codex 0.165.0 · 单轮详情

20261107-codex-0.165.0-t1

测量日 2026-11-07 · source SHA 缺测
schema 缺测
provider: ppio；template: teams-work-v1；cpu: 4；memory_mib: 8192；os: Debian 13；deployment_kind: fresh_isolated

无独立交付快照；本页保留本轮公开指标。

版本 / 试次	测量日	completed / 输入	内容 PASS	回复耗时合计 / 秒	模型响应	工具调用	报告
codex 0.165.0	2026-11-07	14 / 14	9 / 10	629.5	68	72	详情与报告

逐用户输入

输入	原生终态	回复延迟 / 秒	CLI / 观察耗时 / 秒	模型响应	工具调用
C01/U1	completed	30.6	40.6	4	3
C02/U1	completed	46.8	56.8	4	8
C03/U1	completed	29.1	39.2	5	6
C04/U1	completed	52.3	62.3	8	9
C05/U1	completed	93.0	103.0	7	10
C06/U1	completed	23.3	33.4	3	2
C07/U1	completed	26.8	36.8	3	2
C07/U2	completed	47.1	57.1	3	3
C08/U1	completed	29.4	39.4	5	7
C08/U2	completed	49.5	59.5	4	4
C09/U1	completed	63.1	80.1	8	7
C09/U2	completed	90.4	100.5	7	6
C10/U1	completed	23.9	33.9	4	3
C10/U2	completed	24.1	34.2	3	2

逐题内容判分

题目	内容复核	原检查器	回复合计 / 秒
C01 发票去重与对账	PASS	PASS	30.6
C02 日志关联与事故时间线	PASS	PASS	46.8
C03 多文件政策优先级	PASS	PASS	29.1
C04 时区与稳定排序	PASS	PASS	52.3
C05 小模块修复与回归	PASS	PASS	93.0
C06 依赖图与关键路径	PASS	PASS	23.3
C07 采购方案追加预算约束	PASS	PASS	73.9
C08 销售汇总与增量更正	PASS	PASS	78.9
C09 命令行工具连续迭代	PASS	PASS	153.6
C10 发布判断与新证据修正	INCONCLUSIVE	FAIL	48.0

指标与缺测

{
  "artifact_grades": {
    "INCONCLUSIVE": 1,
    "PASS": 9
  },
  "invocation_or_observer": {
    "expected": 14,
    "observed": 14,
    "observed_sum_s": 776.780873712,
    "sum_s": 776.780873712
  },
  "model_id_mismatch": 0,
  "model_id_missing": 0,
  "native_tool_calls": 72,
  "observed_model_ids": {
    "deepseek/deepseek-v4.1-flash-1017": 68
  },
  "observed_model_intervals": {
    "expected": 68,
    "observed": 68,
    "observed_sum_s": 627.6859998703003,
    "sum_s": 627.6859998703003
  },
  "observed_tmp_commands": 4,
  "observed_tool_intervals": {
    "expected": 72,
    "observed": 72,
    "observed_sum_s": 93.06400036811829,
    "sum_s": 93.06400036811829
  },
  "raw_artifact_grades": {
    "FAIL": 1,
    "PASS": 9
  },
  "response_latency": {
    "expected": 14,
    "observed": 14,
    "observed_sum_s": 629.4801471233368,
    "sum_s": 629.4801471233368
  },
  "scope_observed_cases": 10,
  "terminals": {
    "completed": 14
  },
  "tmp_command_cases": [
    "C04",
    "C10"
  ],
  "tool_distribution": {
    "exec_command": 72
  },
  "tool_input_availability": {
    "unmarked": 72
  },
  "tool_missing_arguments": 0,
  "tool_missing_results": 0,
  "tool_output_availability": {
    "unmarked": 72
  },
  "user_turns": 14,
  "visible_model_responses": 68
}

采集与覆盖说明（manifest 原文）

All declared user turns have final reply timestamps
Native model counts do not establish auxiliary/provider request completeness
post-terminal CLI shutdown delay measured separately

口径与限制

n=1；并发争用、模型随机性、跨日基线未受控；不能据此作统计排名或因果结论。输入终态、内容判分、范围遵守与采集完整度分开呈现。

固定基线为2026-11-07 Codex0.165.0 / OpenClaw2026.11.1；欢喜OS各版本单独新增试次。相同10题、14用户输入，liqian deepseek/deepseek-v4.1-flash-1017。每次报告明确区分：输入→原生终态延迟、CLI/观察收尾、模型响应、工具步骤、内容判分、范围遵守与证据缺测。跨日固定基线不是同时随机对照。

Relative to historical .68/.69 (teams-work-v1), this run uses a different teams-work-v2 environment. Main comparison uses same-day .69 v2 environment anchor.
Six versions execute concurrently in separate 4C/8G sandboxes; cross-version provider contention and model nondeterminism are uncontrolled. One trial per case; no statistical ranking or causal attribution.

内容复核口径（comparisons/latest.md）：历史复核：C08 原检查器漏查子目录备份，原判保留、内容复核 PASS；C10 部署清单完整性与 JSON 层级题面有歧义，保留 INCONCLUSIVE。

主耗时不包含CLI终态后退出等待。模型/工具时间线允许重叠；不得相加当整体墙钟。单trial及跨日历史基线只支持探索性比较。