new-api

5.6 KiB

Raw Blame History

allowed-tools	description	name
Bash(wsl -- *)	Command: claude-stats (imported from Claude Code)	cmd-claude-stats

Claude 模型统计

查询 Claude 模型的错误率和缓存率，自动排除 relay-pulse 检测令牌。

参数 `$ARGUMENTS`

根据 $ARGUMENTS 确定日期范围：

传入格式	含义
（空）	今天
`2026-05-09`	指定某一天
`2026-05-01 2026-05-09`	指定日期范围（含首尾两天）
`7d`	最近 7 天

执行步骤

第一步：根据参数计算日期过滤条件

created_at 字段是秒级 Unix 时间戳（bigint），用 UNIX_TIMESTAMP('YYYY-MM-DD') 转换
今天：created_at >= UNIX_TIMESTAMP(CURDATE())
某一天 D：created_at >= UNIX_TIMESTAMP('D') AND created_at < UNIX_TIMESTAMP('D') + 86400
范围 D1 到 D2：created_at >= UNIX_TIMESTAMP('D1') AND created_at < UNIX_TIMESTAMP('D2') + 86400
最近 N 天：created_at >= UNIX_TIMESTAMP(DATE_SUB(CURDATE(), INTERVAL N-1 DAY))

将计算好的条件替换到下面 SQL 中的 {DATE_FILTER} 占位符。

第二步：上传 SQL 文件到服务器

wsl -- bash -c "
cat > /tmp/claude_stats_error.sql << 'ENDSQL'
SELECT
  model_name,
  SUM(CASE WHEN type = 2 THEN 1 ELSE 0 END) as success,
  SUM(CASE WHEN type = 5 THEN 1 ELSE 0 END) as errors,
  COUNT(*) as total,
  ROUND(SUM(CASE WHEN type = 5 THEN 1 ELSE 0 END) * 100.0 / COUNT(*), 2) as error_rate_pct
FROM logs
WHERE {DATE_FILTER}
  AND model_name LIKE '%claude%'
  AND type IN (2, 5)
  AND token_name != 'relay-pulse'
GROUP BY model_name
ORDER BY total DESC;
ENDSQL

cat > /tmp/claude_stats_cache.sql << 'ENDSQL'
SELECT
  model_name,
  COUNT(*) as total_requests,
  SUM(prompt_tokens) as prompt_tokens,
  SUM(completion_tokens) as completion_tokens,
  SUM(CAST(IFNULL(JSON_UNQUOTE(JSON_EXTRACT(other, '$.cache_tokens')), 0) AS UNSIGNED)) as cache_read_tokens,
  SUM(CAST(IFNULL(JSON_UNQUOTE(JSON_EXTRACT(other, '$.cache_creation_tokens')), 0) AS UNSIGNED)) as cache_creation_tokens,
  SUM(CASE WHEN CAST(IFNULL(JSON_UNQUOTE(JSON_EXTRACT(other, '$.cache_tokens')), 0) AS UNSIGNED) > 0 THEN 1 ELSE 0 END) as cache_hit_requests,
  ROUND(
    SUM(CASE WHEN CAST(IFNULL(JSON_UNQUOTE(JSON_EXTRACT(other, '$.cache_tokens')), 0) AS UNSIGNED) > 0 THEN 1 ELSE 0 END) * 100.0 / COUNT(*),
    2
  ) as cache_hit_rate_pct,
  ROUND(
    (SUM(prompt_tokens) * 1.0
     + SUM(CAST(IFNULL(JSON_UNQUOTE(JSON_EXTRACT(other, '$.cache_tokens')), 0) AS UNSIGNED)) * 0.1
     + SUM(CAST(IFNULL(JSON_UNQUOTE(JSON_EXTRACT(other, '$.cache_creation_tokens')), 0) AS UNSIGNED)) * 1.25)
    * 100.0
    / NULLIF(
        SUM(prompt_tokens)
        + SUM(CAST(IFNULL(JSON_UNQUOTE(JSON_EXTRACT(other, '$.cache_tokens')), 0) AS UNSIGNED))
        + SUM(CAST(IFNULL(JSON_UNQUOTE(JSON_EXTRACT(other, '$.cache_creation_tokens')), 0) AS UNSIGNED)),
        0),
    2
  ) as cost_pct_of_full,
  ROUND(
    SUM(CAST(IFNULL(JSON_UNQUOTE(JSON_EXTRACT(other, '$.cache_creation_tokens')), 0) AS UNSIGNED)) * 1.0
    / NULLIF(SUM(CAST(IFNULL(JSON_UNQUOTE(JSON_EXTRACT(other, '$.cache_tokens')), 0) AS UNSIGNED)), 0),
    3
  ) as creation_to_read_ratio
FROM logs
WHERE {DATE_FILTER}
  AND type = 2
  AND model_name LIKE '%claude%'
  AND other IS NOT NULL
  AND token_name != 'relay-pulse'
GROUP BY model_name
ORDER BY total_requests DESC;
ENDSQL

cat > /tmp/claude_stats_errcode.sql << 'ENDSQL'
SELECT
  model_name,
  CAST(JSON_UNQUOTE(JSON_EXTRACT(other, '$.status_code')) AS UNSIGNED) as status_code,
  COUNT(*) as cnt,
  ROUND(COUNT(*) * 100.0 / SUM(COUNT(*)) OVER (PARTITION BY model_name), 1) as pct
FROM logs
WHERE {DATE_FILTER}
  AND type = 5
  AND model_name LIKE '%claude%'
  AND token_name != 'relay-pulse'
  AND other IS NOT NULL
GROUP BY model_name, status_code
ORDER BY model_name, cnt DESC;
ENDSQL

scp /tmp/claude_stats_error.sql root@139.180.189.205:/tmp/claude_stats_error.sql
scp /tmp/claude_stats_cache.sql root@139.180.189.205:/tmp/claude_stats_cache.sql
scp /tmp/claude_stats_errcode.sql root@139.180.189.205:/tmp/claude_stats_errcode.sql
"

第三步：执行查询并展示结果

wsl -- bash -c "ssh -o ConnectTimeout=20 -o ServerAliveInterval=10 root@139.180.189.205 'echo \"=== 错误率 ===\"; docker exec -i mysql mysql -uroot -pLanqi123456 new-api < /tmp/claude_stats_error.sql; echo; echo \"=== 缓存 ===\"; docker exec -i mysql mysql -uroot -pLanqi123456 new-api < /tmp/claude_stats_cache.sql; echo; echo \"=== 错误状态码分布 ===\"; docker exec -i mysql mysql -uroot -pLanqi123456 new-api < /tmp/claude_stats_errcode.sql' 2>&1 | grep -v 'Warning'"

字段说明

错误率表：

success — 正常请求数（type=2）
errors — 错误请求数（type=5）
error_rate_pct — 错误率 %

缓存率表：

cache_hit_rate_pct — 次数缓存率 = 有缓存命中的请求数 / 总请求数（最能反映缓存策略是否生效）
cost_pct_of_full — 实际成本占比 = (prompt × 1 + cache_read × 0.1 + cache_creation × 1.25) / (全量按原价计费)，越低越省钱（Anthropic 标准定价：cache_read 是原价 10%，cache_creation 是原价 125%）
creation_to_read_ratio — 缓存创建/读取比 = cache_creation_tokens / cache_read_tokens，比值高说明缓存频繁失效再重建，命中复用差

错误状态码分布表：

status_code — HTTP 状态码（504=上游超时，429=限流，400=请求参数错误）
cnt — 该状态码的错误次数
pct — 占该模型全部错误的百分比

注意事项

日期以 OV 服务器（UTC+0）的时区为准，北京时间当天 08:00 后才对应同一自然日
relay-pulse 是自动探活令牌，已从统计中排除

5.6 KiB Raw Blame History