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.

cloud/harness2.py

@@ -241,6 +241,56 @@ def parse_harness_files(text: str) -> dict[str, str]:
     return out
 
 
+_DELIVERABLE_FILE_START_RX = re.compile(r"^=== DELIVERABLE FILE:\s*(.+?)\s*===$")
+_DELIVERABLE_FILE_END = "=== END DELIVERABLE FILE ==="
+_DELIVERABLE_FILE_MAX = 8
+_DELIVERABLE_FILE_MAX_CHARS = 65536
+
+
+def _is_safe_deliverable_path(path: str) -> bool:
+    if not path or path.startswith("/") or "\x00" in path:
+        return False
+    return all(p not in ("", ".", "..") for p in path.split("/"))
+
+
+def parse_deliverable_files(text: str) -> dict[str, str]:
+    """解析探针回复里的产物文件块，返回 {相对路径: 内容}。
+
+    与编辑文件块的严格语义不同：这里故意宽容——坏块（未闭合、
+    非法路径、超量、超长）只丢弃该块、不抛错，因为产物文件是
+    交付文本之外的 enrichment，绝不能因为一块坏内容毁掉整轮
+    交付物。超长内容截断并打截断标记，让评审知道它不完整；
+    超过 8 块时只取前 8 块。同一路径出现两次时后一块覆盖前一块。
+    """
+    out: dict[str, str] = {}
+    cur_path: str | None = None
+    cur_lines: list[str] = []
+    for line in (text or "").splitlines():
+        m = _DELIVERABLE_FILE_START_RX.match(line.strip())
+        if m:
+            if cur_path is not None:
+                cur_path = None
+                cur_lines = []
+            path = m.group(1).strip()
+            if _is_safe_deliverable_path(path) and len(out) < _DELIVERABLE_FILE_MAX:
+                cur_path = path
+                cur_lines = []
+            continue
+        if line.strip() == _DELIVERABLE_FILE_END:
+            if cur_path is not None:
+                content = "\n".join(cur_lines) + ("\n" if cur_lines else "")
+                if len(content) > _DELIVERABLE_FILE_MAX_CHARS:
+                    content = (content[:_DELIVERABLE_FILE_MAX_CHARS]
+                               + "\n...[truncated]")
+                out[cur_path] = content
+            cur_path = None
+            cur_lines = []
+            continue
+        if cur_path is not None:
+            cur_lines.append(line)
+    return out
+
+
 MAX_K = 3
 STAGE_ATTEMPTS = 2  # 提议被拒后给一次按反馈修正的机会，再拒就回落
 
@@ -490,6 +540,7 @@ class Rollout:
     trajectory: str = ""
     cost_usd: float = 0.0
     error: str = ""
+    deliverable_files: dict[str, str] = field(default_factory=dict)
 
 
 @dataclass
@@ -589,7 +640,12 @@ def build_context(
             (dd / f"{r.label}.md").write_text(
                 r.deliverable_text or "(no deliverable)", encoding="utf-8"
             )
-        lines.append("context/deliverables/           what each rollout produced")
+            for name, content in (r.deliverable_files or {}).items():
+                dest = dd / r.label / "files" / name
+                dest.parent.mkdir(parents=True, exist_ok=True)
+                dest.write_text(content, encoding="utf-8")
+        lines.append("context/deliverables/           what each rollout produced"
+                     " (+ <label>/files/ when it produced files)")
     write_tree(harness, root / "harness")
     lines.append("context/harness/                the harness tree you may edit, and ONLY this")
     if library_dir and Path(library_dir).is_dir():
@@ -842,6 +898,7 @@ class ModeResult:
                     "delivered": r.delivered,
                     "error": (r.error or "")[:400],
                     "deliverable_text": (r.deliverable_text or "")[:500],
+                    "deliverable_files": sorted(r.deliverable_files or {}),
                 }
                 for lbl, r in self.rollouts.items()
             },

cloud/harness2_tasks.py

@@ -1,6 +1,6 @@
 """Harness² 接线层：把 cloud/harness2.py 的三个协议落到真实 codex 任务上。
 
-只用控制面已有的 API，不新增任何服务端端点：
+控制面 API：
   POST /uploads（multipart + paths JSON）-> upload_id
   POST /tasks（upload_id 绑定 + idempotency_key 幂等）-> 任务排队
   GET  /tasks/{id}        轮询终态（completed / failed / interrupted /
@@ -8,6 +8,8 @@
   GET  /tasks/{id}/trace  轨迹（探针的 trajectory 来源）
   GET  /tasks/{id}/cost   成本（usd_micros / 1e6；无价 token 不计入 USD，
                           这是现有 cost 模型的边界，不是本层漏算）
+  GET  /tasks/{id}/files  读本地任务 worktree 文件（产物资产端点；
+                          Docker/跨机任务回 409，见下）
   POST /tasks/{id}/cancel 超时取消，不留孤儿任务继续烧钱
 认证：X-API-Key 请求头；密钥只由调用方持有，本模块不读、不存、
 不打印密钥。
@@ -18,12 +20,13 @@
   文件会覆盖仓库同名文件——这正是探针要的效果：同一任务在不同
   harness 下跑。prompt = 原任务 prompt + 一段交付脚注（见
   _PROBE_DELIVERABLE_FOOTER），因为评审只看最终回复：agent 必须
-  把交付物完整写在最终回复里，不能只说“文件已写好”。这就是本层
-  最大的适用边界：交付物是最终回复文本的任务；文件型产物（代码
-  diff、报告文件）目前只能靠最终回复复述/摘要，worktree 快照 API
-  只给 diff 截断与未跟踪文件列表、不给完整文件内容，跨主机还读
-  不到——文件产物评审需要新增文件读取端点，另行评审，不在本层
-  假装已解决。
+  把交付物完整写在最终回复里，不能只说“文件已写好”。
+  文件型产物走两层（产物资产）：①本地任务经 GET /tasks/{id}/files
+  读 worktree 新增/改动文件（端点为准）；②所有执行器都可用
+  回复里的 DELIVERABLE FILE 块（agent 自报，端点结果优先）。
+  边界仍在：租户 Docker 下 /app/work 是整项目共享目录，按任务
+  归属读文件会错指，端点对 Docker 任务一律 409——这时只有第
+  ②层可用，且内容以 agent 自报为准（可能截断，截断有标记）。
 - 编辑（CodexEditor.edit）：bundle = context/ 前缀的证据树（task.md /
   rollouts / deliverables / harness / library / pool / decisions）。
   agent 在 worktree 的 context/harness/ 里改文件，最终回复里用
@@ -65,6 +68,7 @@ import json
 import re
 import time
 import urllib.error
+import urllib.parse
 import urllib.request
 import uuid
 from dataclasses import dataclass, field
@@ -77,6 +81,7 @@ from .harness2 import (
     Tree,
     improve_task,
     parse_choice,
+    parse_deliverable_files,
     parse_experience,
     parse_harness_files,
     parse_no_change,
@@ -185,6 +190,11 @@ write the complete deliverable in your final reply itself — do not only say
 "done" or "the file is written". If the task asks for a file, also write the
 file in the worktree, but the final reply must still contain the full content
 a reviewer needs to judge the work.
+If you created deliverable files (a page, a report, an image manifest),
+also paste each file's full content in your final reply as a block:
+=== DELIVERABLE FILE: <relative path> ===
+<complete file content>
+=== END DELIVERABLE FILE ===
 """
 
 _EDITOR_OUTPUT_FORMAT = """
@@ -278,6 +288,12 @@ class TaskClient(Protocol):
         """轨迹渲染文本；读不到时返回空串（调用方用最终回复兜底）。"""
         ...
 
+    def get_task_files(self, task_id: str, path: str = "") -> dict:
+        """GET /tasks/{id}/files：不传 path 列 worktree 文件，传 path
+        读单个文件。服务端对 Docker/跨机任务回 409，调用方须把
+        409 当成“无文件通道”、不得当成探针失败。"""
+        ...
+
     def get_cost_usd(self, task_id: str) -> float:
         """有价成本 USD；读不到时返回 0.0，不炸轮次。"""
         ...
@@ -452,6 +468,10 @@ class HttpTaskClient:
         rendered = "\n".join(lines).strip()
         return rendered[:TRACE_MAX_CHARS] if rendered else ""
 
+    def get_task_files(self, task_id: str, path: str = "") -> dict:
+        q = f"?path={urllib.parse.quote(path)}" if path else ""
+        return self._request("GET", f"/tasks/{task_id}/files{q}")
+
     def get_cost_usd(self, task_id: str) -> float:
         try:
             obj = self._request("GET", f"/tasks/{task_id}/cost")
@@ -596,6 +616,12 @@ class CodexProbeAdapter:
                            error=f"{type(exc).__name__}: {exc}"[:400])
         state = done.state
         delivered = state.status == "completed" and bool(state.result_text.strip())
+        deliverable_files = self._collect_deliverable_files(done.task_id, files)
+        # 回复里的产物文件块是补充通道（所有执行器都可用）；
+        # 端点读到的文件为准，冲突时不覆盖端点结果。
+        for name, content in parse_deliverable_files(
+                state.result_text or "").items():
+            deliverable_files.setdefault(name, content)
         return Rollout(
             label=label, ok=state.status == "completed", delivered=delivered,
             deliverable_text=state.result_text,
@@ -603,8 +629,46 @@ class CodexProbeAdapter:
             cost_usd=done.cost_usd,
             error="" if state.status == "completed" else (
                 state.error_text or f"task ended as {state.status}")[:400],
+            deliverable_files=deliverable_files,
         )
 
+    def _collect_deliverable_files(self, task_id: str,
+                                   seed_files: dict[str, bytes]) -> dict[str, str]:
+        """经 GET /tasks/{id}/files 读探针 worktree 里新增的文件。
+
+        取 seed 里没有的新文件，以及内容被改过的 seed 文件（与
+        seed 字节比对，不一致才算产物）；跳过隐藏文件、最多收
+        8 个、最多读 16 个；端点不可用（Docker/跨机 409、读错）
+        时返回空 dict——文件产物是 enrichment，绝不能因此炸轮次。
+        """
+        try:
+            listing = self.client.get_task_files(task_id)
+        except Exception:
+            return {}
+        out: dict[str, str] = {}
+        reads = 0
+        for entry in listing.get("entries") or []:
+            if not isinstance(entry, dict):
+                continue
+            path = entry.get("path") or ""
+            if not path or path.startswith("."):
+                continue
+            if len(out) >= 8 or reads >= 16:
+                break
+            try:
+                data = self.client.get_task_files(task_id, path)
+            except Exception:
+                continue
+            reads += 1
+            content = data.get("content")
+            if not isinstance(content, str):
+                continue
+            seed = seed_files.get(path)
+            if seed is not None and seed.decode("utf-8", errors="replace") == content:
+                continue
+            out[path] = content
+        return out
+
 
 @dataclass
 class CodexEditor:
@@ -773,6 +837,7 @@ def run_improvement_via_tasks(
     budget_usd: float | None = None,
     max_sessions: int | None = None,
     model: str | None = None,
+    role_models: dict[str, str | None] | None = None,
     effort: str | None = None,
     approval: str | None = None,
     environment_id: str | None = None,
@@ -785,22 +850,37 @@ def run_improvement_via_tasks(
     library_root 为空时不建经验库（只跑改进、不入库）；非空时按
     domain 建库。auto_apply 恒为 False——采纳的树只入库待评审，
     不自动成为后续任务的 base（见 docs/harness2.md 安全一节）。
+
+    分角色模型：role_models 键为 probe / editor / selector /
+    reflector，非空值覆盖该角色的 model；缺键或空值时回落到
+    共享 model；两者都空时不传 model（服务端默认模型）。默认
+    行为与此前完全一致——本参数只提供分流能力，不替用户选
+    模型（有价模型会产生真实成本，见 docs/harness2.md）。
     """
     from .harness2 import Library
 
+    roles = role_models or {}
+
+    def _role_model(role: str) -> str | None:
+        return roles.get(role) or model
+
     adapter = CodexProbeAdapter(
         client=client, project_id=project_id, task_id=task_id, prompt=prompt,
         base=base_harness, run_tag=run_tag,
         source_files=source_files or {}, environment_id=environment_id,
-        executor=executor, model=model, effort=effort, approval=approval,
+        executor=executor, model=_role_model("probe"), effort=effort,
+        approval=approval,
         no_time_limit=no_time_limit, timeout_seconds=timeout_seconds,
         domain=domain)
-    shared = dict(environment_id=environment_id, executor=executor, model=model,
+    shared = dict(environment_id=environment_id, executor=executor,
                   effort=effort, approval=approval, no_time_limit=no_time_limit,
                   timeout_seconds=timeout_seconds)
-    editor = CodexEditor(client=client, project_id=project_id, run_tag=run_tag, **shared)
-    selector = CodexSelector(client=client, project_id=project_id, run_tag=run_tag, **shared)
-    reflector = CodexReflector(client=client, project_id=project_id, run_tag=run_tag, **shared)
+    editor = CodexEditor(client=client, project_id=project_id, run_tag=run_tag,
+                         model=_role_model("editor"), **shared)
+    selector = CodexSelector(client=client, project_id=project_id, run_tag=run_tag,
+                             model=_role_model("selector"), **shared)
+    reflector = CodexReflector(client=client, project_id=project_id, run_tag=run_tag,
+                               model=_role_model("reflector"), **shared)
     library = Library(library_root, domain) if library_root is not None else None
     validate_run_tag(run_tag)
     return improve_task(
@@ -809,10 +889,62 @@ def run_improvement_via_tasks(
         auto_apply=False, budget_usd=budget_usd, max_sessions=max_sessions)
 
 
+def run_sweep(client, *, project_id, runs, work_root, library_root=None):
+    """批量跑多轮改进：共享经验库、单轮失败隔离、汇总报告。
+
+    runs：每轮一个 dict，键为 run_improvement_via_tasks 的参数，
+    run_tag 必填；client / project_id / work_dir / library_root
+    由本函数统管，出现在 spec 里会被忽略（不报错）。每轮
+    work_dir = work_root/<run_tag>。library_root 非空时所有轮
+    共享同一库——已入库的经验会被后续轮读到，即上游 sweep 的
+    library restore 语义；为空时各轮都不入库（仍汇总）。
+
+    单轮抛错只记入报告的 error，不中断后续轮（dead 隔离）。
+    返回报告 dict：runs 逐轮（ok / chosen_label / deliverable
+    前 500 字 / cost / library_entry / error）+ totals（轮数、
+    成功轮数、总成本）。
+    """
+    work_root = Path(work_root)
+    report_runs = []
+    total_cost = 0.0
+    n_ok = 0
+    for spec in runs:
+        spec = dict(spec)
+        run_tag = spec.pop("run_tag", "")
+        for owned in ("client", "project_id", "work_dir", "library_root"):
+            spec.pop(owned, None)
+        entry = {"run_tag": run_tag, "ok": False, "chosen_label": None,
+                 "deliverable_text": "", "total_cost_usd": 0.0,
+                 "library_entry": None, "error": None}
+        try:
+            if not run_tag:
+                raise ValueError("sweep 每轮必须给 run_tag（用作 work_dir 子目录）")
+            validate_run_tag(run_tag)
+            result = run_improvement_via_tasks(
+                client, project_id=project_id, run_tag=run_tag,
+                work_dir=work_root / run_tag,
+                library_root=library_root, **spec)
+            entry["ok"] = bool(result.ok)
+            entry["chosen_label"] = result.selection.chosen_label
+            chosen = result.mode_result.rollouts.get(result.selection.chosen_label)
+            if chosen is not None:
+                entry["deliverable_text"] = (chosen.deliverable_text or "")[:500]
+            entry["total_cost_usd"] = result.mode_result.total_cost_usd
+            entry["library_entry"] = result.library_entry
+            total_cost += result.mode_result.total_cost_usd
+            n_ok += 1 if result.ok else 0
+        except Exception as exc:
+            entry["error"] = f"{type(exc).__name__}: {exc}"[:400]
+        report_runs.append(entry)
+    return {"runs": report_runs,
+            "totals": {"n_runs": len(report_runs), "n_ok": n_ok,
+                       "total_cost_usd": total_cost}}
+
+
 __all__ = [
     "TERMINAL_STATUSES", "DEFAULT_TASK_TIMEOUT_SECONDS", "DEFAULT_POLL_SECONDS",
     "TRACE_MAX_CHARS", "TaskState", "TaskClient", "build_multipart",
     "HttpTaskClient", "CompletedTask", "run_task",
     "CodexProbeAdapter", "CodexEditor", "CodexSelector", "CodexReflector",
-    "run_improvement_via_tasks", "validate_run_tag", "preflight_checks",
+    "run_improvement_via_tasks", "run_sweep", "validate_run_tag", "preflight_checks",
 ]

cloud/server.py

@@ -445,6 +445,22 @@ def _thread_worktree_for_task(server, task):
     return os.path.join(base, safe)
 
 
+def _safe_worktree_path(worktree, rel):
+    """worktree 内相对路径的安全解析：空串返回 worktree 本身；
+    绝对路径、.. 逃逸、符号链接逃逸一律返回 None。realpath 后
+    再校验 commonpath，目录内的符号链接指回内部时放行（读到
+    的仍是 worktree 内的内容），指出去时拦住。"""
+    if not rel:
+        return worktree
+    if os.path.isabs(rel):
+        return None
+    base = os.path.realpath(worktree)
+    target = os.path.realpath(os.path.join(base, rel))
+    if os.path.commonpath([base, target]) != base:
+        return None
+    return target
+
+
 def _extract_thread_list(result):
     """从 thread-list 的 CLI 结果里取 thread 列表：标准形态是
     ThreadListResponse 的 data；threads 键与 result 包裹是
@@ -2152,6 +2168,8 @@ class CloudControlPlaneHandler(BaseHTTPRequestHandler):
                 self._list_followups(user_id, rest[: -len("/followups")])
             elif rest.endswith("/timeline") and method == "GET":
                 self._task_timeline(user_id, rest[: -len("/timeline")])
+            elif rest.endswith("/files") and method == "GET":
+                self._task_files(user_id, rest[: -len("/files")], query)
             elif rest.endswith("/chat-history") and method == "GET":
                 self._task_chat_history(
                     user_id, rest[: -len("/chat-history")])
@@ -4698,6 +4716,87 @@ class CloudControlPlaneHandler(BaseHTTPRequestHandler):
                             for r in cp_rows],
         })
 
+    def _task_files(self, user_id, task_id, query):
+        """GET /tasks/{id}/files：读本地任务 worktree 里的文件。
+
+        只做 local 执行器任务（复用 thread 操作的执行器门）：
+        Docker 任务的 worktree 在容器内、且租户 Docker 下 /app/work
+        是整项目共享目录，按任务归属读文件会错指，一律 409 明
+        说；cloud 任务与跨机 worktree 同理。executor 为空的老
+        任务放行（与 thread 门同款），解析不到 worktree 时 404。
+
+        ?path=<相对路径>：不传列 worktree 顶层（递归、最多 200
+        条，只含文件路径与大小，不跟符号链接）；传目录列该目录
+        （非递归，含 is_dir）；传文件读文本内容（UTF-8、≤256KB，
+        二进制或超大一律 409/413 明说，不截断冒充全文）。路径
+        解析一律走 _safe_worktree_path，防 .. 与符号链接逃逸。
+        读操作不限制任务状态（进行中任务读到的是中间态）。
+        """
+        task, err = self._thread_task(user_id, task_id, "read")
+        if err is not None:
+            self._send_json(err[0], {"error": err[1]})
+            return
+        guard = self._thread_task_guard(task)
+        if guard is not None:
+            self._send_json(guard[0], {"error": guard[1]})
+            return
+        worktree = _thread_worktree_for_task(self.server, task)
+        if not os.path.isdir(worktree):
+            self._send_json(404, {"error": "no local worktree for this task"})
+            return
+        rel = (query.get("path") or [""])[0]
+        target = _safe_worktree_path(worktree, rel)
+        if target is None:
+            self._send_json(400, {"error": "path escapes task worktree"})
+            return
+        if os.path.isdir(target):
+            entries = []
+            if not rel:
+                for root, dirs, files in os.walk(target):
+                    dirs[:] = [d for d in dirs
+                               if not os.path.islink(os.path.join(root, d))]
+                    dirs.sort()
+                    for name in sorted(files):
+                        fpath = os.path.join(root, name)
+                        if os.path.islink(fpath):
+                            continue
+                        entries.append({
+                            "path": os.path.relpath(fpath, worktree),
+                            "size": os.path.getsize(fpath),
+                        })
+                        if len(entries) >= 200:
+                            break
+                    if len(entries) >= 200:
+                        break
+            else:
+                for name in sorted(os.listdir(target)):
+                    fpath = os.path.join(target, name)
+                    is_link = os.path.islink(fpath)
+                    entries.append({
+                        "path": os.path.relpath(fpath, worktree),
+                        "size": os.path.getsize(fpath)
+                        if os.path.isfile(fpath) and not is_link else 0,
+                        "is_dir": os.path.isdir(fpath) and not is_link,
+                    })
+            self._send_json(200, {"task_id": task_id, "path": rel,
+                                  "entries": entries})
+            return
+        if not os.path.isfile(target):
+            self._send_json(404, {"error": "no such file"})
+            return
+        size = os.path.getsize(target)
+        if size > 256 * 1024:
+            self._send_json(413, {"error": "file too large for inline read"})
+            return
+        try:
+            with open(target, encoding="utf-8") as fh:
+                content = fh.read()
+        except (OSError, UnicodeDecodeError):
+            self._send_json(409, {"error": "file is not readable UTF-8 text"})
+            return
+        self._send_json(200, {"task_id": task_id, "path": rel,
+                              "size": size, "content": content})
+
     def _task_chat_history(self, user_id, task_id):
         """GET /tasks/{id}/chat-history：真正的交替历史，一次性
         按时间交错拼好返回，前端不再猜顺序。

cloud/tests/test_harness2_tasks.py

@@ -19,26 +19,26 @@ from pathlib import Path
 try:
     from cloud import harness2_tasks
     from cloud.harness2 import (
-        parse_choice, parse_experience, parse_harness_files, parse_no_change,
-        parse_notes, parse_why,
+        parse_choice, parse_deliverable_files, parse_experience,
+        parse_harness_files, parse_no_change, parse_notes, parse_why,
     )
     from cloud.harness2_tasks import (
         CodexEditor, CodexProbeAdapter, HttpTaskClient, TaskState,
-        build_multipart, preflight_checks, run_improvement_via_tasks, run_task,
-        unwrap_result_text, validate_run_tag,
+        build_multipart, preflight_checks, run_improvement_via_tasks, run_sweep,
+        run_task, unwrap_result_text, validate_run_tag,
     )
     from cloud.uploads import parse_multipart, parse_paths_field
 except ImportError:  # discover -s cloud/tests 时把项目根目录补进 sys.path
     sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
     from cloud import harness2_tasks
     from cloud.harness2 import (
-        parse_choice, parse_experience, parse_harness_files, parse_no_change,
-        parse_notes, parse_why,
+        parse_choice, parse_deliverable_files, parse_experience,
+        parse_harness_files, parse_no_change, parse_notes, parse_why,
     )
     from cloud.harness2_tasks import (
         CodexEditor, CodexProbeAdapter, HttpTaskClient, TaskState,
-        build_multipart, preflight_checks, run_improvement_via_tasks, run_task,
-        unwrap_result_text, validate_run_tag,
+        build_multipart, preflight_checks, run_improvement_via_tasks, run_sweep,
+        run_task, unwrap_result_text, validate_run_tag,
     )
     from cloud.uploads import parse_multipart, parse_paths_field
 
@@ -61,15 +61,22 @@ LEDGER = b"amount\n100.25\n-20.10\n5.00\n"
 class FakeTaskClient:
     """进程内控制面：按 prompt 类型分派到确定性假 worker。"""
 
-    def __init__(self, editor_mode="fix", selector_mode="signed", envelope=False):
+    def __init__(self, editor_mode="fix", selector_mode="signed", envelope=False,
+                 fail_tags=(), endpoint_files=None, endpoint_error=False,
+                 probe_extra=""):
         self.editor_mode = editor_mode
         self.selector_mode = selector_mode
         self.envelope = envelope
+        self.fail_tags = tuple(fail_tags)
+        self.endpoint_files = dict(endpoint_files or {})
+        self.endpoint_error = endpoint_error
+        self.probe_extra = probe_extra
         self.uploads = {}
         self.tasks = {}
         self.by_key = {}
         self.created = 0
         self.cancelled = []
+        self.task_models = {}
 
     def upload(self, files):
         uid = f"u{len(self.uploads) + 1}"
@@ -78,6 +85,8 @@ class FakeTaskClient:
 
     def create_task(self, *, project_id, prompt, upload_id=None,
                     idempotency_key=None, **kwargs):
+        if idempotency_key and any(t in idempotency_key for t in self.fail_tags):
+            raise RuntimeError("injected task failure")
         if idempotency_key and idempotency_key in self.by_key:
             return self.by_key[idempotency_key]
         tid = f"t{len(self.tasks) + 1}"
@@ -94,6 +103,7 @@ class FakeTaskClient:
             })
         self.tasks[tid] = TaskState(task_id=tid, status="completed",
                                     result_text=result)
+        self.task_models[tid] = kwargs.get("model")
         if idempotency_key:
             self.by_key[idempotency_key] = tid
         return tid
@@ -163,11 +173,22 @@ class FakeTaskClient:
                 [sys.executable, str(root / "scripts" / "summarize.py"),
                  str(root / "ledger.csv")],
                 capture_output=True, text=True, timeout=10)
-            return proc.stdout.strip()
+            return proc.stdout.strip() + self.probe_extra
 
     def get_task(self, task_id):
         return self.tasks[task_id]
 
+    def get_task_files(self, task_id, path=""):
+        if self.endpoint_error:
+            raise RuntimeError("409 files unavailable")
+        if not path:
+            return {"entries": [
+                {"path": p, "size": len(c.encode("utf-8"))}
+                for p, c in sorted(self.endpoint_files.items())]}
+        if path in self.endpoint_files:
+            return {"content": self.endpoint_files[path]}
+        raise RuntimeError("no such file")
+
     def get_trace_text(self, task_id):
         return "[ACTION] fake probe\n[OBSERVATION]\n" + self.tasks[task_id].result_text
 
@@ -192,6 +213,22 @@ class TestParseHelpers(unittest.TestCase):
         self.assertIsNone(parse_choice("no choice here"))
         self.assertEqual(parse_why("<why>because</why>"), "because")
         self.assertEqual(parse_notes("<notes>did x</notes>"), "did x")
+
+    def test_parse_deliverable_files_lenient(self):
+        text = ("=== DELIVERABLE FILE: a.md ===\nhello\n"
+                "=== END DELIVERABLE FILE ===\n"
+                "=== DELIVERABLE FILE: ../evil ===\nx\n"
+                "=== END DELIVERABLE FILE ===\n"
+                "=== DELIVERABLE FILE: b.md ===\nunterminated\n")
+        self.assertEqual(parse_deliverable_files(text), {"a.md": "hello\n"})
+        many = "".join(
+            f"=== DELIVERABLE FILE: f{i}.md ===\nx\n"
+            "=== END DELIVERABLE FILE ===\n" for i in range(10))
+        self.assertEqual(len(parse_deliverable_files(many)), 8)
+        big = ("=== DELIVERABLE FILE: big.md ===\n" + "x" * 70000 + "\n"
+               "=== END DELIVERABLE FILE ===\n")
+        content = parse_deliverable_files(big)["big.md"]
+        self.assertTrue(content.endswith("...[truncated]"))
         exp = parse_experience(
             "<experience>\ncue: c\nlesson: l\nevidence: e\nconfidence: high\n</experience>")
         self.assertEqual((exp["cue"], exp["lesson"], exp["evidence"]), ("c", "l", "e"))
@@ -238,6 +275,109 @@ class TestFullRunViaTasks(unittest.TestCase):
         self.assertIsNotNone(result.library_entry)
         self.assertTrue((root / "library" / "accounting" / "INDEX.md").is_file())
 
+    def _models_by_role(self, client):
+        roles = {}
+        for key, tid in client.by_key.items():
+            for role in ("probe", "edit", "select", "reflect"):
+                if f"-{role}-" in key or key.endswith(f"-{role}"):
+                    roles.setdefault(role, set()).add(client.task_models[tid])
+        return roles
+
+    def test_role_models_land_on_each_role(self):
+        client = FakeTaskClient()
+        tmp = tempfile.TemporaryDirectory()
+        self.addCleanup(tmp.cleanup)
+        root = Path(tmp.name)
+        result = run_improvement_via_tasks(
+            client, project_id="p1", task_id="ledger-1",
+            prompt="Sum every signed amount in ledger.csv. Return Total: with two decimals.",
+            base_harness=dict(BASE_TREE), run_tag="roles1",
+            work_dir=root / "work", domain="accounting",
+            source_files={"ledger.csv": LEDGER},
+            k=2, mode="parallel", library_root=root / "library",
+            model="shared-model",
+            role_models={"probe": "probe-model", "editor": "editor-model",
+                         "selector": "selector-model",
+                         "reflector": "reflector-model"})
+        self.assertTrue(result.ok)
+        roles = self._models_by_role(client)
+        self.assertEqual(roles["probe"], {"probe-model"})
+        self.assertEqual(roles["edit"], {"editor-model"})
+        self.assertEqual(roles["select"], {"selector-model"})
+        self.assertEqual(roles["reflect"], {"reflector-model"})
+
+    def test_role_model_falls_back_to_shared_then_default(self):
+        client = FakeTaskClient()
+        tmp = tempfile.TemporaryDirectory()
+        self.addCleanup(tmp.cleanup)
+        root = Path(tmp.name)
+        result = run_improvement_via_tasks(
+            client, project_id="p1", task_id="ledger-1",
+            prompt="Sum every signed amount in ledger.csv. Return Total: with two decimals.",
+            base_harness=dict(BASE_TREE), run_tag="roles2",
+            work_dir=root / "work", domain="accounting",
+            source_files={"ledger.csv": LEDGER},
+            k=1, mode="parallel", library_root=root / "library",
+            model="shared-model",
+            role_models={"editor": "editor-model"})
+        self.assertTrue(result.ok)
+        roles = self._models_by_role(client)
+        self.assertEqual(roles["probe"], {"shared-model"})
+        self.assertEqual(roles["edit"], {"editor-model"})
+        self.assertEqual(roles["reflect"], {"shared-model"})
+        # 两者都空时不传模型（服务端默认），假客户端记为 None。
+        client2 = FakeTaskClient()
+        result2, _, _ = self._run(client2, run_tag="roles3")
+        self.assertTrue(result2.ok)
+        self.assertEqual(set(client2.task_models.values()), {None})
+
+    def test_probe_deliverable_files_from_endpoint(self):
+        client = FakeTaskClient(endpoint_files={
+            "report.md": "# R\n",
+            "ledger.csv": "amount\n999.00\n",
+            "AGENTS.md": BASE_TREE["AGENTS.md"],
+        })
+        result, _, root = self._run(client, run_tag="assets1")
+        self.assertTrue(result.ok)
+        agg = result.mode_result.rollouts["agg1"]
+        # 新增文件与改动过的 seed 文件进产物；内容与 seed 一致的不进。
+        self.assertEqual(agg.deliverable_files,
+                         {"report.md": "# R\n",
+                          "ledger.csv": "amount\n999.00\n"})
+        found = [p for p in root.rglob("report.md")
+                 if "deliverables/agg1/files" in str(p)]
+        self.assertTrue(found)
+        self.assertEqual(found[0].read_text(encoding="utf-8"), "# R\n")
+
+    def test_probe_deliverable_files_from_reply_blocks(self):
+        block = ("\n=== DELIVERABLE FILE: page.html ===\n<h1>hi</h1>\n"
+                 "=== END DELIVERABLE FILE ===\n")
+        client = FakeTaskClient(endpoint_error=True, probe_extra=block)
+        result, _, _ = self._run(client, run_tag="assets2")
+        self.assertTrue(result.ok)
+        self.assertEqual(
+            result.mode_result.rollouts["agg1"].deliverable_files,
+            {"page.html": "<h1>hi</h1>\n"})
+
+    def test_probe_deliverable_files_endpoint_wins_on_collision(self):
+        block = ("\n=== DELIVERABLE FILE: page.html ===\nfrom-reply\n"
+                 "=== END DELIVERABLE FILE ===\n")
+        client = FakeTaskClient(
+            endpoint_files={"page.html": "from-endpoint\n"},
+            probe_extra=block)
+        result, _, _ = self._run(client, run_tag="assets3")
+        self.assertTrue(result.ok)
+        self.assertEqual(
+            result.mode_result.rollouts["agg1"].deliverable_files,
+            {"page.html": "from-endpoint\n"})
+
+    def test_probe_deliverable_files_endpoint_error_is_not_fatal(self):
+        client = FakeTaskClient(endpoint_error=True)
+        result, _, _ = self._run(client, run_tag="assets4")
+        self.assertTrue(result.ok)
+        self.assertEqual(
+            result.mode_result.rollouts["agg1"].deliverable_files, {})
+
     def test_same_run_tag_replays_without_new_tasks(self):
         client = FakeTaskClient()
         result1, _, _ = self._run(client, work_sub="work1")
@@ -563,5 +703,56 @@ class TestHttpClientRetry(unittest.TestCase):
         self.assertEqual(counts["n"], 1)
 
 
+class TestSweep(unittest.TestCase):
+    def _spec(self, tag, **over):
+        spec = {
+            "run_tag": tag, "task_id": "ledger-1",
+            "prompt": "Sum every signed amount in ledger.csv. Return Total: with two decimals.",
+            "base_harness": dict(BASE_TREE),
+            "source_files": {"ledger.csv": LEDGER},
+            "k": 1, "mode": "parallel", "domain": "accounting",
+        }
+        spec.update(over)
+        return spec
+
+    def test_sweep_isolates_failures_and_shares_library(self):
+        client = FakeTaskClient(fail_tags=("bad1",))
+        tmp = tempfile.TemporaryDirectory()
+        self.addCleanup(tmp.cleanup)
+        root = Path(tmp.name)
+        report = run_sweep(
+            client, project_id="p1",
+            runs=[self._spec("good1"), self._spec("bad1"),
+                  self._spec("badk", k=99), self._spec("good2")],
+            work_root=root / "sweep", library_root=root / "library")
+        by_tag = {e["run_tag"]: e for e in report["runs"]}
+        self.assertTrue(by_tag["good1"]["ok"])
+        self.assertEqual(by_tag["good1"]["chosen_label"], "agg1")
+        # bad1：全部任务注入失败 → 轮次 ok=False，但不中断后续轮。
+        self.assertFalse(by_tag["bad1"]["ok"])
+        # badk：k 越界在方法层抛错 → 记入 error，同样不中断。
+        self.assertFalse(by_tag["badk"]["ok"])
+        self.assertIn("ValueError", by_tag["badk"]["error"])
+        self.assertTrue(by_tag["good2"]["ok"])
+        self.assertEqual(report["totals"]["n_runs"], 4)
+        self.assertEqual(report["totals"]["n_ok"], 2)
+        # 共享库：两轮好结果都入库，同一 INDEX 可见。
+        self.assertIsNotNone(by_tag["good1"]["library_entry"])
+        self.assertIsNotNone(by_tag["good2"]["library_entry"])
+        self.assertTrue((root / "library" / "accounting" / "INDEX.md").is_file())
+
+    def test_sweep_requires_run_tag(self):
+        client = FakeTaskClient()
+        tmp = tempfile.TemporaryDirectory()
+        self.addCleanup(tmp.cleanup)
+        root = Path(tmp.name)
+        spec = self._spec("good1")
+        del spec["run_tag"]
+        report = run_sweep(client, project_id="p1", runs=[spec],
+                           work_root=root / "sweep")
+        self.assertFalse(report["runs"][0]["ok"])
+        self.assertIn("run_tag", report["runs"][0]["error"])
+
+
 if __name__ == "__main__":
     unittest.main()

cloud/tests/test_server_task_files.py

@@ -0,0 +1,165 @@
+"""GET /tasks/{id}/files 端点测试：本地 worktree 文件读取。
+
+只用标准库 http.client，不跑真实模型，不联网。
+任务行直接插库（不走 POST /tasks，避免 worker 派发），
+worktree 用 server.host_worktrees_dir 注入到临时目录。
+"""
+
+import http.client
+import json
+import os
+import sqlite3
+import sys
+import tempfile
+import threading
+import unittest
+from pathlib import Path
+from urllib.parse import quote
+
+try:
+    from cloud.auth import AuthStore
+    from cloud.server import create_server
+except ImportError:
+    sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
+    from cloud.auth import AuthStore
+    from cloud.server import create_server
+
+
+class TaskFilesTestCase(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.db_path = os.path.join(self.tmp.name, "control.db")
+        self.auth = AuthStore(self.db_path)
+        self.owner_id, self.owner_key = self.auth.create_user("owner")
+        self.viewer_id, self.viewer_key = self.auth.create_user("viewer")
+        self.outsider_id, self.outsider_key = self.auth.create_user("outsider")
+        self.server = create_server(self.db_path, "127.0.0.1", 0)
+        self.worktrees = os.path.join(self.tmp.name, "host-worktrees")
+        os.makedirs(self.worktrees)
+        self.server.host_worktrees_dir = self.worktrees
+        self.port = self.server.server_address[1]
+        self.thread = threading.Thread(
+            target=self.server.serve_forever, daemon=True)
+        self.thread.start()
+        status, self.project = self._req(
+            "POST", "/projects", {"name": "demo"}, self.owner_key)
+        self.assertEqual(status, 201)
+        self.auth.set_member(
+            self.project["id"], self.owner_id, self.viewer_id, "viewer")
+        self._task("task-files-1", executor=None)
+        wt = os.path.join(self.worktrees, "task-files-1")
+        os.makedirs(os.path.join(wt, "sub"))
+        Path(wt, "report.md").write_text("# Report\n", encoding="utf-8")
+        Path(wt, "sub", "page.html").write_text("<h1>hi</h1>\n",
+                                                encoding="utf-8")
+        # 非法 UTF-8 字节（\x00 是合法 UTF-8，不能用来测二进制拒绝）。
+        Path(wt, "blob.bin").write_bytes(b"\x89PNG\r\n\x1a\n\xff\xfe")
+        Path(wt, "big.txt").write_text("x" * (256 * 1024 + 1),
+                                       encoding="utf-8")
+        os.symlink(os.path.join(self.tmp.name, "outside.txt"),
+                   os.path.join(wt, "escape"))
+        Path(self.tmp.name, "outside.txt").write_text("secret\n",
+                                                       encoding="utf-8")
+        self._task("task-docker-1", executor="docker")
+        self._task("task-nowt-1", executor=None)
+
+    def tearDown(self):
+        self.server.shutdown()
+        self.server.server_close()
+        self.thread.join(timeout=5)
+        try:
+            self.auth.conn.close()
+        except Exception:
+            pass
+        self.tmp.cleanup()
+
+    def _task(self, task_id, executor):
+        conn = sqlite3.connect(self.db_path)
+        try:
+            conn.execute(
+                "INSERT INTO tasks(id, project_id, prompt, status, executor)"
+                " VALUES(?, ?, ?, ?, ?)",
+                (task_id, self.project["id"], "do it", "completed", executor))
+            conn.commit()
+        finally:
+            conn.close()
+
+    def _req(self, method, path, body=None, key=None):
+        conn = http.client.HTTPConnection("127.0.0.1", self.port, timeout=10)
+        hdrs = {}
+        payload = None
+        if body is not None:
+            payload = json.dumps(body).encode("utf-8")
+            hdrs["Content-Type"] = "application/json"
+        if key is not None:
+            hdrs["X-API-Key"] = key
+        try:
+            conn.request(method, path, body=payload, headers=hdrs)
+            resp = conn.getresponse()
+            data = resp.read()
+        finally:
+            conn.close()
+        try:
+            parsed = json.loads(data.decode("utf-8")) if data else None
+        except ValueError:
+            parsed = None
+        return resp.status, parsed
+
+    def _files(self, task_id, path=None, key=None):
+        url = f"/tasks/{task_id}/files"
+        if path is not None:
+            url += "?path=" + quote(path)
+        return self._req("GET", url, key=self.owner_key if key is None else key)
+
+    def test_list_top_level_recursive(self):
+        status, body = self._files("task-files-1")
+        self.assertEqual(status, 200, body)
+        paths = sorted(e["path"] for e in body["entries"])
+        self.assertIn("report.md", paths)
+        self.assertIn("sub/page.html", paths)
+        self.assertNotIn("escape", paths)  # 符号链接不跟进列表
+
+    def test_read_file_and_subdir_listing(self):
+        status, body = self._files("task-files-1", "report.md")
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["content"], "# Report\n")
+        status, body = self._files("task-files-1", "sub")
+        self.assertEqual(status, 200, body)
+        self.assertEqual([e["path"] for e in body["entries"]],
+                         ["sub/page.html"])
+
+    def test_path_escape_blocked(self):
+        for bad in ("../outside.txt", "/etc/passwd", "escape", "sub/../../x"):
+            status, body = self._files("task-files-1", bad)
+            self.assertIn(status, (400, 404), (bad, body))
+        # escape 是指到 worktree 外的符号链接：必须 400，不是读到 secret。
+        status, body = self._files("task-files-1", "escape")
+        self.assertEqual(status, 400, body)
+
+    def test_missing_file_and_missing_worktree(self):
+        status, _ = self._files("task-files-1", "nope.md")
+        self.assertEqual(status, 404)
+        status, _ = self._files("task-nowt-1")
+        self.assertEqual(status, 404)
+        status, _ = self._files("task-unknown")
+        self.assertEqual(status, 404)
+
+    def test_binary_and_oversized(self):
+        status, body = self._files("task-files-1", "blob.bin")
+        self.assertEqual(status, 409, body)
+        status, body = self._files("task-files-1", "big.txt")
+        self.assertEqual(status, 413, body)
+
+    def test_docker_task_409_and_auth(self):
+        status, body = self._files("task-docker-1")
+        self.assertEqual(status, 409, body)
+        status, _ = self._req("GET", "/tasks/task-files-1/files")
+        self.assertEqual(status, 401)
+        status, _ = self._files("task-files-1", key=self.outsider_key)
+        self.assertEqual(status, 403)
+        status, _ = self._files("task-files-1", key=self.viewer_key)
+        self.assertEqual(status, 200)
+
+
+if __name__ == "__main__":
+    unittest.main()