Drop summary.md from outputs; reports switch to report.jsonl (header line + one sample per line, round-trip verified)

Co-Authored-By: Claude <noreply@anthropic.com>

README.md

@@ -127,9 +127,9 @@ evalharness viz show report.json --style errors            # 失败样本下钻
 ```
 <out-dir>/
 ├── summary.xlsx          总表 Excel（Summary/Perf/Categories/Samples 四 sheet，主入口）
-├── summary.md / .csv     同一张总表的 markdown / csv 版
+├── summary.csv           同一张总表的 csv 版
 └── <bench>/              每个 benchmark 一个目录
-    ├── report.json       完整报告（每样本预测、分数明细、token、轨迹；重判分用）
+    ├── report.jsonl      完整报告，流式行格式（首行报告头，之后每行一个样本；grep/tail 友好）
     └── detail.md         该 bench 的 markdown 详情
 ```
 

evalharness/cli.py

@@ -524,7 +524,7 @@ def _cmd_eval_run(args) -> int:
 
                 bench_dir = _P(out_dir) / name
                 bench_dir.mkdir(parents=True, exist_ok=True)
-                report.save(str(bench_dir / 'report.json'))
+                report.save(str(bench_dir / 'report.jsonl'))
                 with open(bench_dir / 'detail.md', 'w', encoding='utf-8') as f:
                     f.write(render(report, style='md'))
             if args.verbose:
@@ -618,22 +618,6 @@ def _cmd_eval_run(args) -> int:
                                    'ttft_mean_s', 'ttft_p90_s', 'ttft_p99_s',
                                    'tpot_mean_s', 'tpot_p90_s', 'tpot_p99_s')] +
                                [cats])
-            with open(f'{out_dir}/summary.md', 'w', encoding='utf-8') as f:
-                import time as _tt
-                model_names = {r.get('model', '') for r in rows if r.get('model')}
-                head = f"# eval run summary\n\n- model: {', '.join(model_names) or '?'}\n"
-                head += f"- created: {_tt.strftime('%Y-%m-%d %H:%M:%S')}\n"
-                head += f"- benchmarks: {sum(1 for r in rows if r['ok'])}/{len(rows)} ok\n\n"
-                f.write(head)
-                f.write('| benchmark | metric | score | n | time | status |\n'
-                        '|---|---|---:|---:|---:|---|\n')
-                for r in rows:
-                    v = _fmt_score(r.get('value'))
-                    h = r.get('hours') or 0
-                    t = f'{h:.2f}h' if h >= 0.995 else f"{r.get('secs', 0):.0f}s"
-                    st = 'ok' if r['ok'] else f"failed: {r.get('err', '')[:60]}"
-                    f.write(f"| {r['name']} | {r['metric']} | {v} | "
-                            f"{r.get('n', '')} | {t} | {st} |\n")
 
     # artifacts notice: tell the user where everything landed (or how to save);
     # rich terminals get clickable file:// links (iTerm2/kitty/WezTerm/WT...)
@@ -649,12 +633,12 @@ def _cmd_eval_run(args) -> int:
                 f'\n{mark} [bold]运行结束[/bold] · {ok_n}/{len(rows)} benchmarks ok\n'
                 f'  结果 {ap}\n'
                 f'  ├─ summary.xlsx (excel 打开总表)\n'
-                f'  ├─ summary.md / summary.csv\n'
-                f'  └─ <bench>/report.json + detail.md (每 benchmark 一个目录)')
+                f'  ├─ summary.csv\n'
+                f'  └─ <bench>/report.jsonl + detail.md (每 benchmark 一个目录)')
         else:
             print(f'\n运行结束 · {ok_n}/{len(rows)} benchmarks ok\n'
                   f'  结果 {ap}\n'
-                  f'  ├─ summary.xlsx\n  └─ <bench>/report.json + detail.md')
+                  f'  ├─ summary.xlsx\n  └─ <bench>/report.jsonl + detail.md')
     elif rows and rows[0]['ok'] and args.out:
         _notice('运行结束 · 结果已保存', args.out)
         print(f'{ok_n}/{len(rows)} benchmarks ok')
@@ -828,8 +812,7 @@ def build_parser() -> argparse.ArgumentParser:
                    help='first N samples PER subset/category (evalscope --limit semantics); '
                         'composable with --limit (intersection)')
     p.add_argument('--out', help='save the EvalReport json here (single dataset)')
-    p.add_argument('--out-dir', help='save reports/<name>.json + viz/<name>.txt + summary.md '
-                                     'here (multi-dataset runs)')
+    p.add_argument('--out-dir', help='output directory: summary.xlsx/csv + one dir per benchmark')
     p.add_argument('--style', default='text', help='result render style (text/md/radar/errors)')
     p.add_argument('--verbose', action='store_true', help='print full render for every dataset')
     _add_override_flags(p)

evalharness/eval/record.py

@@ -65,6 +65,18 @@ class EvalReport(BaseModel):
     def save(self, path) -> None:
         import json
 
+        if str(path).endswith('.jsonl'):
+            # streaming format: first line = report header, then one
+            # sample per line (grep/split/tail friendly)
+            head = self.model_dump(exclude={'samples'})
+            head['type'] = 'report'
+            with open(path, 'w', encoding='utf-8') as f:
+                f.write(json.dumps(head, ensure_ascii=False) + '\n')
+                for smp in self.samples:
+                    row = smp if isinstance(smp, dict) else smp.model_dump()
+                    row['type'] = 'sample'
+                    f.write(json.dumps(row, ensure_ascii=False) + '\n')
+            return
         with open(path, 'w', encoding='utf-8') as f:
             json.dump(self.model_dump(), f, ensure_ascii=False, indent=2)
 
@@ -72,5 +84,20 @@ class EvalReport(BaseModel):
     def load(cls, path) -> 'EvalReport':
         import json
 
+        if str(path).endswith('.jsonl'):
+            head, samples = None, []
+            with open(path, encoding='utf-8') as f:
+                for line in f:
+                    line = line.strip()
+                    if not line:
+                        continue
+                    row = json.loads(line)
+                    if row.pop('type', '') == 'report' or head is None:
+                        head = {k: v for k, v in row.items() if k != 'type'}
+                    else:
+                        samples.append({k: v for k, v in row.items() if k != 'type'})
+            head = head or {}
+            head['samples'] = samples
+            return cls.model_validate(head)
         with open(path, encoding='utf-8') as f:
             return cls.model_validate(json.load(f))