Files
HadTavern/agentui/pipeline/executor.py
2025-10-03 21:55:24 +03:00

4453 lines
207 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

from __future__ import annotations
from typing import Any, Dict, List, Optional, Callable, Awaitable
from urllib.parse import urljoin
import json
import re
import asyncio
import time
import hashlib
from collections import deque
from agentui.providers.http_client import build_client
from agentui.common.vendors import detect_vendor
from agentui.providers.adapters.registry import get_adapter, default_base_url_for as _adapter_default_base_url_for
from agentui.pipeline.templating import (
_OUT_MACRO_RE,
_VAR_MACRO_RE,
_PROMPT_MACRO_RE,
_OUT_SHORT_RE,
_BRACES_RE,
_split_path,
_get_by_path,
_stringify_for_template,
_deep_find_text,
_best_text_from_outputs,
render_template_simple,
eval_condition_expr,
)
from agentui.pipeline.storage import load_var_store, save_var_store, clear_var_store
from agentui.common.cancel import is_cancelled, clear_cancel, get_cancel_mode
# HTTP request registry for manual resend feature (store original untrimmed bodies)
from collections import OrderedDict
_HTTP_REQ_REGISTRY_MAX = 200
_HTTP_REQ_REGISTRY = OrderedDict()
def register_http_request(req_id: str, info: Dict[str, Any]) -> None:
try:
rid = str(req_id or "")
if not rid:
return
# overwrite if exists to refresh order
if rid in _HTTP_REQ_REGISTRY:
try:
_HTTP_REQ_REGISTRY.pop(rid, None)
except Exception:
pass
_HTTP_REQ_REGISTRY[rid] = info or {}
# trim oldest beyond cap
while len(_HTTP_REQ_REGISTRY) > _HTTP_REQ_REGISTRY_MAX:
try:
_HTTP_REQ_REGISTRY.popitem(last=False)
except Exception:
break
except Exception:
pass
def get_http_request(req_id: str) -> Optional[Dict[str, Any]]:
try:
rid = str(req_id or "")
if not rid:
return None
v = _HTTP_REQ_REGISTRY.get(rid)
# return shallow copy to avoid mutation
return dict(v) if isinstance(v, dict) else None
except Exception:
return None
# --- Global helpers: robust auto-decompression for logging/JSON parsing ----------
import gzip
import zlib
def _decode_content_bytes(data: bytes, enc_header: str) -> Optional[bytes]:
"""
Try to decompress raw bytes according to Content-Encoding.
Returns decompressed bytes or None if unable/not needed.
"""
if not data:
return b""
enc = (enc_header or "").lower()
try:
if "br" in enc:
# Try brotli libs
brotli = None
try:
import brotlicffi as brotli # type: ignore
except Exception:
try:
import brotli # type: ignore
except Exception:
brotli = None
if brotli is not None:
try:
return brotli.decompress(data) # type: ignore[attr-defined]
except Exception:
try:
# Some packages expose .decompress as brotli.decompress
return brotli.decompress(data) # type: ignore
except Exception:
return None
return None
if "gzip" in enc:
try:
return gzip.decompress(data)
except Exception:
return None
if "deflate" in enc:
try:
return zlib.decompress(data)
except Exception:
# Some servers send raw deflate without zlib header
try:
return zlib.decompress(data, -zlib.MAX_WBITS)
except Exception:
return None
except Exception:
return None
return None
def _safe_response_text(resp) -> str:
"""
Return response text with best-effort auto-decompression (br/gzip/deflate) for logging.
Falls back to resp.text on failure.
"""
try:
enc = (resp.headers.get("content-encoding") or "").lower()
if enc:
dec = _decode_content_bytes(resp.content, enc)
if dec is not None:
try:
charset = resp.encoding or "utf-8"
return dec.decode(charset, errors="replace")
except Exception:
return dec.decode("utf-8", errors="replace")
# If no encoding or not handled — rely on httpx auto-decoding for gzip/deflate
return resp.text
except Exception:
try:
return resp.text
except Exception:
return "<resp.text decode error>"
def _safe_response_json(resp) -> Any:
"""
Return parsed JSON with auto-decompression fallback.
"""
try:
return resp.json()
except Exception:
t = _safe_response_text(resp)
try:
return json.loads(t)
except Exception:
return {"error": "Failed to decode JSON from upstream", "text": t}
# --- Cooperative cancel/abort helper for in-flight awaits (HTTP, etc.) ---------
async def _await_coro_with_cancel(coro: Awaitable[Any], pipeline_id: str, poll_interval: float = 0.1) -> Any:
"""
Await 'coro' while polling manual cancel flag for the pipeline.
Behavior:
- If cancel mode is 'graceful' -> do NOT interrupt 'coro'; we just continue waiting
and upper layers will stop before scheduling new work.
- If cancel mode is 'abort' -> cancel in-flight task immediately and raise ExecutionError.
Returns the result of 'coro' or raises ExecutionError on abort.
"""
task = asyncio.create_task(coro)
while True:
try:
# Shield from timeout cancellation; we only use timeout to poll flags.
return await asyncio.wait_for(asyncio.shield(task), timeout=poll_interval)
except asyncio.TimeoutError:
# Not done yet — fall through to polling
pass
except asyncio.CancelledError:
# Map in-flight cancellation to ExecutionError on abort; otherwise re-raise
try:
mode = get_cancel_mode(pipeline_id)
except Exception:
mode = "abort"
if mode == "abort":
raise ExecutionError("Cancelled by user (abort)")
raise
# Poll cancel flag
try:
if is_cancelled(pipeline_id):
mode = get_cancel_mode(pipeline_id)
if mode == "abort":
try:
if not task.done():
task.cancel()
await task
except asyncio.CancelledError:
raise ExecutionError("Cancelled by user (abort)")
except Exception as exc: # noqa: BLE001
raise ExecutionError(f"Cancelled by user (abort): {exc}")
# graceful: do not interrupt; keep waiting
except Exception:
# Be defensive: ignore polling failures and keep waiting
pass
# --- Helpers: sanitize base64/data URLs in JSON for logging (Burp-like) ----------
def _is_b64ish_string(s: str) -> bool:
try:
if not isinstance(s, str) or len(s) < 128:
return False
return bool(re.fullmatch(r"[A-Za-z0-9+/=\r\n]+", s))
except Exception:
return False
_DATA_URL_RE = re.compile(r"(?is)^data:([^;]+);base64,([A-Za-z0-9+/=\r\n]+)$")
def _sanitize_b64_for_log(obj: Any, max_len: int = 180) -> Any:
"""
Recursively walk JSON-like obj and trim long base64 strings and data URLs for logging only.
- If key name contains 'data' (case-insensitive) and value looks like base64 → trim
- If value is data URL (data:mime;base64,...) → trim base64 part
Structure is preserved. Only strings are shortened.
"""
def trim_str(val: str, key_hint: Optional[str] = None) -> str:
# data URL?
m = _DATA_URL_RE.match(val or "")
if m:
mime, b64 = m.group(1), m.group(2)
if len(b64) > max_len:
return f"data:{mime};base64,{b64[:max_len]}... (trimmed {len(b64)-max_len})"
return val
# plain base64 by key hint
if key_hint and "data" in key_hint.lower() and _is_b64ish_string(val) and len(val) > max_len:
return val[:max_len] + f"... (trimmed {len(val)-max_len})"
return val
if isinstance(obj, dict):
out: Dict[str, Any] = {}
for k, v in obj.items():
if isinstance(v, str):
out[k] = trim_str(v, str(k))
else:
out[k] = _sanitize_b64_for_log(v, max_len)
return out
if isinstance(obj, list):
return [_sanitize_b64_for_log(v, max_len) for v in obj]
return obj
def _sanitize_json_string_for_log(s: str, max_len: int = 180) -> str:
"""
If 's' is a JSON string, sanitize base64/data URLs inside and dump back.
Otherwise return original string.
"""
try:
j = json.loads(s)
j2 = _sanitize_b64_for_log(j, max_len=max_len)
return json.dumps(j2, ensure_ascii=False, indent=2)
except Exception:
return s
# --- Templating helpers are imported from agentui.pipeline.templating ---
# moved to agentui.pipeline.templating
# moved to agentui.pipeline.templating
# moved to agentui.pipeline.templating
# moved to agentui.pipeline.templating
def _extract_out_node_id_from_ref(s: Any) -> Optional[str]:
"""
Извлекает node_id из строки с макросом [[OUT:nodeId(.path)*]].
Возвращает None, если макрос не найден.
"""
if not isinstance(s, str):
return None
m = _OUT_MACRO_RE.search(s)
if not m:
return None
body = m.group(1).strip()
node_id = body.split(".", 1)[0].strip()
return node_id or None
def _resolve_in_value(source: Any, context: Dict[str, Any], values: Dict[str, Dict[str, Any]]) -> Any:
"""
Разрешает входные связи/макросы в значение для inputs:
- Нестроковые значения возвращаются как есть
- "macro:path" → берёт значение из context по точечному пути
- "[[VAR:path]]" → берёт значение из context
- "[[OUT:nodeId(.path)*]]" → берёт из уже вычисленных выходов ноды
- "nodeId(.path)*" → ссылка на выходы ноды
- Если передан список ссылок — вернёт список разрешённых значений
- Иначе пытается взять из context по пути; если не найдено, оставляет исходную строку
"""
# Поддержка массивов ссылок (для multi-depends или будущих списковых входов)
if isinstance(source, list):
return [_resolve_in_value(s, context, values) for s in source]
if not isinstance(source, str):
return source
s = source.strip()
# macro:path
if s.lower().startswith("macro:"):
path = s.split(":", 1)[1].strip()
return _get_by_path(context, path)
# [[VAR: path]]
m = _VAR_MACRO_RE.fullmatch(s)
if m:
path = m.group(1).strip()
return _get_by_path(context, path)
# [[OUT: nodeId(.path)*]]
m = _OUT_MACRO_RE.fullmatch(s)
if m:
body = m.group(1).strip()
if "." in body:
node_id, rest = body.split(".", 1)
node_val = values.get(node_id)
return _get_by_path(node_val, rest)
node_val = values.get(body)
return node_val
# "nodeId(.path)*"
if "." in s:
node_id, rest = s.split(".", 1)
if node_id in values:
return _get_by_path(values.get(node_id), rest)
if s in values:
return values.get(s)
# fallback: from context by dotted path or raw string
ctx_val = _get_by_path(context, s)
return ctx_val if ctx_val is not None else source
# moved to agentui.pipeline.templating
class ExecutionError(Exception):
pass
class Node:
type_name: str = "Base"
def __init__(self, node_id: str, config: Optional[Dict[str, Any]] = None) -> None:
self.node_id = node_id
self.config = config or {}
async def run(self, inputs: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]: # noqa: D401
"""Execute node with inputs and context. Return dict of outputs."""
raise NotImplementedError
# Регистрация поддерживаемых типов нод (минимальный набор)
NODE_REGISTRY: Dict[str, Any] = {}
class PipelineExecutor:
def __init__(self, pipeline: Dict[str, Any]) -> None:
self.pipeline = pipeline
self.nodes_by_id: Dict[str, Node] = {}
for n in pipeline.get("nodes", []):
node_cls = NODE_REGISTRY.get(n.get("type"))
if not node_cls:
raise ExecutionError(f"Unknown node type: {n.get('type')}")
self.nodes_by_id[n["id"]] = node_cls(n["id"], n.get("config", {}))
# Настройки режима исполнения (с дефолтами из default_pipeline)
try:
self.loop_mode = str(self.pipeline.get("loop_mode", "dag")).lower().strip() or "dag"
except Exception:
self.loop_mode = "dag"
try:
self.loop_max_iters = int(self.pipeline.get("loop_max_iters", 1000))
except Exception:
self.loop_max_iters = 1000
try:
self.loop_time_budget_ms = int(self.pipeline.get("loop_time_budget_ms", 10000))
except Exception:
self.loop_time_budget_ms = 10000
# Идентификатор пайплайна и политика очистки стора переменных
try:
self.pipeline_id = str(self.pipeline.get("id", "pipeline_editor")) or "pipeline_editor"
except Exception:
self.pipeline_id = "pipeline_editor"
try:
self.clear_var_store = bool(self.pipeline.get("clear_var_store", True))
except Exception:
self.clear_var_store = True
# В памяти держим актуальный STORE (доступен в шаблонах через {{ store.* }} и [[STORE:*]])
self._store: Dict[str, Any] = {}
# Локальный журнал выполнения для человекочитаемого трейсинга
self._exec_log: List[str] = []
async def run(
self,
context: Dict[str, Any],
trace: Optional[Callable[[Dict[str, Any]], Awaitable[None]]] = None,
) -> Dict[str, Any]:
"""
Точка входа исполнителя. Переключает режим между:
- "dag": волновое исполнение с барьерами (исходное поведение)
- "iterative": итеративное исполнение с очередью и гейтами
"""
# Инициализация STORE по политике
try:
if self.clear_var_store:
clear_var_store(self.pipeline_id)
self._store = {}
else:
self._store = load_var_store(self.pipeline_id) or {}
except Exception:
self._store = {}
# Перед стартом прогона сбрасываем возможный «старый» флаг отмены
try:
clear_cancel(self.pipeline_id)
except Exception:
pass
mode = (self.loop_mode or "dag").lower()
if mode == "iterative":
res = await self._run_iterative(context, trace)
else:
res = await self._run_dag(context, trace)
# На всякий случай финальная запись стора (если были изменения)
try:
save_var_store(self.pipeline_id, self._store)
except Exception:
pass
return res
async def _exec_node(
self,
node_id: str,
ndef: Dict[str, Any],
values_snapshot: Dict[str, Any],
wave_num: int,
context: Dict[str, Any],
trace: Optional[Callable[[Dict[str, Any]], Awaitable[None]]] = None,
user_vars_snapshot: Optional[Dict[str, Any]] = None,
) -> tuple[str, Dict[str, Any]]:
"""
Общий исполнитель ноды (используется и в DAG, и в iterative режимах)
Поведение идентично прежним локальным exec_one():
- подготовка контекста (OUT, vars, store, meta, _trace)
- разрешение inputs через макросы
- события trace node_start/node_done/node_error
- TRACE-печати начала/окончания с prev_text
"""
if not ndef:
raise ExecutionError(f"Node definition not found: {node_id}")
node = self.nodes_by_id[node_id]
# Сбор контекста
ctx = dict(context)
ctx["OUT"] = values_snapshot
try:
ctx["vars"] = dict(user_vars_snapshot or {})
except Exception:
ctx["vars"] = {}
# STORE доступен в шаблонах
try:
ctx["store"] = dict(self._store)
except Exception:
ctx["store"] = {}
# Pipeline meta (доступно нодам через context.meta)
try:
ctx["meta"] = {
"id": self.pipeline_id,
"loop_mode": self.loop_mode,
"loop_max_iters": self.loop_max_iters,
"loop_time_budget_ms": self.loop_time_budget_ms,
"clear_var_store": self.clear_var_store,
"http_timeout_sec": float(self.pipeline.get("http_timeout_sec", 60) or 60),
# v1: стратегия извлечения текста для [[OUTx]]
"text_extract_strategy": str(self.pipeline.get("text_extract_strategy", "auto") or "auto"),
"text_extract_json_path": str(self.pipeline.get("text_extract_json_path", "") or ""),
"text_join_sep": str(self.pipeline.get("text_join_sep", "\n") or "\n"),
# v2: коллекция пресетов парсинга для выбора в нодах
"text_extract_presets": list(self.pipeline.get("text_extract_presets", []) or []),
}
except Exception:
pass
# Прокидываем traceфункцию в контекст, чтобы ноды могли посылать события (SSE лог)
try:
if trace is not None:
ctx["_trace"] = trace
except Exception:
pass
# Разворачиваем inputs
inputs: Dict[str, Any] = {}
for name, source in (ndef.get("in") or {}).items():
if isinstance(source, list):
inputs[name] = [_resolve_in_value(s, ctx, values_snapshot) for s in source]
else:
inputs[name] = _resolve_in_value(source, ctx, values_snapshot)
# Trace: node_start
if trace is not None:
try:
await trace({
"event": "node_start",
"node_id": ndef["id"],
"node_type": node.type_name,
"wave": wave_num,
"ts": int(time.time() * 1000),
})
except Exception:
pass
started = time.perf_counter()
try:
print(f"TRACE start: {ndef['id']} ({node.type_name})")
except Exception:
pass
# Исполнение
try:
out = await node.run(inputs, ctx)
except BaseException as exc:
err = exc
try:
import asyncio as _asyncio
if isinstance(exc, _asyncio.CancelledError):
err = ExecutionError("Cancelled by user (abort)")
except Exception:
pass
if trace is not None:
try:
await trace({
"event": "node_error",
"node_id": ndef["id"],
"node_type": node.type_name,
"wave": wave_num,
"ts": int(time.time() * 1000),
"error": str(err),
})
except Exception:
pass
raise err
else:
dur_ms = int((time.perf_counter() - started) * 1000)
if trace is not None:
try:
await trace({
"event": "node_done",
"node_id": ndef["id"],
"node_type": node.type_name,
"wave": wave_num,
"ts": int(time.time() * 1000),
"duration_ms": dur_ms,
})
except Exception:
pass
try:
prev_text = ""
try:
from agentui.pipeline.templating import _best_text_from_outputs as _bt # type: ignore
if isinstance(out, dict):
prev_text = _bt(out) or ""
except Exception:
prev_text = ""
if isinstance(prev_text, str) and len(prev_text) > 200:
prev_text = prev_text[:200]
print(f"TRACE done: {ndef['id']} ({node.type_name}) dur={dur_ms}ms text={prev_text!r}")
except Exception:
pass
return node_id, out
async def _run_dag(
self,
context: Dict[str, Any],
trace: Optional[Callable[[Dict[str, Any]], Awaitable[None]]] = None,
) -> Dict[str, Any]:
"""
Исходная волновая модель (DAG) — без изменений поведения.
"""
nodes: List[Dict[str, Any]] = list(self.pipeline.get("nodes", []))
id_set = set(self.nodes_by_id.keys())
# Собираем зависимости: node_id -> set(parent_ids), и обратные связи dependents
deps_map: Dict[str, set] = {n["id"]: set() for n in nodes}
dependents: Dict[str, set] = {n["id"]: set() for n in nodes}
# Гейты для ветвлений: child_id -> List[(parent_id, gate_name)] (gate_name: "true"/"false")
gate_deps: Dict[str, List[tuple[str, str]]] = {n["id"]: [] for n in nodes}
for n in nodes:
nid = n["id"]
for _, source in (n.get("in") or {}).items():
# Разворачиваем массивы ссылок (multi-depends)
sources = source if isinstance(source, list) else [source]
for src in sources:
if not isinstance(src, str):
# Нестрочные значения и массивы констант — зависимостей нет
continue
if src.startswith("macro:"):
# Макросы берутся из контекста, без зависимостей
continue
# [[VAR:...]] — макрос из контекста, зависимостей нет
if re.fullmatch(r"\[\[\s*VAR\s*[:\s]\s*[^\]]+\s*\]\]", src.strip()):
continue
# [[OUT:nodeId(.key)*]] — зависимость от указанной ноды
out_ref_node = _extract_out_node_id_from_ref(src)
if out_ref_node and out_ref_node in id_set:
deps_map[nid].add(out_ref_node)
dependents[out_ref_node].add(nid)
continue
# Ссылки вида "node.outKey" или "node"
src_id = src.split(".", 1)[0] if "." in src else src
if src_id in id_set:
# Если указали конкретный outKey (например, nIf.true / nIf.false),
# трактуем это ТОЛЬКО как гейт (НЕ как топологическую зависимость), чтобы избежать дедлоков.
is_gate = False
if "." in src:
try:
_, out_name = src.split(".", 1)
on = str(out_name).strip().lower()
if on in {"true", "false"}:
gate_deps[nid].append((src_id, on))
# Smart mode v2: If-node may act as a real parent ONLY when the child has no real parents yet
# AND adding this parent won't create an immediate cycle (i.e. If-node does not depend on this child).
if len(deps_map[nid]) == 0:
dep_on_child = False
try:
src_def = next((x for x in nodes if str(x.get("id")) == str(src_id)), None)
if src_def and isinstance(src_def.get("in"), dict):
for _, s2 in (src_def.get("in") or {}).items():
arr = s2 if isinstance(s2, list) else [s2]
for it2 in arr:
if isinstance(it2, str):
base = it2.split(".", 1)[0] if "." in it2 else it2
if str(base) == str(nid):
dep_on_child = True
break
if dep_on_child:
break
except Exception:
dep_on_child = False
if not dep_on_child:
deps_map[nid].add(src_id)
dependents[src_id].add(nid)
is_gate = True
except Exception:
is_gate = False
if not is_gate:
deps_map[nid].add(src_id)
dependents[src_id].add(nid)
# Входящие степени и первая волна
in_degree: Dict[str, int] = {nid: len(deps) for nid, deps in deps_map.items()}
ready: List[str] = [nid for nid, deg in in_degree.items() if deg == 0]
processed: List[str] = []
values: Dict[str, Dict[str, Any]] = {}
last_result: Dict[str, Any] = {}
last_node_id: Optional[str] = None
node_def_by_id: Dict[str, Dict[str, Any]] = {n["id"]: n for n in nodes}
user_vars: Dict[str, Any] = {}
try:
parallel_limit = int(self.pipeline.get("parallel_limit", 8))
except Exception:
parallel_limit = 8
if parallel_limit <= 0:
parallel_limit = 1
async def exec_one(node_id: str, values_snapshot: Dict[str, Any], wave_num: int) -> tuple[str, Dict[str, Any]]:
ndef = node_def_by_id.get(node_id)
return await self._exec_node(node_id, ndef, values_snapshot, wave_num, context, trace, user_vars)
wave_idx = 0
while ready:
# Ручная отмена исполнения (DAG)
try:
if is_cancelled(self.pipeline_id):
if trace is not None:
try:
await trace({
"event": "cancelled",
"node_id": "",
"node_type": "Pipeline",
"wave": wave_idx,
"ts": int(time.time() * 1000),
})
except Exception:
pass
# Снимок и финальный EXEC TRACE
try:
self._commit_snapshot(context, values, last_node_id or "")
except Exception:
pass
try:
summary = " -> ".join(self._exec_log) if getattr(self, "_exec_log", None) else ""
if summary and isinstance(self._store.get("snapshot"), dict):
self._store["snapshot"]["EXEC_TRACE"] = summary
save_var_store(self.pipeline_id, self._store)
except Exception:
pass
return last_result
except Exception:
pass
wave_nodes = list(ready)
ready = []
wave_results: Dict[str, Dict[str, Any]] = {}
values_snapshot = dict(values)
for i in range(0, len(wave_nodes), parallel_limit):
chunk = wave_nodes[i : i + parallel_limit]
results = await asyncio.gather(
*(exec_one(nid, values_snapshot, wave_idx) for nid in chunk),
return_exceptions=False,
)
for nid, out in results:
wave_results[nid] = out
last_result = out
last_node_id = nid
try:
ndef = node_def_by_id.get(nid) or {}
ntype = ndef.get("type", "")
if ntype == "If":
expr = str((ndef.get("config") or {}).get("expr", ""))
resb = False
try:
if isinstance(out, dict):
resb = bool(out.get("result"))
except Exception:
resb = False
self._exec_log.append(f"If (#{nid}) {expr} => {str(resb).lower()}")
else:
self._exec_log.append(f"{nid}({ntype})")
except Exception:
pass
values.update(wave_results)
processed.extend(wave_nodes)
try:
for _nid, out in wave_results.items():
if isinstance(out, dict):
v = out.get("vars")
if isinstance(v, dict):
user_vars.update(v)
except Exception:
pass
# Обновляем и сохраняем STORE, если были новые переменные из SetVars/прочих нод
try:
merged: Dict[str, Any] = {}
for _nid, out in wave_results.items():
if isinstance(out, dict) and isinstance(out.get("vars"), dict):
merged.update(out.get("vars") or {})
if merged:
self._store.update(merged)
try:
save_var_store(self.pipeline_id, self._store)
except Exception:
pass
except Exception:
pass
for done in wave_nodes:
for child in dependents.get(done, ()):
in_degree[child] -= 1
next_ready_candidates = [nid for nid, deg in in_degree.items() if deg == 0 and nid not in processed and nid not in wave_nodes]
def _gates_ok(child_id: str) -> bool:
pairs = gate_deps.get(child_id) or []
if not pairs:
return True
missing_seen = False
for pid, gate in pairs:
v = values.get(pid)
if v is None:
# значение гейта ещё не известно
missing_seen = True
continue
try:
flag = bool(isinstance(v, dict) and v.get(gate))
except Exception:
flag = False
if not flag:
return False
# Семантика "missing gate = pass" как в iterative-режиме:
# если флаги гейта ещё не известны, но у ребёнка есть реальные родители — допускаем первый запуск.
# если реальных родителей нет (только гейты) — ждём появления значения гейта.
if missing_seen:
real_parents = deps_map.get(child_id) or set()
if len(real_parents) == 0:
return False
return True
ready = [nid for nid in next_ready_candidates if _gates_ok(nid)]
wave_idx += 1
if len(processed) != len(nodes):
remaining = [n["id"] for n in nodes if n["id"] not in processed]
raise ExecutionError(f"Cycle detected or unresolved dependencies among nodes: {remaining}")
# Persist snapshot of macro context and outputs into STORE
try:
self._commit_snapshot(context, values, last_node_id or "")
except Exception:
pass
# Построение и сохранение человекочитаемого EXECUTION TRACE
try:
summary = " -> ".join(self._exec_log) if getattr(self, "_exec_log", None) else ""
if summary:
try:
print("===== EXECUTION TRACE =====")
print(summary)
print("===== END TRACE =====")
except Exception:
pass
try:
if isinstance(self._store.get("snapshot"), dict):
self._store["snapshot"]["EXEC_TRACE"] = summary
save_var_store(self.pipeline_id, self._store)
except Exception:
pass
except Exception:
pass
return last_result
async def _run_iterative(
self,
context: Dict[str, Any],
trace: Optional[Callable[[Dict[str, Any]], Awaitable[None]]] = None,
) -> Dict[str, Any]:
"""
Итеративное исполнение:
- Очередь готовых задач
- Повторные пуски допустимы: ребёнок ставится в очередь, когда какой-либо из его родителей обновился после последнего запуска ребёнка
- Гейты If.true/If.false фильтруют постановку в очередь
- Ограничения: loop_max_iters, loop_time_budget_ms
"""
nodes: List[Dict[str, Any]] = list(self.pipeline.get("nodes", []))
id_set = set(self.nodes_by_id.keys())
# Строим зависимости и гейты (как в DAG)
deps_map: Dict[str, set] = {n["id"]: set() for n in nodes}
dependents: Dict[str, set] = {n["id"]: set() for n in nodes}
gate_deps: Dict[str, List[tuple[str, str]]] = {n["id"]: [] for n in nodes}
# Обратная карта для гейтов: кто «слушает» изменения конкретного родителя по гейту
gate_dependents: Dict[str, set] = {n["id"]: set() for n in nodes}
for n in nodes:
nid = n["id"]
for _, source in (n.get("in") or {}).items():
sources = source if isinstance(source, list) else [source]
for src in sources:
if not isinstance(src, str):
continue
if src.startswith("macro:"):
continue
if re.fullmatch(r"\[\[\s*VAR\s*[:\s]\s*[^\]]+\s*\]\]", src.strip()):
continue
out_ref_node = _extract_out_node_id_from_ref(src)
if out_ref_node and out_ref_node in id_set:
# [[OUT:nodeId(.key)*]] — это РЕАЛЬНАЯ топологическая зависимость
deps_map[nid].add(out_ref_node)
dependents[out_ref_node].add(nid)
continue
src_id = src.split(".", 1)[0] if "." in src else src
if src_id in id_set:
# "node.true/false" — ТОЛЬКО гейт (без топологической зависимости)
is_gate = False
if "." in src:
try:
_, out_name = src.split(".", 1)
on = str(out_name).strip().lower()
if on in {"true", "false"}:
gate_deps[nid].append((src_id, on))
# Регистрируем обратную зависимость по гейту,
# чтобы будить ребёнка при изменении флага у родителя
gate_dependents[src_id].add(nid)
# Smart mode v2: делаем If-узел реальным родителем ТОЛЬКО если
# у ребёнка ещё нет реальных родителей И это не создаёт цикл (If не зависит от ребёнка).
if len(deps_map[nid]) == 0:
dep_on_child = False
try:
src_def = next((x for x in nodes if str(x.get("id")) == str(src_id)), None)
if src_def and isinstance(src_def.get("in"), dict):
for _, s2 in (src_def.get("in") or {}).items():
arr = s2 if isinstance(s2, list) else [s2]
for it2 in arr:
if isinstance(it2, str):
base = it2.split(".", 1)[0] if "." in it2 else it2
if str(base) == str(nid):
dep_on_child = True
break
if dep_on_child:
break
except Exception:
dep_on_child = False
if not dep_on_child:
deps_map[nid].add(src_id)
dependents[src_id].add(nid)
is_gate = True
except Exception:
is_gate = False
if not is_gate:
deps_map[nid].add(src_id)
dependents[src_id].add(nid)
# Вспом. структуры состояния
values: Dict[str, Dict[str, Any]] = {}
last_result: Dict[str, Any] = {}
last_node_id: Optional[str] = None
user_vars: Dict[str, Any] = {}
node_def_by_id: Dict[str, Dict[str, Any]] = {n["id"]: n for n in nodes}
# Версии обновлений нод (инкремент при каждом коммите)
version: Dict[str, int] = {n["id"]: 0 for n in nodes}
# Когда ребёнок запускался в последний раз (номер глобального шага)
last_run_step: Dict[str, int] = {n["id"]: -1 for n in nodes}
# Глобальный счётчик шагов
step = 0
# Очередь готовых задач
q: deque[str] = deque()
# Начальные готовые — ноды без зависимостей
in_degree: Dict[str, int] = {nid: len(deps) for nid, deps in deps_map.items()}
for nid, deg in in_degree.items():
if deg == 0:
q.append(nid)
def gates_ok(child_id: str) -> bool:
pairs = gate_deps.get(child_id) or []
if not pairs:
return True
missing_seen = False
for pid, gate in pairs:
v = values.get(pid)
# Если значение гейта ещё не появилось — отмечаем и продолжаем проверку других гейтов
if v is None:
missing_seen = True
continue
try:
flag = bool(isinstance(v, dict) and v.get(gate))
except Exception:
flag = False
if not flag:
return False
# Если были отсутствующие значения гейтов, то:
# - у узла есть реальные (топологические) родители → не блокируем (первый запуск допустим)
# - у узла НЕТ реальных родителей (только гейты) → откладываем до появления значения гейта
if missing_seen:
real_parents = deps_map.get(child_id) or set()
if len(real_parents) == 0:
return False
return True
# Лимиты
total_runs = 0
t0 = time.perf_counter()
try:
parallel_limit = int(self.pipeline.get("parallel_limit", 8))
except Exception:
parallel_limit = 8
if parallel_limit <= 0:
parallel_limit = 1
async def exec_one(node_id: str, snapshot: Dict[str, Any], wave_num: int) -> tuple[str, Dict[str, Any]]:
ndef = node_def_by_id.get(node_id)
return await self._exec_node(node_id, ndef, snapshot, wave_num, context, trace, user_vars)
# Главный цикл
while q:
# Ручная отмена исполнения (iterative)
try:
if is_cancelled(self.pipeline_id):
if trace is not None:
try:
await trace({
"event": "cancelled",
"node_id": "",
"node_type": "Pipeline",
"wave": step,
"ts": int(time.time() * 1000),
})
except Exception:
pass
# Снимок и финальный EXEC TRACE
try:
self._commit_snapshot(context, values, last_node_id or "")
except Exception:
pass
try:
summary = " -> ".join(self._exec_log) if getattr(self, "_exec_log", None) else ""
if summary and isinstance(self._store.get("snapshot"), dict):
self._store["snapshot"]["EXEC_TRACE"] = summary
save_var_store(self.pipeline_id, self._store)
except Exception:
pass
return last_result
except Exception:
pass
# Проверяем лимиты
if total_runs >= self.loop_max_iters:
raise ExecutionError(f"Iterative mode exceeded loop_max_iters={self.loop_max_iters}")
if (time.perf_counter() - t0) * 1000.0 > self.loop_time_budget_ms:
raise ExecutionError(f"Iterative mode exceeded loop_time_budget_ms={self.loop_time_budget_ms}")
# Собираем партию до parallel_limit, с учётом гейтов.
# ВАЖНО: защищаемся от «вечной карусели» в пределах одной партии.
# Обрабатываем не более исходной длины очереди за проход.
batch: List[str] = []
seen: set[str] = set()
spin_guard = len(q)
processed_in_pass = 0
while q and len(batch) < parallel_limit and processed_in_pass < spin_guard:
nid = q.popleft()
processed_in_pass += 1
if nid in seen:
# Уже добавлен в партию — пропускаем без возврата в очередь
continue
# Должны быть удовлетворены зависимости: или нет родителей, или все родители хотя бы раз исполнились
parents = deps_map.get(nid) or set()
deps_ok = all(version[p] > 0 or in_degree.get(nid, 0) == 0 for p in parents) or (len(parents) == 0)
if not deps_ok:
# отложим до лучших времён
q.append(nid)
continue
if not gates_ok(nid):
# гейты пока не открыты — попробуем позже
q.append(nid)
continue
batch.append(nid)
seen.add(nid)
if not batch:
# Очередь «застряла» — либо циклические ожидания без начальных значений, либо гейт не откроется.
# Безопасно выходим.
break
# Снимок на момент партии
snapshot = dict(values)
results = await asyncio.gather(*(exec_one(nid, snapshot, step) for nid in batch), return_exceptions=False)
# Коммитим результаты немедленно и продвигаем версии
produced_vars: Dict[str, Any] = {}
for nid, out in results:
values[nid] = out
version[nid] = version.get(nid, 0) + 1
last_result = out
last_node_id = nid
total_runs += 1
last_run_step[nid] = step
# Человекочитаемый журнал
try:
ndef = node_def_by_id.get(nid) or {}
ntype = ndef.get("type", "")
if ntype == "If":
expr = str((ndef.get("config") or {}).get("expr", ""))
resb = False
try:
if isinstance(out, dict):
resb = bool(out.get("result"))
except Exception:
resb = False
self._exec_log.append(f"If (#{nid}) {expr} => {str(resb).lower()}")
else:
self._exec_log.append(f"{nid}({ntype})")
except Exception:
pass
# Сбор пользовательских переменных
try:
if isinstance(out, dict) and isinstance(out.get("vars"), dict):
produced_vars.update(out.get("vars") or {})
except Exception:
pass
if produced_vars:
user_vars.update(produced_vars)
# Обновляем STORE и сохраняем на диск
try:
self._store.update(produced_vars)
save_var_store(self.pipeline_id, self._store)
except Exception:
pass
# Постановка потомков, у кого изменился хотя бы один родитель
for nid in batch:
# Топологические потомки (реальные зависимости): будим всегда
for child in dependents.get(nid, ()):
if last_run_step.get(child, -1) < step:
q.append(child)
# Потомки по гейтам: будим ТОЛЬКО выбранную ветку
# Логика: если у gchild есть пара gate_deps с текущим nid и нужным gate ('true'/'false'),
# то будим только если соответствующий флаг у родителя истинный. Иначе — не будим.
g_out = values.get(nid) or {}
g_children = gate_dependents.get(nid, ())
for gchild in g_children:
if last_run_step.get(gchild, -1) >= step:
continue
pairs = [p for p in (gate_deps.get(gchild) or []) if p and p[0] == nid]
if not pairs:
# gchild зарегистрирован как зависимый по гейту, но явной пары нет — пропускаем
continue
wake = False
for (_pid, gate_name) in pairs:
try:
want = str(gate_name or "").strip().lower()
if want in ("true", "false"):
if bool(g_out.get(want, False)):
wake = True
break
except Exception:
pass
if wake:
q.append(gchild)
try:
print(f"TRACE wake_by_gate: parent={nid} -> child={gchild}")
except Exception:
pass
# Ранний выход, если встретили Return среди выполненных
try:
for nid in batch:
ndef = node_def_by_id.get(nid) or {}
if (ndef.get("type") or "") == "Return":
# Зафиксируем, что именно Return завершил выполнение
try:
last_node_id = nid
last_result = values.get(nid) or last_result
except Exception:
pass
# Снимок стора до выхода
try:
self._commit_snapshot(context, values, last_node_id or "")
except Exception:
pass
# Финальный EXECUTION TRACE в лог и STORE
try:
summary = " -> ".join(self._exec_log) if getattr(self, "_exec_log", None) else ""
if summary:
try:
print("===== EXECUTION TRACE =====")
print(summary)
print("===== END TRACE =====")
except Exception:
pass
try:
if isinstance(self._store.get("snapshot"), dict):
self._store["snapshot"]["EXEC_TRACE"] = summary
save_var_store(self.pipeline_id, self._store)
except Exception:
pass
except Exception:
pass
return last_result
except Exception:
pass
step += 1
# Финальный снапшот всего контекста и OUT в STORE.snapshot
try:
self._commit_snapshot(context, values, last_node_id or "")
except Exception:
pass
# Построение и сохранение человекочитаемого EXECUTION TRACE
try:
summary = " -> ".join(self._exec_log) if getattr(self, "_exec_log", None) else ""
if summary:
try:
print("===== EXECUTION TRACE =====")
print(summary)
print("===== END TRACE =====")
except Exception:
pass
try:
if isinstance(self._store.get("snapshot"), dict):
self._store["snapshot"]["EXEC_TRACE"] = summary
save_var_store(self.pipeline_id, self._store)
except Exception:
pass
except Exception:
pass
return last_result
def _safe_preview(self, obj: Any, max_bytes: int = 262_144) -> Any:
"""
Отменяем тримминг: сохраняем объект полностью.
Важно: это может сделать STORE очень большим для тяжёлых ответов.
"""
return obj
def _commit_snapshot(self, context: Dict[str, Any], values: Dict[str, Any], last_node_id: str) -> None:
"""
Собирает STORE.snapshot:
- incoming.*, params.*, model, vendor_format, system
- OUT_TEXT.nX — строковая вытяжка (до 16КБ)
- OUT.nX — сырой JSON (с триммингом до ~256КБ через preview)
- Алиасы OUT1/OUT2/... — плоские строки (то же, что [[OUT1]])
- LAST_NODE — последний выполнившийся узел
"""
# OUT_TEXT / OUT
out_text: Dict[str, str] = {}
out_raw: Dict[str, Any] = {}
aliases: Dict[str, Any] = {}
for nid, out in (values or {}).items():
if out is None:
continue
try:
txt = _best_text_from_outputs(out) # type: ignore[name-defined]
except Exception:
txt = ""
if not isinstance(txt, str):
try:
txt = str(txt)
except Exception:
txt = ""
# Без тримминга текста — сохраняем полный text
out_text[nid] = txt
out_raw[nid] = self._safe_preview(out, 262_144)
m = re.match(r"^n(\d+)$", str(nid))
if m:
aliases[f"OUT{int(m.group(1))}"] = txt
snapshot: Dict[str, Any] = {
"incoming": self._safe_preview(context.get("incoming"), 262_144),
"params": self._safe_preview(context.get("params"), 65_536),
"model": context.get("model"),
"vendor_format": context.get("vendor_format"),
"system": context.get("system") or "",
"OUT": out_raw,
"OUT_TEXT": out_text,
"LAST_NODE": last_node_id or "",
**aliases,
}
# Сохраняем под ключом snapshot, пользовательские vars остаются как есть в корне STORE
try:
if not isinstance(self._store, dict):
self._store = {}
self._store["snapshot"] = snapshot
save_var_store(self.pipeline_id, self._store)
except Exception:
pass
class SetVarsNode(Node):
type_name = "SetVars"
def _normalize(self) -> List[Dict[str, Any]]:
raw = self.config.get("variables") or []
if not isinstance(raw, list):
return []
norm: List[Dict[str, Any]] = []
for i, b in enumerate(raw):
if not isinstance(b, dict):
continue
name = str(b.get("name", "")).strip()
mode = str(b.get("mode", "string")).lower().strip()
value = b.get("value", "")
try:
order = int(b.get("order")) if b.get("order") is not None else i
except Exception:
order = i
norm.append({
"id": b.get("id") or f"v{i}",
"name": name,
"mode": "expr" if mode == "expr" else "string",
"value": value,
"order": order,
})
return norm
def _safe_eval_expr(self, expr: str, context: Optional[Dict[str, Any]] = None, out_map: Optional[Dict[str, Any]] = None) -> Any:
"""
Безопасная оценка выражений для SetVars.
Поддержка:
- Литералы: числа/строки/bool/None, списки, кортежи, словари
- JSONлитералы: true/false/null, объекты и массивы (парсятся как Python True/False/None, dict/list)
- Арифметика: + - * / // %, унарные +-
- Логика: and/or, сравнения (== != < <= > >=, цепочки)
- Безопасные функции:
rand(), randint(a,b), choice(list)
jp(value, path, join_sep="\n"), jp_text(value, path, join_sep="\n")
from_json(x)
file_b64(path), data_url(b64, mime), file_data_url(path, mime?)
Примечание: функции jp/jp_text/from_json принимают как обычные значения,
так и строки с макросами ([[...]] или {{ ... }}). Макросы разворачиваются с помощью render_template_simple.
"""
import ast
import operator as op
import random
import base64
import os
import mimetypes
from pathlib import Path
def _resolve_input(x: Any) -> Any:
# Если уже структура — вернуть как есть
if isinstance(x, (dict, list, bool, int, float)) or x is None:
return x
s = str(x)
# Попробуем развернуть макросы/шаблон, если есть контекст
try:
s2 = render_template_simple(s, context or {}, out_map or {})
except Exception:
s2 = s
# Попробуем распарсить JSON
try:
return json.loads(s2)
except Exception:
return s2
def _fn_from_json(x: Any) -> Any:
val = _resolve_input(x)
if isinstance(val, str):
try:
return json.loads(val)
except Exception:
return val
return val
def _fn_jp(value: Any, path: Any, join_sep: Any = "\n") -> Any:
src = _resolve_input(value)
p = str(path or "")
try:
vals = _json_path_extract(src, p)
return vals
except Exception:
return None
def _fn_jp_text(value: Any, path: Any, join_sep: Any = "\n") -> str:
vals = _fn_jp(value, path, join_sep)
try:
return _stringify_join(vals, str(join_sep or "\n"))
except Exception:
# на крайний случай — строковое представление
try:
return str(vals)
except Exception:
return ""
# Новые функции для работы с изображениями
def _fn_file_b64(path_like: Any) -> str:
"""Прочитать файл и вернуть base64-строку (без префикса)."""
p_raw = _resolve_input(path_like)
p = Path(str(p_raw)).expanduser()
data = p.read_bytes()
return base64.b64encode(data).decode("ascii")
def _fn_data_url(b64: Any, mime: Any) -> str:
"""Собрать data URL: data:<mime>;base64,<b64>"""
b64s = str(_resolve_input(b64) if not isinstance(b64, str) else b64)
m = str(_resolve_input(mime) if not isinstance(mime, str) else mime).strip() or "application/octet-stream"
return f"data:{m};base64,{b64s}"
def _fn_file_data_url(path_like: Any, mime_opt: Any = None) -> str:
"""Скомбинировать чтение файла и сборку data URL. mime можно не указывать — попытка угадать по расширению."""
p_raw = _resolve_input(path_like)
p = Path(str(p_raw)).expanduser()
try:
data = p.read_bytes()
except Exception as e:
print(f"DEBUG: file_data_url error reading file {p_raw}: {e}")
raise
size = len(data)
print(f"DEBUG: file_data_url loaded file {p!s} size={size} bytes")
b64s = base64.b64encode(data).decode("ascii")
mime = None
if mime_opt is not None:
mime = str(_resolve_input(mime_opt))
if not mime:
mt, _ = mimetypes.guess_type(str(p))
mime = mt or "application/octet-stream"
return f"data:{mime};base64,{b64s}"
# 0) Попытаться распознать чистый JSONлитерал (включая true/false/null, объекты/массивы/числа/строки).
try:
s = str(expr).strip()
return json.loads(s)
except Exception:
pass
allowed_bin = {
ast.Add: op.add, ast.Sub: op.sub, ast.Mult: op.mul, ast.Div: op.truediv,
ast.FloorDiv: op.floordiv, ast.Mod: op.mod,
}
allowed_unary = {ast.UAdd: lambda x: +x, ast.USub: lambda x: -x}
allowed_cmp = {
ast.Eq: op.eq, ast.NotEq: op.ne, ast.Lt: op.lt, ast.LtE: op.le, ast.Gt: op.gt, ast.GtE: op.ge,
}
def eval_node(node: ast.AST) -> Any:
if isinstance(node, ast.Expression):
return eval_node(node.body)
if isinstance(node, ast.Constant):
return node.value
if isinstance(node, ast.Tuple):
return tuple(eval_node(e) for e in node.elts)
if isinstance(node, ast.List):
return [eval_node(e) for e in node.elts]
if isinstance(node, ast.Dict):
return {eval_node(k): eval_node(v) for k, v in zip(node.keys, node.values)}
if isinstance(node, ast.UnaryOp) and type(node.op) in allowed_unary:
return allowed_unary[type(node.op)](eval_node(node.operand))
if isinstance(node, ast.BinOp) and type(node.op) in allowed_bin:
return allowed_bin[type(node.op)](eval_node(node.left), eval_node(node.right))
if isinstance(node, ast.BoolOp):
vals = [eval_node(v) for v in node.values]
if isinstance(node.op, ast.And):
res = True
for v in vals:
res = res and bool(v)
return res
if isinstance(node.op, ast.Or):
res = False
for v in vals:
res = res or bool(v)
return res
if isinstance(node, ast.Compare):
left = eval_node(node.left)
for opnode, comparator in zip(node.ops, node.comparators):
if type(opnode) not in allowed_cmp:
raise ExecutionError("Unsupported comparison operator")
right = eval_node(comparator)
if not allowed_cmp[type(opnode)](left, right):
return False
left = right
return True
# Разрешённые вызовы: rand(), randint(a,b), choice(list), jp(), jp_text(), from_json(), file_b64(), data_url(), file_data_url()
if isinstance(node, ast.Call):
# Никаких kwargs, *args
if node.keywords or isinstance(getattr(node, "starargs", None), ast.AST) or isinstance(getattr(node, "kwargs", None), ast.AST):
raise ExecutionError("Call with kwargs/starargs is not allowed")
fn = node.func
if not isinstance(fn, ast.Name):
raise ExecutionError("Only simple function calls are allowed")
name = fn.id
if name == "rand":
if len(node.args) != 0:
raise ExecutionError("rand() takes no arguments")
return random.random()
if name == "randint":
if len(node.args) != 2:
raise ExecutionError("randint(a,b) requires two arguments")
a = eval_node(node.args[0])
b = eval_node(node.args[1])
try:
return random.randint(int(a), int(b))
except Exception as exc: # noqa: BLE001
raise ExecutionError(f"randint invalid arguments: {exc}")
if name == "choice":
if len(node.args) != 1:
raise ExecutionError("choice(list) requires one argument")
seq = eval_node(node.args[0])
if not isinstance(seq, (list, tuple)):
raise ExecutionError("choice() expects list or tuple")
if not seq:
raise ExecutionError("choice() on empty sequence")
return random.choice(seq)
if name == "jp":
if len(node.args) < 2 or len(node.args) > 3:
raise ExecutionError("jp(value, path, [join_sep]) requires 2 or 3 args")
v = eval_node(node.args[0])
p = eval_node(node.args[1])
return _fn_jp(v, p, eval_node(node.args[2]) if len(node.args) == 3 else "\n")
if name == "jp_text":
if len(node.args) < 2 or len(node.args) > 3:
raise ExecutionError("jp_text(value, path, [join_sep]) requires 2 or 3 args")
v = eval_node(node.args[0])
p = eval_node(node.args[1])
return _fn_jp_text(v, p, eval_node(node.args[2]) if len(node.args) == 3 else "\n")
if name == "from_json":
if len(node.args) != 1:
raise ExecutionError("from_json(x) requires one argument")
v = eval_node(node.args[0])
return _fn_from_json(v)
if name == "file_b64":
if len(node.args) != 1:
raise ExecutionError("file_b64(path) requires one argument")
return _fn_file_b64(eval_node(node.args[0]))
if name == "data_url":
if len(node.args) != 2:
raise ExecutionError("data_url(b64, mime) requires two arguments")
b64v = eval_node(node.args[0])
mimev = eval_node(node.args[1])
return _fn_data_url(b64v, mimev)
if name == "file_data_url":
if len(node.args) < 1 or len(node.args) > 2:
raise ExecutionError("file_data_url(path, [mime]) requires 1 or 2 args")
p = eval_node(node.args[0])
mv = eval_node(node.args[1]) if len(node.args) == 2 else None
return _fn_file_data_url(p, mv)
raise ExecutionError(f"Function {name} is not allowed")
# Запрещаем всё остальное (Name/Attribute/Subscript/IfExp/Comprehensions и пр.)
raise ExecutionError("Expression not allowed")
try:
tree = ast.parse(str(expr), mode="eval")
except Exception as exc:
raise ExecutionError(f"SetVars expr parse error: {exc}") from exc
return eval_node(tree)
async def run(self, inputs: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]: # noqa: D401
out_map = context.get("OUT") or {}
result: Dict[str, Any] = {}
import re as _re
for v in sorted(self._normalize(), key=lambda x: x.get("order", 0)):
name = v.get("name") or ""
if not _re.fullmatch(r"[A-Za-z_][A-Za-z0-9_]*", name or ""):
raise ExecutionError(f"SetVars invalid variable name: {name!r}")
mode = v.get("mode", "string")
raw_val = v.get("value", "")
if mode == "expr":
resolved = self._safe_eval_expr(str(raw_val), context, out_map)
else:
resolved = render_template_simple(str(raw_val or ""), context, out_map)
result[name] = resolved
# Событие «vars_set» в SSEлог (+ превью значений)
try:
trace_fn = context.get("_trace")
if trace_fn:
def _pv(val: Any) -> str:
try:
if isinstance(val, str):
s = val
else:
s = json.dumps(val, ensure_ascii=False)
except Exception:
try:
s = str(val)
except Exception:
s = "<unrepresentable>"
s = s if isinstance(s, str) else str(s)
return (s[:300] + "") if len(s) > 300 else s
values_preview = {k: _pv(v) for k, v in (result or {}).items()}
await trace_fn({
"event": "vars_set",
"node_id": self.node_id,
"node_type": self.type_name,
"vars": list(result.keys()),
"count": len(result),
"values_preview": values_preview,
})
except Exception:
pass
return {"vars": result}
# --- v1: Нормализованный экстрактор текста для OUTx -----------------------------
def _json_path_extract(obj: Any, path: str) -> Any:
if not path:
return None
nodes = [obj]
for raw_seg in str(path).split("."):
seg = (raw_seg or "").strip()
if not seg:
continue
next_nodes = []
# Поддержка "*": разворачиваем dict.values() или элементы списка
if seg == "*":
for n in nodes:
if isinstance(n, dict):
next_nodes.extend(list(n.values()))
elif isinstance(n, list):
next_nodes.extend(list(n))
nodes = next_nodes
if not nodes:
break
continue
# Числовой индекс для списков
idx = None
try:
idx = int(seg)
except Exception:
idx = None
for n in nodes:
if idx is not None and isinstance(n, list):
if 0 <= idx < len(n):
next_nodes.append(n[idx])
elif idx is None and isinstance(n, dict):
if seg in n:
next_nodes.append(n[seg])
nodes = next_nodes
if not nodes:
break
if not nodes:
return None
return nodes if len(nodes) > 1 else nodes[0]
def _stringify_join(values: Any, join_sep: str = "\n") -> str:
def _flatten(x):
if isinstance(x, list):
for it in x:
yield from _flatten(it)
else:
yield x
arr = list(_flatten(values if isinstance(values, list) else [values]))
out: List[str] = []
for v in arr:
if isinstance(v, str):
if v:
out.append(v)
elif isinstance(v, (dict, list)):
try:
t = _deep_find_text(v) # type: ignore[name-defined]
if isinstance(t, str) and t:
out.append(t)
except Exception:
pass
else:
try:
sv = str(v)
if sv:
out.append(sv)
except Exception:
pass
if not out:
return ""
return (join_sep or "\n").join(out)
def _extract_text_for_out(data: Any, strategy: str = "auto", provider_hint: str = "", json_path: str = "", join_sep: str = "\n") -> str:
s = (strategy or "auto").lower().strip() or "auto"
prov = (provider_hint or "").lower().strip()
# Принудительные стратегии
if s == "deep":
try:
t = _deep_find_text(data) # type: ignore[name-defined]
return t if isinstance(t, str) else (str(t) if t is not None else "")
except Exception:
return ""
if s == "jsonpath":
try:
vals = _json_path_extract(data, json_path or "")
return _stringify_join(vals, join_sep)
except Exception:
return ""
if s in {"openai", "gemini", "claude"}:
try:
if s == "openai" and isinstance(data, dict):
msg = (data.get("choices") or [{}])[0].get("message") or {}
c = msg.get("content")
return c if isinstance(c, str) else (str(c) if c is not None else "")
if s == "gemini" and isinstance(data, dict):
cands = data.get("candidates") or []
texts: List[str] = []
for cand in cands:
try:
content = cand.get("content") or {}
parts = content.get("parts") or []
for p in parts:
if isinstance(p, dict):
t = p.get("text")
if isinstance(t, str) and t.strip():
texts.append(t.strip())
except Exception:
continue
if texts:
return (join_sep or "\n").join(texts)
return ""
if s == "claude" and isinstance(data, dict):
blocks = data.get("content") or []
texts = [b.get("text") for b in blocks if isinstance(b, dict) and isinstance(b.get("text"), str)]
if texts:
return "\n".join(texts)
return ""
except Exception:
return ""
return ""
# auto: если есть подсказка — пробуем сначала её ветку, потом общий
if prov in {"openai", "gemini", "claude"}:
forced = _extract_text_for_out(data, prov, prov, json_path, join_sep)
if forced:
return forced
# Универсальный best-effort: знает openai/gemini/claude
try:
t = _best_text_from_outputs({"result": data}) # type: ignore[name-defined]
if isinstance(t, str) and t:
return t
if t is not None:
t2 = str(t)
if t2:
return t2
except Exception:
pass
# Последний шанс — глубокий поиск
try:
t = _deep_find_text(data) # type: ignore[name-defined]
return t if isinstance(t, str) else (str(t) if t is not None else "")
except Exception:
return ""
class ProviderCallNode(Node):
type_name = "ProviderCall"
# --- Prompt Manager helpers -------------------------------------------------
def _get_blocks(self) -> List[Dict[str, Any]]:
"""Return normalized list of prompt blocks from config."""
raw = self.config.get("blocks") or self.config.get("prompt_blocks") or []
if not isinstance(raw, list):
return []
norm: List[Dict[str, Any]] = []
for i, b in enumerate(raw):
if not isinstance(b, dict):
continue
role = str(b.get("role", "user")).lower().strip()
if role not in {"system", "user", "assistant", "tool"}:
role = "user"
# order fallback: keep original index if not provided/correct
try:
order = int(b.get("order")) if b.get("order") is not None else i
except Exception: # noqa: BLE001
order = i
norm.append(
{
"id": b.get("id") or f"b{i}",
"name": b.get("name") or f"Block {i+1}",
"role": role,
"prompt": b.get("prompt") or "",
"enabled": bool(b.get("enabled", True)),
"order": order,
}
)
return norm
def _render_blocks_to_unified(self, context: Dict[str, Any]) -> List[Dict[str, Any]]:
"""
Filter+sort+render blocks to unified messages:
[{role, content, name?}]
Расширение: поддержка изображений в Markdown вида ![alt](URL|DATA_URL).
Если в тексте встречаются такие конструкции — content превращается в список частей:
- {"type":"text","text":"..."}
- {"type":"image_url","url":"..."}
"""
import re as _re
out_map = context.get("OUT") or {}
blocks = [b for b in self._get_blocks() if b.get("enabled", True)]
blocks.sort(key=lambda x: x.get("order", 0))
def _parse_markdown_parts(s: str) -> List[Dict[str, Any]]:
print(f"DEBUG: _parse_markdown_parts input={s!r}")
pat = _re.compile(r"!\[[^\]]*\]\(([^)]+)\)")
parts: List[Dict[str, Any]] = []
pos = 0
for m in pat.finditer(s):
print(f"DEBUG: _parse_markdown_parts match at span={m.span()} url={m.group(1)!r}")
if m.start() > pos:
txt = s[pos:m.start()]
if txt:
parts.append({"type": "text", "text": txt})
url = (m.group(1) or "").strip()
if url:
parts.append({"type": "image_url", "url": url})
pos = m.end()
if pos < len(s):
tail = s[pos:]
if tail:
parts.append({"type": "text", "text": tail})
return parts
messages: List[Dict[str, Any]] = []
for b in blocks:
content_raw = render_template_simple(str(b.get("prompt") or ""), context, out_map)
parts = _parse_markdown_parts(content_raw or "")
# Если нет картинок — оставляем строкой для обратной совместимости
content_final: Any = parts if any(p.get("type") == "image_url" for p in parts) else content_raw
msg = {"role": b["role"], "content": content_final}
messages.append(msg)
return messages
def _messages_to_payload(self, provider: str, messages: List[Dict[str, Any]], context: Dict[str, Any]) -> Dict[str, Any]:
"""Convert unified messages (with optional text+image parts) to provider-specific request payload."""
import re as _re
params = context.get("params") or {}
model = context.get("model") or ""
def _is_data_url(u: str) -> bool:
return isinstance(u, str) and u.strip().lower().startswith("data:")
def _split_data_url(u: str) -> tuple[str, str]:
"""
Возвращает (mime, b64) для data URL.
Поддерживаем форму: data:<mime>;base64,<b64>
"""
try:
header, b64 = u.split(",", 1)
mime = "application/octet-stream"
if header.startswith("data:"):
header2 = header[5:]
if ";base64" in header2:
mime = header2.split(";base64", 1)[0] or mime
elif ";" in header2:
mime = header2.split(";", 1)[0] or mime
elif header2:
mime = header2
return mime, b64
except Exception:
return "application/octet-stream", ""
provider = (provider or "").lower()
if provider == "openai":
def _map_openai_msg(m: Dict[str, Any]) -> Dict[str, Any]:
c = m.get("content")
if isinstance(c, list):
parts = []
for p in c:
if isinstance(p, dict) and p.get("type") == "text":
parts.append({"type": "text", "text": str(p.get("text") or "")})
elif isinstance(p, dict) and p.get("type") in {"image_url", "image"}:
url = str(p.get("url") or "")
parts.append({"type": "image_url", "image_url": {"url": url}})
return {"role": m.get("role", "user"), "content": parts}
# строка — как было
return {"role": m.get("role", "user"), "content": str(c or "")}
payload: Dict[str, Any] = {
"model": model,
"messages": [_map_openai_msg(m) for m in messages],
"temperature": params.get("temperature", 0.7),
}
if params.get("max_tokens") is not None:
payload["max_tokens"] = params.get("max_tokens")
if params.get("top_p") is not None:
payload["top_p"] = params.get("top_p")
if params.get("stop") is not None:
payload["stop"] = params.get("stop")
return payload
if provider == "gemini_image":
def _map_gemini_image_msg(m: Dict[str, Any]) -> Dict[str, Any]:
c = m.get("content")
parts: List[Dict[str, Any]] = []
if isinstance(c, list):
for p in c:
if isinstance(p, dict) and p.get("type") == "text":
parts.append({"text": str(p.get("text") or "")})
elif isinstance(p, dict) and p.get("type") in {"image_url", "image"}:
url = str(p.get("url") or "")
if _is_data_url(url):
mime, b64 = _split_data_url(url)
parts.append({"inline_data": {"mime_type": mime, "data": b64}})
else:
parts.append({"text": url})
else:
parts.append({"text": str(c or "")})
return {"parts": parts}
payload: Dict[str, Any] = {"model": model, "contents": [_map_gemini_image_msg(m) for m in messages]}
return payload
if provider == "gemini":
def _map_claude_msg(m: Dict[str, Any]) -> Dict[str, Any]:
role_raw = str(m.get("role") or "user")
role = role_raw if role_raw in {"user", "assistant"} else "user"
c = m.get("content")
blocks: List[Dict[str, Any]] = []
if isinstance(c, list):
for p in c:
if not isinstance(p, dict):
continue
if p.get("type") == "text":
blocks.append({"type": "text", "text": str(p.get("text") or "")})
elif p.get("type") in {"image_url", "image"}:
url = str(p.get("url") or "")
if _is_data_url(url):
mime, b64 = _split_data_url(url)
blocks.append({"type": "image", "source": {"type": "base64", "media_type": mime, "data": b64}})
else:
blocks.append({"type": "image", "source": {"type": "url", "url": url}})
else:
blocks.append({"type": "text", "text": str(c or "")})
return {"role": role, "content": blocks}
# system — текстовая склейка (без картинок)
sys_text = "\n\n".join(
[str(m.get("content") or "") if not isinstance(m.get("content"), list)
else "\n".join([str(p.get("text") or "") for p in m.get("content") if isinstance(p, dict) and p.get("type") == "text"])
for m in messages if m.get("role") == "system"]
).strip()
msgs = [_map_claude_msg(m) for m in messages if m.get("role") != "system"]
payload: Dict[str, Any] = {
"model": model,
"messages": msgs,
}
if sys_text:
payload["system"] = sys_text
if params.get("temperature") is not None:
payload["temperature"] = params.get("temperature")
if params.get("max_tokens") is not None:
payload["max_tokens"] = params.get("max_tokens")
if params.get("top_p") is not None:
payload["top_p"] = params.get("top_p")
if params.get("stop") is not None:
payload["stop"] = params.get("stop")
return payload
return {}
def _blocks_struct_for_template(self, provider: str, messages: List[Dict[str, Any]], context: Dict[str, Any]) -> Dict[str, Any]:
"""
Сформировать структуру для вставки в шаблон (template) из Prompt Blocks.
Возвращает provider-специфичные ключи, которые можно вставлять в JSON:
- openai: { "messages": [...] , "system_text": "..." }
- gemini: { "contents": [...], "systemInstruction": {...}, "system_text": "..." }
- claude: { "system_text": "...", "system": "...", "messages": [...] }
Расширение: поддержка текст+изображения, как и в _messages_to_payload.
"""
provider = (provider or "openai").lower()
msgs = messages or []
def _is_data_url(u: str) -> bool:
return isinstance(u, str) and u.strip().lower().startswith("data:")
def _split_data_url(u: str) -> tuple[str, str]:
try:
header, b64 = u.split(",", 1)
mime = "application/octet-stream"
if header.startswith("data:"):
header2 = header[5:]
if ";base64" in header2:
mime = header2.split(";base64", 1)[0] or mime
elif ";" in header2:
mime = header2.split(";", 1)[0] or mime
elif header2:
mime = header2
return mime, b64
except Exception:
return "application/octet-stream", ""
if provider == "openai":
def _map(m: Dict[str, Any]) -> Dict[str, Any]:
c = m.get("content")
if isinstance(c, list):
parts = []
for p in c:
if isinstance(p, dict) and p.get("type") == "text":
parts.append({"type": "text", "text": str(p.get("text") or "")})
elif isinstance(p, dict) and p.get("type") in {"image_url", "image"}:
url = str(p.get("url") or "")
parts.append({"type": "image_url", "image_url": {"url": url}})
return {"role": m.get("role", "user"), "content": parts}
return {"role": m.get("role", "user"), "content": str(c or "")}
sys_text = "\n\n".join(
[str(m.get("content") or "") if not isinstance(m.get("content"), list)
else "\n".join([str(p.get("text") or "") for p in m.get("content") if isinstance(p, dict) and p.get("type") == "text"])
for m in msgs if m.get("role") == "system"]
).strip()
return {
"messages": [_map(m) for m in msgs],
"system_text": sys_text,
}
if provider in {"gemini", "gemini_image"}:
def _text_from_msg(m: Dict[str, Any]) -> str:
c = m.get("content")
if isinstance(c, list):
texts = [str(p.get("text") or "") for p in c if isinstance(p, dict) and p.get("type") == "text"]
return "\n".join([t for t in texts if t])
return str(c or "")
sys_text = "\n\n".join([_text_from_msg(m) for m in msgs if m.get("role") == "system"]).strip()
contents = []
for m in msgs:
if m.get("role") == "system":
continue
role = "model" if m.get("role") == "assistant" else "user"
c = m.get("content")
parts: List[Dict[str, Any]] = []
if isinstance(c, list):
for p in c:
if not isinstance(p, dict):
continue
if p.get("type") == "text":
parts.append({"text": str(p.get("text") or "")})
elif p.get("type") in {"image_url", "image"}:
url = str(p.get("url") or "")
if _is_data_url(url):
mime, b64 = _split_data_url(url)
parts.append({"inline_data": {"mime_type": mime, "data": b64}})
else:
parts.append({"text": url})
else:
parts.append({"text": str(c or "")})
contents.append({"role": role, "parts": parts})
d: Dict[str, Any] = {
"contents": contents,
"system_text": sys_text,
}
if sys_text:
d["systemInstruction"] = {"parts": [{"text": sys_text}]}
return d
if provider == "claude":
# Системные сообщения как текст
sys_msgs = []
for m in msgs:
if m.get("role") == "system":
c = m.get("content")
if isinstance(c, list):
sys_msgs.append("\n".join([str(p.get("text") or "") for p in c if isinstance(p, dict) and p.get("type") == "text"]))
else:
sys_msgs.append(str(c or ""))
sys_text = "\n\n".join([s for s in sys_msgs if s]).strip()
out_msgs = []
for m in msgs:
if m.get("role") == "system":
continue
role = m.get("role")
role = role if role in {"user", "assistant"} else "user"
c = m.get("content")
blocks: List[Dict[str, Any]] = []
if isinstance(c, list):
for p in c:
if not isinstance(p, dict):
continue
if p.get("type") == "text":
blocks.append({"type": "text", "text": str(p.get("text") or "")})
elif p.get("type") in {"image_url", "image"}:
url = str(p.get("url") or "")
if _is_data_url(url):
mime, b64 = _split_data_url(url)
blocks.append({"type": "image", "source": {"type": "base64", "media_type": mime, "data": b64}})
else:
blocks.append({"type": "image", "source": {"type": "url", "url": url}})
else:
blocks.append({"type": "text", "text": str(c or "")})
out_msgs.append({"role": role, "content": blocks})
# Optional Claude mode: omit top-level "system" and prepend it as a user message
claude_no_system = False
try:
claude_no_system = bool((self.config or {}).get("claude_no_system", False))
except Exception:
claude_no_system = False
if claude_no_system:
if sys_text:
out_msgs = [{"role": "user", "content": [{"type": "text", "text": sys_text}]}] + out_msgs
return {
"messages": out_msgs,
"system_text": sys_text,
}
d = {
"system_text": sys_text,
"messages": out_msgs,
}
if sys_text:
# Prefer system as a plain string (proxy compatibility)
d["system"] = sys_text
return d
# По умолчанию ничего, но это валидный JSON
return {"messages": []}
async def run(self, inputs: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]:
# While-loop wrapper: if while_expr is set or ignore_errors enabled, route through helper
try:
cfg = self.config or {}
except Exception:
cfg = {}
try:
_while_expr = str(cfg.get("while_expr") or "").strip()
except Exception:
_while_expr = ""
try:
_ignore = bool(cfg.get("ignore_errors", False))
except Exception:
_ignore = False
if not (context.get("_in_while") or False) and (_while_expr or _ignore):
return await _providercall_run_with_while(self, inputs, context)
provider = (self.config.get("provider") or "openai").lower()
# Optional sleep before execution (UI-configured, milliseconds)
try:
sleep_ms = int(self.config.get("sleep_ms") or 0)
except Exception:
sleep_ms = 0
if sleep_ms > 0:
# Emit SSE event to highlight node as "sleeping" on UI
try:
trace_fn = context.get("_trace")
if trace_fn:
await trace_fn({
"event": "node_sleep",
"node_id": self.node_id,
"node_type": self.type_name,
"sleep_ms": int(sleep_ms),
"ts": int(time.time() * 1000),
})
except Exception:
pass
# Non-blocking pause
try:
await asyncio.sleep(max(0.0, sleep_ms / 1000.0))
except Exception:
pass
# Support provider-specific configs stored in UI as provider_configs.{provider}
prov_cfg: Dict[str, Any] = {}
try:
cfgs = self.config.get("provider_configs") or {}
if isinstance(cfgs, dict):
prov_cfg = cfgs.get(provider) or {}
except Exception: # noqa: BLE001
prov_cfg = {}
base_url = prov_cfg.get("base_url") or self.config.get("base_url")
if not base_url:
raise ExecutionError(f"Node {self.node_id} ({self.type_name}) requires 'base_url' in config")
if not str(base_url).startswith(("http://", "https://")):
base_url = "http://" + str(base_url)
endpoint_tmpl: str = prov_cfg.get("endpoint") or self.config.get("endpoint") or ""
template: str = prov_cfg.get("template") or self.config.get("template") or "{}"
headers_json: str = prov_cfg.get("headers") or self.config.get("headers") or "{}"
# Default endpoints if not set
if not endpoint_tmpl:
_ad = None
try:
_ad = get_adapter(provider)
except Exception:
_ad = None
if _ad is not None:
try:
endpoint_tmpl = _ad.default_endpoint(str(context.get("model") or ""))
except Exception:
endpoint_tmpl = ""
if not endpoint_tmpl:
if provider == "openai":
endpoint_tmpl = "/v1/chat/completions"
elif provider == "gemini":
endpoint_tmpl = "/v1beta/models/{{ model }}:generateContent"
elif provider == "gemini_image":
# Специальный провайдер для генерации/превью изображений Gemini (generateContent)
endpoint_tmpl = "/v1beta/models/{{ model }}:generateContent"
elif provider == "claude":
endpoint_tmpl = "/v1/messages"
# Default template for gemini_image if none provided (inject [[PROMPT]])
try:
_tmpl_src = (prov_cfg.get("template") or self.config.get("template") or "").strip()
except Exception:
_tmpl_src = ""
if provider == "gemini_image" and (not _tmpl_src or _tmpl_src == "{}"):
# Минимальный валидный JSON с [[PROMPT]] для contents/systemInstruction
template = '{ "model": "{{ model }}", [[PROMPT]] }'
# Подготовим Prompt Blocks + pm-структуру для шаблона
unified_msgs = self._render_blocks_to_unified(context)
adapter = None
try:
adapter = get_adapter(provider)
except Exception:
adapter = None
if adapter:
blocks_struct = adapter.blocks_struct_for_template(unified_msgs, context, self.config or {})
else:
blocks_struct = self._blocks_struct_for_template(provider, unified_msgs, context)
pm_struct = dict(blocks_struct)
# Расширяем контекст для рендера шаблонов
render_ctx = dict(context)
render_ctx["pm"] = pm_struct
# Прокинем конфиг ноды в контекст для адаптеров (например, claude_no_system)
try:
render_ctx["_node_config"] = dict(self.config or {})
except Exception:
render_ctx["_node_config"] = {}
# Node-local: track VAR paths overridden by prompt_preprocess (no extra syntax)
pre_var_paths = set()
# prompt_preprocess (pre-merge DSL): парсим строки до prompt_combine и готовим «пред‑сегменты»
# Синтаксис строки:
# SEGMENT [delKeyContains "needle"] [delpos=prepend|append|N|-1] [pruneEmpty] [case=ci|cs]
# По умолчанию: case=ci, pruneEmpty=false, без delpos → append
# SEGMENT поддерживает [[...]] и {{ ... }}
pre_segments_raw: List[Dict[str, Any]] = []
try:
pre_raw = str((self.config or {}).get("prompt_preprocess") or "").strip()
except Exception:
pre_raw = ""
if pre_raw:
lines = [ln.strip() for ln in pre_raw.splitlines() if str(ln or "").strip()]
import re as _repp
# Collect local overrides for plain [[VAR:path]] segments after filtering
pre_var_overrides: Dict[str, Any] = {}
def _pp_try_json(s: str) -> Any:
try:
obj = json.loads(s)
except Exception:
try:
obj = json.loads(s, strict=False) # type: ignore[call-arg]
except Exception:
return None
for _ in range(2):
if isinstance(obj, str):
st = obj.strip()
if (st.startswith("{") and st.endswith("}")) or (st.startswith("[") and st.endswith("]")):
try:
obj = json.loads(st)
continue
except Exception:
break
break
return obj
def _norm(s: Any, ci: bool) -> str:
try:
ss = s if isinstance(s, str) else str(s)
except Exception:
ss = ""
return ss.lower() if ci else ss
def _delkeys_by_val_contains(x: Any, needles: List[str], ci: bool, prune_empty: bool, is_root: bool = False) -> tuple[Any, int]:
"""
Рекурсивно удаляет ключи словаря, если строковое представление их значения содержит needle.
- needles: список подстрок
- ci: регистронезависимый поиск (по умолчанию True)
- prune_empty: удалять пустые {} / [] из родителей (кроме корня)
Возвращает (новое_значение_или_None, удалённые_ключи_суммарно)
"""
removed = 0
if isinstance(x, dict):
out: Dict[str, Any] = {}
for k, v in x.items():
v2, rem2 = _delkeys_by_val_contains(v, needles, ci, prune_empty, False)
removed += rem2
sv = _stringify_for_template(v2)
cond = False
try:
nsv = _norm(sv, ci)
for nd in needles:
nds = _norm(nd, ci)
if nds and (nds in nsv):
cond = True
break
except Exception:
cond = False
if cond:
removed += 1
continue
out[k] = v2
if prune_empty and not is_root and len(out) == 0:
return None, removed # type: ignore[return-value]
return out, removed
if isinstance(x, list):
arr: List[Any] = []
for it in x:
it2, rem2 = _delkeys_by_val_contains(it, needles, ci, prune_empty, False)
removed += rem2
if it2 is None and prune_empty:
continue
arr.append(it2)
if prune_empty and not is_root and len(arr) == 0:
return None, removed # type: ignore[return-value]
return arr, removed
# scalar
return x, removed
pre_used = 0
pre_removed_total = 0
for ln in lines:
try:
# Опции
needles = [m.group(1) for m in _repp.finditer(r'(?is)\bdelKeyContains\s+"([^"]*)"', ln)]
mpos = _repp.search(r'(?is)\bdelpos\s*=\s*(prepend|append|-?\d+)\b', ln)
pos_spec = mpos.group(1).strip().lower() if mpos else None
mcase = _repp.search(r'(?is)\bcase\s*=\s*(ci|cs)\b', ln)
ci = True if not mcase else (mcase.group(1).strip().lower() == "ci")
prune = bool(_repp.search(r'(?is)\bpruneEmpty\b', ln))
# Очищаем директивы из текста строки → остаётся сам SEGMENT
s2 = _repp.sub(r'(?is)\bdelKeyContains\s+"[^"]*"', "", ln)
s2 = _repp.sub(r'(?is)\bdelpos\s*=\s*(prepend|append|-?\d+)\b', "", s2)
s2 = _repp.sub(r'(?is)\bcase\s*=\s*(ci|cs)\b', "", s2)
s2 = _repp.sub(r'(?is)\bpruneEmpty\b', "", s2)
seg = s2.strip()
if not seg:
continue
# Try to detect plain [[VAR:path]] to support node-local override without extra syntax
var_path = None
try:
mvar = _VAR_MACRO_RE.fullmatch(seg)
if mvar:
var_path = (mvar.group(1) or "").strip()
except Exception:
var_path = None
# Макросы и попытка распарсить JSON
resolved = render_template_simple(seg, render_ctx, render_ctx.get("OUT") or {})
obj = _pp_try_json(resolved)
base = obj if obj is not None else resolved
# Удаление ключей по contains, если задано
if needles:
try:
base2, remcnt = _delkeys_by_val_contains(base, needles, ci, prune, True)
except Exception:
base2, remcnt = base, 0
else:
base2, remcnt = base, 0
# If the segment was a pure [[VAR:path]] and we had filters,
# locally override this VAR for the rest of the node (so prompt_combine sees the filtered value)
try:
if var_path and needles:
pre_var_overrides[var_path] = base2
except Exception:
pass
pre_segments_raw.append({"obj": base2, "pos": pos_spec})
pre_used += 1
pre_removed_total += int(remcnt or 0)
except Exception:
continue
# Apply local VAR overrides onto render_ctx (node-local, no extra syntax)
try:
import copy as _copy
def _safe_deepcopy(x: Any) -> Any:
try:
return _copy.deepcopy(x)
except Exception:
# Fallback deep copy for dict/list; otherwise return as-is
try:
if isinstance(x, dict):
return {k: _safe_deepcopy(v) for k, v in x.items()}
if isinstance(x, list):
return [_safe_deepcopy(i) for i in x]
except Exception:
pass
try:
return json.loads(json.dumps(x))
except Exception:
return x
def _set_by_path(obj: Any, path: str, value: Any) -> None:
cur = obj
parts = [p.strip() for p in str(path).split(".") if p.strip()]
for i, part in enumerate(parts):
# list index?
idx = None
try:
idx = int(part)
except Exception:
idx = None
last = (i == len(parts) - 1)
if idx is not None:
if not isinstance(cur, list) or idx < 0 or idx >= len(cur):
return
if last:
cur[idx] = value
return
cur = cur[idx]
continue
# dict key
if not isinstance(cur, dict):
return
if last:
cur[part] = value
return
if part not in cur or not isinstance(cur[part], (dict, list)):
cur[part] = {}
cur = cur[part]
# Deep-copy only the top-level roots we are going to mutate (e.g., 'incoming' for 'incoming.*')
roots_to_copy: set = set()
for _p in (pre_var_overrides or {}).keys():
if "." in str(_p):
roots_to_copy.add(str(_p).split(".", 1)[0].strip())
for _root in roots_to_copy:
try:
if _root in render_ctx:
render_ctx[_root] = _safe_deepcopy(render_ctx[_root])
except Exception:
pass
for _p, _v in (pre_var_overrides or {}).items():
_set_by_path(render_ctx, _p, _v)
pre_var_paths = set(pre_var_overrides.keys())
except Exception:
pre_var_paths = set()
# SSE: prompt_preprocess summary
try:
trace_fn = context.get("_trace")
if trace_fn:
await trace_fn({
"event": "prompt_preprocess",
"node_id": self.node_id,
"node_type": self.type_name,
"provider": provider,
"lines": len(lines),
"used": pre_used,
"removed_keys": pre_removed_total,
"ts": int(time.time() * 1000),
})
except Exception:
pass
else:
pre_segments_raw = []
# prompt_combine (DSL "&"): комбинируем сегменты в заданном порядке.
# Расширения:
# - [[PROMPT]] как специальный сегмент (позиционно вставляет pm_struct)
# - Директивы @pos=prepend|append|<idx> (управление позицией в массиве)
# - Фильтрация пустых сообщений/частей
try:
cfg = self.config or {}
combine_raw = str(cfg.get("prompt_combine") or "").strip()
except Exception:
combine_raw = ""
# Быстрая ветка: если есть адаптер — выполняем merge сегментов через него и коротко-замыкаем легаси-блок ниже
raw_segs_for_adapter = [s.strip() for s in combine_raw.split("&") if str(s or "").strip()]
if adapter and (raw_segs_for_adapter or (pre_segments_raw and len(pre_segments_raw) > 0)):
try:
pm_struct = adapter.combine_segments(
blocks_struct=blocks_struct,
pre_segments_raw=pre_segments_raw,
raw_segs=raw_segs_for_adapter,
render_ctx=render_ctx,
pre_var_paths=pre_var_paths,
render_template_simple_fn=render_template_simple,
var_macro_fullmatch_re=_VAR_MACRO_RE,
detect_vendor_fn=detect_vendor,
)
# обновим pm в контексте
try:
render_ctx["pm"] = pm_struct
except Exception:
pass
except Exception as _e:
try:
print(f"TRACE adapter_combine_error: node={self.node_id} provider={provider} err={_e}")
except Exception:
pass
else:
# коротко-замкнём легаси-блок prompt_combine ниже
pre_segments_raw = []
combine_raw = ""
if combine_raw or (pre_segments_raw and len(pre_segments_raw) > 0):
raw_segs = [s.strip() for s in combine_raw.split("&") if str(s or "").strip()]
def _try_json(s: str) -> Any:
"""
Try to parse JSON from a string.
Unwrap double-encoded JSON strings up to two passes and use a permissive fallback.
Returns parsed object (dict/list/primitive) or None on failure.
"""
try:
obj = json.loads(s)
except Exception:
# permissive fallback for odd control characters
try:
obj = json.loads(s, strict=False) # type: ignore[call-arg]
except Exception:
return None
# If decoded into a string that itself looks like JSON — try to parse again up to 2 times.
for _ in range(2):
if isinstance(obj, str):
st = obj.strip()
if (st.startswith("{") and st.endswith("}")) or (st.startswith("[") and st.endswith("]")):
try:
obj = json.loads(st)
continue
except Exception:
break
break
return obj
# Нормализация в формат провайдера
def _as_openai_messages(x: Any) -> List[Dict[str, Any]]:
msgs: List[Dict[str, Any]] = []
try:
# Dict inputs
if isinstance(x, dict):
if isinstance(x.get("messages"), list):
# Already OpenAI-style
return list(x.get("messages") or [])
if isinstance(x.get("contents"), list):
# Gemini -> OpenAI (text-only join)
for c in (x.get("contents") or []):
if not isinstance(c, dict):
continue
role_raw = str(c.get("role") or "user")
role = "assistant" if role_raw == "model" else ("user" if role_raw not in {"user","assistant"} else role_raw)
parts = c.get("parts") or []
text = "\n".join([str(p.get("text")) for p in parts if isinstance(p, dict) and isinstance(p.get("text"), str)]).strip()
msgs.append({"role": role, "content": text})
return msgs
# List inputs
if isinstance(x, list):
# Gemini contents list -> OpenAI messages
if all(isinstance(c, dict) and "parts" in c for c in x):
for c in x:
role_raw = str(c.get("role") or "user")
role = "assistant" if role_raw == "model" else ("user" if role_raw not in {"user","assistant"} else role_raw)
parts = c.get("parts") or []
text = "\n".join([str(p.get("text")) for p in parts if isinstance(p, dict) and isinstance(p.get("text"), str)]).strip()
msgs.append({"role": role, "content": text})
return msgs
# OpenAI messages list already — normalize parts if needed
if all(isinstance(m, dict) and "content" in m for m in x):
out: List[Dict[str, Any]] = []
for m in x:
role = m.get("role", "user")
cont = m.get("content")
if isinstance(cont, str):
out.append({"role": role, "content": cont})
elif isinstance(cont, list):
parts2: List[Dict[str, Any]] = []
for p in cont:
if not isinstance(p, dict):
continue
if p.get("type") == "text":
parts2.append({"type": "text", "text": str(p.get("text") or "")})
elif p.get("type") in {"image_url", "image"}:
# try OpenAI image_url shape
url = ""
if isinstance(p.get("image_url"), dict):
url = str((p.get("image_url") or {}).get("url") or "")
elif "url" in p:
url = str(p.get("url") or "")
if url:
parts2.append({"type": "image_url", "image_url": {"url": url}})
out.append({"role": role, "content": parts2 if parts2 else ""})
return out
# Fallback: dump JSON as a single user message
return [{"role":"user","content": json.dumps(x, ensure_ascii=False)}]
# Primitive inputs or embedded JSON string
if isinstance(x, str):
try_obj = _try_json(x)
if try_obj is not None:
return _as_openai_messages(try_obj)
return [{"role":"user","content": x}]
return [{"role":"user","content": json.dumps(x, ensure_ascii=False)}]
except Exception:
return [{"role":"user","content": str(x)}]
def _as_gemini_contents(x: Any) -> List[Dict[str, Any]]:
cnts: List[Dict[str, Any]] = []
try:
if isinstance(x, dict):
if isinstance(x.get("contents"), list):
return list(x.get("contents") or [])
if isinstance(x.get("messages"), list):
# OpenAI → Gemini
for m in (x.get("messages") or []):
if not isinstance(m, dict):
continue
role_raw = str(m.get("role") or "user")
role = "model" if role_raw == "assistant" else "user"
cont = m.get("content")
parts: List[Dict[str, Any]] = []
if isinstance(cont, str):
parts = [{"text": cont}]
elif isinstance(cont, list):
for p in cont:
if not isinstance(p, dict):
continue
if p.get("type") == "text":
parts.append({"text": str(p.get("text") or "")})
elif p.get("type") in {"image_url", "image"}:
# Gemini doesn't accept external image URLs directly — keep as text fallback
url = ""
if isinstance(p.get("image_url"), dict):
url = str((p.get("image_url") or {}).get("url") or "")
elif "url" in p:
url = str(p.get("url") or "")
if url:
parts.append({"text": url})
else:
parts = [{"text": json.dumps(cont, ensure_ascii=False)}]
cnts.append({"role": role, "parts": parts})
return cnts
if isinstance(x, list):
# Gemini contents list already
if all(isinstance(c, dict) and "parts" in c for c in x):
return list(x)
# OpenAI messages list -> Gemini
if all(isinstance(m, dict) and "content" in m for m in x):
out: List[Dict[str, Any]] = []
for m in x:
role_raw = str(m.get("role") or "user")
role = "model" if role_raw == "assistant" else "user"
cont = m.get("content")
parts: List[Dict[str, Any]] = []
if isinstance(cont, str):
parts = [{"text": cont}]
elif isinstance(cont, list):
for p in cont:
if not isinstance(p, dict):
continue
if p.get("type") == "text":
parts.append({"text": str(p.get("text") or "")})
elif p.get("type") in {"image_url", "image"}:
url = ""
if isinstance(p.get("image_url"), dict):
url = str((p.get("image_url") or {}).get("url") or "")
elif "url" in p:
url = str(p.get("url") or "")
if url:
parts.append({"text": url})
else:
parts = [{"text": json.dumps(cont, ensure_ascii=False)}]
out.append({"role": role, "parts": parts})
return out
# Fallback
return [{"role":"user","parts":[{"text": json.dumps(x, ensure_ascii=False)}]}]
if isinstance(x, str):
try_obj = _try_json(x)
if try_obj is not None:
return _as_gemini_contents(try_obj)
return [{"role":"user","parts":[{"text": x}]}]
return [{"role":"user","parts":[{"text": json.dumps(x, ensure_ascii=False)}]}]
except Exception:
return [{"role":"user","parts":[{"text": str(x)}]}]
def _as_claude_messages(x: Any) -> List[Dict[str, Any]]:
msgs: List[Dict[str, Any]] = []
try:
if isinstance(x, dict):
# Dict with messages (OpenAI-like)
if isinstance(x.get("messages"), list):
x = x.get("messages") or []
# fallthrough to list mapping below
elif isinstance(x.get("contents"), list):
# Gemini -> Claude
for c in (x.get("contents") or []):
if not isinstance(c, dict):
continue
role_raw = str(c.get("role") or "user")
role = "assistant" if role_raw == "model" else ("user" if role_raw not in {"user","assistant"} else role_raw)
parts = c.get("parts") or []
text = "\n".join([str(p.get("text")) for p in parts if isinstance(p, dict) and isinstance(p.get("text"), str)]).strip()
msgs.append({"role": role, "content":[{"type":"text","text": text}]})
return msgs
if isinstance(x, list):
# Gemini contents list -> Claude messages
if all(isinstance(c, dict) and "parts" in c for c in x):
for c in x:
role_raw = str(c.get("role") or "user")
role = "assistant" if role_raw == "model" else ("user" if role_raw not in {"user","assistant"} else role_raw)
blocks: List[Dict[str, Any]] = []
for p in (c.get("parts") or []):
if isinstance(p, dict) and isinstance(p.get("text"), str):
txt = p.get("text").strip()
if txt:
blocks.append({"type":"text","text": txt})
msgs.append({"role": role, "content": blocks or [{"type":"text","text": ""}]})
return msgs
# OpenAI messages list -> Claude
if all(isinstance(m, dict) and "content" in m for m in x):
out: List[Dict[str, Any]] = []
for m in x:
role = m.get("role", "user")
cont = m.get("content")
blocks: List[Dict[str, Any]] = []
if isinstance(cont, str):
blocks.append({"type":"text","text": cont})
elif isinstance(cont, list):
for p in cont:
if not isinstance(p, dict):
continue
if p.get("type") == "text":
blocks.append({"type":"text","text": str(p.get("text") or "")})
elif p.get("type") in {"image_url", "image"}:
url = ""
if isinstance(p.get("image_url"), dict):
url = str((p.get("image_url") or {}).get("url") or "")
elif "url" in p:
url = str(p.get("url") or "")
if url:
blocks.append({"type":"image","source":{"type":"url","url": url}})
else:
blocks.append({"type":"text","text": json.dumps(cont, ensure_ascii=False)})
out.append({"role": role if role in {"user","assistant"} else "user", "content": blocks})
return out
# Fallback
return [{"role":"user","content":[{"type":"text","text": json.dumps(x, ensure_ascii=False)}]}]
if isinstance(x, str):
try_obj = _try_json(x)
if try_obj is not None:
return _as_claude_messages(try_obj)
return [{"role":"user","content":[{"type":"text","text": x}]}]
return [{"role":"user","content":[{"type":"text","text": json.dumps(x, ensure_ascii=False)}]}]
except Exception:
return [{"role":"user","content":[{"type":"text","text": str(x)}]}]
# Helper: extract system text from mixed-provider fragments
def _extract_sys_text_from_obj(x: Any) -> Optional[str]:
"""
Extract textual 'system' from provider-like fragments:
- Claude: top-level 'system' (str or list of {type:'text', text:'...'})
- Gemini: 'systemInstruction' with parts[].text
- OpenAI: messages[*] with role == 'system' (string content or parts[].text)
Returns a string or None.
"""
try:
# 1) Dict objects
if isinstance(x, dict):
# Gemini systemInstruction
if "systemInstruction" in x:
si = x.get("systemInstruction")
def _parts_to_text(siobj: Any) -> str:
try:
parts = siobj.get("parts") or []
texts = [
str(p.get("text") or "")
for p in parts
if isinstance(p, dict) and isinstance(p.get("text"), str) and p.get("text").strip()
]
return "\n".join([t for t in texts if t]).strip()
except Exception:
return ""
if isinstance(si, dict):
t = _parts_to_text(si)
if t:
return t
if isinstance(si, list):
texts = []
for p in si:
if isinstance(p, dict) and isinstance(p.get("text"), str) and p.get("text").strip():
texts.append(p.get("text").strip())
t = "\n".join(texts).strip()
if t:
return t
if isinstance(si, str) and si.strip():
return si.strip()
# Claude system (string or blocks)
if "system" in x and not ("messages" in x and isinstance(x.get("messages"), list)):
sysv = x.get("system")
if isinstance(sysv, str) and sysv.strip():
return sysv.strip()
if isinstance(sysv, list):
texts = [
str(b.get("text") or "")
for b in sysv
if isinstance(b, dict)
and (b.get("type") == "text")
and isinstance(b.get("text"), str)
and b.get("text").strip()
]
t = "\n".join([t for t in texts if t]).strip()
if t:
return t
# OpenAI messages with role=system
if isinstance(x.get("messages"), list):
sys_msgs = []
for m in (x.get("messages") or []):
try:
if (str(m.get("role") or "").lower().strip() == "system"):
cont = m.get("content")
if isinstance(cont, str) and cont.strip():
sys_msgs.append(cont.strip())
elif isinstance(cont, list):
for p in cont:
if isinstance(p, dict) and p.get("type") == "text" and isinstance(p.get("text"), str) and p.get("text").strip():
sys_msgs.append(p.get("text").strip())
except Exception:
continue
if sys_msgs:
return "\n\n".join(sys_msgs).strip()
# 2) List objects
if isinstance(x, list):
# OpenAI messages list with role=system
if all(isinstance(m, dict) and "role" in m for m in x):
sys_msgs = []
for m in x:
try:
if (str(m.get("role") or "").lower().strip() == "system"):
cont = m.get("content")
if isinstance(cont, str) and cont.strip():
sys_msgs.append(cont.strip())
elif isinstance(cont, list):
for p in cont:
if isinstance(p, dict) and p.get("type") == "text" and isinstance(p.get("text"), str) and p.get("text").strip():
sys_msgs.append(p.get("text").strip())
except Exception:
continue
if sys_msgs:
return "\n\n".join(sys_msgs).strip()
# Gemini 'contents' list: try to read systemInstruction from incoming JSON snapshot
if all(isinstance(c, dict) and "parts" in c for c in x):
try:
inc = (render_ctx.get("incoming") or {}).get("json") or {}
si = inc.get("systemInstruction")
if si is not None:
return _extract_sys_text_from_obj({"systemInstruction": si})
except Exception:
pass
return None
except Exception:
return None
# Фильтр пустых сообщений/частей
def _filter_gemini(arr: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
out: List[Dict[str, Any]] = []
for it in (arr or []):
if not isinstance(it, dict):
continue
parts = it.get("parts") or []
norm_parts = []
for p in parts:
if isinstance(p, dict):
t = p.get("text")
if isinstance(t, str) and t.strip():
norm_parts.append({"text": t})
elif "inline_data" in p or "inlineData" in p:
norm_parts.append(p) # изображения пропускаем как есть
if norm_parts:
out.append({"role": it.get("role","user"), "parts": norm_parts})
return out
def _filter_openai(arr: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
out: List[Dict[str, Any]] = []
for m in (arr or []):
if not isinstance(m, dict):
continue
c = m.get("content")
if isinstance(c, str) and c.strip():
out.append({"role": m.get("role","user"), "content": c})
elif isinstance(c, list):
parts = []
for p in c:
if isinstance(p, dict) and p.get("type") == "text":
txt = str(p.get("text") or "")
if txt.strip():
parts.append({"type":"text","text": txt})
if parts:
out.append({"role": m.get("role","user"), "content": parts})
return out
def _filter_claude(arr: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
out: List[Dict[str, Any]] = []
for m in (arr or []):
if not isinstance(m, dict):
continue
blocks = m.get("content")
if isinstance(blocks, list):
norm = []
for b in blocks:
if isinstance(b, dict) and b.get("type") == "text":
txt = str(b.get("text") or "")
if txt.strip():
norm.append({"type":"text","text": txt})
if norm:
out.append({"role": m.get("role","user"), "content": norm})
return out
# Построим массивы по сегментам в заданном порядке.
# Обработка [[PROMPT]] как спец.сегмента сохранив роль/части из pm_struct.
# Поддержка @pos=... только внутри текущего собираемого массива (messages/contents).
seg_items: List[Any] = []
# Снимок blocks как provider-структуры
blocks_struct = dict(pm_struct)
def _insert_items(base: List[Any], items: List[Any], pos_spec: Optional[str]) -> List[Any]:
if not items:
return base
if not pos_spec or pos_spec.lower() == "append":
base.extend(items)
return base
if pos_spec.lower() == "prepend":
return list(items) + base
# индекс
try:
idx = int(pos_spec)
if idx < 0:
idx = len(base) + idx
if idx < 0:
idx = 0
if idx > len(base): idx = len(base)
return base[:idx] + list(items) + base[idx:]
except Exception:
base.extend(items)
return base
# парсинг директивы @pos=
import re as _re2
def _split_pos_spec(s: str) -> tuple[str, Optional[str]]:
m = _re2.search(r"@pos\s*=\s*(prepend|append|-?\d+)\s*$", s, flags=_re2.IGNORECASE)
if not m:
return s.strip(), None
body = s[:m.start()].strip()
return body, m.group(1).strip().lower()
if provider in {"gemini", "gemini_image"}:
built: List[Dict[str, Any]] = []
sys_texts: List[str] = []
# Preprocess-inserted segments (prompt_preprocess)
for _pre in pre_segments_raw:
try:
_obj = _pre.get("obj")
items = _as_gemini_contents(_obj)
items = _filter_gemini(items)
built = _insert_items(built, items, _pre.get("pos"))
try:
sx = _extract_sys_text_from_obj(_obj)
if isinstance(sx, str) and sx.strip():
sys_texts.append(sx.strip())
except Exception:
pass
except Exception:
pass
for raw_seg in raw_segs:
body_seg, pos_spec = _split_pos_spec(raw_seg)
if body_seg == "[[PROMPT]]":
items = _filter_gemini(list(blocks_struct.get("contents", []) or []))
built = _insert_items(built, items, pos_spec)
continue
m_pre = _VAR_MACRO_RE.fullmatch(body_seg)
if m_pre:
_p = (m_pre.group(1) or "").strip()
try:
if _p in pre_var_paths:
# Skip duplicate var segment - already inserted via prompt_preprocess (filtered)
continue
except Exception:
pass
resolved = render_template_simple(body_seg, render_ctx, render_ctx.get("OUT") or {})
obj = _try_json(resolved)
# provider guess + system extract for cross-provider combine
try:
pg = detect_vendor(obj if isinstance(obj, dict) else {})
print(f"DEBUG: prompt_combine seg provider_guess={pg} -> target=gemini pos={pos_spec}")
except Exception:
pass
items = _as_gemini_contents(obj if obj is not None else resolved)
items = _filter_gemini(items)
built = _insert_items(built, items, pos_spec)
try:
sx = _extract_sys_text_from_obj(obj) if obj is not None else None
if isinstance(sx, str) and sx.strip():
sys_texts.append(sx.strip())
except Exception:
pass
# Если ни один сегмент не был [[PROMPT]] и массив пуст — оставим исходные blocks
if not built:
built = _filter_gemini(list(blocks_struct.get("contents", []) or []))
# Merge systemInstruction: PROMPT blocks + gathered sys_texts
existing_si = blocks_struct.get("systemInstruction")
parts = []
if isinstance(existing_si, dict) and isinstance(existing_si.get("parts"), list):
parts = list(existing_si.get("parts") or [])
for s in sys_texts:
parts.append({"text": s})
new_si = {"parts": parts} if parts else existing_si
pm_struct = {"contents": built, "systemInstruction": new_si, "system_text": blocks_struct.get("system_text")}
elif provider in {"openai"}:
built2: List[Dict[str, Any]] = []
sys_texts: List[str] = []
# Preprocess-inserted segments (prompt_preprocess)
for _pre in pre_segments_raw:
try:
_obj = _pre.get("obj")
items = _as_openai_messages(_obj)
items = _filter_openai(items)
built2 = _insert_items(built2, items, _pre.get("pos"))
try:
sx = _extract_sys_text_from_obj(_obj)
if isinstance(sx, str) and sx.strip():
sys_texts.append(sx.strip())
except Exception:
pass
except Exception:
pass
for raw_seg in raw_segs:
body_seg, pos_spec = _split_pos_spec(raw_seg)
if body_seg == "[[PROMPT]]":
items = _filter_openai(list(blocks_struct.get("messages", []) or []))
built2 = _insert_items(built2, items, pos_spec)
continue
m_pre = _VAR_MACRO_RE.fullmatch(body_seg)
if m_pre:
_p = (m_pre.group(1) or "").strip()
try:
if _p in pre_var_paths:
# Skip duplicate var segment - already inserted via prompt_preprocess (filtered)
continue
except Exception:
pass
resolved = render_template_simple(body_seg, render_ctx, render_ctx.get("OUT") or {})
obj = _try_json(resolved)
try:
pg = detect_vendor(obj if isinstance(obj, dict) else {})
print(f"DEBUG: prompt_combine seg provider_guess={pg} -> target=openai pos={pos_spec}")
except Exception:
pass
items = _as_openai_messages(obj if obj is not None else resolved)
items = _filter_openai(items)
built2 = _insert_items(built2, items, pos_spec)
# collect only top-level system/systemInstruction (not role=system in messages)
try:
sx = _extract_sys_text_from_obj(obj) if obj is not None else None
if isinstance(sx, str) and sx.strip():
sys_texts.append(sx.strip())
except Exception:
pass
if not built2:
built2 = _filter_openai(list(blocks_struct.get("messages", []) or []))
# Prepend system messages derived from sys_texts
if sys_texts:
sys_msgs = [{"role": "system", "content": s} for s in sys_texts if s]
if sys_msgs:
built2 = sys_msgs + built2
# keep system_text for UI/debug
st0 = blocks_struct.get("system_text") or ""
st = "\n\n".join([t for t in [st0] + sys_texts if isinstance(t, str) and t.strip()])
pm_struct = {"messages": built2, "system_text": st}
else: # claude
built3: List[Dict[str, Any]] = []
sys_texts: List[str] = []
# Preprocess-inserted segments (prompt_preprocess)
for _pre in pre_segments_raw:
try:
_obj = _pre.get("obj")
items = _as_claude_messages(_obj)
items = _filter_claude(items)
built3 = _insert_items(built3, items, _pre.get("pos"))
try:
sx = _extract_sys_text_from_obj(_obj)
if isinstance(sx, str) and sx.strip():
sys_texts.append(sx.strip())
except Exception:
pass
except Exception:
pass
for raw_seg in raw_segs:
body_seg, pos_spec = _split_pos_spec(raw_seg)
if body_seg == "[[PROMPT]]":
items = _filter_claude(list(blocks_struct.get("messages", []) or []))
built3 = _insert_items(built3, items, pos_spec)
continue
m_pre = _VAR_MACRO_RE.fullmatch(body_seg)
if m_pre:
_p = (m_pre.group(1) or "").strip()
try:
if _p in pre_var_paths:
# Skip duplicate var segment - already inserted via prompt_preprocess (filtered)
continue
except Exception:
pass
resolved = render_template_simple(body_seg, render_ctx, render_ctx.get("OUT") or {})
obj = _try_json(resolved)
try:
pg = detect_vendor(obj if isinstance(obj, dict) else {})
print(f"DEBUG: prompt_combine seg provider_guess={pg} -> target=claude pos={pos_spec}")
except Exception:
pass
items = _as_claude_messages(obj if obj is not None else resolved)
items = _filter_claude(items)
built3 = _insert_items(built3, items, pos_spec)
try:
sx = _extract_sys_text_from_obj(obj) if obj is not None else None
if isinstance(sx, str) and sx.strip():
sys_texts.append(sx.strip())
except Exception:
pass
if not built3:
built3 = _filter_claude(list(blocks_struct.get("messages", []) or []))
# Merge system blocks from PROMPT blocks + gathered sys_texts
existing_sys = blocks_struct.get("system") or []
sys_blocks = []
if isinstance(existing_sys, list):
sys_blocks.extend(existing_sys)
st0 = blocks_struct.get("system_text") or ""
# Ensure PROMPT system_text from blocks is included as a Claude system block
if isinstance(st0, str) and st0.strip():
sys_blocks.append({"type": "text", "text": st0})
for s in sys_texts:
sys_blocks.append({"type": "text", "text": s})
st = "\n\n".join([t for t in [st0] + sys_texts if isinstance(t, str) and t.strip()])
# New: optional Claude mode to omit top-level "system"
claude_no_system = False
try:
claude_no_system = bool((self.config or {}).get("claude_no_system", False))
except Exception:
claude_no_system = False
if claude_no_system:
# Prepend system text as a user message instead of top-level system
if st:
built3 = [{"role": "user", "content": [{"type": "text", "text": st}]}] + built3
pm_struct = {"messages": built3, "system_text": st}
else:
# Prefer top-level system as plain string (proxy compatibility)
pm_struct = {"messages": built3, "system_text": st}
# Prefer array of system blocks when possible; fallback to single text block
if sys_blocks:
pm_struct["system"] = sys_blocks
elif st:
pm_struct["system"] = [{"type": "text", "text": st}]
# SSE метрика
try:
trace_fn = context.get("_trace")
if trace_fn:
after_cnt = (len(pm_struct.get("messages", [])) if provider in {"openai","claude"} else len(pm_struct.get("contents", [])))
await trace_fn({
"event": "prompt_merge",
"node_id": self.node_id,
"node_type": self.type_name,
"provider": provider,
"segments": len(raw_segs),
"before": (len(blocks_struct.get("messages", [])) if provider in {"openai","claude"} else len(blocks_struct.get("contents", []))),
"after": after_cnt,
"ts": int(time.time()*1000),
})
except Exception:
pass
# Единый JSON-фрагмент PROMPT для шаблонов: [[PROMPT]]
prompt_fragment = ""
try:
if adapter:
prompt_fragment = adapter.prompt_fragment(pm_struct, self.config or {})
else:
if provider == "openai":
prompt_fragment = '"messages": ' + json.dumps(pm_struct.get("messages", []), ensure_ascii=False)
elif provider in {"gemini", "gemini_image"}:
parts = []
contents = pm_struct.get("contents")
if contents is not None:
parts.append('"contents": ' + json.dumps(contents, ensure_ascii=False))
sysi = pm_struct.get("systemInstruction")
if sysi is not None:
parts.append('"systemInstruction": ' + json.dumps(sysi, ensure_ascii=False))
prompt_fragment = ", ".join(parts)
elif provider == "claude":
parts = []
# Учитываем флаг совместимости: при claude_no_system не добавляем top-level "system"
claude_no_system = False
try:
claude_no_system = bool((self.config or {}).get("claude_no_system", False))
except Exception:
claude_no_system = False
if not claude_no_system:
# Предпочитаем массив блоков system, если он есть; иначе строковый system_text
sys_val = pm_struct.get("system", None)
if sys_val is None:
sys_val = pm_struct.get("system_text")
if sys_val:
parts.append('"system": ' + json.dumps(sys_val, ensure_ascii=False))
msgs = pm_struct.get("messages")
if msgs is not None:
parts.append('"messages": ' + json.dumps(msgs, ensure_ascii=False))
prompt_fragment = ", ".join(parts)
except Exception: # noqa: BLE001
prompt_fragment = ""
render_ctx["PROMPT"] = prompt_fragment
# Render helper с поддержкой [[VAR]], [[OUT]] и {{ ... }}
def render(s: str) -> str:
return render_template_simple(s or "", render_ctx, render_ctx.get("OUT") or {})
# Рендер endpoint с макросами/шаблонами
endpoint = render(endpoint_tmpl)
# Формируем тело ТОЛЬКО из template/[[PROMPT]] (без сырого payload/входов).
# Больше НИКАКОГО фоллбэка на unified-построение: если шаблон невалиден — это ошибка ноды.
try:
rendered = render(template)
# DEBUG: печать отрендеренного шаблона с номерами строк для точной диагностики JSONDecodeError
try:
_lines = rendered.splitlines()
_preview = "\n".join(f"{i+1:03d}: {_lines[i]}" for i in range(min(len(_lines), 120)))
print(f"DEBUG: ProviderCallNode rendered_template node={self.node_id} provider={provider}\\n{_preview}")
except Exception:
try:
print(f"DEBUG: ProviderCallNode rendered_template(node={self.node_id}, provider={provider}) len={len(rendered)}")
except Exception:
pass
payload = json.loads(rendered)
except Exception as exc: # noqa: BLE001
raise ExecutionError(f"ProviderCall template invalid JSON: {exc}")
# Заголовки — полностью из редактируемого JSON с макросами
try:
headers_src = render(headers_json) if headers_json else "{}"
headers = json.loads(headers_src) if headers_src else {}
if not isinstance(headers, dict):
raise ValueError("headers must be a JSON object")
# Drop headers with empty values (proxy compatibility, e.g. "anthropic-beta": "")
try:
headers = {k: v for k, v in headers.items() if v is not None and (not isinstance(v, str) or v.strip() != "")}
except Exception:
pass
except Exception as exc: # noqa: BLE001
raise ExecutionError(f"ProviderCall headers invalid JSON: {exc}")
# Итоговый URL
if not base_url.startswith(("http://", "https://")):
base_url = "http://" + base_url
url = endpoint if endpoint.startswith("http") else urljoin(base_url.rstrip('/') + '/', endpoint.lstrip('/'))
# Debug logs to validate config selection and payload
# Brute request/response logging (request body sanitized like Burp)
try:
final_headers = {"Content-Type": "application/json", **headers}
print("===== ProviderCall REQUEST BEGIN =====")
print(f"node={self.node_id} type={self.type_name} provider={provider}")
print(f"URL: {url}")
try:
print("Headers:")
print(json.dumps(final_headers, ensure_ascii=False, indent=2))
except Exception:
print(f"Headers(raw): {final_headers}")
try:
san_payload = _sanitize_b64_for_log(payload, max_len=180)
print("Body JSON:")
print(json.dumps(san_payload, ensure_ascii=False, indent=2))
except Exception:
print(f"Body(raw): {payload}")
print("===== ProviderCall REQUEST END =====")
except Exception:
pass
# SSE: http_req (как в Burp)
req_id = f"{self.node_id}-{int(time.time()*1000)}"
# Register original (untrimmed) request details for manual resend
try:
pipeline_id = str((context.get("meta") or {}).get("id", "pipeline_editor"))
except Exception:
pipeline_id = "pipeline_editor"
try:
register_http_request(req_id, {
"pipeline_id": pipeline_id,
"node_id": self.node_id,
"node_type": self.type_name,
"provider": provider,
"method": "POST",
"url": url,
"headers": dict(final_headers),
"body_json": payload, # untrimmed original payload
})
except Exception:
pass
try:
trace_fn = context.get("_trace")
if trace_fn:
await trace_fn({
"event": "http_req",
"node_id": self.node_id,
"node_type": self.type_name,
"provider": provider,
"req_id": req_id,
"method": "POST",
"url": url,
"headers": final_headers,
"body_text": json.dumps(_sanitize_b64_for_log(payload, max_len=180), ensure_ascii=False, indent=2),
"ts": int(time.time() * 1000),
})
except Exception:
pass
# Timeout from run-meta (seconds); fallback to 60
try:
timeout_sec = float((context.get("meta") or {}).get("http_timeout_sec", 60) or 60)
except Exception:
timeout_sec = 60.0
st: Optional[int] = None
async with build_client(timeout=timeout_sec) as client:
body_bytes = json.dumps(payload, ensure_ascii=False).encode("utf-8")
# Cooperative cancel pre-check (avoid starting new HTTP on abort)
try:
pipeline_id = str((context.get("meta") or {}).get("id", "pipeline_editor"))
except Exception:
pipeline_id = "pipeline_editor"
try:
if is_cancelled(pipeline_id) and get_cancel_mode(pipeline_id) == "abort":
try:
print(f"TRACE http_cancel_pre: {self.node_id} abort before request")
except Exception:
pass
raise ExecutionError("Cancelled by user (abort)")
except Exception:
pass
async def _do_post():
return await client.post(url, content=body_bytes, headers=final_headers)
# Await HTTP with cooperative cancel/abort handling
resp = await _await_coro_with_cancel(_do_post(), pipeline_id)
# Do not raise_for_status: keep body/logs on 4xx/5xx
try:
print("===== ProviderCall RESPONSE BEGIN =====")
print(f"node={self.node_id} type={self.type_name} provider={provider}")
print(f"Status: {resp.status_code}")
try:
print("Headers:")
print(json.dumps(dict(resp.headers), ensure_ascii=False, indent=2))
except Exception:
try:
print(f"Headers(raw): {dict(resp.headers)}")
except Exception:
print("Headers(raw): <unavailable>")
# Печатаем JSON тела с ТРИММИНГОМ ТОЛЬКО base64 строк, оставляя остальную структуру нетронутой
try:
raw_obj = _safe_response_json(resp)
import re as _re
def _is_b64ish(s: str) -> bool:
try:
if not isinstance(s, str) or len(s) < 128:
return False
# быстрый тест на base64 (разрешаем переводы строк)
return bool(_re.fullmatch(r"[A-Za-z0-9+/=\r\n]+", s))
except Exception:
return False
def _trim_b64_in_obj(x):
if isinstance(x, dict):
out = {}
for k, v in x.items():
if isinstance(v, str) and "data" in str(k).lower() and _is_b64ish(v) and len(v) > 256:
out[k] = v[:180] + f"... (trimmed {len(v)-180})"
else:
out[k] = _trim_b64_in_obj(v)
return out
if isinstance(x, list):
return [_trim_b64_in_obj(i) for i in x]
return x
sanitized_obj = _trim_b64_in_obj(raw_obj if isinstance(raw_obj, (dict, list)) else {})
body_text = json.dumps(sanitized_obj if sanitized_obj else raw_obj, ensure_ascii=False, indent=2)
except Exception:
body_text = _safe_response_text(resp)
print("Body JSON:")
print(body_text)
print("===== ProviderCall RESPONSE END =====")
except Exception:
body_text = "<resp.text decode error>"
pass
try:
data = _safe_response_json(resp)
except Exception:
data = {"error": "Failed to decode JSON from upstream", "text": _safe_response_text(resp)}
# Построим данные для панели Data:
# - images_data: полный список inline картинок (mime+base64, НЕ обрезанные)
# - data_preview: краткая сводка, если картинок нет или UI не может их отрисовать
try:
def _collect_images_http(obj: Any) -> List[Dict[str, str]]:
imgs: List[Dict[str, str]] = []
def walk(o: Any) -> None:
if isinstance(o, dict):
# Gemini: inlineData / inline_data
if isinstance(o.get("inlineData"), dict):
idat = o.get("inlineData") or {}
b64 = str(idat.get("data") or "")
mime = str(idat.get("mimeType") or idat.get("mime_type") or "image/png")
if b64:
imgs.append({"mime": mime, "data": b64})
if isinstance(o.get("inline_data"), dict):
idat = o.get("inline_data") or {}
b64 = str(idat.get("data") or "")
mime = str(idat.get("mimeType") or idat.get("mime_type") or "image/png")
if b64:
imgs.append({"mime": mime, "data": b64})
# Claude style
if (o.get("type") == "image") and isinstance(o.get("source"), dict):
src = o.get("source") or {}
if (src.get("type") or src.get("source_type")) in {"base64", "inline"}:
b64 = str(src.get("data") or "")
mime = str(src.get("media_type") or src.get("mediaType") or "image/png")
if b64:
imgs.append({"mime": mime, "data": b64})
for v in o.values():
walk(v)
elif isinstance(o, list):
for v in o:
walk(v)
walk(obj)
return imgs
def _collect_image_summaries_http(obj: Any) -> List[str]:
info: List[str] = []
def walk(o: Any, path: str = "") -> None:
if isinstance(o, dict):
if isinstance(o.get("inlineData"), dict):
idat = o.get("inlineData") or {}
b64 = str(idat.get("data") or "")
mime = str(idat.get("mimeType") or idat.get("mime_type") or "")
info.append(f"{path}/inlineData mime={mime or '?'} b64_len={len(b64)}")
if isinstance(o.get("inline_data"), dict):
idat = o.get("inline_data") or {}
b64 = str(idat.get("data") or "")
mime = str(idat.get("mimeType") or idat.get("mime_type") or "")
info.append(f"{path}/inline_data mime={mime or '?'} b64_len={len(b64)}")
if (o.get("type") == "image") and isinstance(o.get("source"), dict):
src = o.get("source") or {}
if (src.get("type") or src.get("source_type")) in {"base64", "inline"}:
b64 = str(src.get("data") or "")
mime = str(src.get("media_type") or src.get("mediaType") or "")
info.append(f"{path}/source mime={mime or '?'} b64_len={len(b64)}")
for k, v in o.items():
walk(v, path + (("." + k) if path else k))
elif isinstance(o, list):
for i, v in enumerate(o):
walk(v, path + f"[{i}]")
walk(obj, "")
return info
images_data = _collect_images_http(data)
_img_lines_http = _collect_image_summaries_http(data)
data_preview = ""
if _img_lines_http:
data_preview = "images: " + str(len(_img_lines_http)) + "\n" + "\n".join(_img_lines_http[:12])
except Exception:
images_data = []
data_preview = ""
# SSE: http_resp
try:
trace_fn = context.get("_trace")
if trace_fn:
await trace_fn({
"event": "http_resp",
"node_id": self.node_id,
"node_type": self.type_name,
"provider": provider,
"req_id": req_id,
"status": int(resp.status_code),
"headers": dict(resp.headers),
"body_text": body_text if isinstance(body_text, str) else str(body_text),
"data_preview": data_preview,
"images": images_data,
"ts": int(time.time() * 1000),
})
except Exception:
pass
try:
st = int(resp.status_code)
except Exception:
st = None
# Извлечение текста: приоритет — пресет ноды -> кастомн. поля ноды -> глобальные meta
try:
meta_opts = (context.get("meta") or {})
presets = meta_opts.get("text_extract_presets") or []
cfg = self.config or {}
strategy = None
json_path = None
join_sep = None
# 1) Пресет по id
pid = str(cfg.get("text_extract_preset_id") or "").strip()
if pid:
try:
pr = next((x for x in presets if isinstance(x, dict) and str(x.get("id", "")) == pid), None)
except Exception:
pr = None
if pr:
strategy = str(pr.get("strategy") or "auto")
json_path = str(pr.get("json_path") or "")
join_sep = str(pr.get("join_sep") or "\n")
# 2) Кастомные поля ноды (если задано что-то из набора)
if not strategy and (cfg.get("text_extract_strategy") is not None or cfg.get("text_extract_json_path") is not None or cfg.get("text_join_sep") is not None):
strategy = str(cfg.get("text_extract_strategy") or "auto")
json_path = str(cfg.get("text_extract_json_path") or "")
join_sep = str(cfg.get("text_join_sep") or "\n")
# 3) Глобальные метаданные запуска
if not strategy:
strategy = str(meta_opts.get("text_extract_strategy", "auto") or "auto")
json_path = str(meta_opts.get("text_extract_json_path", "") or "")
join_sep = str(meta_opts.get("text_join_sep", "\n") or "\n")
except Exception:
strategy, json_path, join_sep = "auto", "", "\n"
text = _extract_text_for_out(data, strategy, provider, json_path, join_sep)
# Если текст пуст (например, пришли только inlineData/base64-картинки) — сформируем краткое резюме картинок,
# чтобы панель Data не оставалась пустой.
if not text:
try:
def _collect_image_summaries(obj: Any) -> List[str]:
info: List[str] = []
def walk(o: Any, path: str = "") -> None:
if isinstance(o, dict):
# Gemini: inlineData / inline_data
if isinstance(o.get("inlineData"), dict):
idat = o.get("inlineData") or {}
b64 = str(idat.get("data") or "")
mime = str(idat.get("mimeType") or idat.get("mime_type") or "")
info.append(f"{path}/inlineData mime={mime or '?'} b64_len={len(b64)}")
if isinstance(o.get("inline_data"), dict):
idat = o.get("inline_data") or {}
b64 = str(idat.get("data") or "")
mime = str(idat.get("mimeType") or idat.get("mime_type") or "")
info.append(f"{path}/inline_data mime={mime or '?'} b64_len={len(b64)}")
# Claude: {"type":"image","source":{"type":"base64","media_type":..., "data":...}}
if (o.get("type") == "image") and isinstance(o.get("source"), dict):
src = o.get("source") or {}
if (src.get("type") or src.get("source_type")) in {"base64", "inline"}:
b64 = str(src.get("data") or "")
mime = str(src.get("media_type") or src.get("mediaType") or "")
info.append(f"{path}/source mime={mime or '?'} b64_len={len(b64)}")
for k, v in o.items():
walk(v, path + (("." + k) if path else k))
elif isinstance(o, list):
for i, v in enumerate(o):
walk(v, path + f"[{i}]")
walk(obj, "")
return info
img_lines = _collect_image_summaries(data)
if img_lines:
text = "images: " + str(len(img_lines)) + "\n" + "\n".join(img_lines[:12])
except Exception:
pass
# Подробный SSE-лог завершения ProviderCall
try:
trace_fn = context.get("_trace")
if trace_fn:
# Compute destination macros for extracted text
try:
to_path = f"OUT.{self.node_id}.response_text"
macro_braces = f"{{{{ OUT.{self.node_id}.response_text }}}}"
alias_macro = ""
m = re.match(r"^n(\d+)$", str(self.node_id))
if m:
alias_macro = f"[[OUT{int(m.group(1))}]]"
except Exception:
to_path = f"OUT.{self.node_id}.response_text"
macro_braces = f"{{{{ OUT.{self.node_id}.response_text }}}}"
alias_macro = ""
await trace_fn({
"event": "provider_done",
"node_id": self.node_id,
"node_type": self.type_name,
"provider": provider,
"url": url,
"req_id": req_id,
"status": int(st) if st is not None else None,
"text_len": int(len(text or "")),
"extracted_text": text or "",
"strategy": strategy,
"json_path": json_path or "",
"join_sep": join_sep or "\n",
"to_path": to_path,
"to_macro_outx": alias_macro,
"to_macro_braces": macro_braces,
"ts": int(time.time() * 1000),
})
except Exception:
pass
return {"result": data, "response_text": text or ""}
class RawForwardNode(Node):
type_name = "RawForward"
async def run(self, inputs: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]:
"""
Прямой форвард входящего запроса на апстрим:
- base_url/override_path/extra_headers поддерживают макросы [[...]] и {{ ... }}
- По умолчанию пробрасываем входящие заголовки (кроме Host/Content-Length), затем применяем extra_headers поверх.
- Тело берём из incoming.json (если есть), иначе пытаемся использовать текст/байты.
- Эмитим SSE события http_req/http_resp для панели логов.
"""
# While-loop wrapper: if while_expr is set or ignore_errors enabled, route through helper (single-run when expr empty)
try:
cfg = self.config or {}
except Exception:
cfg = {}
try:
_while_expr = str(cfg.get("while_expr") or "").strip()
except Exception:
_while_expr = ""
try:
_ignore = bool(cfg.get("ignore_errors", False))
except Exception:
_ignore = False
if not (context.get("_in_while") or False) and (_while_expr or _ignore):
return await _rawforward_run_with_while(self, inputs, context)
incoming = context.get("incoming", {}) or {}
raw_payload = incoming.get("json")
# Optional sleep before forwarding (UI-configured, milliseconds)
try:
sleep_ms = int(self.config.get("sleep_ms") or 0)
except Exception:
sleep_ms = 0
if sleep_ms > 0:
# Notify UI to show "sleep" state for this node
try:
trace_fn = context.get("_trace")
if trace_fn:
await trace_fn({
"event": "node_sleep",
"node_id": self.node_id,
"node_type": self.type_name,
"sleep_ms": int(sleep_ms),
"ts": int(time.time() * 1000),
})
except Exception:
pass
try:
await asyncio.sleep(max(0.0, sleep_ms / 1000.0))
except Exception:
pass
# Конфиг с поддержкой макросов
base_url: Optional[str] = self.config.get("base_url")
override_path: Optional[str] = self.config.get("override_path")
out_map_for_macros = context.get("OUT") or {}
if base_url:
base_url = render_template_simple(str(base_url), context, out_map_for_macros)
if override_path:
override_path = render_template_simple(str(override_path), context, out_map_for_macros)
# Автодетекция вендора для базового URL если base_url не задан
if not base_url:
vendor = detect_vendor(raw_payload)
base_url = None
try:
base_url = _adapter_default_base_url_for(vendor)
except Exception:
base_url = None
if not base_url:
if vendor == "openai":
base_url = "https://api.openai.com"
elif vendor == "claude":
base_url = "https://api.anthropic.com"
elif vendor == "gemini":
base_url = "https://generativelanguage.googleapis.com"
if not base_url:
raise ExecutionError(
f"Node {self.node_id} ({self.type_name}): 'base_url' is not configured and vendor could not be detected."
)
if not str(base_url).startswith(("http://", "https://")):
base_url = "http://" + str(base_url)
path = override_path or incoming.get("path") or "/"
query = incoming.get("query")
path_with_qs = f"{path}?{query}" if query else str(path)
url = urljoin(str(base_url).rstrip("/") + "/", str(path_with_qs).lstrip("/"))
# Заголовки: passthrough + extra_headers(JSON)
passthrough_headers: bool = bool(self.config.get("passthrough_headers", True))
extra_headers_json: str = self.config.get("extra_headers") or "{}"
try:
extra_headers_src = render_template_simple(extra_headers_json, context, out_map_for_macros) if extra_headers_json else "{}"
extra_headers = json.loads(extra_headers_src) if extra_headers_src else {}
if not isinstance(extra_headers, dict):
raise ValueError("extra_headers must be an object")
except Exception as exc: # noqa: BLE001
raise ExecutionError(f"RawForward extra_headers invalid JSON: {exc}")
headers: Dict[str, str] = {}
if passthrough_headers:
inc_headers = incoming.get("headers") or {}
for k, v in inc_headers.items():
# Не пробрасываем управляющие заголовки, которые рассчитает httpx/сервер
if k and k.lower() not in {"host", "content-length"}:
headers[k] = v
# extra_headers перекрывают проброс
try:
headers.update(extra_headers)
except Exception:
pass
# Нормализация: защитимся от повторного наличия host/content-length после update
for k in list(headers.keys()):
kl = k.lower()
if kl in {"host", "content-length"}:
try:
headers.pop(k, None)
except Exception:
pass
# Метод
method = str(incoming.get("method") or "POST").upper().strip() or "POST"
# Тело запроса
body_text = ""
body_bytes: bytes = b""
if raw_payload is not None:
# JSON — сериализуем явно как UTF8
try:
body_text = json.dumps(raw_payload, ensure_ascii=False, indent=2)
except Exception:
try:
body_text = str(raw_payload)
except Exception:
body_text = ""
try:
body_bytes = json.dumps(raw_payload, ensure_ascii=False).encode("utf-8")
except Exception:
body_bytes = (body_text or "").encode("utf-8", errors="ignore")
# Если Content-Type не задан — выставим JSON
if "content-type" not in {k.lower() for k in headers.keys()}:
headers["Content-Type"] = "application/json"
else:
# Попробуем текст/байты из incoming
raw_bytes = incoming.get("body_bytes") or incoming.get("bytes")
raw_text = incoming.get("body_text") or incoming.get("text")
if isinstance(raw_bytes, (bytes, bytearray)):
body_bytes = bytes(raw_bytes)
try:
body_text = body_bytes.decode("utf-8", errors="ignore")
except Exception:
body_text = ""
elif raw_text is not None:
body_text = str(raw_text)
try:
body_bytes = body_text.encode("utf-8")
except Exception:
body_bytes = (body_text or "").encode("utf-8", errors="ignore")
else:
body_text = ""
body_bytes = b""
# Таймаут из метаданных запуска
try:
timeout_sec = float((context.get("meta") or {}).get("http_timeout_sec", 60) or 60)
except Exception:
timeout_sec = 60.0
# Диагностический вывод полной заявки/ответа (request body sanitized)
try:
print("===== RawForward REQUEST BEGIN =====")
print(f"node={self.node_id} type={self.type_name}")
print(f"Method: {method}")
print(f"URL: {url}")
try:
print("Headers:")
print(json.dumps(headers, ensure_ascii=False, indent=2))
except Exception:
print(f"Headers(raw): {headers}")
print("Body:")
try:
san_body_text = _sanitize_json_string_for_log(body_text, max_len=180)
except Exception:
san_body_text = body_text
print(san_body_text)
print("===== RawForward REQUEST END =====")
except Exception:
pass
# SSE: http_req
req_id = f"{self.node_id}-{int(time.time()*1000)}"
# Register original (untrimmed) request details for manual resend
try:
pipeline_id = str((context.get("meta") or {}).get("id", "pipeline_editor"))
except Exception:
pipeline_id = "pipeline_editor"
try:
reg_info: Dict[str, Any] = {
"pipeline_id": pipeline_id,
"node_id": self.node_id,
"node_type": self.type_name,
"method": method,
"url": url,
"headers": dict(headers),
}
if raw_payload is not None:
reg_info["body_json"] = raw_payload # original JSON body
# Always keep original textual/bytes preview for non-JSON
reg_info["body_text"] = body_text
try:
reg_info["body_bytes_len"] = int(len(body_bytes or b""))
except Exception:
pass
register_http_request(req_id, reg_info)
except Exception:
pass
try:
trace_fn = context.get("_trace")
if trace_fn:
await trace_fn({
"event": "http_req",
"node_id": self.node_id,
"node_type": self.type_name,
"req_id": req_id,
"method": method,
"url": url,
"headers": headers,
"body_text": san_body_text if isinstance(san_body_text, str) else body_text,
"ts": int(time.time() * 1000),
})
except Exception:
pass
# Отправка
async with build_client(timeout=timeout_sec) as client:
# Для GET/HEAD обычно не отправляем body
send_content = None if method in {"GET", "HEAD"} else body_bytes
# Cooperative cancel pre-check (avoid starting new HTTP on abort)
try:
pipeline_id = str((context.get("meta") or {}).get("id", "pipeline_editor"))
except Exception:
pipeline_id = "pipeline_editor"
try:
if is_cancelled(pipeline_id) and get_cancel_mode(pipeline_id) == "abort":
try:
print(f"TRACE http_cancel_pre: {self.node_id} abort before request")
except Exception:
pass
raise ExecutionError("Cancelled by user (abort)")
except Exception:
pass
async def _do_req():
return await client.request(method, url, headers=headers, content=send_content)
# Await HTTP with cooperative cancel/abort handling
resp = await _await_coro_with_cancel(_do_req(), pipeline_id)
# Ответ: лог/печать
try:
print("===== RawForward RESPONSE BEGIN =====")
print(f"node={self.node_id} type={self.type_name}")
print(f"Status: {resp.status_code}")
try:
print("Headers:")
print(json.dumps(dict(resp.headers), ensure_ascii=False, indent=2))
except Exception:
try:
print(f"Headers(raw): {dict(resp.headers)}")
except Exception:
print("Headers(raw): <unavailable>")
try:
resp_text = _safe_response_text(resp)
except Exception:
resp_text = "<resp.text decode error>"
print("Body Text:")
print(resp_text)
print("===== RawForward RESPONSE END =====")
except Exception:
resp_text = "<resp.text decode error>"
# SSE: http_resp
try:
trace_fn = context.get("_trace")
if trace_fn:
await trace_fn({
"event": "http_resp",
"node_id": self.node_id,
"node_type": self.type_name,
"req_id": req_id,
"status": int(resp.status_code),
"headers": dict(resp.headers),
"body_text": resp_text if isinstance(resp_text, str) else str(resp_text),
"ts": int(time.time() * 1000),
})
except Exception:
pass
# Разбор результата
try:
data = _safe_response_json(resp)
except Exception:
data = {"error": "Failed to decode JSON from upstream", "text": resp_text}
# Извлечение текста: приоритет — пресет ноды -> кастомные поля ноды -> глобальные meta.
try:
meta_opts = (context.get("meta") or {})
presets = meta_opts.get("text_extract_presets") or []
cfg = self.config or {}
strategy = None
json_path = None
join_sep = None
# 1) Пресет по id, выбранный в ноде
pid = str(cfg.get("text_extract_preset_id") or "").strip()
if pid:
try:
pr = next((x for x in presets if isinstance(x, dict) and str(x.get("id", "")) == pid), None)
except Exception:
pr = None
if pr:
strategy = str(pr.get("strategy") or "auto")
json_path = str(pr.get("json_path") or "")
join_sep = str(pr.get("join_sep") or "\n")
# 2) Кастомные поля ноды (если что-то задано)
if not strategy and (cfg.get("text_extract_strategy") is not None or cfg.get("text_extract_json_path") is not None or cfg.get("text_join_sep") is not None):
strategy = str(cfg.get("text_extract_strategy") or "auto")
json_path = str(cfg.get("text_extract_json_path") or "")
join_sep = str(cfg.get("text_join_sep") or "\n")
# 3) Глобальные метаданные запуска (дефолт)
if not strategy:
strategy = str(meta_opts.get("text_extract_strategy", "auto") or "auto")
json_path = str(meta_opts.get("text_extract_json_path", "") or "")
join_sep = str(meta_opts.get("text_join_sep", "\n") or "\n")
except Exception:
strategy, json_path, join_sep = "auto", "", "\n"
# Подсказка о провайдере: используем detect_vendor(data) как hint для auto
try:
prov_hint = detect_vendor(data if isinstance(data, dict) else {}) # type: ignore[arg-type]
except Exception:
prov_hint = "unknown"
best_text = _extract_text_for_out(data, strategy, prov_hint, json_path, join_sep)
# Подробный SSE-лог завершения RawForward
try:
trace_fn = context.get("_trace")
if trace_fn:
# Compute destination macros for extracted text and include req_id
try:
to_path = f"OUT.{self.node_id}.response_text"
macro_braces = f"{{{{ OUT.{self.node_id}.response_text }}}}"
alias_macro = ""
m = re.match(r"^n(\d+)$", str(self.node_id))
if m:
alias_macro = f"[[OUT{int(m.group(1))}]]"
except Exception:
to_path = f"OUT.{self.node_id}.response_text"
macro_braces = f"{{{{ OUT.{self.node_id}.response_text }}}}"
alias_macro = ""
await trace_fn({
"event": "rawforward_done",
"node_id": self.node_id,
"node_type": self.type_name,
"method": method,
"url": url,
"req_id": req_id,
"status": int(resp.status_code) if 'resp' in locals() else None,
"text_len": int(len(best_text or "")),
"extracted_text": best_text or "",
"strategy": strategy,
"json_path": json_path or "",
"join_sep": join_sep or "\n",
"to_path": to_path,
"to_macro_outx": alias_macro,
"to_macro_braces": macro_braces,
"ts": int(time.time() * 1000),
})
except Exception:
pass
return {"result": data, "response_text": best_text or ""}
class ReturnNode(Node):
type_name = "Return"
async def run(self, inputs: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]: # noqa: D401
# Определяем целевой формат
cfg = self.config or {}
target = str(cfg.get("target_format", "auto")).lower().strip() or "auto"
if target == "auto":
target = str(context.get("vendor_format") or "openai").lower().strip() or "openai"
# Рендерим текст из шаблона (по умолчанию берём [[OUT1]])
out_map = context.get("OUT") or {}
template = cfg.get("text_template")
if template is None or template == "":
template = "[[OUT1]]"
try:
text = render_template_simple(str(template), context, out_map)
except Exception:
text = ""
model = str(context.get("model") or "")
# Форматтеры под провайдеры (как в execute_pipeline_echo)
def fmt_openai(t: str) -> Dict[str, Any]:
return {
"id": "ret_mock_123",
"object": "chat.completion",
"model": model,
"choices": [
{
"index": 0,
"message": {"role": "assistant", "content": t},
"finish_reason": "stop",
}
],
"usage": {"prompt_tokens": 0, "completion_tokens": len((t or "").split()), "total_tokens": 0},
}
def fmt_gemini(t: str) -> Dict[str, Any]:
return {
"candidates": [
{
"content": {
"role": "model",
"parts": [{"text": t}],
},
"finishReason": "STOP",
"index": 0,
}
],
"modelVersion": model,
}
def fmt_claude(t: str) -> Dict[str, Any]:
return {
"id": "msg_ret_123",
"type": "message",
"model": model,
"role": "assistant",
"content": [
{"type": "text", "text": t}
],
"stop_reason": "end_turn",
}
if target == "openai":
result = fmt_openai(text)
elif target == "gemini":
result = fmt_gemini(text)
elif target == "claude":
result = fmt_claude(text)
else:
# неизвестное значение — безопасный дефолт
result = fmt_openai(text)
# Подробный SSE-лог Return
try:
trace_fn = context.get("_trace")
if trace_fn:
await trace_fn({
"event": "return_detail",
"node_id": self.node_id,
"node_type": self.type_name,
"target": target,
"text_len": int(len(text or "")),
"template_used": str(template),
"ts": int(time.time() * 1000),
})
except Exception:
pass
return {"result": result, "response_text": text}
class IfNode(Node):
type_name = "If"
async def run(self, inputs: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]: # noqa: D401
expr = str(self.config.get("expr") or "").strip()
out_map = context.get("OUT") or {}
# Для логов: отображаем развёрнутое выражение (с подставленными макросами)
try:
expanded = render_template_simple(expr, context, out_map)
except Exception:
expanded = expr
try:
res = bool(eval_condition_expr(expr, context, out_map)) if expr else False
except Exception as exc: # noqa: BLE001
# Расширенный лог, чтобы быстрее найти ошибку парсинга/оценки выражения
try:
print(f"TRACE if_error: {self.node_id} expr={expr!r} expanded={expanded!r} error={exc}")
except Exception:
pass
raise ExecutionError(f"If expr error: {exc}")
try:
print(f"TRACE if: {self.node_id} expr={expr!r} expanded={expanded!r} result={str(res).lower()}")
except Exception:
pass
# Подробный SSE-лог по If
try:
trace_fn = context.get("_trace")
if trace_fn:
await trace_fn({
"event": "if_result",
"node_id": self.node_id,
"node_type": self.type_name,
"expr": expr,
"expanded": expanded,
"result": bool(res),
"ts": int(time.time() * 1000),
})
except Exception:
pass
return {"result": res, "true": res, "false": (not res)}
NODE_REGISTRY.update({
SetVarsNode.type_name: SetVarsNode,
ProviderCallNode.type_name: ProviderCallNode,
RawForwardNode.type_name: RawForwardNode,
ReturnNode.type_name: ReturnNode,
IfNode.type_name: IfNode,
})
# --- While-loop helpers for ProviderCall and RawForward (node-scoped) -----------
async def _providercall_run_with_while(self, inputs, context):
"""
Execute ProviderCall node in a while loop using the same If-expression parser.
- Config keys (optional):
while_expr: string expression (same grammar as If: &&, ||, !, contains, [[...]], {{ ... }})
while_max_iters: safety cap (default: 50)
ignore_errors: bool; if True, HTTP/client errors do not crash pipeline; result carries {"error": "..."}
- Local variables within loop:
[[cycleindex]]: 0-based iteration index (scoped to this node call)
[[WAS_ERROR]]: True only for current iteration when an exception occurred (scoped to this node call)
These locals are injected into context.vars for the inner single-run only and do not leak globally.
- Outputs (global user vars for downstream nodes):
[[WAS_ERROR__nX]]: bool — final iteration error flag for this node
[[CYCLEINDEX__nX]]: int — index of the last executed iteration for this node (0 if none executed)
"""
try:
cfg = self.config or {}
except Exception:
cfg = {}
# Expression and safety cap
try:
expr = str(cfg.get("while_expr") or "").strip()
except Exception:
expr = ""
try:
max_iters = int(cfg.get("while_max_iters") or 50)
except Exception:
max_iters = 50
if max_iters <= 0:
max_iters = 50
ignore = False
try:
ignore = bool(cfg.get("ignore_errors", False))
except Exception:
ignore = False
out_map = context.get("OUT") or {}
last_out = {"result": {}, "response_text": ""}
last_was_error = False
last_idx = -1
for i in range(max_iters):
# Cancel check before starting next iteration
try:
pid = str((context.get("meta") or {}).get("id", "pipeline_editor"))
except Exception:
pid = "pipeline_editor"
try:
if is_cancelled(pid):
mode = get_cancel_mode(pid)
try:
print(f"TRACE while_cancel: {self.node_id} mode={mode} at i={i} (pre)")
except Exception:
pass
break
except Exception:
pass
# Build loop-local context with cycleindex and WAS_ERROR
ctx2 = dict(context or {})
try:
vmap = dict(ctx2.get("vars") or {})
except Exception:
vmap = {}
vmap["cycleindex"] = i
vmap["WAS_ERROR"] = False # local flag for this iteration only
ctx2["vars"] = vmap
# mark reentry to avoid recursive while wrapping in self.run()
ctx2["_in_while"] = True
# Do-while semantics:
# - First iteration always runs.
# - For subsequent iterations, evaluate expression with last_out injected into OUT map.
expanded = ""
if not expr:
# No expression -> perform exactly one iteration
cond = (i == 0)
else:
if i == 0:
cond = True
else:
# Inject previous iteration error flag so [[WAS_ERROR]] in while_expr refers to the last iteration
try:
vmap2 = dict(ctx2.get("vars") or {})
except Exception:
vmap2 = {}
vmap2["WAS_ERROR"] = bool(last_was_error)
ctx2["vars"] = vmap2
# Augment OUT with the last output of this node so [[OUTn]] / [[OUT:nX...]] can see it
try:
out_aug = dict(out_map or {})
out_aug[self.node_id] = dict(last_out or {})
except Exception:
out_aug = out_map
try:
expanded = render_template_simple(expr, ctx2, out_aug)
except Exception:
expanded = expr
try:
cond = bool(eval_condition_expr(expr, ctx2, out_aug))
except Exception as exc: # noqa: BLE001
cond = False
try:
print(f"TRACE while_error: {self.node_id} expr={expr!r} expanded={expanded!r} error={exc}")
except Exception:
pass
# SSE: while_result (analogous to if_result)
try:
trace_fn = context.get("_trace")
if trace_fn:
await trace_fn({
"event": "while_result",
"node_id": self.node_id,
"node_type": self.type_name,
"expr": expr,
"expanded": expanded,
"result": bool(cond),
"index": i,
"ts": int(time.time() * 1000),
})
except Exception:
pass
try:
print(f"TRACE while: {self.node_id} expr={expr!r} expanded={expanded!r} index={i} result={str(cond).lower()}")
except Exception:
pass
if not cond:
break
# Single iteration run; reenter node.run with _in_while flag set
try:
inner_out = await self.run(inputs, ctx2)
except BaseException as exc:
if ignore:
inner_out = {"result": {"error": str(exc)}, "response_text": ""}
try:
print(f"TRACE suppressed_error(inner): {self.node_id} {exc}")
except Exception:
pass
else:
raise
# Detect error by sentinel on result
was_err = False
try:
r = inner_out.get("result")
was_err = bool(isinstance(r, dict) and ("error" in r))
except Exception:
was_err = False
last_out = dict(inner_out or {})
last_was_error = was_err
last_idx = i
# Finalize outputs with per-node global vars
try:
cyc = int(last_idx if last_idx >= 0 else 0)
except Exception:
cyc = 0
try:
last_out = dict(last_out or {})
except Exception:
last_out = {"result": {}, "response_text": ""}
try:
last_out["vars"] = {
f"WAS_ERROR__{self.node_id}": bool(last_was_error),
f"CYCLEINDEX__{self.node_id}": cyc,
}
except Exception:
# ensure at least WAS_ERROR var
last_out["vars"] = {f"WAS_ERROR__{self.node_id}": bool(last_was_error)}
return last_out
async def _rawforward_run_with_while(self, inputs, context):
"""
Execute RawForward node in a while loop using the same If-expression parser.
Conventions identical to _providercall_run_with_while() regarding config and variables.
"""
try:
cfg = self.config or {}
except Exception:
cfg = {}
try:
expr = str(cfg.get("while_expr") or "").strip()
except Exception:
expr = ""
try:
max_iters = int(cfg.get("while_max_iters") or 50)
except Exception:
max_iters = 50
if max_iters <= 0:
max_iters = 50
ignore = False
try:
ignore = bool(cfg.get("ignore_errors", False))
except Exception:
ignore = False
out_map = context.get("OUT") or {}
last_out = {"result": {}, "response_text": ""}
last_was_error = False
last_idx = -1
for i in range(max_iters):
# Cancel check before starting next iteration
try:
pid = str((context.get("meta") or {}).get("id", "pipeline_editor"))
except Exception:
pid = "pipeline_editor"
try:
if is_cancelled(pid):
mode = get_cancel_mode(pid)
try:
print(f"TRACE while_cancel: {self.node_id} mode={mode} at i={i} (pre)")
except Exception:
pass
break
except Exception:
pass
ctx2 = dict(context or {})
try:
vmap = dict(ctx2.get("vars") or {})
except Exception:
vmap = {}
vmap["cycleindex"] = i
vmap["WAS_ERROR"] = False
ctx2["vars"] = vmap
ctx2["_in_while"] = True
# Do-while semantics + last_out in OUT for subsequent checks
expanded = ""
if not expr:
cond = (i == 0)
else:
if i == 0:
cond = True
else:
# Inject previous iteration error flag so [[WAS_ERROR]] in while_expr refers to the last iteration
try:
vmap2 = dict(ctx2.get("vars") or {})
except Exception:
vmap2 = {}
vmap2["WAS_ERROR"] = bool(last_was_error)
ctx2["vars"] = vmap2
try:
out_aug = dict(out_map or {})
out_aug[self.node_id] = dict(last_out or {})
except Exception:
out_aug = out_map
try:
expanded = render_template_simple(expr, ctx2, out_aug)
except Exception:
expanded = expr
try:
cond = bool(eval_condition_expr(expr, ctx2, out_aug))
except Exception as exc: # noqa: BLE001
cond = False
try:
print(f"TRACE while_error: {self.node_id} expr={expr!r} expanded={expanded!r} error={exc}")
except Exception:
pass
# SSE: while_result (analogous to if_result)
try:
trace_fn = context.get("_trace")
if trace_fn:
await trace_fn({
"event": "while_result",
"node_id": self.node_id,
"node_type": self.type_name,
"expr": expr,
"expanded": expanded,
"result": bool(cond),
"index": i,
"ts": int(time.time() * 1000),
})
except Exception:
pass
try:
print(f"TRACE while: {self.node_id} expr={expr!r} expanded={expanded!r} index={i} result={str(cond).lower()}")
except Exception:
pass
if not cond:
break
try:
inner_out = await self.run(inputs, ctx2)
except BaseException as exc:
if ignore:
inner_out = {"result": {"error": str(exc)}, "response_text": ""}
try:
print(f"TRACE suppressed_error(inner): {self.node_id} {exc}")
except Exception:
pass
else:
raise
was_err = False
try:
r = inner_out.get("result")
was_err = bool(isinstance(r, dict) and ("error" in r))
except Exception:
was_err = False
last_out = dict(inner_out or {})
last_was_error = was_err
last_idx = i
try:
cyc = int(last_idx if last_idx >= 0 else 0)
except Exception:
cyc = 0
try:
last_out = dict(last_out or {})
except Exception:
last_out = {"result": {}, "response_text": ""}
try:
last_out["vars"] = {
f"WAS_ERROR__{self.node_id}": bool(last_was_error),
f"CYCLEINDEX__{self.node_id}": cyc,
}
except Exception:
last_out["vars"] = {f"WAS_ERROR__{self.node_id}": bool(last_was_error)}
return last_out