Repository navigation
Expand file tree
/
Copy pathllm.py
More file actions
680 lines (598 loc) · 21.7 KB
/
Copy pathllm.py
File metadata and controls
680 lines (598 loc) · 21.7 KB
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
from __future__ import annotations
import json
from dataclasses import dataclass
from typing import Any, Iterator
import anthropic
import httpx
_OPENAI_IMPORT_ERROR: Exception | None = None
try:
from openai import OpenAI
import openai
except Exception as exc: # pragma: no cover - exercised in tests via stubs
OpenAI = None # type: ignore[assignment]
openai = None # type: ignore[assignment]
_OPENAI_IMPORT_ERROR = exc
ProviderName = str
_ANTHROPIC_PROVIDER = "anthropic"
_OPENAI_PROVIDER = "openai"
_VALID_PROVIDERS = {_ANTHROPIC_PROVIDER, _OPENAI_PROVIDER}
MODEL_CONTEXT_WINDOW_DEFAULT = 200_000
def get_context_window_for_model(model: str) -> int:
"""Return the context window size for a given model."""
lowered = model.lower()
if "[1m]" in lowered or "1m" in lowered.split("-"):
return 1_000_000
return MODEL_CONTEXT_WINDOW_DEFAULT
# Upper limit for max output tokens — used for auto-escalation scenarios.
_MODEL_MAX_OUTPUT_UPPER: dict[str, int] = {
"claude-opus-4-6": 64000,
"claude-sonnet-4-6": 64000,
"claude-opus-4-5": 64000,
"claude-sonnet-4-5": 64000,
"claude-sonnet-4": 64000,
"claude-opus-4-1": 64000,
"claude-opus-4": 64000,
"claude-3-7-sonnet": 64000,
"claude-3-5-sonnet": 8192,
"claude-3-5-haiku": 8192,
}
def get_max_output_tokens_upper(model: str) -> int | None:
"""Return the upper limit of output tokens for a model, if known."""
for prefix, limit in _MODEL_MAX_OUTPUT_UPPER.items():
if model.startswith(prefix):
return limit
return None
@dataclass
class LLMUsage:
input_tokens: int = 0
output_tokens: int = 0
cache_read_input_tokens: int = 0
cache_creation_input_tokens: int = 0
advisor_input_tokens: int = 0
advisor_output_tokens: int = 0
@dataclass
class LLMMessage:
content: list[dict[str, Any]]
usage: LLMUsage | None = None
stop_reason: str | None = None # "end_turn", "max_tokens", "tool_use", etc.
def validate_provider(provider: str | None) -> ProviderName:
normalized = (provider or _ANTHROPIC_PROVIDER).strip().lower()
if normalized not in _VALID_PROVIDERS:
raise ValueError(f"Unsupported provider: {provider}")
return normalized
def default_model_for_provider(provider: str) -> str:
provider = validate_provider(provider)
if provider == _OPENAI_PROVIDER:
return "gpt-5.1-codex"
return "claude-sonnet-4-6"
def default_companion_model(provider: str, model: str) -> str:
provider = validate_provider(provider)
if provider == _OPENAI_PROVIDER:
return model
return "claude-haiku-4-5-20251001"
def default_max_tokens_for_provider(provider: str) -> int:
provider = validate_provider(provider)
if provider == _OPENAI_PROVIDER:
return 8192
return 32000
def supports_reasoning_effort(provider: str, model: str) -> bool:
provider = validate_provider(provider)
if provider != _OPENAI_PROVIDER:
return False
lowered = model.lower()
return lowered.startswith(("gpt-5", "o1", "o3", "o4"))
class LLMClient:
def __init__(
self,
provider: str = _ANTHROPIC_PROVIDER,
api_key: str | None = None,
base_url: str | None = None,
):
self.provider = validate_provider(provider)
self._api_key = api_key
self._base_url = base_url
if self.provider == _OPENAI_PROVIDER:
if OpenAI is None:
message = "OpenAI support requires the `openai` package to be installed."
if _OPENAI_IMPORT_ERROR is not None:
message += f" Import failed: {_OPENAI_IMPORT_ERROR}"
raise ValueError(message)
self._client = OpenAI(api_key=api_key, base_url=base_url)
else:
self._client = anthropic.Anthropic(
api_key=api_key,
base_url=base_url,
timeout=httpx.Timeout(600.0, connect=30.0),
)
def create_message(
self,
*,
model: str,
max_tokens: int,
messages: list[dict[str, Any]],
system: str | None = None,
tools: list[dict[str, Any]] | None = None,
effort: str | None = None,
) -> LLMMessage:
if self.provider == _OPENAI_PROVIDER:
return self._openai_create_message(
model=model,
max_tokens=max_tokens,
messages=messages,
system=system,
tools=tools,
effort=effort,
)
return self._anthropic_create_message(
model=model,
max_tokens=max_tokens,
messages=messages,
system=system,
tools=tools,
)
def stream_messages(
self,
*,
model: str,
max_tokens: int,
messages: list[dict[str, Any]],
8000
system: str | None = None,
tools: list[dict[str, Any]] | None = None,
effort: str | None = None,
):
if self.provider == _OPENAI_PROVIDER:
return _OpenAIStream(
client=self._client,
model=model,
max_tokens=max_tokens,
messages=messages,
system=system,
tools=tools or [],
effort=effort,
)
return _AnthropicStream(
client=self._client,
model=model,
max_tokens=max_tokens,
messages=messages,
system=system,
tools=tools or [],
)
def is_authentication_error(self, exc: Exception) -> bool:
if self.provider == _OPENAI_PROVIDER:
return openai is not None and isinstance(exc, openai.AuthenticationError)
return isinstance(exc, anthropic.AuthenticationError)
def is_retryable_error(self, exc: Exception) -> bool:
if isinstance(exc, (httpx.RemoteProtocolError, httpx.ReadError, httpx.ConnectError)):
return True
if self.provider == _OPENAI_PROVIDER:
return openai is not None and isinstance(
exc,
(
openai.RateLimitError,
openai.APIConnectionError,
openai.InternalServerError,
),
)
return isinstance(
exc,
(
anthropic.RateLimitError,
anthropic.APIConnectionError,
anthropic.InternalServerError,
),
)
def is_api_error(self, exc: Exception) -> bool:
if self.provider == _OPENAI_PROVIDER:
return openai is not None and isinstance(exc, openai.APIError)
return isinstance(exc, anthropic.APIError)
@staticmethod
def error_message(exc: Exception) -> str:
return str(getattr(exc, "message", None) or exc)
def _anthropic_create_message(
self,
*,
model: str,
max_tokens: int,
messages: list[dict[str, Any]],
system: str | None,
tools: list[dict[str, Any]] | None,
) -> LLMMessage:
kwargs: dict[str, Any] = dict(
model=model,
max_tokens=max_tokens,
messages=messages,
)
if system:
kwargs["system"] = system
if tools:
kwargs["tools"] = tools
response = self._client.messages.create(**kwargs)
usage = _usage_from_anthropic(getattr(response, "usage", None))
return LLMMessage(
content=_normalize_anthropic_content(getattr(response, "content", [])),
usage=usage,
stop_reason=getattr(response, "stop_reason", None),
)
def _openai_create_message(
self,
*,
model: str,
max_tokens: int,
messages: list[dict[str, Any]],
system: str | None,
tools: list[dict[str, Any]] | None,
effort: str | None,
) -> LLMMessage:
params = _build_openai_request(
model=model,
max_tokens=max_tokens,
system=system,
messages=messages,
tools=tools or [],
effort=effort,
stream=False,
)
response = self._client.chat.completions.create(**params)
choice = response.choices[0] if response.choices else None
finish_reason = _value(choice, "finish_reason") if choice else None
message = choice.message if choice else None
usage = _usage_from_openai(getattr(response, "usage", None))
return LLMMessage(
content=_normalize_openai_message(message),
usage=usage,
stop_reason=_normalize_openai_stop_reason(finish_reason),
)
class _AnthropicStream:
def __init__(
self,
*,
client: Any,
model: str,
max_tokens: int,
messages: list[dict[str, Any]],
system: str | None,
tools: list[dict[str, Any]],
):
self._raw = client.messages.stream(
model=model,
max_tokens=max_tokens,
system=system,
tools=tools,
messages=messages,
)
self._ctx = None
self.text_stream: Iterator[str] = iter(())
def __enter__(self):
self._ctx = self._raw.__enter__()
self.text_stream = self._ctx.text_stream
return self
def __exit__(self, exc_type, exc, tb):
return self._raw.__exit__(exc_type, exc, tb)
def close(self) -> None:
"""Signal cancellation — best-effort close of the underlying stream."""
try:
self._raw.close()
except Exception:
pass
def get_final_message(self) -> LLMMessage:
final = self._ctx.get_final_message()
return LLMMessage(
content=_normalize_anthropic_content(getattr(final, "content", [])),
usage=_usage_from_anthropic(getattr(final, "usage", None)),
stop_reason=getattr(final, "stop_reason", None),
)
class _OpenAIStream:
def __init__(
self,
*,
client: Any,
model: str,
max_tokens: int,
messages: list[dict[str, Any]],
system: str | None,
tools: list[dict[str, Any]],
effort: str | None,
):
self._client = client
self._params = _build_openai_request(
model=model,
max_tokens=max_tokens,
system=system,
messages=messages,
tools=tools,
effort=effort,
stream=True,
)
self._stream = None
self._text_parts: list[str] = []
self._tool_calls: dict[int, dict[str, Any]] = {}
self._usage: LLMUsage | None = None
self._finish_reason: str | None = None
self.text_stream: Iterator[str] = iter(())
def __enter__(self):
self._stream = self._client.chat.completions.create(**self._params)
self.text_stream = self._iter_text()
return self
def __exit__(self, exc_type, exc, tb):
self.close()
return False
def close(self) -> None:
if self._stream is not None and hasattr(self._stream, "close"):
self._stream.close()
def _iter_text(self) -> Iterator[str]:
for chunk in self._stream:
usage = getattr(chunk, "usage", None)
if usage is not None:
self._usage = _usage_from_openai(usage)
for choice in _value(chunk, "choices", []) or []:
finish_reason = _value(choice, "finish_reason")
if finish_reason:
self._finish_reason = finish_reason
delta = _value(choice, "delta", {}) or {}
content = _value(delta, "content")
if content:
self._text_parts.append(content)
yield content
for tool_call in _value(delta, "tool_calls", []) or []:
index = int(_value(tool_call, "index", 0) or 0)
entry = self._tool_calls.setdefault(index, {
"id": "",
"name": "",
"arguments": "",
})
tool_id = _value(tool_call, "id")
if tool_id:
entry["id"] = tool_id
function = _value(tool_call, "function", {}) or {}
name = _value(function, "name")
if name:
entry["name"] = name
arguments = _value(function, "arguments")
if arguments:
entry["arguments"] += arguments
def get_final_message(self) -> LLMMessage:
content: list[dict[str, Any]] = []
text = "".join(self._text_parts)
if text:
content.append({"type": "text", "text": text})
for index in sorted(self._tool_calls):
tool_call = self._tool_calls[index]
raw_args = tool_call.get("arguments", "").strip()
parsed_args: Any = {}
if raw_args:
try:
parsed_args = json.loads(raw_args)
except json.JSONDecodeError:
parsed_args = {}
content.append({
"type": "tool_use",
"id": tool_call.get("id", ""),
"name": tool_call.get("name", ""),
"input": parsed_args if isinstance(parsed_args, dict) else {},
})
return LLMMessage(
content=content,
usage=self._usage,
stop_reason=_normalize_openai_stop_reason(self._finish_reason),
)
def _normalize_openai_stop_reason(reason: str | None) -> str | None:
"""Map OpenAI finish_reason to Anthropic stop_reason equivalents."""
if reason is None:
return None
_MAPPING = {
"stop": "end_turn",
"length": "max_tokens",
"tool_calls": "tool_use",
}
return _MAPPING.get(reason, reason)
def _normalize_anthropic_content(content: Any) -> list[dict[str, Any]]:
blocks: list[dict[str, Any]] = []
for block in content or []:
normalized = _normalize_anthropic_block(block)
if normalized is not None:
blocks.append(normalized)
return blocks
def _normalize_anthropic_block(block: Any) -> dict[str, Any] | None:
block_type = _value(block, "type")
if block_type == "text":
return {"type": "text", "text": _value(block, "text", "")}
if block_type == "tool_use":
return {
"type": "tool_use",
"id": _value(block, "id", ""),
"name": _value(block, "name", ""),
"input": _value(block, "input", {}) or {},
}
if block_type == "tool_result":
normalized = {
"type": "tool_result",
"tool_use_id": _value(block, "tool_use_id", ""),
"content": _value(block, "content", ""),
}
is_error = _value(block, "is_error")
if is_error is not None:
normalized["is_error"] = bool(is_error)
return normalized
if block_type == "image":
return {
"type": "image",
"source": _value(block, "source", {}),
}
if isinstance(block, dict):
return dict(block)
if hasattr(block, "model_dump"):
return block.model_dump()
return None
def _normalize_openai_message(message: Any) -> list[dict[str, Any]]:
if message is None:
return []
content: list[dict[str, Any]] = []
text = _extract_openai_text(_value(message, "content"))
if text:
content.append({"type": "text", "text": text})
for tool_call in _value(message, "tool_calls", []) or []:
arguments = _value(_value(tool_call, "function", {}) or {}, "arguments", "") or ""
parsed_args: Any = {}
if arguments:
try:
parsed_args = json.loads(arguments)
except json.JSONDecodeError:
parsed_args = {}
content.append({
"type": "tool_use",
"id": _value(tool_call, "id", ""),
"name": _value(_value(tool_call, "function", {}) or {}, "name", ""),
"input": parsed_args if isinstance(parsed_args, dict) else {},
})
return content
def _extract_openai_text(content: Any) -> str:
if content is None:
return ""
if isinstance(content, str):
return content
parts: list[str] = []
for item in content:
item_type = _value(item, "type")
if item_type == "text":
text = _value(item, "text")
if isinstance(text, str):
parts.append(text)
elif isinstance(text, dict):
parts.append(str(text.get("value", "")))
return "".join(parts)
def _usage_from_anthropic(usage: Any) -> LLMUsage | None:
if usage is None:
return None
return LLMUsage(
input_tokens=int(_value(usage, "input_tokens", 0) or 0),
output_tokens=int(_value(usage, "output_tokens", 0) or 0),
cache_read_input_tokens=int(_value(usage, "cache_read_input_tokens", 0) or 0),
cache_creation_input_tokens=int(_value(usage, "cache_creation_input_tokens", 0) or 0),
advisor_input_tokens=int(_value(usage, "advisor_input_tokens", 0) or 0),
advisor_output_tokens=int(_value(usage, "advisor_output_tokens", 0) or 0),
)
def _usage_from_openai(usage: Any) -> LLMUsage | None:
if usage is None:
return None
return LLMUsage(
input_tokens=int(_value(usage, "prompt_tokens", 0) or 0),
output_tokens=int(_value(usage, "completion_tokens", 0) or 0),
)
def _build_openai_request(
*,
model: str,
max_tokens: int,
system: str | None,
messages: list[dict[str, Any]],
tools: list[dict[str, Any]],
effort: str | None,
stream: bool,
) -> dict[str, Any]:
params: dict[str, Any] = {
"model": model,
"messages": _to_openai_messages(system, messages),
"max_tokens": max_tokens,
"stream": stream,
}
if tools:
params["tools"] = [_tool_schema_to_openai(tool) for tool in tools]
if effort and supports_reasoning_effort(_OPENAI_PROVIDER, model):
params["reasoning_effort"] = effort
return params
def _to_openai_messages(system: str | None, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
out: list[dict[str, Any]] = []
if system:
out.append({"role": "system", "content": system})
for message in messages:
role = message.get("role")
content = message.get("content", "")
if role == "user" and isinstance(content, list):
tool_results = [
block for block in content
if isinstance(block, dict) and block.get("type") == "tool_result"
]
if tool_results and len(tool_results) == len(content):
for block in tool_results:
out.append({
"role": "tool",
"tool_call_id": block.get("tool_use_id", ""),
"content": _tool_result_to_text(block.get("content", "")),
})
continue
out.append({
"role": "user",
"content": _user_content_blocks_to_openai(content),
})
continue
if role == "assistant" and isinstance(content, list):
text_parts: list[str] = []
tool_calls: list[dict[str, Any]] = []
for block in content:
if not isinstance(block, dict):
continue
block_type = block.get("type")
if block_type == "text":
text_parts.append(block.get("text", ""))
elif block_type == "tool_use":
tool_calls.append({
"id": block.get("id", ""),
"type": "function",
"function": {
"name": block.get("name", ""),
"arguments": json.dumps(block.get("input", {}), ensure_ascii=False),
},
})
assistant_message: dict[str, Any] = {
"role": "assistant",
"content": "".join(text_parts) or None,
}
if tool_calls:
assistant_message["tool_calls"] = tool_calls
out.append(assistant_message)
continue
out.append({
"role": role,
"content": content,
})
return out
def _user_content_blocks_to_openai(content: list[Any]) -> list[dict[str, Any]]:
parts: list[dict[str, Any]] = []
for block in content:
if not isinstance(block, dict):
continue
block_type = block.get("type")
if block_type == "text":
parts.append({"type": "text", "text": block.get("text", "")})
elif block_type == "image":
source = block.get("source", {})
media_type = source.get("media_type", "image/png")
data = source.get("data", "")
parts.append({
"type": "image_url",
"image_url": {"url": f"data:{media_type};base64,{data}"},
})
if not parts:
return [{"type": "text", "text": ""}]
return parts
def _tool_schema_to_openai(tool: dict[str, Any]) -> dict[str, Any]:
return {
"type": "function",
"function": {
"name": tool.get("name", ""),
"description": tool.get("description", ""),
"parameters": tool.get("input_schema", {}),
},
}
def _tool_result_to_text(content: Any) -> str:
if isinstance(content, str):
return content
if content is None:
return ""
return json.dumps(content, ensure_ascii=False)
def _value(obj: Any, key: str, default: Any = None) -> Any:
if obj is None:
return default
if isinstance(obj, dict):
return obj.get(key, default)
return getattr(obj, key, default)