honcho/src/llm/types.py

309 lines
13 KiB
Python

"""Public response/stream/iteration types for the LLM API.
These used to live in src/utils/clients.py and have been moved here as part
of the migration toward src/llm/ owning all non-embedding LLM orchestration.
"""
from __future__ import annotations
import asyncio
import logging
from collections.abc import AsyncIterator, Callable
from dataclasses import dataclass, field
from typing import TYPE_CHECKING, Any, Generic, Literal, TypeVar
from anthropic import AsyncAnthropic
from google import genai
from openai import AsyncOpenAI
from pydantic import BaseModel, Field
if TYPE_CHECKING:
from src.llm.capture import CapturedMessage
logger = logging.getLogger(__name__)
T = TypeVar("T")
# OpenAI GPT-5 specific reasoning levels.
ReasoningEffortType = (
Literal["none", "minimal", "low", "medium", "high", "xhigh", "max"] | None
)
VerbosityType = Literal["low", "medium", "high"] | None
# Raw SDK client union used by the provider-selection layer.
ProviderClient = AsyncAnthropic | AsyncOpenAI | genai.Client
@dataclass
class IterationData:
"""Data passed to iteration callbacks after each tool execution loop iteration."""
iteration: int
"""1-indexed iteration number."""
tool_calls: list[str]
"""List of tool names called in this iteration."""
input_tokens: int
"""Input tokens used in this iteration's LLM call."""
output_tokens: int
"""Output tokens generated in this iteration's LLM call."""
cache_read_tokens: int = 0
"""Tokens read from cache in this iteration."""
cache_creation_tokens: int = 0
"""Tokens written to cache in this iteration."""
@dataclass
class LLMTelemetryContext:
"""Context threaded through honcho_llm_call → honcho_llm_call_inner so the
LLMCallCompletedEvent emitter (and AgentIterationEvent emitter)
can attribute calls to the right workspace / agent / iteration without
re-deriving any of it from ambient state.
Iteration is mutable: tool_loop updates this field before each inner call.
NOT read from set_current_iteration ContextVar — that fires after the LLM
call returns, so reading it from the executor would yield stale values.
"""
workspace_name: str | None = None
# call_purpose carries the same string as src.telemetry.events.llm.CallPurpose values.
# Stored as str rather than importing the enum here to keep src/llm/ free of
# telemetry imports — the emitter validates against the enum.
call_purpose: str | None = None
parent_category: str | None = None
run_id: str | None = None
iteration: int | None = None
# OpenTelemetry-style span-tree correlation.
trace_id: str | None = None
span_id: str | None = None
parent_span_id: str | None = None
# Monotonic executor-call ordinal WITHIN a span (total ordering of its
# steps).
step_seq: int = 0
# Retry/fallback attempt within an iteration.
attempt: int = 1
# Optional peer context (dream agents pass observer/observed; dialectic
# passes peer_name). Kept here so AgentIterationEvent can populate
# them without a separate threading path.
observer: str | None = None
observed: str | None = None
peer_name: str | None = None
# Used to group traces (should not use session_name because it is not unique)
session_id: str | None = None
# Tool-related context: agent_type is the human-readable identifier of the
# agent — dialectic/deduction/induction. Used by agent iteration
# event and tool call event.
agent_type: str | None = None
# Human-readable name for the Langfuse trace + per-call generation
# (e.g. "Dialectic Agent", "Minimal Deriver"). Sole home for this name —
# callers set it here; `honcho_llm_call` no longer takes a separate kwarg.
# Also used to label the sentry `ai_track` decorator and as the source for
# the run-level `langfuse_agent_run` label.
track_name: str | None = None
# Per-span memo for O(N) message capture in CapturedLLMCall
hash_memo: dict[int, CapturedMessage] | None = field(
default=None, compare=False, repr=False
)
def span_identity(self) -> str | None:
"""Effective span id: the new `span_id`, falling back to legacy `run_id`."""
return self.span_id or self.run_id
def exported_parent_span_id(self) -> str | None:
"""`parent_span_id` for EXPORT, collapsing the self-parent sentinel to None."""
pid = self.parent_span_id
return None if pid is not None and pid == self.span_id else pid
IterationCallback = Callable[[IterationData], None]
class HonchoLLMCallResponse(BaseModel, Generic[T]):
"""Response object for LLM calls.
Note:
Uncached input tokens = input_tokens - cache_read_input_tokens
+ cache_creation_input_tokens
(cache_creation costs 25% more, cache_read costs 90% less)
"""
content: T
input_tokens: int = 0
output_tokens: int
cache_creation_input_tokens: int = 0
cache_read_input_tokens: int = 0
finish_reasons: list[str]
tool_calls_made: list[dict[str, Any]] = Field(default_factory=list)
iterations: int = 0
"""Number of LLM calls made in the tool execution loop."""
thinking_content: str | None = None
# Full thinking blocks with signatures for multi-turn replay (Anthropic only).
thinking_blocks: list[dict[str, Any]] = Field(default_factory=list)
# OpenRouter reasoning_details for Gemini models — must be preserved across turns.
reasoning_details: list[dict[str, Any]] = Field(default_factory=list)
# True when the original input exceeded `max_input_tokens` — covers
# both "messages were dropped" and "couldn't drop the last unit and
# remaining tokens still exceeded the cap" (the deriver's prompt-only
# case). Maps 1:1 to `RepresentationCompletedEvent.hit_input_token_cap`
# and `DialecticCompletedEvent.hit_input_token_cap`.
hit_input_token_cap: bool = False
class HonchoLLMCallStreamChunk(BaseModel):
"""A single chunk in a streaming LLM response."""
content: str
is_done: bool = False
finish_reasons: list[str] = Field(default_factory=list)
output_tokens: int | None = None
class StreamingResponseWithMetadata:
"""Streaming response wrapper carrying metadata from a completed tool loop.
Lets callers read tool_calls_made / token counts / thinking_content from
the tool-execution phase while still iterating the final streamed answer.
`output_tokens` is updated AS THE STREAM DRAINS — `__aiter__` wraps the
underlying iterator and accumulates the latest non-None `output_tokens`
value reported by chunk usage. Providers like OpenAI (with
`stream_options.include_usage`) and Anthropic emit a final usage chunk
with the cumulative count, so the post-drain `output_tokens` value
reflects tool-loop output + final-stream output. Callers that read
`output_tokens` AFTER fully iterating the stream get the true total;
callers that read it before drain see only the tool-loop portion.
`langfuse_run_handle` (optional) is the run-level Langfuse span handle
transferred from `honcho_llm_call` when streaming. The wrapper owns it
after construction: on drain, the accumulated streamed text is stamped
as the run span's output and the span is closed. Without this transfer,
streaming traces would show blank output because the synchronous return
happens before any chunks arrive.
`capture_finalizer` (optional) closes the replay-grade content capture for
a streamed call. The synchronous return happens before any chunks arrive,
so the streamed text only exists once the stream drains — the wrapper calls
the finalizer with `(accumulated_text, finish_reason)` in its `finally`.
A partial/aborted stream still finalizes, with `finish_reason` =
"cancelled"/"error".
"""
_stream: AsyncIterator[HonchoLLMCallStreamChunk]
tool_calls_made: list[dict[str, Any]]
input_tokens: int
output_tokens: int
cache_creation_input_tokens: int
cache_read_input_tokens: int
thinking_content: str | None
iterations: int
hit_input_token_cap: bool
_langfuse_run_handle: Any | None
_capture_finalizer: Callable[[str, str], None] | None
def __init__(
self,
stream: AsyncIterator[HonchoLLMCallStreamChunk],
tool_calls_made: list[dict[str, Any]],
input_tokens: int,
output_tokens: int,
cache_creation_input_tokens: int,
cache_read_input_tokens: int,
thinking_content: str | None = None,
iterations: int = 0,
hit_input_token_cap: bool = False,
langfuse_run_handle: Any | None = None,
capture_finalizer: Callable[[str, str], None] | None = None,
):
self._stream = stream
self.tool_calls_made = tool_calls_made
self.input_tokens = input_tokens
self.output_tokens = output_tokens
self.cache_creation_input_tokens = cache_creation_input_tokens
self.cache_read_input_tokens = cache_read_input_tokens
self.thinking_content = thinking_content
self.iterations = iterations
self.hit_input_token_cap = hit_input_token_cap
self._langfuse_run_handle = langfuse_run_handle
self._capture_finalizer = capture_finalizer
def __aiter__(self) -> AsyncIterator[HonchoLLMCallStreamChunk]:
# Wrap the underlying iterator to capture final-stream output_tokens
# from chunks as they arrive. Providers emit a usage chunk at end-of-
# stream with the cumulative output_tokens count; we fold it into
# self.output_tokens (which carries the tool-loop running total at
# construction) so the post-drain value reflects the true cost.
return self._iterate_with_usage_capture()
async def _iterate_with_usage_capture(
self,
) -> AsyncIterator[HonchoLLMCallStreamChunk]:
final_stream_output_tokens = 0
# Accumulate the streamed text when either consumer needs it: the
# Langfuse run span (stamped as output on drain) or the content-capture
# finalizer.
accumulate = (
self._langfuse_run_handle is not None or self._capture_finalizer is not None
)
accumulated_text: list[str] = []
last_finish_reason: str | None = None
stream_error: BaseException | None = None
try:
async for chunk in self._stream:
if chunk.output_tokens is not None:
# Take the LATEST value, not the sum — providers report
# the cumulative usage in the final chunk, not deltas.
final_stream_output_tokens = chunk.output_tokens
if chunk.finish_reasons:
last_finish_reason = chunk.finish_reasons[-1]
if accumulate and chunk.content:
accumulated_text.append(chunk.content)
yield chunk
# Stream drained — fold the final-stream output tokens into the
# tool-loop totals so DialecticCompletedEvent / downstream readers
# see the true cost.
if final_stream_output_tokens > 0:
self.output_tokens += final_stream_output_tokens
except BaseException as exc:
stream_error = exc
raise
finally:
text = "".join(accumulated_text)
# Close the run span once, stamping the streamed text as its
# output. In `finally` so an early-exit caller still closes
# the span rather than leaking it.
handle = self._langfuse_run_handle
if handle is not None:
self._langfuse_run_handle = None
handle.end(output=text or None)
# Finalize the content capture with the full streamed text. Even a
# partial/aborted stream captures, tagged with the right outcome.
finalizer = self._capture_finalizer
if finalizer is not None:
self._capture_finalizer = None
finish_reason = (
(last_finish_reason or "stop")
if stream_error is None
else (
"cancelled"
if isinstance(stream_error, asyncio.CancelledError)
else "error"
)
)
try:
finalizer(text, finish_reason)
except Exception: # pragma: no cover - best-effort telemetry
logger.debug("Stream capture finalizer failed", exc_info=True)
__all__ = [
"HonchoLLMCallResponse",
"HonchoLLMCallStreamChunk",
"IterationCallback",
"IterationData",
"LLMTelemetryContext",
"ProviderClient",
"ReasoningEffortType",
"StreamingResponseWithMetadata",
"T",
"VerbosityType",
]