File size: 63,998 Bytes
b48ae75 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 | #!/usr/bin/env python3
"""Local protocol bridge and run-scoped provider usage observation for native agents.
One process serves one protocol against an OpenAI-compatible Chat Completions
upstream:
* ``--protocol anthropic``: POST /v1/messages -> /chat/completions
* ``--protocol responses``: POST /v1/responses -> /chat/completions
The Responses path also handles Codex namespace/custom tools; the Anthropic
path translates Claude Code's text/tool_use/tool_result blocks. Both paths
stream SSE, forward usage where available, and never log prompts or credentials.
Each POST is submitted exactly once.
"""
from __future__ import annotations
import argparse
import hmac
from contextlib import contextmanager
from collections import OrderedDict
import hashlib
import json
import os
import queue
import re
import sys
import threading
import time
import uuid
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
from pathlib import Path
from typing import Any
from urllib.parse import urlsplit, parse_qs
import requests
sys.path.insert(0, str(Path(__file__).resolve().parents[2] / "eval" / "native_harness"))
from native_bench.usage import TransportMeter, provider_usage
def usage_endpoint(handler):
route = urlsplit(handler.path)
if route.path != "/__native_usage":
return False
send = getattr(handler, "send_json", None) or handler._json
token = os.environ.get("NATIVE_BENCH_USAGE_TOKEN", "")
if not token or not hmac.compare_digest(handler.headers.get("Authorization", ""), "Bearer " + token):
send(403, {"error": "invalid usage observer credential"})
return True
try:
if handler.command == "POST":
size = int(handler.headers.get("Content-Length", "0"))
if size > 8192 or size < 0:
raise ValueError("invalid reset body")
handler.rfile.read(size)
payload = handler.server.usage_meter.reset()
else:
value = parse_qs(route.query).get("at_ns", [None])[0]
payload = handler.server.usage_meter.snapshot(int(value) if value is not None else None)
send(200, payload)
except (ValueError, KeyError, TypeError):
send(409, {"error": "invalid usage snapshot or active requests"})
return True
class ObservedResponse:
def __init__(self, response, meter, record):
self.response, self.meter, self.record = response, meter, record
self.raw_usage = {}
self.kind = "openai"
self.final = False
def __getattr__(self, name):
return getattr(self.response, name)
def observe(self, payload):
if not isinstance(payload, dict):
return
typ = payload.get("type", "")
response_payload = payload.get("response")
response_payload = response_payload if isinstance(response_payload, dict) else {}
if typ == "message_start":
self.kind = "anthropic"
self.raw_usage.update(payload.get("message", {}).get("usage") or {})
elif typ == "message_delta":
self.kind = "anthropic"
self.raw_usage.update(payload.get("usage") or {})
delta = payload.get("delta")
stop_reason = delta.get("stop_reason") if isinstance(delta, dict) else None
if isinstance(stop_reason, str) and stop_reason and isinstance(payload.get("usage"), dict):
self.final = provider_usage(self.raw_usage, self.kind) is not None
elif typ == "message_stop":
self.final = True
elif typ == "message" and isinstance(payload.get("usage"), dict):
self.kind = "anthropic"
self.raw_usage.update(payload["usage"])
self.final = True
elif isinstance(response_payload.get("usage"), dict):
self.raw_usage.update(response_payload["usage"])
self.final = typ in ("response.completed", "response.incomplete", "response.failed")
elif isinstance(payload.get("usage"), dict):
self.raw_usage.update(payload["usage"])
self.final = bool("choices" in payload or payload.get("object") == "response"
or payload.get("object") == "response.compaction" or bool(payload.get("error")))
if self.final and self.record["finished_ns"] is None:
self.meter.finish(self.record, provider_usage(self.raw_usage, self.kind))
@property
def content(self):
raw = self.response.content
try:
self.observe(json.loads(raw))
except (ValueError, TypeError):
pass
return raw
def json(self, **kwargs):
payload = self.response.json(**kwargs)
self.observe(payload)
return payload
def iter_lines(self, **kwargs):
pending = []
for line in self.response.iter_lines(**kwargs):
raw = line.decode("utf-8", "replace") if isinstance(line, bytes) else line
if raw.startswith("data:"):
value = raw[5:].strip()
if value != "[DONE]":
pending.append(value)
try:
payload = json.loads("\n".join(pending))
except ValueError:
pass
else:
self.observe(payload)
pending = []
elif not raw.strip():
pending = []
yield line
@contextmanager
def metered_post(server, session, *args, **kwargs):
record = server.usage_meter.begin()
try:
with session.post(*args, **kwargs) as response:
yield ObservedResponse(response, server.usage_meter, record)
finally:
if record["finished_ns"] is None:
server.usage_meter.finish(record, None)
class PassthroughHandler(BaseHTTPRequestHandler):
protocol_version = "HTTP/1.1"
def log_message(self, *args):
pass
def send_json(self, status, payload):
raw = json.dumps(payload).encode()
self.send_response(status)
self.send_header("Content-Type", "application/json")
self.send_header("Content-Length", str(len(raw)))
self.end_headers()
self.wfile.write(raw)
def do_GET(self):
if not usage_endpoint(self):
self.forward()
def do_POST(self):
if not usage_endpoint(self):
self.forward()
def forward(self):
route = urlsplit(self.path)
if route.scheme or route.netloc or route.path.startswith("//"):
self.send_json(400, {"error": "invalid upstream path"})
return
try:
size = int(self.headers.get("Content-Length", "0"))
if size < 0 or size > 100 * 1024 * 1024:
raise ValueError("invalid request size")
raw = self.rfile.read(size) if size else b""
body = json.loads(raw) if raw else {}
stream = bool(body.get("stream"))
if stream and route.path.rstrip("/").endswith("/chat/completions"):
body.setdefault("stream_options", {})["include_usage"] = True
raw = json.dumps(body, ensure_ascii=False).encode("utf-8")
headers = {k: v for k, v in self.headers.items() if k.lower() not in
{"host", "content-length", "connection", "transfer-encoding", "proxy-authorization"}}
cfg = self.server.config
if not any(k.lower() in {"authorization", "x-api-key"} for k in headers) and cfg["api_key"]:
headers["Authorization"] = "Bearer " + cfg["api_key"]
base_path = urlsplit(cfg["upstream"]).path.rstrip("/")
suffix = self.path
if base_path and (route.path == base_path or route.path.startswith(base_path + "/")):
suffix = route.path[len(base_path):] + ("?" + route.query if route.query else "")
url = cfg["upstream"].rstrip("/") + suffix
generation = self.command == "POST" and route.path.rstrip("/").endswith(
("/responses", "/responses/compact", "/messages", "/chat/completions"))
with requests.Session() as session:
kwargs = dict(data=raw or None, headers=headers, stream=stream,
timeout=(15, cfg["timeout"]), allow_redirects=False)
context = metered_post(self.server, session, url, **kwargs) if generation else session.request(self.command, url, **kwargs)
with context as response:
self.send_response(response.status_code)
for key, value in response.headers.items():
if key.lower() not in {"content-length", "content-encoding", "transfer-encoding", "connection"}:
self.send_header(key, value)
self.send_header("Connection", "close")
self.end_headers()
self.close_connection = True
if stream:
for line in response.iter_lines(chunk_size=1, decode_unicode=False):
self.wfile.write(line + b"\n")
self.wfile.flush()
else:
if generation:
try:
response.json()
except ValueError:
pass
self.wfile.write(response.content)
except (BrokenPipeError, ConnectionResetError):
self.close_connection = True
except (ValueError, TypeError, requests.RequestException):
self.close_connection = True
def passthrough_main(argv=None):
parser = argparse.ArgumentParser()
parser.add_argument("--host", default="127.0.0.1")
parser.add_argument("--port", type=int, required=True)
parser.add_argument("--upstream", required=True)
parser.add_argument("--name", default="")
parser.add_argument("--timeout", type=float, default=3600)
args = parser.parse_args(argv)
parsed = urlsplit(args.upstream)
if parsed.scheme not in ("http", "https") or not parsed.netloc or parsed.query or parsed.fragment:
parser.error("invalid upstream base URL")
server = ThreadingHTTPServer((args.host, args.port), PassthroughHandler)
server.usage_meter = TransportMeter()
server.config = {"upstream": args.upstream, "api_key": os.environ.get("PROXY_API_KEY", ""), "timeout": args.timeout}
start_parent_watch()
try:
server.serve_forever()
finally:
server.server_close()
def new_id(prefix: str) -> str:
return prefix + uuid.uuid4().hex
def start_parent_watch() -> None:
"""Exit if the launching bridge process disappears (including SIGKILL)."""
parent = int(os.environ.get("PROXY_PARENT_PID", "0") or 0)
if parent <= 0:
return
def watch() -> None:
while True:
time.sleep(2)
try:
os.kill(parent, 0)
except OSError:
os._exit(0)
threading.Thread(target=watch, name="parent-watch", daemon=True).start()
class BadRequest(ValueError):
pass
class ToolTypes(dict):
"""Kinds plus reversible Responses namespace / Chat name mappings."""
def __init__(self):
super().__init__()
self.names = {}
self.aliases = {}
def wire_name(self, name: str, namespace: str | None = None) -> str:
qualified = namespace + "." + name if namespace else name
if qualified in self.aliases:
return self.aliases[qualified]
if name in self:
return name
matches = [alias for alias, original in self.names.items() if original["name"] == name]
if len(matches) == 1:
return matches[0]
if len(matches) > 1:
raise BadRequest("Ambiguous unqualified historical tool name.")
return safe_tool_name(qualified.replace(".", "__"))
def safe_tool_name(name: str) -> str:
safe = re.sub(r"[^a-zA-Z0-9_-]", "_", name)
if len(safe) > 64:
safe = safe[:51] + "_" + hashlib.sha256(name.encode()).hexdigest()[:12]
return safe
def convert_tools(definitions: list) -> tuple[list, ToolTypes]:
kinds = ToolTypes()
translated = []
def add(tool: dict, namespace: str | None = None, namespace_description: str = "") -> None:
kind = tool.get("type")
if kind == "namespace":
if namespace:
raise BadRequest("Nested namespace tools are not supported.")
if not tool.get("name") or not isinstance(tool.get("tools"), list):
raise BadRequest("Namespace tools require a name and a tools list.")
for child in tool["tools"]:
add(child, tool["name"], tool.get("description", ""))
return
if kind not in ("function", "custom"):
raise BadRequest("Unsupported Responses tool type: " + str(kind))
name = tool.get("name")
if not isinstance(name, str) or not name:
raise BadRequest("Tools require a name.")
qualified = namespace + "." + name if namespace else name
alias = safe_tool_name((namespace + "__" + name) if namespace else name)
if alias in kinds or qualified in kinds.aliases:
raise BadRequest("Tool names collide after namespace translation.")
kinds[alias] = kind
kinds.names[alias] = {"name": name, "namespace": namespace}
kinds.aliases[qualified] = alias
description = tool.get("description", "")
if namespace_description:
description = namespace_description + "\n" + description
fn = {"name": alias, "description": description}
if kind == "custom":
if tool.get("format"):
fn["description"] += "\nInput format: " + json.dumps(tool["format"], ensure_ascii=False)
fn["parameters"] = {"type": "object", "properties": {"input": {"type": "string", "description": "The complete raw input for this tool."}}, "required": ["input"], "additionalProperties": False}
else:
fn["parameters"] = tool.get("parameters") or {"type": "object", "properties": {}}
if "strict" in tool:
fn["strict"] = tool["strict"]
translated.append({"type": "function", "function": fn})
for definition in definitions:
add(definition)
return translated, kinds
def text_content(value: Any) -> str:
if isinstance(value, str):
return value
if value is None:
return ""
return json.dumps(value, ensure_ascii=False, separators=(",", ":"))
def message_content(content: Any) -> Any:
if isinstance(content, str):
return content
if content is None:
return ""
if not isinstance(content, list):
raise BadRequest("Message content must be a string or a list.")
result = []
for part in content:
kind = part.get("type")
if kind in ("input_text", "output_text", "text"):
result.append({"type": "text", "text": part.get("text", "")})
elif kind == "input_image":
image = part.get("image_url")
if not isinstance(image, str):
raise BadRequest("Only image_url input images are supported.")
image_data = {"url": image}
if part.get("detail"):
image_data["detail"] = part["detail"]
result.append({"type": "image_url", "image_url": image_data})
elif kind == "refusal":
result.append({"type": "text", "text": part.get("refusal", "")})
else:
raise BadRequest("Unsupported message content type: " + str(kind))
if all(part["type"] == "text" for part in result):
return "".join(part["text"] for part in result)
return result
def reasoning_text(item: dict) -> str:
parts = item.get("content") or item.get("summary") or []
return "".join(part.get("text", "") for part in parts if isinstance(part, dict) and part.get("type") in ("summary_text", "reasoning_text", "text"))
def convert_request(body: dict, model_name: str, reasoning_lookup=None) -> tuple[dict, dict]:
if body.get("previous_response_id"):
raise BadRequest("previous_response_id is not supported; send the complete input history with store=false.")
if body.get("background"):
raise BadRequest("Background Responses requests are not supported.")
tools, tool_types = convert_tools(body.get("tools") or [])
messages = []
if body.get("instructions"):
messages.append({"role": "system", "content": body["instructions"]})
inputs = body.get("input", [])
if isinstance(inputs, str):
inputs = [{"role": "user", "content": inputs}]
if not isinstance(inputs, list):
raise BadRequest("input must be a string or a list.")
pending_reasoning = ""
for item in inputs:
if not isinstance(item, dict):
raise BadRequest("Every input item must be an object.")
kind = item.get("type", "message")
if kind == "message":
role = item.get("role", "user")
if role == "developer":
role = "system"
if role not in ("system", "user", "assistant", "tool"):
raise BadRequest("Unsupported message role: " + str(role))
message = {"role": role, "content": message_content(item.get("content", ""))}
if role == "assistant" and pending_reasoning:
message["reasoning_content"] = pending_reasoning
pending_reasoning = ""
if role == "tool":
message["tool_call_id"] = item.get("tool_call_id")
messages.append(message)
elif kind in ("function_call", "custom_tool_call"):
call_id = item.get("call_id")
if not call_id or not item.get("name"):
raise BadRequest("Tool calls require call_id and name.")
arguments = item.get("arguments", "") if kind == "function_call" else json.dumps({"input": item.get("input", "")}, ensure_ascii=False)
tool_call = {"id": call_id, "type": "function", "function": {"name": tool_types.wire_name(item["name"], item.get("namespace")), "arguments": text_content(arguments)}}
if messages and messages[-1]["role"] == "assistant":
messages[-1].setdefault("tool_calls", []).append(tool_call)
else:
messages.append({"role": "assistant", "content": None, "tool_calls": [tool_call]})
prior_reasoning = pending_reasoning or (reasoning_lookup(call_id) if reasoning_lookup else "")
if prior_reasoning and not messages[-1].get("reasoning_content"):
messages[-1]["reasoning_content"] = prior_reasoning
pending_reasoning = ""
elif kind in ("function_call_output", "custom_tool_call_output"):
if not item.get("call_id"):
raise BadRequest("Tool outputs require call_id.")
output = item.get("output", "")
if isinstance(output, list):
output = message_content(output)
messages.append({"role": "tool", "tool_call_id": item["call_id"], "content": text_content(output)})
elif kind == "reasoning":
restored = reasoning_text(item)
if item.get("encrypted_content") and not restored:
raise BadRequest("Opaque encrypted reasoning cannot be translated; replay this bridge's reasoning summary instead.")
pending_reasoning += restored
elif kind == "compaction":
raise BadRequest("Opaque compaction items cannot be translated; send complete text/tool history.")
else:
raise BadRequest("Unsupported input item type: " + str(kind))
if not messages:
raise BadRequest("The request contains no supported input messages.")
upstream = {"model": model_name, "messages": messages, "stream": bool(body.get("stream", False))}
if tools:
upstream["tools"] = tools
choice = body.get("tool_choice")
if isinstance(choice, dict):
if choice.get("type") not in ("function", "custom") or not choice.get("name"):
raise BadRequest("Only named function/custom tool_choice is supported.")
upstream["tool_choice"] = {"type": "function", "function": {"name": tool_types.wire_name(choice["name"], choice.get("namespace"))}}
elif choice is not None:
upstream["tool_choice"] = choice
for field in ("temperature", "top_p", "parallel_tool_calls", "user"):
if field in body and body[field] is not None:
upstream[field] = body[field]
if body.get("max_output_tokens") is not None:
upstream["max_tokens"] = body["max_output_tokens"]
if isinstance(body.get("reasoning"), dict) and body["reasoning"].get("effort"):
upstream["reasoning_effort"] = body["reasoning"]["effort"]
if upstream["stream"]:
upstream["stream_options"] = {"include_usage": True}
return upstream, tool_types
def convert_usage(usage: Any) -> dict | None:
if not isinstance(usage, dict):
return None
converted = {}
for original, target in (("prompt_tokens", "input_tokens"), ("completion_tokens", "output_tokens"), ("total_tokens", "total_tokens")):
if original in usage:
converted[target] = usage[original]
prompt_details = usage.get("prompt_tokens_details") or {}
completion_details = usage.get("completion_tokens_details") or {}
if "cached_tokens" in prompt_details:
converted["input_tokens_details"] = {"cached_tokens": prompt_details["cached_tokens"]}
if "reasoning_tokens" in completion_details:
converted["output_tokens_details"] = {"reasoning_tokens": completion_details["reasoning_tokens"]}
return converted or None
def response_object(response_id: str, model: str, created: int, status: str, output: list, usage: dict | None = None, error: dict | None = None, incomplete: dict | None = None) -> dict:
return {"id": response_id, "object": "response", "created_at": created, "status": status, "error": error, "incomplete_details": incomplete, "model": model, "output": output, "usage": usage, "store": False, "parallel_tool_calls": True}
def tool_item(call: dict, tool_types: dict, status: str) -> dict:
name = call.get("name", "")
if not name:
raise ValueError("Upstream returned an unnamed tool call.")
custom = tool_types.get(name) == "custom"
item = {"id": call.get("item_id") or new_id("ctc_" if custom else "fc_"), "type": "custom_tool_call" if custom else "function_call", "call_id": call.get("call_id") or new_id("call_"), "name": name, "status": status}
if isinstance(tool_types, ToolTypes) and name in tool_types.names:
original = tool_types.names[name]
item["name"] = original["name"]
if original["namespace"]:
item["namespace"] = original["namespace"]
arguments = call.get("arguments", "")
if custom:
try:
decoded = json.loads(arguments)
if not isinstance(decoded, dict) or not isinstance(decoded.get("input"), str):
raise ValueError("Custom tool arguments require a string input field.")
item["input"] = decoded["input"]
except (ValueError, TypeError):
if status == "incomplete":
item["input"] = ""
else:
raise ValueError("Upstream returned invalid custom tool arguments.") from None
else:
item["arguments"] = arguments
return item
class BridgeServer(ThreadingHTTPServer):
daemon_threads = True
def __init__(self, address: tuple, config: dict):
self.config = config
self.usage_meter = TransportMeter()
self.reasoning_cache = OrderedDict()
self.reasoning_cache_bytes = 0
self.reasoning_lock = threading.Lock()
super().__init__(address, ResponsesHandler)
def remember_reasoning(self, call_id: str, reasoning: str) -> None:
if not call_id or not reasoning:
return
with self.reasoning_lock:
previous = self.reasoning_cache.pop(call_id, "")
self.reasoning_cache_bytes -= len(previous.encode("utf-8"))
self.reasoning_cache[call_id] = reasoning
self.reasoning_cache_bytes += len(reasoning.encode("utf-8"))
while len(self.reasoning_cache) > 128 or self.reasoning_cache_bytes > 32 * 1024 * 1024:
_, removed = self.reasoning_cache.popitem(last=False)
self.reasoning_cache_bytes -= len(removed.encode("utf-8"))
def recall_reasoning(self, call_id: str) -> str:
with self.reasoning_lock:
return self.reasoning_cache.get(call_id, "")
class ResponsesHandler(BaseHTTPRequestHandler):
protocol_version = "HTTP/1.1"
server_version = "ResponsesChatBridge/1.0"
def log_message(self, fmt: str, *args: Any) -> None:
# Deliberately exclude arbitrary URLs, request bodies, and credentials.
sys.stderr.write("%s %s %s\n" % (time.strftime("%Y-%m-%dT%H:%M:%S%z"), self.command, urlsplit(self.path).path))
def send_json(self, status: int, payload: dict) -> None:
raw = json.dumps(payload, ensure_ascii=False).encode("utf-8")
self.send_response(status)
self.send_header("Content-Type", "application/json; charset=utf-8")
self.send_header("Content-Length", str(len(raw)))
self.end_headers()
self.wfile.write(raw)
def error_json(self, status: int, message: str, code: str = "invalid_request_error") -> None:
self.send_json(status, {"error": {"type": code, "code": code, "message": message}})
def do_GET(self) -> None:
if usage_endpoint(self):
return
route = urlsplit(self.path).path.rstrip("/")
if route in ("/health", "/v1/health"):
self.send_json(200, {"status": "ok", "model": self.server.config["name"], "stateless": True})
elif route in ("/models", "/v1/models"):
self.send_json(200, {"object": "list", "data": [{"id": self.server.config["name"], "object": "model", "owned_by": "catalog"}]})
else:
self.error_json(404, "Unknown route.")
def do_POST(self) -> None:
if usage_endpoint(self):
return
if urlsplit(self.path).path.rstrip("/") not in ("/responses", "/v1/responses"):
self.error_json(404, "Only /v1/responses is supported.")
return
try:
size = int(self.headers.get("Content-Length", "0"))
if size <= 0 or size > 100 * 1024 * 1024:
raise BadRequest("A JSON request body of at most 100 MiB is required.")
body = json.loads(self.rfile.read(size))
if not isinstance(body, dict):
raise BadRequest("The request must be a JSON object.")
upstream, tool_types = convert_request(body, self.server.config["name"], self.server.recall_reasoning)
if "max_tokens" not in upstream and self.server.config.get("max_output") is not None:
upstream["max_tokens"] = self.server.config["max_output"]
except (ValueError, TypeError, KeyError, IndexError):
# BadRequest messages are controlled strings, never body excerpts.
exc = sys.exc_info()[1]
self.error_json(400, str(exc) if isinstance(exc, BadRequest) else "Invalid JSON request.")
return
try:
with requests.Session() as session:
if self.server.config.get("proxy") == "never":
session.trust_env = False
headers = {"Content-Type": "application/json", "Accept": "text/event-stream" if upstream["stream"] else "application/json"}
if self.server.config.get("api_key"):
headers["Authorization"] = "Bearer " + self.server.config["api_key"]
# No retry adapter and no redirects: a generation POST is never replayed.
with metered_post(self.server, session, self.server.config["url"], json=upstream, headers=headers, stream=upstream["stream"], timeout=(15, self.server.config["timeout"]), allow_redirects=False) as result:
if result.status_code != 200:
raw = result.content
self.send_response(result.status_code)
self.send_header("Content-Type", result.headers.get("Content-Type", "application/json"))
self.send_header("Content-Length", str(len(raw)))
self.end_headers()
self.wfile.write(raw)
return
if upstream["stream"]:
self.stream_result(result, tool_types)
else:
self.nonstream_result(result.json(), tool_types)
except (BrokenPipeError, ConnectionResetError):
self.close_connection = True
except requests.RequestException:
self.error_json(502, "Upstream connection failed or timed out; the POST was not retried.", "upstream_connection_error")
except (ValueError, TypeError, KeyError):
self.error_json(502, "Upstream returned an invalid Chat Completions response.", "upstream_protocol_error")
def nonstream_result(self, result: dict, tool_types: dict) -> None:
choice = result["choices"][0]
finish = choice.get("finish_reason")
if finish not in ("stop", "tool_calls", "length", "content_filter"):
raise ValueError("Missing or unsupported finish reason.")
status = "incomplete" if finish in ("length", "content_filter") else "completed"
message = choice["message"]
output = []
reasoning = message.get("reasoning_content") or ""
if reasoning:
output.append({"id": new_id("rs_"), "type": "reasoning", "summary": [{"type": "summary_text", "text": reasoning}], "content": [{"type": "reasoning_text", "text": reasoning}], "status": status})
if message.get("content"):
output.append({"id": new_id("msg_"), "type": "message", "status": status, "role": "assistant", "content": [{"type": "output_text", "text": message["content"], "annotations": []}]})
if message.get("refusal"):
output.append({"id": new_id("msg_"), "type": "message", "status": status, "role": "assistant", "content": [{"type": "refusal", "refusal": message["refusal"]}]})
for call in message.get("tool_calls") or []:
self.server.remember_reasoning(call.get("id"), reasoning)
output.append(tool_item({"call_id": call.get("id"), "name": call["function"].get("name"), "arguments": call["function"].get("arguments", "")}, tool_types, status))
incomplete = {"reason": "max_output_tokens" if finish == "length" else "content_filter"} if status == "incomplete" else None
self.send_json(200, response_object(new_id("resp_"), self.server.config["name"], int(time.time()), status, output, convert_usage(result.get("usage")), incomplete=incomplete))
def stream_result(self, upstream: requests.Response, tool_types: dict) -> None:
self.send_response(200)
self.send_header("Content-Type", "text/event-stream; charset=utf-8")
self.send_header("Cache-Control", "no-cache")
self.send_header("Connection", "close")
self.send_header("X-Accel-Buffering", "no")
self.end_headers()
self.close_connection = True
sequence = 0
response_id = new_id("resp_")
created = int(time.time())
output = []
message = None
reasoning_item = None
message_index = None
reasoning_index = None
usage = None
finish = None
calls = {}
events = queue.Queue(maxsize=256)
stop = threading.Event()
def emit(kind: str, **payload: Any) -> None:
nonlocal sequence
payload = {"type": kind, "sequence_number": sequence, **payload}
sequence += 1
raw = "event: " + kind + "\ndata: " + json.dumps(payload, ensure_ascii=False, separators=(",", ":")) + "\n\n"
self.wfile.write(raw.encode("utf-8"))
self.wfile.flush()
def put_event(event: tuple) -> None:
while not stop.is_set():
try:
events.put(event, timeout=0.5)
return
except queue.Full:
pass
def read_stream() -> None:
try:
# chunk_size=1 avoids requests buffering a quiet, small SSE stream.
data_lines = []
for raw in upstream.iter_lines(chunk_size=1, decode_unicode=False):
if stop.is_set():
return
line = raw.decode("utf-8")
if not line:
if data_lines:
data = "\n".join(data_lines)
data_lines = []
if data == "[DONE]":
put_event(("end", None))
return
put_event(("data", json.loads(data)))
elif line.startswith("data:"):
data_lines.append(line[5:].lstrip(" "))
if data_lines:
data = "\n".join(data_lines)
if data == "[DONE]":
put_event(("end", None))
return
put_event(("data", json.loads(data)))
put_event(("end_of_file", None))
except Exception:
put_event(("error", None))
try:
initial = response_object(response_id, self.server.config["name"], created, "in_progress", [])
emit("response.created", response=initial)
emit("response.in_progress", response=initial)
threading.Thread(target=read_stream, name="upstream-sse-reader", daemon=True).start()
while True:
try:
kind, payload = events.get(timeout=10)
except queue.Empty:
self.wfile.write(b": keepalive\n\n")
self.wfile.flush()
continue
if kind in ("error", "end_of_file"):
raise ValueError("Upstream stream failed or was interrupted.")
if kind == "end":
break
if payload.get("error"):
raise ValueError("Upstream sent an error inside the stream.")
if payload.get("usage"):
usage = convert_usage(payload["usage"])
choices = payload.get("choices") or []
if not choices:
continue
choice = choices[0]
if choice.get("finish_reason") is not None:
finish = choice["finish_reason"]
delta = choice.get("delta") or {}
if delta.get("refusal"):
raise ValueError("Upstream refused this generation.")
if delta.get("reasoning_content"):
if not isinstance(delta["reasoning_content"], str):
raise ValueError("Unexpected upstream reasoning type.")
if reasoning_item is None:
reasoning_index = len(output)
reasoning_item = {"id": new_id("rs_"), "type": "reasoning", "summary": [], "status": "in_progress"}
output.append(reasoning_item)
emit("response.output_item.added", output_index=reasoning_index, item=dict(reasoning_item))
part = {"type": "summary_text", "text": ""}
reasoning_item["summary"].append(part)
emit("response.reasoning_summary_part.added", item_id=reasoning_item["id"], output_index=reasoning_index, summary_index=0, part=dict(part))
reasoning_item["summary"][0]["text"] += delta["reasoning_content"]
emit("response.reasoning_summary_text.delta", item_id=reasoning_item["id"], output_index=reasoning_index, summary_index=0, delta=delta["reasoning_content"])
if delta.get("content"):
if not isinstance(delta["content"], str):
raise ValueError("Unexpected upstream content type.")
if message is None:
message_index = len(output)
message = {"id": new_id("msg_"), "type": "message", "status": "in_progress", "role": "assistant", "content": []}
output.append(message)
emit("response.output_item.added", output_index=message_index, item=dict(message))
part = {"type": "output_text", "text": "", "annotations": []}
message["content"].append(part)
emit("response.content_part.added", item_id=message["id"], output_index=message_index, content_index=0, part=dict(part))
message["content"][0]["text"] += delta["content"]
emit("response.output_text.delta", item_id=message["id"], output_index=message_index, content_index=0, delta=delta["content"], logprobs=[])
for part in delta.get("tool_calls") or []:
index = part.get("index", 0)
call = calls.setdefault(index, {"call_id": "", "name": "", "arguments": ""})
if part.get("id"):
call["call_id"] = part["id"]
fn = part.get("function") or {}
call["name"] += fn.get("name") or ""
call["arguments"] += fn.get("arguments") or ""
if finish not in ("stop", "tool_calls", "length", "content_filter"):
raise ValueError("Upstream stream ended without a supported finish reason.")
status = "incomplete" if finish in ("length", "content_filter") else "completed"
# Validate all complete custom wrappers before emitting tool completion.
converted_calls = [tool_item(calls[index], tool_types, status) for index in sorted(calls)]
if reasoning_item is not None:
reasoning_item["status"] = status
part = reasoning_item["summary"][0]
# Preserve the provider's raw reasoning for full-history replay;
# the summary events also make it visible to existing clients.
reasoning_item["content"] = [{"type": "reasoning_text", "text": part["text"]}]
emit("response.reasoning_summary_text.done", item_id=reasoning_item["id"], output_index=reasoning_index, summary_index=0, text=part["text"])
emit("response.reasoning_summary_part.done", item_id=reasoning_item["id"], output_index=reasoning_index, summary_index=0, part=part)
emit("response.output_item.done", output_index=reasoning_index, item=reasoning_item)
for item in converted_calls:
self.server.remember_reasoning(item["call_id"], part["text"])
if message is not None:
message["status"] = status
part = message["content"][0]
emit("response.output_text.done", item_id=message["id"], output_index=message_index, content_index=0, text=part["text"], logprobs=[])
emit("response.content_part.done", item_id=message["id"], output_index=message_index, content_index=0, part=part)
emit("response.output_item.done", output_index=message_index, item=message)
for item in converted_calls:
index = len(output)
output.append(item)
custom = item["type"] == "custom_tool_call"
field = "input" if custom else "arguments"
added = dict(item, status="in_progress", **{field: ""})
emit("response.output_item.added", output_index=index, item=added)
event = "response.custom_tool_call_input" if custom else "response.function_call_arguments"
emit(event + ".delta", item_id=item["id"], output_index=index, delta=item[field])
emit(event + ".done", item_id=item["id"], output_index=index, **{field: item[field]})
emit("response.output_item.done", output_index=index, item=item)
incomplete = {"reason": "max_output_tokens" if finish == "length" else "content_filter"} if status == "incomplete" else None
final = response_object(response_id, self.server.config["name"], created, status, output, usage, incomplete=incomplete)
emit("response.incomplete" if status == "incomplete" else "response.completed", response=final)
except (BrokenPipeError, ConnectionResetError):
pass
except Exception:
error = {"code": "upstream_protocol_error", "message": "Upstream stream failed, was interrupted, or contained invalid tool arguments; the POST was not retried."}
try:
emit("response.failed", response=response_object(response_id, self.server.config["name"], created, "failed", output, usage, error=error))
except (BrokenPipeError, ConnectionResetError):
pass
finally:
stop.set()
upstream.close()
def responses_main(argv=None) -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--catalog", type=Path,
default=Path(__file__).resolve().parents[1] / "configs" / "op_catalog.json")
parser.add_argument("--model", default="glm-5.3-bigmodel", help="Upstream catalog model key.")
parser.add_argument("--upstream", default=None,
help="OpenAI-compatible base URL; when set, --catalog/--model are not read.")
parser.add_argument("--name", default=None, help="Model name returned to the client (direct mode).")
parser.add_argument("--api-key", default=os.environ.get("PROXY_API_KEY") or os.environ.get("MODEL_PROXY_API_KEY", ""))
parser.add_argument("--max-output", type=int, default=None)
parser.add_argument("--host", default="127.0.0.1")
parser.add_argument("--port", type=int, default=18008)
parser.add_argument("--timeout", type=float, default=900, help="Upstream read timeout in seconds.")
args = parser.parse_args(argv)
try:
if args.upstream:
base = args.upstream.rstrip("/")
config = {"name": args.name or args.model, "api_key": args.api_key, "proxy": "never",
"url": base + "/chat/completions", "timeout": args.timeout,
"max_output": args.max_output}
else:
catalog = json.loads(args.catalog.read_text(encoding="utf-8"))
selected = catalog["models"][args.model]
base = selected["openai_base"].rstrip("/")
config = {"name": selected["name"], "api_key": selected.get("api_key", ""), "proxy": selected.get("proxy"), "url": base + "/chat/completions", "timeout": args.timeout, "max_output": selected.get("max_output")}
if urlsplit(base).scheme not in ("http", "https"):
raise ValueError("Invalid upstream URL scheme.")
except (OSError, ValueError, KeyError, TypeError):
parser.error("Cannot load the requested model configuration from the catalog.")
server = BridgeServer((args.host, args.port), config)
start_parent_watch()
print("Responses bridge listening on %s:%d; model=%s; stateless full-history requests only" % (args.host, args.port, config["name"]), file=sys.stderr, flush=True)
try:
server.serve_forever()
except KeyboardInterrupt:
pass
finally:
server.server_close()
def _id(prefix: str) -> str:
return prefix + uuid.uuid4().hex
def _text_of(content: Any) -> str:
if isinstance(content, str):
return content
if content is None:
return ""
if isinstance(content, list):
parts = []
for block in content:
if isinstance(block, str):
parts.append(block)
elif isinstance(block, dict) and block.get("type") == "text":
parts.append(str(block.get("text", "")))
elif isinstance(block, dict) and block.get("type") == "tool_result":
parts.append(_text_of(block.get("content")))
return "\n".join(p for p in parts if p)
return json.dumps(content, ensure_ascii=False)
def _image_part(block: dict) -> dict | None:
source = block.get("source") or {}
if source.get("type") == "base64":
media = source.get("media_type") or "image/png"
data = source.get("data") or ""
return {"type": "image_url", "image_url": {"url": f"data:{media};base64,{data}"}}
if source.get("type") == "url" and source.get("url"):
return {"type": "image_url", "image_url": {"url": source["url"]}}
return None
def _user_message(content: Any) -> list[dict]:
"""Convert one Anthropic user message into OpenAI messages.
tool_result blocks become role=tool messages; text/image blocks become a
following role=user message. Order matters: OpenAI requires tool results
immediately after the assistant tool_calls message.
"""
if isinstance(content, str):
return [{"role": "user", "content": content}]
if not isinstance(content, list):
return [{"role": "user", "content": _text_of(content)}]
tool_messages: list[dict] = []
user_parts: list[dict] = []
for block in content:
if not isinstance(block, dict):
user_parts.append({"type": "text", "text": str(block)})
continue
kind = block.get("type")
if kind == "tool_result":
text = _text_of(block.get("content"))
if block.get("is_error"):
text = "Error: " + text
tool_messages.append({
"role": "tool",
"tool_call_id": block.get("tool_use_id") or "",
"content": text,
})
elif kind == "text":
user_parts.append({"type": "text", "text": str(block.get("text", ""))})
elif kind == "image":
part = _image_part(block)
if part:
user_parts.append(part)
# thinking / redacted_thinking and other beta blocks are ignored.
messages = tool_messages
if user_parts:
if len(user_parts) == 1 and user_parts[0].get("type") == "text":
messages.append({"role": "user", "content": user_parts[0]["text"]})
else:
messages.append({"role": "user", "content": user_parts})
return messages
def _assistant_message(content: Any) -> dict:
if isinstance(content, str):
return {"role": "assistant", "content": content}
if not isinstance(content, list):
return {"role": "assistant", "content": _text_of(content)}
texts: list[str] = []
tool_calls: list[dict] = []
for block in content:
if not isinstance(block, dict):
texts.append(str(block))
elif block.get("type") == "text":
texts.append(str(block.get("text", "")))
elif block.get("type") == "tool_use":
tool_calls.append({
"id": block.get("id") or _id("call_"),
"type": "function",
"function": {
"name": block.get("name") or "",
"arguments": json.dumps(block.get("input") or {}, ensure_ascii=False),
},
})
message: dict[str, Any] = {"role": "assistant"}
text = "\n".join(t for t in texts if t)
if text:
message["content"] = text
else:
message["content"] = None
if tool_calls:
message["tool_calls"] = tool_calls
return message
def anthropic_convert_request(body: dict, max_tokens_cap: int | None) -> dict:
messages: list[dict] = []
system = body.get("system")
if system:
messages.append({"role": "system", "content": _text_of(system)})
for item in body.get("messages") or []:
if not isinstance(item, dict):
continue
role = item.get("role")
content = item.get("content")
if role == "assistant":
messages.append(_assistant_message(content))
elif role in ("user", "system"):
messages.extend(_user_message(content) if role == "user" else
[{"role": "system", "content": _text_of(content)}])
else:
messages.append({"role": role or "user", "content": _text_of(content)})
upstream: dict[str, Any] = {
"model": body.get("model"),
"messages": messages,
"stream": bool(body.get("stream")),
}
max_tokens = body.get("max_tokens")
if max_tokens_cap:
max_tokens = min(int(max_tokens or max_tokens_cap), max_tokens_cap)
if max_tokens:
upstream["max_tokens"] = max_tokens
for key in ("temperature", "top_p", "stop", "seed"):
if body.get(key) is not None:
upstream[key] = body[key]
tools = []
for tool in body.get("tools") or []:
if not isinstance(tool, dict):
continue
name = tool.get("name")
if not name:
continue
tools.append({
"type": "function",
"function": {
"name": name,
"description": tool.get("description") or "",
"parameters": tool.get("input_schema") or {"type": "object", "properties": {}},
},
})
if tools:
upstream["tools"] = tools
choice = body.get("tool_choice")
if isinstance(choice, dict):
kind = choice.get("type")
if kind == "tool" and choice.get("name"):
upstream["tool_choice"] = {
"type": "function",
"function": {"name": choice["name"]},
}
elif kind == "any":
upstream["tool_choice"] = "required"
elif kind == "none":
upstream["tool_choice"] = "none"
else:
upstream["tool_choice"] = "auto"
elif choice:
upstream["tool_choice"] = choice
if upstream["stream"]:
upstream["stream_options"] = {"include_usage": True}
return upstream
def _usage(usage: dict | None, fallback_output: int = 0) -> dict:
usage = usage or {}
prompt = int(usage.get("prompt_tokens") or 0)
completion = int(usage.get("completion_tokens") or fallback_output)
return {"input_tokens": prompt, "output_tokens": completion}
def _stop_reason(finish: str | None, has_tool: bool) -> str:
if has_tool or finish == "tool_calls":
return "tool_use"
if finish == "length":
return "max_tokens"
if finish == "content_filter":
return "end_turn"
return "end_turn"
def anthropic_convert_response(result: dict, model: str) -> dict:
choice = (result.get("choices") or [{}])[0]
message = choice.get("message") or {}
content: list[dict] = []
text = message.get("content")
if text:
content.append({"type": "text", "text": text})
for call in message.get("tool_calls") or []:
raw = (call.get("function") or {}).get("arguments") or "{}"
try:
parsed = json.loads(raw)
except (TypeError, ValueError):
parsed = {"raw": raw}
content.append({
"type": "tool_use",
"id": call.get("id") or _id("toolu_"),
"name": (call.get("function") or {}).get("name") or "",
"input": parsed if isinstance(parsed, dict) else {"value": parsed},
})
if not content:
content.append({"type": "text", "text": ""})
return {
"id": _id("msg_"),
"type": "message",
"role": "assistant",
"model": model,
"content": content,
"stop_reason": _stop_reason(choice.get("finish_reason"), bool(message.get("tool_calls"))),
"stop_sequence": None,
"usage": _usage(result.get("usage")),
}
class Handler(BaseHTTPRequestHandler):
protocol_version = "HTTP/1.1"
server_version = "anthropic-proxy/1.0"
@property
def cfg(self) -> argparse.Namespace:
return self.server.cfg # type: ignore[attr-defined]
def log_message(self, fmt: str, *args: Any) -> None:
sys.stderr.write("[%s] %s\n" % (time.strftime("%H:%M:%S"), fmt % args))
def _json(self, status: int, payload: dict) -> None:
raw = json.dumps(payload, ensure_ascii=False).encode()
self.send_response(status)
self.send_header("Content-Type", "application/json")
self.send_header("Content-Length", str(len(raw)))
self.end_headers()
self.wfile.write(raw)
def _error(self, status: int, kind: str, message: str) -> None:
self._json(status, {"type": "error", "error": {"type": kind, "message": message}})
def do_GET(self) -> None:
if usage_endpoint(self):
return
route = urlsplit(self.path).path.rstrip("/")
if route in ("/health", "/v1/health"):
self._json(200, {"status": "ok", "upstream": self.cfg.upstream})
elif route in ("/models", "/v1/models"):
self._json(200, {
"object": "list",
"data": [{"id": self.cfg.model or "model", "object": "model", "owned_by": "proxy"}],
})
else:
self._error(404, "not_found_error", "Unknown route.")
def do_POST(self) -> None:
if usage_endpoint(self):
return
if urlsplit(self.path).path.rstrip("/") not in ("/v1/messages", "/messages"):
self._error(404, "not_found_error", "Only /v1/messages is supported.")
return
try:
size = int(self.headers.get("Content-Length", "0"))
if size <= 0 or size > 100 * 1024 * 1024:
raise ValueError("A JSON request body of at most 100 MiB is required.")
body = json.loads(self.rfile.read(size))
if not isinstance(body, dict):
raise ValueError("The request must be a JSON object.")
upstream_body = anthropic_convert_request(body, self.cfg.max_tokens)
except (ValueError, TypeError, KeyError) as exc:
self._error(400, "invalid_request_error", str(exc) or "Invalid request.")
return
url = self.cfg.upstream.rstrip("/") + "/chat/completions"
headers = {"Content-Type": "application/json"}
if self.cfg.api_key:
headers["Authorization"] = "Bearer " + self.cfg.api_key
try:
with requests.Session() as session:
session.trust_env = False
with metered_post(self.server, session,
url, json=upstream_body, headers=headers,
stream=upstream_body["stream"],
timeout=(15, self.cfg.timeout), allow_redirects=False,
) as result:
if result.status_code != 200:
raw = result.content[:2000]
try:
detail = json.loads(raw).get("error", raw.decode("utf-8", "replace"))
except Exception:
detail = raw.decode("utf-8", "replace")
self._error(502, "upstream_error", str(detail))
return
if upstream_body["stream"]:
self._stream(result, body.get("model") or "")
else:
self._json(200, anthropic_convert_response(result.json(), body.get("model") or ""))
except requests.RequestException as exc:
self._error(502, "upstream_error", f"Upstream connection failed: {exc.__class__.__name__}")
def _stream(self, result: requests.Response, model: str) -> None:
self.send_response(200)
self.send_header("Content-Type", "text/event-stream; charset=utf-8")
self.send_header("Cache-Control", "no-cache")
self.send_header("Connection", "close")
self.send_header("X-Accel-Buffering", "no")
self.end_headers()
self.close_connection = True
def event(kind: str, payload: dict) -> None:
data = json.dumps(payload, ensure_ascii=False)
self.wfile.write(f"event: {kind}\ndata: {data}\n\n".encode())
self.wfile.flush()
started = False
text_started = False
text_index = 0
next_index = 1
calls: dict[int, dict] = {}
finish: str | None = None
usage: dict | None = None
def start_message(prompt_tokens: int = 0) -> None:
nonlocal started
if started:
return
started = True
event("message_start", {
"type": "message_start",
"message": {
"id": _id("msg_"),
"type": "message",
"role": "assistant",
"model": model,
"content": [],
"stop_reason": None,
"stop_sequence": None,
"usage": {"input_tokens": prompt_tokens, "output_tokens": 0},
},
})
try:
for line in result.iter_lines(decode_unicode=True):
if not line or not line.startswith("data:"):
continue
data = line[5:].strip()
if data == "[DONE]":
break
chunk = json.loads(data)
if chunk.get("usage"):
usage = chunk["usage"]
choices = chunk.get("choices") or []
if not choices:
continue
choice = choices[0]
delta = choice.get("delta") or {}
if not started:
start_message(int((usage or {}).get("prompt_tokens") or 0))
text = delta.get("content")
if text:
if not text_started:
text_started = True
event("content_block_start", {
"type": "content_block_start",
"index": text_index,
"content_block": {"type": "text", "text": ""},
})
event("content_block_delta", {
"type": "content_block_delta",
"index": text_index,
"delta": {"type": "text_delta", "text": text},
})
for call in delta.get("tool_calls") or []:
index = int(call.get("index") or 0)
slot = calls.setdefault(index, {"id": "", "name": "", "arguments": ""})
if call.get("id"):
slot["id"] = call["id"]
function = call.get("function") or {}
if function.get("name"):
slot["name"] = function["name"]
if function.get("arguments"):
slot["arguments"] += function["arguments"]
if choice.get("finish_reason"):
finish = choice["finish_reason"]
if not started:
start_message(int((usage or {}).get("prompt_tokens") or 0))
if text_started:
event("content_block_stop", {"type": "content_block_stop", "index": text_index})
for _, slot in sorted(calls.items()):
event("content_block_start", {
"type": "content_block_start",
"index": next_index,
"content_block": {
"type": "tool_use",
"id": slot["id"] or _id("toolu_"),
"name": slot["name"],
"input": {},
},
})
event("content_block_delta", {
"type": "content_block_delta",
"index": next_index,
"delta": {"type": "input_json_delta", "partial_json": slot["arguments"] or "{}"},
})
event("content_block_stop", {"type": "content_block_stop", "index": next_index})
next_index += 1
event("message_delta", {
"type": "message_delta",
"delta": {"stop_reason": _stop_reason(finish, bool(calls)), "stop_sequence": None},
"usage": _usage(usage),
})
event("message_stop", {"type": "message_stop"})
except (BrokenPipeError, ConnectionResetError):
self.close_connection = True
except Exception as exc: # noqa: BLE001 - report a protocol error to the client
try:
event("error", {
"type": "error",
"error": {"type": "api_error", "message": f"Proxy stream error: {exc.__class__.__name__}"},
})
except (BrokenPipeError, ConnectionResetError):
self.close_connection = True
def anthropic_main(argv=None) -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--listen", "--host", dest="listen", default="127.0.0.1")
parser.add_argument("--port", type=int, default=18008)
parser.add_argument("--upstream", required=True, help="OpenAI-compatible base, e.g. http://host:8008/v1")
parser.add_argument("--api-key", default=os.environ.get("PROXY_API_KEY") or os.environ.get("ANTHROPIC_PROXY_API_KEY", ""))
parser.add_argument("--max-tokens", type=int, default=16384,
help="Clamp Anthropic max_tokens to this value (0 disables).")
parser.add_argument("--timeout", type=int, default=3600)
parser.add_argument("--model", default="", help="Optional model id returned by GET /v1/models.")
args = parser.parse_args(argv)
server = ThreadingHTTPServer((args.listen, args.port), Handler)
server.cfg = args # type: ignore[attr-defined]
server.usage_meter = TransportMeter()
start_parent_watch()
print(f"anthropic proxy listening on http://{args.listen}:{args.port} -> {args.upstream}", flush=True)
try:
server.serve_forever()
except KeyboardInterrupt:
pass
finally:
server.server_close()
def main() -> None:
"""Select the protocol implementation and start one local server."""
if len(sys.argv) > 1 and sys.argv[1] in ("-h", "--help"):
print("usage: proxy.py --protocol {anthropic,responses,passthrough} [options]")
print(" --protocol anthropic expose POST /v1/messages")
print(" --protocol responses expose POST /v1/responses")
print(" --protocol passthrough preserve native HTTP protocol and observe provider usage")
print("Run with --protocol <name> --help for protocol-specific options.")
return
parser = argparse.ArgumentParser(add_help=False)
parser.add_argument("--protocol", choices=("anthropic", "responses", "passthrough"), required=True)
parser.add_argument("--parent-pid", type=int, default=0)
args, rest = parser.parse_known_args()
if args.parent_pid > 0:
os.environ["PROXY_PARENT_PID"] = str(args.parent_pid)
if args.protocol == "passthrough":
passthrough_main(rest)
elif args.protocol == "anthropic":
anthropic_main(rest)
else:
responses_main(rest)
if __name__ == "__main__":
main()
|