Files
OpenOPC/opc/llm/provider.py
T
LZH-YS1998 d14f3920e0 fix(company): stop/resume identity truth, failure-path closure, quota park
OBS-11 — stop/resume killed pure-native runs over a phantom external pin.
Role templates' preferred_external_agent leaked into execution identity even
when the user requested native and execution actually ran native; on resume
the availability gate trusted the pin and failed every non-terminal item.
Root fixes across the whole chain:
- Staffing card per-role defaults are now the RESOLVED backend (explicit
  session agent choice > runnable template preference > native), never a
  hardcoded external default; seat enrichment and the dispatch selector's
  locked branch downgrade provably unavailable externals to native and
  record the wish in execution_agent_unavailable.
- The resume availability gate fails closed only when a resumable external
  session actually exists; a bare pin heals to native (snapshot AND task
  durable identity) and the run resumes — mirroring dispatch fallback.
- Suspend-checkpoint replies: force_resume (chat/headless spelling) is
  recognized alongside ui_force_resume, and bare continuation tokens
  (English and Chinese spellings) take the plain-resume path instead of
  being routed to the final decider as content, which reopened the
  already-approved intake card.

OBS-5 — failed runs never closed and dropped new input. The dispatcher's
convergence exit now settles terminally-failed runs (status=failed,
lifecycle=closed_failed, run_failure metadata) and emits a
company_run_failure_review card whose replies never swallow messages:
dismiss acknowledges, content falls through so normal routing starts a
fresh run. _maybe_resume_existing_company_runtime no longer re-executes a
terminally-failed tree: control replies get an honest closed status,
content-bearing input starts a new run.

OBS-6 — provider quota exhaustion terminally failed work items. Rate-limit
rejections are classified (LLMProvider.is_rate_limit_error, covering
status codes, exception types, and English/Chinese provider error text),
the agent runtime raises typed ProviderQuotaExhaustedError instead of
burning conversation-feedback retries, and the company dispatcher parks:
the item returns to READY (attempt interrupted, no terminal failure), the
member session idles, and claiming backs off exponentially (60s doubling
to a 900s cap; a quiet 30min resets the streak) before resuming
automatically.

Verified end-to-end on the real minimax-m3 campaign: same goal, same 300s
stop point, same run shape that previously killed the whole tree within
90s now resumes cleanly and completes with all items approved; staffing
defaults native for all 11 roles.

Tests: test_stop_resume_native_pin (10), test_run_failure_settlement (6),
test_provider_quota_park (9); attempt-ledger, recruiter, and
suspend-resume suites updated to the new contracts (their old assertions
pinned the defective behaviors).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-28 17:03:40 +08:00

995 lines
37 KiB
Python

"""LLM provider layer built on LiteLLM for unified model access."""
from __future__ import annotations
import base64
import hashlib
import json
import os
from pathlib import Path
from typing import Any, AsyncIterator
from urllib.parse import urlparse
from opc.core.windows_ssl import sanitize_windows_sslkeylogfile
sanitize_windows_sslkeylogfile()
import litellm
from loguru import logger
from opc.core.attachment_content import attachment_suffix
from opc.core.attachment_store import AttachmentRef
from opc.core.config import LLMConfig
from opc.core.models import ModelCapabilitySet, RuntimeLLMEvent
litellm.suppress_debug_info = True
litellm.drop_params = True
_MULTIMODAL_MODEL_HINTS = (
"gpt-4.1",
"gpt-4o",
"gpt-4.5",
"gpt-5",
"o1",
"o3",
"o4",
"claude-3",
"claude-sonnet-4",
"claude-opus-4",
"gemini",
"pixtral",
"llava",
"qwen-vl",
"qwen2-vl",
"qwen2.5-vl",
"internvl",
"minicpm-v",
"glm-4v",
)
_DOCUMENT_MODEL_HINTS = (
"gpt-4.1",
"gpt-4o",
"gpt-4.5",
"gpt-5",
"claude-3",
"claude-sonnet-4",
"claude-opus-4",
"gemini",
)
_VIDEO_MODEL_HINTS = (
"gemini",
"veo",
"video",
)
_TOOL_PROTOCOL_ERROR_HINTS = (
"no tool output found for function call",
"no tool output found",
"messages with role 'tool' must be a response to a preceding message with 'tool_calls'",
"assistant message with tool_calls",
"tool_calls must be followed by tool messages",
"missing tool response",
"missing tool output",
"tool_call_id",
)
def _normalized_model_name(model: str) -> str:
if "/" in model:
return model.split("/", 1)[1].strip().lower()
return model.strip().lower()
# Used when neither user config nor litellm can supply a window. Conservative
# enough for modern models so compaction still has a real denominator.
_CONTEXT_WINDOW_FALLBACK = 128_000
_context_window_fallback_warned: set[str] = set()
_max_tokens_clamp_warned: set[str] = set()
def _clamp_max_tokens(model: str, requested: int) -> int:
"""Cap the requested output tokens at the model's known output limit.
Providers disagree on how to handle an oversized max_tokens: some clamp
silently, others (e.g. DeepSeek) reject the request outright. Clamping
here keeps a generous config default (32768) safe on small-cap models.
Unknown models pass through unchanged.
"""
try:
info = litellm.get_model_info(model)
cap = info.get("max_output_tokens") or info.get("max_tokens")
except Exception:
return requested
if not cap or requested <= int(cap):
return requested
if model not in _max_tokens_clamp_warned:
_max_tokens_clamp_warned.add(model)
logger.info(
"max_tokens {} exceeds output limit {} of model={}; clamping.",
requested,
cap,
model,
)
return int(cap)
_CONTEXT_WINDOW_OVERRIDES: tuple[tuple[str, int], ...] = (
("gpt-5.4-pro", 1_050_000),
("gpt-5.4-mini", 400_000),
("gpt-5.4-nano", 400_000),
("gpt-5.4", 1_050_000),
("gpt-5-pro", 400_000),
("gpt-5", 400_000),
)
_POE_CONTEXT_WINDOW_OVERRIDES: tuple[tuple[str, int], ...] = (
("claude-sonnet-4.5", 64_000),
("claude-sonnet-4-5", 64_000),
)
def _context_window_override(model: str) -> int | None:
normalized = _normalized_model_name(model)
for prefix, window in _CONTEXT_WINDOW_OVERRIDES:
if normalized == prefix or normalized.startswith(f"{prefix}-"):
return window
return None
def _poe_context_window_override(model: str) -> int | None:
normalized = _normalized_model_name(model)
for prefix, window in _POE_CONTEXT_WINDOW_OVERRIDES:
if normalized == prefix or normalized.startswith(f"{prefix}-"):
return window
return None
def _is_official_openai_base(api_base: str | None) -> bool:
normalized = str(api_base or "").strip()
if not normalized:
return True
try:
parsed = urlparse(normalized)
except Exception:
return False
hostname = (parsed.hostname or "").strip().lower()
return hostname in {"api.openai.com", "openai.com"}
def _is_poe_base(api_base: str | None) -> bool:
normalized = str(api_base or "").strip()
if not normalized:
return False
try:
parsed = urlparse(normalized)
except Exception:
return False
hostname = (parsed.hostname or "").strip().lower()
return hostname == "api.poe.com"
def _looks_like_multimodal_model(model: str) -> bool:
normalized = _normalized_model_name(model)
if any(hint in normalized for hint in _MULTIMODAL_MODEL_HINTS):
return True
return normalized.endswith("-vl") or "-vl-" in normalized or "_vl" in normalized
def _looks_like_document_capable_model(model: str) -> bool:
normalized = _normalized_model_name(model)
return any(hint in normalized for hint in _DOCUMENT_MODEL_HINTS)
def _looks_like_video_capable_model(model: str) -> bool:
normalized = _normalized_model_name(model)
return any(hint in normalized for hint in _VIDEO_MODEL_HINTS)
def _parse_tool_arguments(tool_name: str, arguments: Any) -> tuple[Any, str | None, str | None]:
"""Parse tool-call arguments and preserve failures for downstream recovery."""
if not isinstance(arguments, str):
return arguments, None, None
raw = arguments
try:
parsed = json.loads(raw)
return parsed, raw, None
except json.JSONDecodeError as e:
snippet = raw[:500].replace("\n", "\\n")
error = f"Invalid tool arguments JSON for `{tool_name}`: {e.msg} at char {e.pos}"
logger.warning(f"{error}. Raw snippet: {snippet}")
return raw, raw, error
class ProviderQuotaExhaustedError(RuntimeError):
"""The provider rejected the request for quota/rate-limit reasons.
Raised by the agent runtime instead of retrying in place: replaying the
same payload against an exhausted quota can only fail, so the company
dispatcher parks the work (returns the item to READY and backs off)
rather than failing it terminally (OBS-6).
"""
class LLMProvider:
"""Unified LLM interface via LiteLLM supporting tool calls."""
# Well-known provider API-key env vars that litellm reads directly when no
# explicit api_key is passed. Used only by ``has_credentials()`` to avoid a
# false "no credentials" verdict for env-based setups. Missing a provider
# here just preserves the old behavior (a real LLM attempt), never a wrong
# skip of a working key.
_CREDENTIAL_ENV_VARS = (
"OPENAI_API_KEY",
"ANTHROPIC_API_KEY",
"OPENROUTER_API_KEY",
"AZURE_API_KEY",
"AZURE_OPENAI_API_KEY",
"GEMINI_API_KEY",
"GOOGLE_API_KEY",
"MISTRAL_API_KEY",
"GROQ_API_KEY",
"DEEPSEEK_API_KEY",
"TOGETHERAI_API_KEY",
"ARK_API_KEY",
)
def __init__(self, config: LLMConfig, opc_home: Path | None = None) -> None:
self.config = config
self.opc_home = opc_home
self._total_tokens_in = 0
self._total_tokens_out = 0
self._total_cost = 0.0
self._api_key = config.api_key or (
os.environ.get(config.api_key_env) if config.api_key_env else None
) or None
self._api_base = config.api_base or None
def has_credentials(self) -> bool:
"""Whether an LLM call can plausibly authenticate.
True when a key is configured (``api_key`` / ``api_key_env``) or a
well-known provider env var is present. False only when no credential
is found anywhere — callers use that to skip LLM work that would
certainly fail (e.g. native agent selection when an external agent can
run the task instead). A False at worst degrades to rule-based behavior,
which stays functional; it never blocks execution.
"""
if self._api_key:
return True
return any(os.environ.get(var) for var in self._CREDENTIAL_ENV_VARS)
@property
def stats(self) -> dict[str, Any]:
return {
"tokens_in": self._total_tokens_in,
"tokens_out": self._total_tokens_out,
"estimated_cost": self._total_cost,
}
def _select_model(self, task_type: str | None = None) -> str:
if task_type and task_type in self.config.routing:
return self.config.routing[task_type]
return self.config.default_model
def _config_context_window_override(self, model: str) -> int | None:
"""User-configured context window for models litellm cannot map.
Per-model overrides win over the scalar default. Both come from
``LLMConfig`` so proxy/self-hosted models (doubao, minimax, glm, …)
can report a real context window to the usage ring and compaction.
"""
per_model = getattr(self.config, "context_window_overrides", None) or {}
normalized = _normalized_model_name(model)
for key, window in per_model.items():
candidate = _normalized_model_name(str(key))
if candidate and (normalized == candidate or normalized.startswith(f"{candidate}-")):
try:
value = int(window)
except (TypeError, ValueError):
continue
if value > 0:
return value
try:
scalar = int(getattr(self.config, "context_window", 0) or 0)
except (TypeError, ValueError):
scalar = 0
return scalar if scalar > 0 else None
def get_context_window(self, task_type: str | None = None, model: str | None = None) -> int | None:
resolved_model = model or self._select_model(task_type)
config_override = self._config_context_window_override(resolved_model)
if config_override is not None:
return config_override
poe_override = _poe_context_window_override(resolved_model) if _is_poe_base(self._api_base) else None
if poe_override is not None:
return poe_override
override = _context_window_override(resolved_model) if _is_official_openai_base(self._api_base) else None
if override is not None:
return override
try:
# max_input_tokens is the context window; litellm.get_max_tokens()
# returns the "max_tokens" map entry, which for many models (e.g.
# deepseek) is the OUTPUT cap and would wildly under-report here.
info = litellm.get_model_info(resolved_model)
limit = info.get("max_input_tokens") or info.get("max_tokens")
if limit:
return int(limit)
reason = "model is not mapped in litellm"
except Exception as e:
reason = str(e)
if resolved_model not in _context_window_fallback_warned:
_context_window_fallback_warned.add(resolved_model)
logger.warning(
"Unable to resolve context window for model={} ({}); assuming {} tokens. "
"Set llm.context_window or llm.context_window_overrides in llm_config.yaml "
"to use the model's real window.",
resolved_model,
reason,
_CONTEXT_WINDOW_FALLBACK,
)
return _CONTEXT_WINDOW_FALLBACK
def count_input_tokens(
self,
messages: list[dict[str, Any]],
tools: list[dict[str, Any]] | None = None,
task_type: str | None = None,
model: str | None = None,
) -> int | None:
resolved_model = model or self._select_model(task_type)
try:
return int(litellm.token_counter(
model=resolved_model,
messages=messages,
tools=tools,
))
except Exception as e:
logger.warning(f"Unable to count prompt tokens for model={resolved_model}: {e}")
return None
def count_text_tokens(
self,
text: str,
task_type: str | None = None,
model: str | None = None,
) -> int | None:
resolved_model = model or self._select_model(task_type)
try:
return int(litellm.token_counter(
model=resolved_model,
text=text,
))
except Exception as e:
logger.warning(f"Unable to count text tokens for model={resolved_model}: {e}")
return None
def get_capabilities(
self,
task_type: str | None = None,
model: str | None = None,
) -> ModelCapabilitySet:
resolved_model = model or self._select_model(task_type)
normalized = _normalized_model_name(resolved_model)
provider_family = resolved_model.split("/", 1)[0].strip().lower() if "/" in resolved_model else ""
supports_thinking = any(hint in normalized for hint in ("o1", "o3", "o4", "gpt-5", "claude", "reason"))
return ModelCapabilitySet(
model=resolved_model,
supports_streaming=True,
supports_tool_calling=True,
supports_streaming_tool_calls=True,
supports_thinking=supports_thinking,
supports_multimodal=_looks_like_multimodal_model(resolved_model),
supports_documents=_looks_like_document_capable_model(resolved_model),
supports_video=_looks_like_video_capable_model(resolved_model),
provider_family=provider_family,
metadata={
"api_base": self._api_base or "",
},
)
def build_cache_fingerprint(
self,
*,
messages: list[dict[str, Any]],
tools: list[dict[str, Any]] | None = None,
task_type: str | None = None,
model: str | None = None,
extra: dict[str, Any] | None = None,
) -> str:
resolved_model = model or self._select_model(task_type)
payload = {
"model": resolved_model,
"messages": messages,
"tools": tools or [],
"extra": extra or {},
}
raw = json.dumps(payload, ensure_ascii=False, sort_keys=True, default=str)
return hashlib.sha256(raw.encode("utf-8")).hexdigest()
def is_rate_limit_error(self, error: Exception) -> bool:
"""Classify provider quota / rate-limit rejections.
These never produce model output and never get better by replaying
the identical payload, so callers must park/back off instead of
burning conversation-feedback retries (OBS-6). Classification is by
exception type when available and by error text otherwise — the
streaming path re-raises provider errors as plain RuntimeError with
only the message preserved.
"""
if isinstance(error, litellm.exceptions.RateLimitError):
return True
if "ratelimit" in type(error).__name__.lower():
return True
if getattr(error, "status_code", None) == 429:
return True
message = str(error).lower()
keywords = (
"rate limit",
"rate_limit",
"ratelimit",
"too many requests",
"insufficient_quota",
"quota exceeded",
"exceeded your quota",
"quota exhausted",
"error code: 429",
"status code: 429",
"http 429",
# Chinese-provider spellings of the same rejection (Volces/DeepSeek
# and other domestic endpoints return localized error text).
"请求过于频繁",
"配额已用完",
"配额耗尽",
"触发限流",
)
return any(keyword in message for keyword in keywords)
def is_context_overflow_error(self, error: Exception) -> bool:
if isinstance(error, litellm.exceptions.ContextWindowExceededError):
return True
message = str(error).lower()
keywords = (
"context window",
"context length",
"maximum context length",
"prompt is too long",
"too many tokens",
"context_length_exceeded",
"token limit exceeded",
)
return any(keyword in message for keyword in keywords)
def is_tool_protocol_error(self, error: Exception) -> bool:
message = str(error).lower()
return any(hint in message for hint in _TOOL_PROTOCOL_ERROR_HINTS)
@staticmethod
def sanitize_tool_call_history(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
"""Drop incomplete or stray assistant/tool tool-call transcripts."""
sanitized: list[dict[str, Any]] = []
pending_ids: list[str] = []
buffered_block: list[dict[str, Any]] = []
def _copy_message(message: dict[str, Any]) -> dict[str, Any]:
cloned = dict(message)
if isinstance(message.get("tool_calls"), list):
cloned["tool_calls"] = [dict(item) for item in message["tool_calls"]]
return cloned
for message in messages:
role = str(message.get("role", "") or "").strip()
if not pending_ids:
if role == "assistant" and isinstance(message.get("tool_calls"), list) and message["tool_calls"]:
ids = [
str(item.get("id", "") or "").strip()
for item in message["tool_calls"]
if isinstance(item, dict) and str(item.get("id", "") or "").strip()
]
if not ids:
continue
buffered_block = [_copy_message(message)]
pending_ids = ids
continue
if role == "tool":
continue
sanitized.append(_copy_message(message))
continue
if role == "tool":
tool_call_id = str(message.get("tool_call_id", "") or "").strip()
if tool_call_id and tool_call_id in pending_ids:
buffered_block.append(_copy_message(message))
pending_ids = [item for item in pending_ids if item != tool_call_id]
if not pending_ids:
sanitized.extend(buffered_block)
buffered_block = []
continue
if role == "assistant" and isinstance(message.get("tool_calls"), list) and message["tool_calls"]:
ids = [
str(item.get("id", "") or "").strip()
for item in message["tool_calls"]
if isinstance(item, dict) and str(item.get("id", "") or "").strip()
]
buffered_block = [_copy_message(message)] if ids else []
pending_ids = ids
continue
buffered_block = []
pending_ids = []
sanitized.append(_copy_message(message))
return sanitized
def prepare_user_message_content(
self,
content: str,
*,
attachment_refs: list[dict[str, Any]] | None = None,
task_type: str | None = None,
) -> str | list[dict[str, Any]]:
text = str(content or "")
refs = list(attachment_refs or [])
if not refs:
return text
model = self._select_model(task_type)
parts = self._build_direct_attachment_parts(model, refs)
if not parts:
return text
content_parts: list[dict[str, Any]] = []
if text:
content_parts.append({"type": "text", "text": text})
content_parts.extend(parts)
return content_parts
async def chat(
self,
messages: list[dict[str, Any]],
tools: list[dict[str, Any]] | None = None,
task_type: str | None = None,
temperature: float | None = None,
max_tokens: int | None = None,
**kwargs: Any,
) -> dict[str, Any]:
model = self._select_model(task_type)
temp = temperature if temperature is not None else self.config.temperature
max_tok = _clamp_max_tokens(model, max_tokens if max_tokens is not None else self.config.max_tokens)
call_kwargs: dict[str, Any] = {
"model": model,
"messages": messages,
"temperature": temp,
"max_tokens": max_tok,
**kwargs,
}
if self._api_base:
call_kwargs["api_base"] = self._api_base
if self._api_key:
call_kwargs["api_key"] = self._api_key
if tools:
call_kwargs["tools"] = tools
call_kwargs["tool_choice"] = "auto"
logger.debug(f"LLM call: model={model}, base={self._api_base or 'default'}, msgs={len(messages)}, tools={len(tools or [])}")
try:
response = await litellm.acompletion(**call_kwargs)
except Exception as e:
logger.error(f"LLM call failed: {e}")
raise
usage = getattr(response, "usage", None)
cost = 0.0
if usage:
self._total_tokens_in += getattr(usage, "prompt_tokens", 0)
self._total_tokens_out += getattr(usage, "completion_tokens", 0)
try:
cost = litellm.completion_cost(completion_response=response)
self._total_cost += cost
except Exception:
pass
choice = response.choices[0]
message = choice.message
result: dict[str, Any] = {
"content": message.content or "",
"tool_calls": [],
"finish_reason": choice.finish_reason,
"model": model,
"cost": cost,
"usage": {
"prompt_tokens": getattr(usage, "prompt_tokens", 0) if usage else 0,
"completion_tokens": getattr(usage, "completion_tokens", 0) if usage else 0,
},
}
if message.tool_calls:
for tc in message.tool_calls:
args, raw_args, parse_error = _parse_tool_arguments(tc.function.name, tc.function.arguments)
tool_call = {
"id": tc.id,
"function": tc.function.name,
"arguments": args,
}
if raw_args is not None:
tool_call["arguments_raw"] = raw_args
if parse_error:
tool_call["arguments_parse_error"] = parse_error
result["tool_calls"].append(tool_call)
return result
def normalize_stream_event(
self,
chunk: Any,
*,
model: str,
) -> list[RuntimeLLMEvent]:
events: list[RuntimeLLMEvent] = []
usage = getattr(chunk, "usage", None)
if usage:
events.append(RuntimeLLMEvent(
event_type="usage",
model=model,
payload={
"prompt_tokens": getattr(usage, "prompt_tokens", 0),
"completion_tokens": getattr(usage, "completion_tokens", 0),
"context_window": self.get_context_window(model=model),
},
))
choices = getattr(chunk, "choices", None) or []
if not choices:
return events
choice = choices[0]
delta = getattr(choice, "delta", None)
if delta is not None:
text = getattr(delta, "content", None)
if text:
events.append(RuntimeLLMEvent(
event_type="assistant_delta",
model=model,
payload={"text": text},
))
thinking_text = (
getattr(delta, "reasoning", None)
or getattr(delta, "reasoning_content", None)
or getattr(delta, "thinking", None)
)
if thinking_text:
events.append(RuntimeLLMEvent(
event_type="thinking_delta",
model=model,
payload={"text": str(thinking_text)},
))
tool_calls = getattr(delta, "tool_calls", None) or []
for tool_call in tool_calls:
function = getattr(tool_call, "function", None)
events.append(RuntimeLLMEvent(
event_type="tool_call_delta",
model=model,
payload={
"index": getattr(tool_call, "index", 0),
"id": getattr(tool_call, "id", ""),
"name": getattr(function, "name", "") if function else "",
"arguments": getattr(function, "arguments", "") if function else "",
},
))
finish_reason = getattr(choice, "finish_reason", None)
if finish_reason:
events.append(RuntimeLLMEvent(
event_type="message_stop",
model=model,
payload={"finish_reason": finish_reason},
))
return events
async def chat_stream(
self,
messages: list[dict[str, Any]],
tools: list[dict[str, Any]] | None = None,
task_type: str | None = None,
temperature: float | None = None,
max_tokens: int | None = None,
**kwargs: Any,
) -> AsyncIterator[RuntimeLLMEvent]:
model = self._select_model(task_type)
temp = temperature if temperature is not None else self.config.temperature
max_tok = _clamp_max_tokens(model, max_tokens if max_tokens is not None else self.config.max_tokens)
call_kwargs: dict[str, Any] = {
"model": model,
"messages": messages,
"temperature": temp,
"max_tokens": max_tok,
"stream": True,
**kwargs,
}
if self._api_base:
call_kwargs["api_base"] = self._api_base
if self._api_key:
call_kwargs["api_key"] = self._api_key
if tools:
call_kwargs["tools"] = tools
call_kwargs["tool_choice"] = "auto"
logger.debug(
f"LLM stream call: model={model}, base={self._api_base or 'default'}, msgs={len(messages)}, tools={len(tools or [])}"
)
last_usage = {"prompt_tokens": 0, "completion_tokens": 0}
yield RuntimeLLMEvent(event_type="message_start", model=model, payload={"model": model})
try:
stream = await litellm.acompletion(**call_kwargs)
if hasattr(stream, "__aiter__"):
async for chunk in stream:
for event in self.normalize_stream_event(chunk, model=model):
if event.event_type == "usage":
total_prompt = int(event.payload.get("prompt_tokens", 0) or 0)
total_completion = int(event.payload.get("completion_tokens", 0) or 0)
delta_prompt = max(0, total_prompt - last_usage["prompt_tokens"])
delta_completion = max(0, total_completion - last_usage["completion_tokens"])
last_usage["prompt_tokens"] = total_prompt
last_usage["completion_tokens"] = total_completion
cost = 0.0
try:
prompt_cost, completion_cost = litellm.cost_per_token(
model=model,
prompt_tokens=delta_prompt,
completion_tokens=delta_completion,
)
cost = float(prompt_cost or 0.0) + float(completion_cost or 0.0)
except Exception:
cost = 0.0
self._total_tokens_in += delta_prompt
self._total_tokens_out += delta_completion
self._total_cost += cost
event.payload = {
**dict(event.payload),
"prompt_tokens": delta_prompt,
"completion_tokens": delta_completion,
"prompt_tokens_total": total_prompt,
"completion_tokens_total": total_completion,
"estimated_cost_delta": cost,
"estimated_cost_total": self._total_cost,
"context_window": event.payload.get("context_window") or self.get_context_window(model=model),
"model": model,
}
yield event
else:
# Provider fallback: treat the response as a single non-streaming completion.
choice = stream.choices[0]
message = choice.message
if getattr(message, "content", None):
yield RuntimeLLMEvent(
event_type="assistant_delta",
model=model,
payload={"text": message.content},
)
for tc in getattr(message, "tool_calls", None) or []:
yield RuntimeLLMEvent(
event_type="tool_call_delta",
model=model,
payload={
"index": 0,
"id": getattr(tc, "id", ""),
"name": getattr(getattr(tc, "function", None), "name", ""),
"arguments": getattr(getattr(tc, "function", None), "arguments", ""),
},
)
usage = getattr(stream, "usage", None)
if usage:
cost = 0.0
prompt_tokens = int(getattr(usage, "prompt_tokens", 0) or 0)
completion_tokens = int(getattr(usage, "completion_tokens", 0) or 0)
try:
prompt_cost, completion_cost = litellm.cost_per_token(
model=model,
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
)
cost = float(prompt_cost or 0.0) + float(completion_cost or 0.0)
except Exception:
cost = 0.0
self._total_tokens_in += prompt_tokens
self._total_tokens_out += completion_tokens
self._total_cost += cost
yield RuntimeLLMEvent(
event_type="usage",
model=model,
payload={
"prompt_tokens": prompt_tokens,
"completion_tokens": completion_tokens,
"prompt_tokens_total": prompt_tokens,
"completion_tokens_total": completion_tokens,
"estimated_cost_delta": cost,
"estimated_cost_total": self._total_cost,
"context_window": self.get_context_window(model=model),
"model": model,
},
)
yield RuntimeLLMEvent(
event_type="message_stop",
model=model,
payload={"finish_reason": getattr(choice, "finish_reason", "stop")},
)
except Exception as e:
logger.error(f"LLM stream failed: {e}")
yield RuntimeLLMEvent(
event_type="error",
model=model,
payload={"message": str(e)},
)
raise
async def simple_chat(
self,
prompt: str,
system: str | None = None,
task_type: str | None = None,
) -> str:
messages: list[dict[str, Any]] = []
if system:
messages.append({"role": "system", "content": system})
messages.append({"role": "user", "content": prompt})
result = await self.chat(messages, task_type=task_type)
return result["content"]
def get_tool_definitions(self, tools: list[dict[str, Any]]) -> list[dict[str, Any]]:
"""Convert internal tool definitions to OpenAI function-calling format."""
formatted = []
for tool in tools:
formatted.append({
"type": "function",
"function": {
"name": tool["name"],
"description": tool.get("description", ""),
"parameters": tool.get("parameters", {"type": "object", "properties": {}}),
},
})
return formatted
def _build_direct_attachment_parts(
self,
model: str,
attachment_refs: list[dict[str, Any]],
) -> list[dict[str, Any]]:
capabilities = self._attachment_capabilities(model)
if not capabilities["enabled"]:
return []
parts: list[dict[str, Any]] = []
for ref_dict in attachment_refs:
try:
ref = AttachmentRef.from_dict(ref_dict)
data_url = self._attachment_data_url(ref)
except Exception as exc:
logger.warning(f"Skipping direct attachment payload: {exc}")
continue
if not data_url:
continue
if ref.mime_type.startswith("image/") and capabilities["image_mode"]:
parts.append(self._build_attachment_part("image_url", ref, data_url))
elif ref.mime_type == "application/pdf" and capabilities["pdf_mode"]:
parts.append(self._build_attachment_part(str(capabilities["pdf_mode"]), ref, data_url))
elif ref.mime_type.startswith("video/") and capabilities["video_mode"]:
parts.append(self._build_attachment_part(str(capabilities["video_mode"]), ref, data_url))
return parts
def _build_attachment_part(
self,
mode: str,
ref: AttachmentRef,
data_url: str,
) -> dict[str, Any]:
if mode == "image_url":
return {
"type": "image_url",
"image_url": {"url": data_url},
}
if mode == "video_url":
return {
"type": "video_url",
"video_url": {"url": data_url},
}
if mode == "file":
return {
"type": "file",
"file": {
"file_data": data_url,
"filename": ref.filename,
},
}
raise ValueError(f"Unsupported attachment transport mode: {mode}")
def _attachment_capabilities(self, model: str) -> dict[str, Any]:
provider = model.split("/", 1)[0].strip().lower()
api_base = (self._api_base or "").lower()
image_mode: str | None = None
pdf_mode: str | None = None
video_mode: str | None = None
if "api.poe.com" in api_base:
image_mode = "image_url"
pdf_mode = "file"
if _looks_like_multimodal_model(model):
video_mode = "file"
elif "openrouter.ai" in api_base:
image_mode = "image_url"
pdf_mode = "file"
if _looks_like_video_capable_model(model):
video_mode = "video_url"
elif provider in {"openai", "azure", "anthropic"}:
image_mode = "image_url"
pdf_mode = "file"
elif provider in {"google", "gemini", "vertex_ai", "vertex"}:
image_mode = "image_url"
pdf_mode = "file"
if _looks_like_video_capable_model(model):
video_mode = "video_url"
elif _looks_like_multimodal_model(model):
image_mode = "image_url"
pdf_mode = "file" if _looks_like_document_capable_model(model) else None
if _looks_like_video_capable_model(model):
video_mode = "video_url"
return {
"enabled": bool(image_mode or pdf_mode or video_mode),
"image_mode": image_mode,
"pdf_mode": pdf_mode,
"video_mode": video_mode,
}
def _attachment_data_url(self, ref: AttachmentRef) -> str | None:
path = self._resolve_attachment_path(ref)
if path is None:
return None
payload = base64.b64encode(path.read_bytes()).decode("ascii")
mime_type = ref.mime_type or _guess_mime_from_filename(ref.filename)
return f"data:{mime_type};base64,{payload}"
def _resolve_attachment_path(self, ref: AttachmentRef) -> Path | None:
if not self.opc_home or not ref.disk_path:
return None
resolved = (self.opc_home / ref.disk_path).resolve()
opc_home_resolved = self.opc_home.resolve()
if not str(resolved).startswith(str(opc_home_resolved)):
raise ValueError(f"Attachment path escapes OPC home: {ref.disk_path}")
return resolved
def _guess_mime_from_filename(filename: str) -> str:
suffix = attachment_suffix(filename)
if suffix == ".pdf":
return "application/pdf"
if suffix == ".docx":
return "application/vnd.openxmlformats-officedocument.wordprocessingml.document"
if suffix == ".xlsx":
return "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet"
if suffix == ".pptx":
return "application/vnd.openxmlformats-officedocument.presentationml.presentation"
if suffix == ".mp4":
return "video/mp4"
if suffix in {".mpeg", ".mpg"}:
return "video/mpeg"
if suffix == ".mov":
return "video/quicktime"
if suffix == ".webm":
return "video/webm"
return "application/octet-stream"