3v324v23's picture
Add 102 active scenarios dataset with prompts and fixtures
10bfe83
Raw
History Blame Contribute Delete
2.37 kB
benchmark_core: false
benchmark_group: coverage
benchmark_status: active
signal_source: workspace_live
custom_check: error_recovery_07_partial_success_live_v2.py
difficulty: expert
dimension: error_recovery
execution_mode: live
expected_outcome: 'recovery_plan.md exists with partial success recovery and retry
strategy grounded in the seeded workspace evidence.
'
expected_tools:
- read
- write
- exec
id: error_recovery_07_partial_success_live
name: 部分成功恢复
optimal_steps: 12
pass_threshold: 0.6
prompt: |
一个批量数据处理任务在末段出现数据库超时,但并不是所有文件都需要同样的恢复动作。
工作区根目录里已经放好了只读证据:
- `process.log`
- `processed_files.csv`:`processed_files` 数据库表导出
- `file_list.txt`
重要约束:
- 不要修改这些输入文件
- 不要补造或重写缺失数据
- 只创建 `recovery_plan.md`
- 不需要把 300 个候选文件完整展开到 markdown
任务:
1. 找出真实根因
2. 交叉核对三份输入,区分:
- 已成功且不应重跑的文件
- 失败但可直接重跑的文件
- 已部分写入、必须先清理再重跑的文件
3. 给出一个保守的重试策略和验证步骤
`recovery_plan.md` 顶部必须先放一个 fenced `json` 代码块,字段必须完整:
```json
{
"root_cause": "db_connection_pool_exhaustion",
"successful_files": 0,
"failed_files": 0,
"partial_files": 0,
"retry_files_total": 0,
"cleanup_files_total": 0,
"cleanup_examples": ["...", "...", "..."],
"retry_examples": ["...", "...", "...", "...", "..."],
"batch_size": 0,
"max_retries": 0,
"backoff_seconds": [0, 0, 0]
}
```
额外要求:
- `cleanup_examples` 必须是需要清理文件按字母序排序后的前 3
- `retry_examples` 必须是需要重试文件按字母序排序后的前 5
- `backoff_seconds` 必须严格递增
- JSON 代码块之后再写简短说明,覆盖 evidence、cleanup、retry strategy、verification
- 不要输出额外文件
tags:
- error_recovery
- partial-success
- retry-strategy
- expert
- benchmark-intelligence
- benchmark-core
- new-expert
timeout_seconds: 900
tools:
- read
- write
- exec
workspace_seed_dir: ../../datasets/error_recovery_07_partial_success/fixtures
weight: 2.0