} id={getattr(llm_node,'model',None) or getattr(llm_node,'_model',None)}")
+
+        tools_node = self.tools_base if (node_rt.allow_local_tools or node_rt.allow_remote_tools) else []
+        if self.dbg.log_tools:
+            self.logger.debug(f"[tools] count={len(tools_node)} names={self._tool_names(tools_node)}")
+
+        # Build input (baton/memory)
+        user_msg_text, chat_history_msgs, src = self._build_input_for_node(current_id)
+        if self.dbg.log_inputs:
+            self.logger.debug(
+                f"[input] src={src} chat_hist={len(chat_history_msgs)} "
+                f"user='{ellipsize(user_msg_text, self.dbg.preview_chars)}'"
+            )
+
+        # Build agent (chat_history/max_iterations in ctor – best practice)
+        built = self.factory.build(
+            node=node,
+            node_runtime=node_rt,
+            llm=llm_node,
+            tools=tools_node,
+            friendly_map=self._friendly_map(),
+            chat_history=chat_history_msgs,
+            max_iterations=self.max_iterations,
+        )
+        agent = built.instance
+        multi_output = built.multi_output
+        allowed_routes = built.allowed_routes
+
+        if self.dbg.log_routes:
+            self.logger.debug(f"[routing] multi_output={multi_output} routes={allowed_routes} mode={self.router_stream_mode}")
+
+        display_text = ""
+        next_id: Optional[str] = None
+
+        # Execute (single LI agent doesn't token-stream; Workflow emits blocks)
+        try:
+            t0 = perf_counter()
+            if self.dbg.step_timeout_sec > 0:
+                ret = await asyncio.wait_for(agent.run(user_msg=user_msg_text), timeout=self.dbg.step_timeout_sec)
+            else:
+                ret = await agent.run(user_msg=user_msg_text)
+            dt_ms = (perf_counter() - t0) * 1000.0
+            if self.dbg.timeit_agent_run:
+                self.logger.debug(f"[time] agent.run took {dt_ms:.1f} ms")
+        except asyncio.TimeoutError:
+            self.logger.error(f"[error] agent.run timeout after {self.dbg.step_timeout_sec}s on node={current_id}")
+            ret = None
+        except Exception as e:
+            self.logger.error(f"[error] agent.run failed on node={current_id}: {e}")
+            ret = None
+
+        # Extract and sanitize text
+        raw_text = extract_agent_text(ret) if ret is not None else ""
+        raw_text_clean = strip_role_prefixes(raw_text)
+
+        if self.dbg.log_outputs:
+            self.logger.debug(f"[out.raw] len={len(raw_text)} preview='{ellipsize(raw_text, self.dbg.preview_chars)}'")
+            self.logger.debug(f"[out.cln] len={len(raw_text_clean)} preview='{ellipsize(raw_text_clean, self.dbg.preview_chars)}'")
+
+        if multi_output:
+            decision = parse_route_output(raw_text_clean or "", allowed_routes)
+            display_text = decision.content or ""
+            if display_text:
+                await self._emit_agent_text(ctx, display_text, agent_name=(node.name or current_id))
+            if self.dbg.log_memory_dump:
+                self.logger.debug(f"[mem.prep] node={current_id} save user='{ellipsize(user_msg_text, self.dbg.preview_chars)}' "
+                                  f"assist='{ellipsize(display_text, self.dbg.preview_chars)}'")
+            await self._update_memory_after_step(current_id, user_msg_text, display_text)
+            next_id = decision.route if decision.valid else (allowed_routes[0] if allowed_routes else None)
+            if self.dbg.log_routes:
+                self.logger.debug(
+                    f"[route] node={current_id} valid={decision.valid} next={next_id} content_len={len(display_text)}"
+                )
+        else:
+            display_text = raw_text_clean or ""
+            if display_text:
+                await self._emit_agent_text(ctx, display_text, agent_name=(node.name or current_id))
+            if self.dbg.log_memory_dump:
+                self.logger.debug(f"[mem.prep] node={current_id} save user='{ellipsize(user_msg_text, self.dbg.preview_chars)}' "
+                                  f"assist='{ellipsize(display_text, self.dbg.preview_chars)}'")
+            await self._update_memory_after_step(current_id, user_msg_text, display_text)
+            outs = self.g.get_next(current_id)
+            next_id = outs[0] if outs else self.g.first_connected_end(current_id)
+            if self.dbg.log_routes:
+                self.logger.debug(f"[route] node={current_id} next={next_id} (first edge or END)")
+
+        if self.dbg.log_outputs:
+            self.logger.debug(f"[output] preview='{ellipsize(display_text, self.dbg.preview_chars)}'")
+
+        # Update baton and next
+        self._first_dispatch_done = True
+        self._last_plain_output = display_text
+
+        if isinstance(next_id, str) and next_id.lower() == "end":
+            end_id = self.g.first_connected_end(current_id) or (self.g.end_nodes[0] if self.g.end_nodes else None)
+            self._current_ids = [end_id] if end_id else []
+            if not self._current_ids:
+                self.logger.info("[step] next=END (no concrete END node); stopping now.")
+                return FlowStopEvent(final_answer=self._last_plain_output or "")
+        elif next_id:
+            self._current_ids = [next_id]
+        else:
+            end_id = self.g.first_connected_end(current_id)
+            if end_id:
+                self._current_ids = [end_id]
+            else:
+                self._current_ids = []
+                self.logger.info("[step] no next, no END; stopping.")
+                return FlowStopEvent(final_answer=self._last_plain_output or "")
+
+        return FlowTickEvent()

pygpt_net/core/agents/custom/llama_index/stream.py

@@ -0,0 +1,56 @@
+#!/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
+import io
+from typing import Tuple
+
+from pygpt_net.core.agents.bridge import ConnectionContext
+from pygpt_net.item.ctx import CtxItem
+
+
+class LIStreamHandler:
+    """
+    Minimal streaming helper for LlamaIndex events, API-compatible with your usage:
+    - reset(), new(), to_buffer(text)
+    - begin flag
+    - handle_token(delta, ctx) -> returns (buffer, None)
+    """
+    def __init__(self, bridge: ConnectionContext):
+        self.bridge = bridge
+        self._buf = io.StringIO()
+        self.begin = True
+
+    @property
+    def buffer(self) -> str:
+        return self._buf.getvalue()
+
+    def reset(self):
+        self._buf = io.StringIO()
+
+    def new(self):
+        self.reset()
+        self.begin = True
+
+    def to_buffer(self, text: str):
+        if text:
+            self._buf.write(text)
+
+    def handle_token(self, delta: str, ctx: CtxItem, flush: bool = True, buffer: bool = True) -> Tuple[str, None]:
+        if not delta:
+            return self.buffer, None
+        ctx.stream = delta
+        if flush:
+            self.bridge.on_step(ctx, self.begin)
+        if buffer:
+            self._buf.write(delta)
+        self.begin = False
+        return self.buffer, None