harness2: recursive harness optimization wired to real codex tasks

Port Google Research's Harness^2 method layer (model weights frozen,
only the surrounding harness is optimized): Propose-Probe-Compose with
INV/DOM/AGG editors, parallel/sequential modes, per-domain experience
library, anonymous-product selection, k<=3 hard cap and budget gate.
Execution goes only through the existing control-plane API (upload ->
task with idempotency key -> poll -> trace/cost -> cancel on timeout);
no new server endpoints, no codex flags composed, tasks inherit the
normal executor/sandbox/quota/audit path.

Wiring lessons from four real rounds against production (project
wwww, ledger task): HttpTaskClient retries transient disconnects and
502/503/429 (3 attempts); editors get a worked file-block example, an
explicit <no-change> marker and one feedback retry (.fmt2) instead of
silently accepting notes-only replies; run_task unwraps the server's
result_text JSON envelope (text field holds the real reply) — without
this, line-anchored file-block parsing could never see blocks the
agents were actually emitting. Final round (real4) fully green:
all stages landed, selection non-fallback, experience入库, probe
delivers the correct signed total 85.15.

Tests: 35 harness tests (17 method + 18 wiring incl. fake-control-plane
full round, idempotent replay, retry, format retry, envelope round).
Deliberate deviations from upstream are documented in docs/harness2.md
(no network-unrestricted/sandbox-disabled defaults, auto_apply off,
no benchmark/scorer, probe deliverable = final reply text).

cloud/harness2_tasks.py

@@ -0,0 +1,754 @@
+"""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：
+  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 /
+                          merged / merge-failed）
+  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）
+  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。
+
+三类任务是怎么跑的：
+- 探针（CodexProbeAdapter.execute）：bundle = 源文件 + harness 树。
+  Worker 派发时把 bundle 经 host.py --seed-dir 拷进 worktree，harness
+  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。
+- 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
+  文件块只输出改动的文件（没改的省略即继承），外加 <notes> 块。
+  接线层解析文件块、落到该阶段的 fallback 树上，再走核心校验。
+  最终回复里既没有文件块也没有 <notes>，或文件块路径越界、未
+  闭合，都算该阶段失败：核心循环给一次修正机会，再失败回落到
+  fallback 树——绝不把“没看懂输出格式”静默当成“无需改动”。
+- 选择（CodexSelector.select）与反思（CodexReflector.reflect）：同款
+  bundle + 最终回复解析（```choice + <why>；<experience> + <edits>），
+  不读 worktree 文件。选择失败时核心循环回落到已交付终版；反思
+  三项不全时返回 {}，不入库。
+
+幂等与断点续跑：每个任务都带 idempotency_key =
+f"h2-{run_tag}-{role}-{label}"。同一 run_tag 重跑时服务端直接返回
+已有任务（200 而非 201），已终态的任务直接复用结果——这是断点
+续跑语义，不是 bug；想从头重跑必须换 run_tag。注意：换了 harness
+内容但沿用旧 run_tag 时，旧任务结果会被复用，改进轮次就白跑了，
+所以 run_tag 必须随“任务 + harness 基线 + 参数”一起变，调用方
+（见 examples/harness2_run.py）默认用时间戳生成，显式 --run-tag
+只用于“接着上次断点跑”。
+
+成本与审批：本层不拼任何 codex 参数、不改任务默认值。model /
+effort / approval / environment_id / executor 全部由调用方透传，
+缺省时走 Worker 现有默认值；探针与编辑任务走与普通任务完全相同
+的执行器、沙箱、租户隔离、配额门与审计。长任务走 #133 的自动续
+跑；接线层单任务超时默认 3600 秒，超时先取消任务再记失败（探针）
+或抛错走回落（编辑/选择/反思），不留任务在后台继续烧钱。
+
+本层是同步阻塞的：improve_task 每一步都要等任务终态。k=1
+parallel 一轮 = 1 base + 2 候选 + 1 终版探针 + 3 编辑 + 1 选择 +
+1 反思 ≈ 9 个任务串行等待，每个都可能几分钟——这是方法本身的
+成本（上游同款：4k 次执行 + 3k 次改进会话），不是本层低效。
+"""
+
+from __future__ import annotations
+
+import json
+import time
+import urllib.error
+import urllib.request
+import uuid
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Protocol
+
+from .harness2 import (
+    EditResult,
+    Rollout,
+    Tree,
+    improve_task,
+    parse_choice,
+    parse_experience,
+    parse_harness_files,
+    parse_no_change,
+    parse_notes,
+    parse_why,
+    read_tree,
+)
+
+TERMINAL_STATUSES = frozenset(
+    {"completed", "failed", "interrupted", "merged", "merge-failed"}
+)
+DEFAULT_TASK_TIMEOUT_SECONDS = 3600
+DEFAULT_POLL_SECONDS = 10
+TRACE_MAX_CHARS = 50000
+
+
+def unwrap_result_text(raw: str) -> str:
+    """把服务端 result_text 的 JSON 信封拆开，只取 agent 的最终回复。
+
+    2026-10-29 诊断的真根因：控制面返回的 result_text 不是 agent
+    原文，而是一层 JSON 信封（task / thread_id / cwd / status /
+    text / usage …），text 字段里才是最终回复。接线层此前原样把
+    信封当回复用：notes / choice / experience 的正则不锚定行首，
+    在信封里照样命中，所以三轮里只有行首锚定的文件块解析一直
+    看不见块——编辑明明发了块，却被判成 notes-only、无改动、
+    格式错误。拆信封放在 run_task 这一个 choke point：探针、
+    编辑、选择、反思四路共用它，一处拆全受益；拆不开（非
+    JSON、无 text 字段、text 非字符串）时原样返回，不丢内容。
+    """
+    if not raw or not raw.lstrip().startswith("{"):
+        return raw
+    try:
+        obj = json.loads(raw)
+    except ValueError:
+        return raw
+    if isinstance(obj, dict) and isinstance(obj.get("text"), str):
+        return obj["text"]
+    return raw
+
+_PROBE_DELIVERABLE_FOOTER = """
+
+[DELIVERABLE — READ THIS BEFORE YOU FINISH]
+Your final reply IS the deliverable that will be judged. When you finish,
+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.
+"""
+
+_EDITOR_OUTPUT_FORMAT = """
+
+# OUTPUT FORMAT — YOU MUST FOLLOW THIS EXACTLY
+Edit the files under context/harness/ in the worktree. Then, in your final
+reply, output ONLY the files you changed, one block per file, in this format
+(files you did not change are inherited automatically — do NOT re-emit them):
+
+=== HARNESS FILE: <relative path on the editable surface> ===
+<the file's COMPLETE new content>
+=== END HARNESS FILE ===
+
+Example (a one-line file changed in full — your blocks must contain the
+COMPLETE new content, not a diff and not a description of the change):
+
+=== HARNESS FILE: scripts/summarize.py ===
+print("Total: 85.15")
+=== END HARNESS FILE ===
+
+Then a <notes>...</notes> block: what you changed, and the exact trajectory
+moment that justifies each change.
+
+HARD RULES — READ CAREFULLY, THEY ARE MACHINE-CHECKED:
+1. A change described only in <notes> is NOT a change. If you changed a
+   file, its COMPLETE new content MUST appear in a file block above, or the
+   change is discarded and this attempt counts as FAILED.
+2. If nothing in the evidence justifies a change, you MUST emit an explicit
+   <no-change>...</no-change> block with the reason, and no file blocks.
+   Saying "no change needed" only in <notes> is NOT enough — without the
+   marker the attempt counts as FAILED.
+3. File blocks and a <no-change> marker in the same reply contradict each
+   other and the attempt counts as FAILED.
+"""
+
+_SELECTOR_OUTPUT_FORMAT = """
+
+# OUTPUT FORMAT — YOU MUST FOLLOW THIS EXACTLY
+Reply with a ```choice block containing only the chosen candidate id, and a
+<why>...</why> block explaining the decision in one paragraph.
+"""
+
+_REFLECTOR_OUTPUT_FORMAT = """
+
+# EVIDENCE NOTE
+context/harness/ is the adopted harness; context/base_harness/ is the
+unmodified base harness; context/pool/<label>/ holds each candidate harness
+plus its rollout. Compare them before writing.
+
+# OUTPUT FORMAT — YOU MUST FOLLOW THIS EXACTLY
+Reply with an <edits>...</edits> block (one row per changed component, with
+its observed effect and a verdict), then an <experience>...</experience>
+block with exactly these lines: `cue: ...`, `lesson: ...`, `evidence: ...`,
+`confidence: ...`. If no lesson is supported by an observed contrast, reply
+with the single line NO_ENTRY instead of inventing one.
+"""
+
+
+# ---------------------------------------------------------------------------
+# 客户端协议与 HTTP 实现
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class TaskState:
+    task_id: str
+    status: str
+    result_text: str = ""
+    error_text: str = ""
+
+
+class TaskClient(Protocol):
+    """控制面客户端协议：HTTP 实现见 HttpTaskClient，测试用假实现。"""
+
+    def upload(self, files: dict[str, bytes]) -> str:
+        """上传一包文件（相对路径 -> 内容），返回 upload_id。"""
+        ...
+
+    def create_task(self, *, project_id: str, prompt: str, upload_id: str | None = None,
+                    idempotency_key: str | None = None, model: str | None = None,
+                    effort: str | None = None, approval: str | None = None,
+                    environment_id: str | None = None, executor: str | None = None,
+                    no_time_limit: bool = False) -> str:
+        """建任务，返回任务 id；同幂等键返回已有任务 id，不建重复。"""
+        ...
+
+    def get_task(self, task_id: str) -> TaskState:
+        ...
+
+    def get_trace_text(self, task_id: str) -> str:
+        """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_cost_usd(self, task_id: str) -> float:
+        """有价成本 USD；读不到时返回 0.0，不炸轮次。"""
+        ...
+
+    def cancel_task(self, task_id: str) -> None:
+        ...
+
+
+def build_multipart(files: dict[str, bytes]) -> tuple[bytes, str]:
+    """按服务端 cloud/uploads.py 的解析器能读懂的格式编码 multipart。
+
+    paths 字段是 JSON 字符串数组，与 file part 一一对应；part 名用
+    file（服务端只按顺序取 files，不认 part 名）。编码结果可用
+    cloud.uploads.parse_multipart + parse_paths_field 原样解回，
+    见 test_harness2_tasks.py 的往返测试。
+    """
+    boundary = "----harness2-" + uuid.uuid4().hex
+    paths = sorted(files)
+    parts: list[bytes] = []
+    parts.append(
+        b"--" + boundary.encode() + b"\r\n"
+        b'Content-Disposition: form-data; name="paths"\r\n\r\n'
+        + json.dumps(paths).encode() + b"\r\n"
+    )
+    for path in paths:
+        parts.append(
+            b"--" + boundary.encode() + b"\r\n"
+            + f'Content-Disposition: form-data; name="file"; filename="{path}"\r\n'.encode()
+            + b"Content-Type: application/octet-stream\r\n\r\n"
+            + files[path] + b"\r\n"
+        )
+    parts.append(b"--" + boundary.encode() + b"--\r\n")
+    return b"".join(parts), f"multipart/form-data; boundary={boundary}"
+
+
+class HttpTaskClient:
+    """基于标准库 urllib 的控制面客户端（零第三方依赖，与项目风格一致）。
+
+    瞬断重试（2026-10-29 真实首跑教训）：VM 到公网控制面的连接会
+    偶发中途断开（RemoteDisconnected / IncompleteRead），一次断开
+    就让整轮提前收尾，尽管服务端任务还在跑。_request 对传输错误与
+    502 / 503 / 429 最多重试 3 次（退避 1s、3s），其它 HTTP 错误
+    （401 / 403 / 400 等服务端裁决）不重试。重试安全的前提是调用
+    方都带 idempotency_key（本层所有建任务调用都带）与“上传包
+    未绑定即孤儿包”的语义：重试建任务时服务端返回已有任务，不
+    重复建；重试上传时最多多出一个 24 小时后过期的孤儿包。
+    """
+
+    def __init__(self, base_url: str, api_key: str, *, timeout: int = 60,
+                 max_attempts: int = 3):
+        self.base_url = base_url.rstrip("/")
+        self.api_key = api_key
+        self.timeout = timeout
+        self.max_attempts = max(1, max_attempts)
+
+    def _request(self, method: str, path: str, body: bytes | None = None,
+                 content_type: str | None = None) -> dict:
+        req = urllib.request.Request(
+            self.base_url + path, data=body, method=method)
+        req.add_header("X-API-Key", self.api_key)
+        if content_type:
+            req.add_header("Content-Type", content_type)
+        elif body is not None:
+            req.add_header("Content-Type", "application/json")
+        last_exc: Exception | None = None
+        for attempt in range(1, self.max_attempts + 1):
+            try:
+                with urllib.request.urlopen(req, timeout=self.timeout) as resp:
+                    raw = resp.read()
+            except urllib.error.HTTPError as exc:
+                if exc.code in (502, 503, 429) and attempt < self.max_attempts:
+                    last_exc = exc
+                    time.sleep(1 if attempt == 1 else 3)
+                    continue
+                detail = exc.read().decode("utf-8", errors="replace")[:500]
+                raise RuntimeError(
+                    f"{method} {path} -> HTTP {exc.code}: {detail}") from exc
+            except (urllib.error.URLError, ConnectionError, TimeoutError,
+                    OSError) as exc:
+                # RemoteDisconnected / IncompleteRead 都落在这里（前者是
+                # ConnectionResetError / http.client 子类，后者是
+                # http.client.HTTPException，而 HTTPException 不是
+                # OSError——见下方 except。
+                if attempt < self.max_attempts:
+                    last_exc = exc
+                    time.sleep(1 if attempt == 1 else 3)
+                    continue
+                raise RuntimeError(
+                    f"{method} {path} -> transport error: {exc}") from exc
+            except Exception as exc:  # http.client.HTTPException 等非 OSError
+                if exc.__class__.__name__ in (
+                        "RemoteDisconnected", "IncompleteRead", "BadStatusLine",
+                        "HTTPException") and attempt < self.max_attempts:
+                    last_exc = exc
+                    time.sleep(1 if attempt == 1 else 3)
+                    continue
+                raise
+            break
+        else:
+            raise RuntimeError(
+                f"{method} {path} -> transport error: {last_exc}")
+        try:
+            obj = json.loads(raw.decode("utf-8"))
+        except ValueError as exc:
+            raise RuntimeError(f"{method} {path} -> non-JSON response") from exc
+        if not isinstance(obj, dict):
+            raise RuntimeError(f"{method} {path} -> unexpected response shape")
+        return obj
+
+    def upload(self, files: dict[str, bytes]) -> str:
+        if not files:
+            raise ValueError("upload requires at least one file")
+        body, content_type = build_multipart(files)
+        obj = self._request("POST", "/uploads", body=body, content_type=content_type)
+        upload_id = obj.get("upload_id")
+        if not upload_id:
+            raise RuntimeError(f"/uploads did not return an upload_id: {obj!r}"[:300])
+        return str(upload_id)
+
+    def create_task(self, *, project_id: str, prompt: str, upload_id: str | None = None,
+                    idempotency_key: str | None = None, model: str | None = None,
+                    effort: str | None = None, approval: str | None = None,
+                    environment_id: str | None = None, executor: str | None = None,
+                    no_time_limit: bool = False) -> str:
+        payload: dict = {"project_id": project_id, "prompt": prompt}
+        if upload_id is not None:
+            payload["upload_id"] = upload_id
+        if idempotency_key is not None:
+            payload["idempotency_key"] = idempotency_key
+        if model is not None:
+            payload["model"] = model
+        if effort is not None:
+            payload["effort"] = effort
+        if approval is not None:
+            payload["approval"] = approval
+        if environment_id is not None:
+            payload["environment_id"] = environment_id
+        if executor is not None:
+            payload["executor"] = executor
+        if no_time_limit:
+            payload["no_time_limit"] = True
+        obj = self._request("POST", "/tasks", body=json.dumps(payload).encode())
+        task_id = obj.get("id")
+        if not task_id:
+            raise RuntimeError(f"/tasks did not return an id: {obj!r}"[:300])
+        return str(task_id)
+
+    def get_task(self, task_id: str) -> TaskState:
+        obj = self._request("GET", f"/tasks/{task_id}")
+        return TaskState(
+            task_id=task_id,
+            status=str(obj.get("status") or ""),
+            result_text=str(obj.get("result_text") or ""),
+            error_text=str(obj.get("error_text") or ""),
+        )
+
+    def get_trace_text(self, task_id: str) -> str:
+        try:
+            obj = self._request("GET", f"/tasks/{task_id}/trace")
+        except Exception:
+            return ""
+        entries = obj.get("entries") or []
+        lines: list[str] = []
+        for entry in entries:
+            if not isinstance(entry, dict):
+                continue
+            kind = entry.get("kind") or entry.get("type") or "entry"
+            text = entry.get("text") or entry.get("message") or ""
+            if not text:
+                text = json.dumps(entry, ensure_ascii=False)[:1000]
+            lines.append(f"[{kind}] {text}")
+        rendered = "\n".join(lines).strip()
+        return rendered[:TRACE_MAX_CHARS] if rendered else ""
+
+    def get_cost_usd(self, task_id: str) -> float:
+        try:
+            obj = self._request("GET", f"/tasks/{task_id}/cost")
+        except Exception:
+            return 0.0
+        try:
+            return float(obj.get("usd_micros") or 0) / 1_000_000
+        except (TypeError, ValueError):
+            return 0.0
+
+    def cancel_task(self, task_id: str) -> None:
+        try:
+            self._request("POST", f"/tasks/{task_id}/cancel", body=b"{}")
+        except Exception:
+            pass  # 终态任务取消会 400，本来就不用再取消；其它错误也不炸轮次
+
+
+@dataclass
+class CompletedTask:
+    task_id: str
+    state: TaskState
+    cost_usd: float = 0.0
+    trace_text: str = ""
+
+
+def run_task(
+    client: TaskClient,
+    *,
+    project_id: str,
+    prompt: str,
+    files: dict[str, bytes] | None = None,
+    idempotency_key: str | None = None,
+    model: str | None = None,
+    effort: str | None = None,
+    approval: str | None = None,
+    environment_id: str | None = None,
+    executor: str | None = None,
+    no_time_limit: bool = False,
+    timeout_seconds: int = DEFAULT_TASK_TIMEOUT_SECONDS,
+    poll_seconds: int = DEFAULT_POLL_SECONDS,
+) -> CompletedTask:
+    """建任务、轮询到终态、读成本与轨迹。超时先取消再抛 TimeoutError。"""
+    upload_id = client.upload(files) if files else None
+    task_id = client.create_task(
+        project_id=project_id, prompt=prompt, upload_id=upload_id,
+        idempotency_key=idempotency_key, model=model, effort=effort,
+        approval=approval, environment_id=environment_id, executor=executor,
+        no_time_limit=no_time_limit)
+    deadline = time.time() + timeout_seconds
+    state = client.get_task(task_id)
+    while state.status not in TERMINAL_STATUSES:
+        if time.time() >= deadline:
+            client.cancel_task(task_id)
+            raise TimeoutError(
+                f"task {task_id} did not reach a terminal status within "
+                f"{timeout_seconds}s (last status: {state.status})")
+        time.sleep(poll_seconds)
+        state = client.get_task(task_id)
+    # 拆信封的唯一位置：run_task 是探针、编辑、选择、反思四路的
+    # 共同 choke point，在这里拆一次，全路受益；各 TaskClient 的
+    # get_task 保持原样返回服务端原文。
+    state = TaskState(
+        task_id=state.task_id, status=state.status,
+        result_text=unwrap_result_text(state.result_text),
+        error_text=state.error_text,
+    )
+    return CompletedTask(
+        task_id=task_id, state=state,
+        cost_usd=client.get_cost_usd(task_id),
+        trace_text=client.get_trace_text(task_id),
+    )
+
+
+def _flatten_dir(root: Path, prefix: str = "") -> dict[str, bytes]:
+    """把一个目录展平成 {相对路径: 内容}；符号链接与 __pycache__ 跳过。"""
+    out: dict[str, bytes] = {}
+    root = Path(root)
+    if not root.is_dir():
+        return out
+    for p in sorted(root.rglob("*")):
+        if not p.is_file() or p.is_symlink() or "__pycache__" in p.parts:
+            continue
+        rel = str(p.relative_to(root))
+        out[(prefix + rel) if prefix else rel] = p.read_bytes()
+    return out
+
+
+def _tree_to_files(tree: Tree) -> dict[str, bytes]:
+    return {rel: content.encode("utf-8") for rel, content in tree.items()}
+
+
+# ---------------------------------------------------------------------------
+# 三个协议的真实任务实现
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class CodexProbeAdapter:
+    """探针 adapter：每个 label 跑一个真实 codex 任务。
+
+    source_files：原任务需要的源文件（如 ledger.csv），会随 harness
+    一起 seed 进 worktree；与 harness 路径冲突时 harness 优先（探针
+    的意义就是让 harness 覆盖仓库默认值）。调用方应在 source_files
+    里避免与 harness 路径重名——重名时源文件会被静默覆盖，本层
+    不为此报错（与 seed“用户文件优先”的语义一致）。
+    """
+
+    client: TaskClient
+    project_id: str
+    task_id: str
+    prompt: str
+    base: Tree
+    run_tag: str
+    source_files: dict[str, bytes] = field(default_factory=dict)
+    environment_id: str | None = None
+    executor: str | None = None
+    model: str | None = None
+    effort: str | None = None
+    approval: str | None = None
+    no_time_limit: bool = False
+    timeout_seconds: int = DEFAULT_TASK_TIMEOUT_SECONDS
+    domain: str = "general"
+
+    def base_harness(self) -> Tree:
+        return dict(self.base)
+
+    def execute(self, label: str, harness: Tree) -> Rollout:
+        files = dict(self.source_files)
+        files.update(_tree_to_files(harness))
+        try:
+            done = run_task(
+                self.client, project_id=self.project_id,
+                prompt=self.prompt + _PROBE_DELIVERABLE_FOOTER,
+                files=files,
+                idempotency_key=f"h2-{self.run_tag}-probe-{label}",
+                model=self.model, effort=self.effort, approval=self.approval,
+                environment_id=self.environment_id, executor=self.executor,
+                no_time_limit=self.no_time_limit,
+                timeout_seconds=self.timeout_seconds)
+        except Exception as exc:
+            return Rollout(label=label, ok=False, delivered=False,
+                           error=f"{type(exc).__name__}: {exc}"[:400])
+        state = done.state
+        delivered = state.status == "completed" and bool(state.result_text.strip())
+        return Rollout(
+            label=label, ok=state.status == "completed", delivered=delivered,
+            deliverable_text=state.result_text,
+            trajectory=done.trace_text or state.result_text,
+            cost_usd=done.cost_usd,
+            error="" if state.status == "completed" else (
+                state.error_text or f"task ended as {state.status}")[:400],
+        )
+
+
+@dataclass
+class CodexEditor:
+    """编辑器：一个 codex 任务，证据 seed 进 worktree，改动以文件块回读。"""
+
+    client: TaskClient
+    project_id: str
+    run_tag: str
+    environment_id: str | None = None
+    executor: str | None = None
+    model: str | None = None
+    effort: str | None = None
+    approval: str | None = None
+    no_time_limit: bool = False
+    timeout_seconds: int = DEFAULT_TASK_TIMEOUT_SECONDS
+
+    def edit(self, mission: str, context_dir: Path, label: str) -> EditResult:
+        """编辑任务：格式错误给一次带反馈的纠正机会，再错才如实回落。
+
+        背景（2026-10-29 两轮真实轮次）：编辑在 <notes> 里声称改
+        动、却不发文件块，旧契约把 notes-only 当“无需改动”采纳，
+        改动无声丢失；且方法层的阶段循环里编辑异常是直接 break（不
+        重试），所以纠正必须发生在本层。格式重试用独立幂等键
+        （…-edit-{label}.fmt2），不复用首个任务、不重复计费到同
+        一任务；任务本身失败（非 completed）仍直接抛错、不重试。
+        """
+        files = _flatten_dir(context_dir, prefix="context/")
+        files["context/MISSION.md"] = mission.encode("utf-8")
+        last_err = ""
+        total_cost = 0.0
+        for attempt in (1, 2):
+            prompt = mission + _EDITOR_OUTPUT_FORMAT
+            if last_err:
+                prompt += (
+                    "\n\n# YOUR PREVIOUS REPLY WAS REJECTED\n\n"
+                    f"{last_err}\n\nFix the output format and reply again: "
+                    "changed files MUST appear as === HARNESS FILE === blocks "
+                    "with their COMPLETE new content, or, if truly no change "
+                    "is justified, emit <no-change>...</no-change>.")
+            done = run_task(
+                self.client, project_id=self.project_id,
+                prompt=prompt,
+                files=files,
+                idempotency_key=f"h2-{self.run_tag}-edit-{label}"
+                + ("" if attempt == 1 else ".fmt2"),
+                model=self.model, effort=self.effort, approval=self.approval,
+                environment_id=self.environment_id, executor=self.executor,
+                no_time_limit=self.no_time_limit,
+                timeout_seconds=self.timeout_seconds)
+            state = done.state
+            if state.status != "completed":
+                raise RuntimeError(
+                    f"editor task {done.task_id} ended as {state.status}: "
+                    f"{state.error_text}"[:400])
+            total_cost += done.cost_usd
+            text = state.result_text or ""
+            try:
+                changed = parse_harness_files(text)
+                notes = parse_notes(text)
+                no_change = parse_no_change(text)
+            except ValueError as exc:
+                last_err = f"output-format error: {exc}"
+                continue
+            if changed and no_change:
+                last_err = ("output-format error: reply contains both harness "
+                            "file blocks and a <no-change> marker")
+                continue
+            if not changed and not no_change:
+                last_err = ("output-format error: reply contains neither "
+                            "harness file blocks nor a <no-change> marker; "
+                            "a change described only in <notes> is discarded")
+                continue
+            base_tree = read_tree(Path(context_dir) / "harness")
+            merged = dict(base_tree)
+            merged.update(changed)
+            return EditResult(tree=merged, notes=notes, cost_usd=total_cost)
+        raise ValueError(
+            f"editor reply failed output-format validation twice: {last_err}")
+
+
+@dataclass
+class CodexSelector:
+    """选择器：一个 codex 任务，候选产物 seed 进 worktree，只解析最终回复。"""
+
+    client: TaskClient
+    project_id: str
+    run_tag: str
+    environment_id: str | None = None
+    executor: str | None = None
+    model: str | None = None
+    effort: str | None = None
+    approval: str | None = None
+    no_time_limit: bool = False
+    timeout_seconds: int = DEFAULT_TASK_TIMEOUT_SECONDS
+
+    def select(self, mission: str, context_dir: Path) -> tuple[str, str, float]:
+        files = _flatten_dir(context_dir, prefix="context/")
+        files["context/MISSION.md"] = mission.encode("utf-8")
+        done = run_task(
+            self.client, project_id=self.project_id,
+            prompt=mission + _SELECTOR_OUTPUT_FORMAT,
+            files=files,
+            idempotency_key=f"h2-{self.run_tag}-select-{Path(context_dir).name}",
+            model=self.model, effort=self.effort, approval=self.approval,
+            environment_id=self.environment_id, executor=self.executor,
+            no_time_limit=self.no_time_limit,
+            timeout_seconds=self.timeout_seconds)
+        state = done.state
+        if state.status != "completed":
+            raise RuntimeError(
+                f"selector task {done.task_id} ended as {state.status}: "
+                f"{state.error_text}"[:400])
+        choice = parse_choice(state.result_text or "")
+        if not choice:
+            raise ValueError("selector task returned no ```choice block")
+        return choice, parse_why(state.result_text or ""), done.cost_usd
+
+
+@dataclass
+class CodexReflector:
+    """反思器：一个 codex 任务，解析 <experience>；三项不全返回 {}。"""
+
+    client: TaskClient
+    project_id: str
+    run_tag: str
+    environment_id: str | None = None
+    executor: str | None = None
+    model: str | None = None
+    effort: str | None = None
+    approval: str | None = None
+    no_time_limit: bool = False
+    timeout_seconds: int = DEFAULT_TASK_TIMEOUT_SECONDS
+
+    def reflect(self, mission: str, context_dir: Path) -> dict:
+        files = _flatten_dir(context_dir, prefix="context/")
+        files["context/MISSION.md"] = mission.encode("utf-8")
+        done = run_task(
+            self.client, project_id=self.project_id,
+            prompt=mission + _REFLECTOR_OUTPUT_FORMAT,
+            files=files,
+            idempotency_key=f"h2-{self.run_tag}-reflect-{Path(context_dir).name}",
+            model=self.model, effort=self.effort, approval=self.approval,
+            environment_id=self.environment_id, executor=self.executor,
+            no_time_limit=self.no_time_limit,
+            timeout_seconds=self.timeout_seconds)
+        state = done.state
+        if state.status != "completed":
+            return {}
+        return parse_experience(state.result_text or "")
+
+
+def run_improvement_via_tasks(
+    client: TaskClient,
+    *,
+    project_id: str,
+    task_id: str,
+    prompt: str,
+    base_harness: Tree,
+    run_tag: str,
+    work_dir: Path,
+    domain: str = "general",
+    source_files: dict[str, bytes] | None = None,
+    k: int = 1,
+    mode: str = "parallel",
+    library_root: Path | None = None,
+    budget_usd: float | None = None,
+    model: str | None = None,
+    effort: str | None = None,
+    approval: str | None = None,
+    environment_id: str | None = None,
+    executor: str | None = None,
+    no_time_limit: bool = False,
+    timeout_seconds: int = DEFAULT_TASK_TIMEOUT_SECONDS,
+):
+    """一站式：用真实 codex 任务跑一轮完整改进，返回 ImprovementResult。
+
+    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）。
+    """
+    from .harness2 import Library
+
+    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,
+        no_time_limit=no_time_limit, timeout_seconds=timeout_seconds,
+        domain=domain)
+    shared = dict(environment_id=environment_id, executor=executor, model=model,
+                  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)
+    library = Library(library_root, domain) if library_root is not None else None
+    return improve_task(
+        adapter, editor, selector=selector, reflector=reflector,
+        work_dir=work_dir, k=k, mode=mode, library=library,
+        auto_apply=False, budget_usd=budget_usd)
+
+
+__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",
+]