mirror of
https://github.com/kuhyx/testsAndMisc.git
synced 2026-11-10 21:47:02 +01:00
Krzysztof kuhy Rudnicki
f7e8c59e8f
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
84 lines
3.3 KiB
Python
84 lines
3.3 KiB
Python
"""Estimate what images cost *after* the turn that loaded them.
|
|
|
|
An image does not leave the context once read. Every subsequent request in the
|
|
session re-sends it as cache-read, so a screenshot loaded at turn 20 of a
|
|
1000-turn session is paid for ~980 more times. At the cache-read weight of 0.1
|
|
that turns ~2.5k tokens into ~250k, which is why images are worth tracking as
|
|
their own axis rather than lumping them in with text tool results.
|
|
|
|
Measured over a real 7-day window this amplification was ~1.6% of weighted
|
|
spend — real, but an order of magnitude below session length. An earlier
|
|
version of this module reported ~27%, because it sized images by their base64
|
|
payload length; see :func:`parse._result_tokens` for why that is ~20x too high.
|
|
The lesson is kept here deliberately: an estimate that is never checked against
|
|
billed tokens can invent a headline finding out of nothing.
|
|
|
|
Two honest limits on the number this produces:
|
|
|
|
* It is an **estimate**, derived from result sizes rather than billed tokens.
|
|
It is deliberately never folded into the reconciled totals, which come only
|
|
from the API's own usage figures.
|
|
* Compaction discards history. An image loaded before a compaction stops being
|
|
re-sent, so charging it forever would roughly double the figure and yields
|
|
the impossible result of sessions costing more than they measurably did.
|
|
Compaction is detected from a sharp drop in context size.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from typing import TYPE_CHECKING
|
|
|
|
from python_pkg.token_audit.model import (
|
|
COMPACTION_DROP_RATIO,
|
|
WEIGHTS,
|
|
Session,
|
|
ToolCall,
|
|
Turn,
|
|
)
|
|
|
|
if TYPE_CHECKING:
|
|
from collections.abc import Iterable
|
|
|
|
# Images are re-sent from cache, so each later turn charges them at cache-read.
|
|
_REREAD_WEIGHT = WEIGHTS["cache_read_input_tokens"]
|
|
|
|
|
|
def _is_compaction(context: int, previous: int | None) -> bool:
|
|
"""Whether this turn's context implies history was just discarded.
|
|
|
|
Context grows monotonically as a session accumulates messages, so the only
|
|
way it falls sharply is compaction (or a fresh branch). A ratio test rather
|
|
than an absolute threshold keeps this valid for both small and huge
|
|
sessions.
|
|
"""
|
|
return previous is not None and context < previous * COMPACTION_DROP_RATIO
|
|
|
|
|
|
def image_cost(events: Iterable[tuple[str, object]]) -> float:
|
|
"""Return cost-weighted tokens spent re-sending images across turns.
|
|
|
|
``events`` must preserve transcript order — the whole calculation is about
|
|
how many turns followed each image before compaction cleared it.
|
|
"""
|
|
live: list[int] = []
|
|
previous: int | None = None
|
|
total = 0.0
|
|
for kind, event in events:
|
|
if kind == "tool":
|
|
if isinstance(event, ToolCall) and event.is_image:
|
|
live.append(event.result_tokens)
|
|
elif isinstance(event, Turn):
|
|
if _is_compaction(event.context, previous):
|
|
live.clear()
|
|
# Charge every still-live image once for this turn: it was part of
|
|
# the context this request re-sent.
|
|
total += sum(live) * _REREAD_WEIGHT
|
|
previous = event.context
|
|
return total
|
|
|
|
|
|
def annotate(session: Session, events: Iterable[tuple[str, object]]) -> Session:
|
|
"""Attach the image amplification estimate to a session."""
|
|
session.image_cost = image_cost(events)
|
|
return session
|