harness2: per-role models, sweep batch, task file assets endpoint

- Role models: run_improvement_via_tasks gains role_models
  (probe/editor/selector/reflector) plus --model-<role> CLI flags;
  missing/empty falls back to shared --model, both empty means
  server default, so default behavior is unchanged.
- Sweep: run_sweep + examples/harness2_sweep.py manifest CLI with
  shared library_root (library restore), per-run failure isolation
  and sweep-report.json summary; preflight gates task creation.
- Deliverable assets: GET /tasks/{id}/files for local-executor
  tasks (safe path resolution, UTF-8 text <=256KB, listing <=200;
  docker/cloud/cross-host 409 with explicit reason, since tenant
  Docker shares one /app/work per project and per-task attribution
  would misattribute); probe rollout deliverable_files from the
  endpoint plus DELIVERABLE FILE reply blocks (endpoint wins on
  collision, bad blocks dropped), surfaced in
  deliverables/<label>/files/ context and full_result summary.
- Docs: harness2.md comparison table updated for all three items.

examples/harness2_run.py

@@ -69,6 +69,10 @@ def main():
     ap.add_argument("--output", default="")
     ap.add_argument("--library-root", default="")
     ap.add_argument("--model", default=None)
+    ap.add_argument("--model-probe", default=None)
+    ap.add_argument("--model-editor", default=None)
+    ap.add_argument("--model-selector", default=None)
+    ap.add_argument("--model-reflector", default=None)
     ap.add_argument("--effort", default=None)
     ap.add_argument("--approval", default=None)
     ap.add_argument("--environment-id", default=None)
@@ -148,7 +152,11 @@ def main():
         k=args.k, mode=args.mode,
         library_root=Path(args.library_root) if args.library_root else None,
         budget_usd=args.budget, max_sessions=args.max_sessions,
-        model=args.model, effort=args.effort,
+        model=args.model,
+        role_models={"probe": args.model_probe, "editor": args.model_editor,
+                     "selector": args.model_selector,
+                     "reflector": args.model_reflector},
+        effort=args.effort,
         approval=args.approval, environment_id=args.environment_id,
         executor=args.executor, no_time_limit=args.no_time_limit,
         timeout_seconds=args.timeout)

examples/harness2_sweep.py

@@ -0,0 +1,176 @@
+"""多轮 Harness² 改进批量跑（sweep，接线层 CLI）。
+
+读一个 JSON manifest，顺序跑多轮 run_improvement_via_tasks：所有轮
+共享同一个 library_root（已入库的经验会被后续轮读到，即上游 sweep
+的 library restore 语义）；单轮失败只记入报告、不中断后续轮
+（dead 隔离）；最后写 sweep-report.json 并汇总打印。
+
+这会创建真实任务、花真实模型配额——所以默认只打印计划，不建
+任务：先跑一遍不带 --yes 看计划，确认无误后，再带 --yes 重跑。
+密钥不走命令行参数（ps 可见）：用环境变量 CODEX_API_KEY。
+
+manifest 形状：
+{
+  "runs": [
+    {"run_tag": "sw1", "task_id": "ledger-1",
+     "prompt": "Sum every signed amount in ledger.csv.",
+     "base_harness_dir": "./base-harness", "source_dir": "./task-source",
+     "domain": "accounting", "k": 1, "mode": "parallel",
+     "budget": 5, "max_sessions": null, "model": null,
+     "role_models": null, "effort": null, "approval": null,
+     "environment_id": null, "executor": null,
+     "no_time_limit": false, "timeout": 3600}
+  ]
+}
+prompt 也可用 prompt_file（相对 manifest 所在目录）；base_harness_dir
+与 source_dir 同样相对 manifest 目录解析（绝对路径原样用）。
+"""
+
+import argparse
+import json
+import os
+import sys
+import tempfile
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+
+from cloud.harness2 import read_tree  # noqa: E402
+from cloud.harness2_tasks import (  # noqa: E402
+    HttpTaskClient, preflight_checks, run_sweep, validate_run_tag,
+)
+
+
+def _flatten(root: Path) -> dict[str, bytes]:
+    out: dict[str, bytes] = {}
+    if not root or not root.is_dir():
+        return out
+    for p in sorted(root.rglob("*")):
+        if p.is_file() and not p.is_symlink() and "__pycache__" not in p.parts:
+            out[str(p.relative_to(root))] = p.read_bytes()
+    return out
+
+
+def _resolve(manifest_dir: Path, value):
+    if not value:
+        return value
+    p = Path(value)
+    return p if p.is_absolute() else manifest_dir / p
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--manifest", required=True, help="JSON manifest 路径")
+    ap.add_argument("--base-url", required=True, help="控制面地址")
+    ap.add_argument("--project-id", required=True)
+    ap.add_argument("--output", default="",
+                    help="sweep 工作根目录（默认临时目录）")
+    ap.add_argument("--library-root", default="",
+                    help="共享经验库根目录（空 = 各轮不入库）")
+    ap.add_argument("--yes", action="store_true", help="确认创建真实任务并消费配额")
+    args = ap.parse_args()
+
+    manifest_path = Path(args.manifest)
+    try:
+        manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
+    except (OSError, ValueError) as exc:
+        print(f"manifest 读不到或不是合法 JSON：{exc}", file=sys.stderr)
+        return 2
+    runs = manifest.get("runs")
+    if not isinstance(runs, list) or not runs:
+        print("manifest 里 runs 必须是非空列表", file=sys.stderr)
+        return 2
+    manifest_dir = manifest_path.resolve().parent
+
+    specs = []
+    seen_tags = set()
+    for i, raw in enumerate(runs):
+        if not isinstance(raw, dict):
+            print(f"runs[{i}] 不是对象", file=sys.stderr)
+            return 2
+        tag = raw.get("run_tag", "")
+        try:
+            if not tag:
+                raise ValueError("每轮必须给 run_tag")
+            validate_run_tag(tag)
+        except ValueError as exc:
+            print(f"runs[{i}] run_tag 非法：{exc}", file=sys.stderr)
+            return 2
+        if tag in seen_tags:
+            print(f"runs[{i}] run_tag 重复：{tag}", file=sys.stderr)
+            return 2
+        seen_tags.add(tag)
+        prompt = raw.get("prompt", "")
+        if raw.get("prompt_file"):
+            try:
+                prompt = _resolve(manifest_dir, raw["prompt_file"]).read_text(
+                    encoding="utf-8")
+            except OSError as exc:
+                print(f"runs[{i}] prompt_file 读不到：{exc}", file=sys.stderr)
+                return 2
+        base_dir = _resolve(manifest_dir, raw.get("base_harness_dir", ""))
+        errors, warnings = preflight_checks(
+            base_harness_dir=base_dir,
+            source_dir=_resolve(manifest_dir, raw.get("source_dir")) or None,
+            prompt=prompt, run_tag=tag, base_url=args.base_url)
+        for e in errors:
+            print(f"FAIL runs[{i}] {tag}: {e}", file=sys.stderr)
+        for w in warnings:
+            print(f"WARN runs[{i}] {tag}: {w}")
+        if errors:
+            return 2
+        source_dir = _resolve(manifest_dir, raw.get("source_dir"))
+        specs.append({
+            "run_tag": tag,
+            "task_id": raw.get("task_id", f"h2-task-{tag}"),
+            "prompt": prompt,
+            "base_harness": read_tree(base_dir),
+            "source_files": _flatten(source_dir) if source_dir else None,
+            "domain": raw.get("domain", "general"),
+            "k": raw.get("k", 1),
+            "mode": raw.get("mode", "parallel"),
+            "budget_usd": raw.get("budget"),
+            "max_sessions": raw.get("max_sessions"),
+            "model": raw.get("model"),
+            "role_models": raw.get("role_models"),
+            "effort": raw.get("effort"),
+            "approval": raw.get("approval"),
+            "environment_id": raw.get("environment_id"),
+            "executor": raw.get("executor"),
+            "no_time_limit": bool(raw.get("no_time_limit", False)),
+            "timeout_seconds": raw.get("timeout", 3600),
+        })
+
+    print(f"计划：项目 {args.project_id}，共 {len(specs)} 轮，"
+          f"tags={sorted(seen_tags)}，"
+          f"共享 library={'是' if args.library_root else '否（不入库）'}")
+    if not args.yes:
+        print("未带 --yes：只打印计划，不创建任何任务。确认无误后加 --yes 重跑。")
+        return 0
+    api_key = os.environ.get("CODEX_API_KEY")
+    if not api_key:
+        print("缺少 CODEX_API_KEY 环境变量（密钥不接受命令行参数）", file=sys.stderr)
+        return 2
+
+    work_root = Path(args.output) if args.output else Path(
+        tempfile.mkdtemp(prefix="h2sweep-"))
+    client = HttpTaskClient(args.base_url, api_key)
+    report = run_sweep(
+        client, project_id=args.project_id, runs=specs, work_root=work_root,
+        library_root=Path(args.library_root) if args.library_root else None)
+    report_path = work_root / "sweep-report.json"
+    report_path.write_text(json.dumps(report, ensure_ascii=False, indent=2),
+                           encoding="utf-8")
+    for entry in report["runs"]:
+        status = "ok" if entry["ok"] else f"FAIL {entry['error']}"
+        print(f"{entry['run_tag']}: {status} "
+              f"chosen={entry['chosen_label']} "
+              f"cost={entry['total_cost_usd']}")
+    totals = report["totals"]
+    print(f"汇总：{totals['n_ok']}/{totals['n_runs']} 轮成功，"
+          f"总成本 {totals['total_cost_usd']}，报告 {report_path}")
+    return 0 if totals["n_ok"] == totals["n_runs"] else 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())