Skip to content

Official Codebase of 3DCodeVerse

Python

models: openai_usage in pricing is the one OpenAI token rule — the SDK adapter and the codex backend each split reasoning out of completion and priced it by hand

gaoypeng3ae96d449d7b
diff --git a/harness/codeverse3d/agents/backends.py b/harness/codeverse3d/agents/backends.py
index cd8d2054..cdc6e315 100644
--- a/harness/codeverse3d/agents/backends.py
+++ b/harness/codeverse3d/agents/backends.py
@@ -71,7 +71,7 @@ from codeverse3d.agents.materialize import MCP_SERVER_NAME, MCP_TOOL_TIMEOUT_MS,
from codeverse3d.config import get_settings
from codeverse3d.contracts.agent import AgentJob, AgentResult
from codeverse3d.contracts.common import Usage
-from codeverse3d.models.pricing import cache_write_surcharge, estimate_cost
+from codeverse3d.models.pricing import cache_write_surcharge, estimate_cost, openai_usage
from codeverse3d.models.retry import KeyPool, KeyPoolExhausted
from codeverse3d.proc import read_jsonl_lenient, run_subprocess
@@ -861,21 +861,10 @@ class CodexEvents:
self.errors.append(str(ev.get("message") or ev))
def usage(self, model: str) -> Usage:
- # `output_tokens` from the Responses API already CONTAINS
- # `reasoning_output_tokens` (measured: output − reasoning tracks the answer
- # length across every recorded cell). `estimate_cost` bills output +
- # thoughts, so the reasoning share is subtracted here instead of being
- # charged twice — it stays visible in `thoughts_tokens`.
- reasoning = self.usage_raw["reasoning_output_tokens"]
- u = Usage(
- backend="codex", model=model,
- input_tokens=self.usage_raw["input_tokens"],
- output_tokens=max(0, self.usage_raw["output_tokens"] - reasoning),
- cached_tokens=self.usage_raw["cached_input_tokens"], tool_calls=self.tool_calls,
- thoughts_tokens=reasoning,
- )
- u.cost_usd = estimate_cost("openai", model, u)
- return u
+ r = self.usage_raw # the Responses API's output_tokens CONTAIN reasoning_output_tokens
+ return openai_usage("codex", model, prompt=r["input_tokens"], cached=r["cached_input_tokens"],
+ completion=r["output_tokens"], reasoning=r["reasoning_output_tokens"],
+ tool_calls=self.tool_calls)
def parse_codex_jsonl(stdout: str) -> CodexEvents:
diff --git a/harness/codeverse3d/models/openai.py b/harness/codeverse3d/models/openai.py
index 099d5379..90851025 100644
--- a/harness/codeverse3d/models/openai.py
+++ b/harness/codeverse3d/models/openai.py
@@ -29,7 +29,7 @@ from codeverse3d.models.parts import (
retry_budget_s,
retry_one_key,
)
-from codeverse3d.models.pricing import estimate_cost
+from codeverse3d.models.pricing import openai_usage
from codeverse3d.models.retry import cause_for
from codeverse3d.models.schema_utils import (
JsonParseError,
@@ -236,23 +236,9 @@ class OpenAIModel(SdkModel):
return ChatResponse(text=text, parsed=parsed, finish_reason=finish, usage=usage, raw=raw)
def _usage(self, u: Any, latency_ms: int) -> Usage:
- if u is None:
- usage = Usage(backend="openai", model=self.model, latency_ms=latency_ms)
- else:
- ptd = getattr(u, "prompt_tokens_details", None)
- ctd = getattr(u, "completion_tokens_details", None)
- cached = int(getattr(ptd, "cached_tokens", 0) or 0) if ptd else 0
- reasoning = int(getattr(ctd, "reasoning_tokens", 0) or 0) if ctd else 0
- completion = int(u.completion_tokens or 0)
- usage = Usage(
- backend="openai",
- model=self.model,
- input_tokens=int(u.prompt_tokens or 0),
- # completion_tokens already includes reasoning; split so the sum stays exact
- output_tokens=max(0, completion - reasoning),
- cached_tokens=cached,
- thoughts_tokens=reasoning,
- latency_ms=latency_ms,
- )
- usage.cost_usd = estimate_cost("openai", self.model, usage)
- return usage
+ ptd = getattr(u, "prompt_tokens_details", None)
+ ctd = getattr(u, "completion_tokens_details", None)
+ return openai_usage("openai", self.model, prompt=int(getattr(u, "prompt_tokens", 0) or 0),
+ cached=int(getattr(ptd, "cached_tokens", 0) or 0),
+ completion=int(getattr(u, "completion_tokens", 0) or 0),
+ reasoning=int(getattr(ctd, "reasoning_tokens", 0) or 0), latency_ms=latency_ms)
diff --git a/harness/codeverse3d/models/pricing.py b/harness/codeverse3d/models/pricing.py
index 4d8d52c7..a74ac75a 100644
--- a/harness/codeverse3d/models/pricing.py
+++ b/harness/codeverse3d/models/pricing.py
@@ -19,6 +19,7 @@ from __future__ import annotations
import logging
import threading
from dataclasses import dataclass
+from typing import Any
from codeverse3d.contracts.common import Usage
@@ -322,6 +323,18 @@ def estimate_cost(provider: str, model: str, usage: Usage) -> float:
return cost
+def openai_usage(backend: str, model: str, *, prompt: int, cached: int, completion: int, reasoning: int,
+ **extra: Any) -> Usage:
+ """OpenAI-family tokens (the SDK and codex alike) as a priced ``Usage``. ``completion``
+ already CONTAINS ``reasoning`` (measured on codex: output − reasoning tracks the answer
+ length), so the reasoning share moves to ``thoughts_tokens`` — still visible, and billed
+ once by :func:`estimate_cost` (output + thoughts), never twice."""
+ usage = Usage(backend=backend, model=model, input_tokens=prompt, output_tokens=max(0, completion - reasoning),
+ cached_tokens=cached, thoughts_tokens=reasoning, **extra)
+ usage.cost_usd = estimate_cost("openai", model, usage)
+ return usage
+
+
def cache_write_surcharge(provider: str, model: str, cache_write_tokens: int) -> float:
"""Extra USD (on top of the normal input rate already charged by
``estimate_cost``) for explicit cache creation, e.g. Anthropic's