@hunter-harness/workflow-harness 0.2.73 → 0.2.75

package/harness/bundles/java/claude-code/harness-submit/SKILL.md

@@ -21,7 +21,7 @@ disallowed-tools:
   - Bash(awk *)
   - Bash(curl *)
 ---
-<!-- generated by harness_deploy.py; core=984aee382a36595c; overlay=java; agent=claude-code; do not edit -->
+<!-- generated by harness_deploy.py; core=4f3a290efefe6a68; overlay=java; agent=claude-code; do not edit -->
 # harness-submit — 最终提交（含 worktree 合并）
 
 ## Purpose
@@ -47,6 +47,7 @@ disallowed-tools:
 
 ## 统一读取协议
 
+0. **脚本在 `<skills-root>/scripts/` 共享，不在每个 skill 子目录下** — 实际形态是 `.codebuddy/skills/scripts/harness_*.py`（`.claude`/`.cursor`/`.codex` 同理），**没有** `.../skills/harness-<phase>/scripts/`。plan、run/test、archive 三份执行日志里都先猜成后者、报 `No such file` 再靠 Search 找回来；照第一种写法直接用
 1. **`.harness/changes/<change-name>/` 是唯一真相源** — 所有输入从该目录读取，产物写入对应子目录
 2. **change-name 优先从 frontmatter 读取** — `plans/*-design.md`、`spec/*-design.md`、`plans/*-plan.md` 的 YAML `change-name`
 3. **frontmatter 缺失时兼容旧格式** — 从路径推断，标记 `🟡 legacy-plan`，不失败
@@ -72,10 +73,11 @@ worktree 合并前必须运行 `harness_change.py integration-lock acquire --run
 
 ### 提交流程（步骤 0–7）
 
-0. **启动准备** — `harness_context.py prepare --phase submit --executor <tool> [--change <id>] --json` 确定唯一变更与 executionRoot；`harness_context.py begin --phase submit --change <id> --executor <tool> --json` 校验 review→submit receipt；**`harness_gate.py begin --phase submit --change <id>`**；读 ledger，以 `harness_ledger.py diff-hash --repo <executionRoot> --base <baseCommit> --change-dir ".harness/changes/<change-name>" --json` 计算 diffHash + post-test 7 类分类（`executionRoot` 必须直接取准备回执；禁止手写 ledger / 手工 phase.end）
+0. **启动准备** — `harness_context.py prepare --project . --change <id> --phase submit --executor <tool> --json`（`--project` 必填）确定唯一变更与 executionRoot；`harness_context.py begin --project . --change <id> --phase submit --executor <tool> --json` 校验 review→submit receipt；**`harness_gate.py begin --phase submit --change <id>`**；读 ledger，以 `harness_ledger.py diff-hash --repo <executionRoot> --base <baseCommit> --change-dir ".harness/changes/<change-name>" --json` 计算 diffHash + post-test 7 类分类（`executionRoot` 必须直接取准备回执；禁止手写 ledger / 手工 phase.end）
 1. **合并最新代码** — 主目录与 worktree 均**不在业务工作区 stash/pull**；远端同步由合并段 integration transaction 在隔离 integration worktree 内完成（见「worktree 合并流程」）；**正常路径禁止 `git stash` / `stash pop`**
 2. **最终验证** — ledger 复用优先；提交前只调用 `can-reuse --project . --profile-input unitTestFull --command <profile 规范命令>`。不得传入另一套 `--files`，不得把 runner 包装说明写进 `command`。`reuse=true` 时禁止重跑；只有真实输入、依赖、工具链或环境身份变化时才执行一次同一 profile 验证，并只登记一条结果。不要读取 ledger/archive 实现源码或临时编写散列脚本排查参数。
    - 无远端 CI 且 gate-policy 未强制 remote provider 时，验证通过/复用后运行 `harness_archive.py certify-local --change-dir ... --project . --json`，从同一 ledger 生成 `local-reproducible` 产品候选收据；该命令不执行测试。
+   - **被与本变更无关的预存失败卡住时的正规出路**：用 `harness_preflight.py record-quirk --project . --action skip-not-block --pattern "<具体错误签名，≥8 字符>" --reason "<为什么与本变更无关>"` 把它声明进 build-profile 的 `knownPreexistingErrors`，certify-local 会在**该失败的 ledger 证据里确实出现该签名**时放行，并把 `{validation, pattern, reason}` 写进收据的 `verification.preexistingExemptions` 留痕。**不得**改 gate-policy 的 `candidateVerification.requiredValidations` 来绕过——那是把门禁本身拆掉。声明了但证据对不上、或签名短到能匹配一切，仍然阻断。
    - gate-policy 要求 `remote-attested` 时不得降级成本地收据，等待远端 attestation。
 3. **.gitignore + 精确暂存** ⚠️ — 检查 `.harness/` 在 `.gitignore`；**禁止 `git add -A`**。若存在 `evidence/test-tracking.json`，先执行 `python <skills-root>/scripts/harness_test_guard.py stage --project . --change-dir ".harness/changes/<change-name>" --json`；失败即硬停止。无 manifest 时不使用 `-f`。manifest 之外的文件按精确业务路径正常暂存，**禁止全局 force-add**。
 4. **提交方式** — 主目录：blocking user confirmation 三选项（commit+push / 仅本地 / 取消）；**worktree：固定仅本地 commit**

package/harness/bundles/java/claude-code/harness-sync/SKILL.md

@@ -19,7 +19,7 @@ disallowed-tools:
   - Bash(awk *)
   - Bash(curl *)
 ---
-<!-- generated by harness_deploy.py; core=984aee382a36595c; overlay=java; agent=claude-code; do not edit -->
+<!-- generated by harness_deploy.py; core=4f3a290efefe6a68; overlay=java; agent=claude-code; do not edit -->
 # harness-sync
 
 ## Purpose

package/harness/bundles/java/claude-code/harness-test/SKILL.md

@@ -23,7 +23,7 @@ disallowed-tools:
   - Bash(node *)
   - Bash(codegraph *)
 ---
-<!-- generated by harness_deploy.py; core=984aee382a36595c; overlay=java; agent=claude-code; do not edit -->
+<!-- generated by harness_deploy.py; core=4f3a290efefe6a68; overlay=java; agent=claude-code; do not edit -->
 # harness-test — 测试执行
 
 ## Purpose
@@ -51,6 +51,7 @@ disallowed-tools:
 
 ## 统一读取协议
 
+0. **脚本在 `<skills-root>/scripts/` 共享，不在每个 skill 子目录下** — 实际形态是 `.codebuddy/skills/scripts/harness_*.py`（`.claude`/`.cursor`/`.codex` 同理），**没有** `.../skills/harness-<phase>/scripts/`。plan、run/test、archive 三份执行日志里都先猜成后者、报 `No such file` 再靠 Search 找回来；照第一种写法直接用
 1. **`.harness/changes/<change-name>/` 是唯一真相源** — 所有输入从该目录读取，产物写入对应子目录
 2. **change-name 优先从 frontmatter 读取** — `plans/*-design.md`、`spec/*-design.md`、`plans/*-plan.md` 的 YAML `change-name`
 3. **frontmatter 缺失时兼容旧格式** — 从路径推断，标记 `🟡 legacy-plan`，不失败
@@ -105,7 +106,7 @@ Runner 强制同项目单实例、低调度优先级、逐命令超时、正常
 
 ### Phase 0：环境准备（主会话执行，需要交互确认）
 
-先 `harness_context.py prepare --phase test --executor <tool> [--change <id>] --json`，再 `harness_context.py begin --phase test --change <id> --executor <tool> --json` 校验最新 run→test receipt 的 artifact/hash/HEAD；然后 **`harness_gate.py begin --phase test --change <id>`**（禁止手工 phase.start / 手写 ledger）。执行各项强制环境检查 + **命令执行模式 preflight (0.1)**；只有首选执行器不可用时，才执行 fallback 执行器探测。
+先 `harness_context.py prepare --project . --change <id> --phase test --executor <tool> --json`（`--project` 必填，漏了直接 argparse 报错），再 `harness_context.py begin --project . --change <id> --phase test --executor <tool> --json` 校验最新 run→test receipt 的 artifact/hash/HEAD；然后 **`harness_gate.py begin --phase test --change <id>`**（禁止手工 phase.start / 手写 ledger）。执行各项强制环境检查 + **命令执行模式 preflight (0.1)**；只有首选执行器不可用时，才执行 fallback 执行器探测。
 
 验证写入**仅**允许 `harness_ledger.py record` / `can-reuse`；禁止 Write/Edit `verification-ledger.json`。测试跟踪：gate begin → 执行（可选 `harness_test_guard.py mark stale-test-repair`）→ 单次 `harness_gate.py close`。`--to-phase` 取实际阶段计划的后继；Fixback 返回 run，普通流程可直接进入 Review、Submit 或 Archive。Fixback 只失效与改动文件相交的验证目标。
 

package/harness/bundles/java/claude-code/scripts/harness_archive.py

@@ -991,6 +991,62 @@ def migrate_legacy_candidate_evidence(
     return receipt
 
 
+# 声明必须是具体的错误签名，不能是能匹配一切的短串。这个下限不是"安全边界"
+# ——审计线索才是——但它挡住最省事的滥用写法。
+_PREEXISTING_MIN_PATTERN = 8
+
+
+def _known_preexisting_patterns(
+    project_root: Path,
+) -> tuple[list[dict[str, str]], list[dict[str, str]]]:
+    """build-profile 里声明的预存失败签名，按"是否足够具体"分成两组。
+
+    `knownPreexistingErrors` 由 `harness_preflight.py record-quirk --action
+    skip-not-block` 写入，此前没有任何消费方——于是预存失败在 certify-local 处
+    成为死结：要么去 gate-policy 降门禁，要么顺手改范围外的产品 bug。
+    """
+    path = project_root / ".harness" / "config" / "build-profile.json"
+    if not path.is_file():
+        return [], []
+    try:
+        profile = json.loads(path.read_text(encoding="utf-8-sig"))
+    except (OSError, ValueError):
+        return [], []
+    declared = profile.get("knownPreexistingErrors") if isinstance(profile, dict) else None
+    if not isinstance(declared, list):
+        return [], []
+    usable: list[dict[str, str]] = []
+    vague: list[dict[str, str]] = []
+    for item in declared:
+        if not isinstance(item, dict) or item.get("action") != "skip-not-block":
+            continue
+        pattern = str(item.get("pattern") or "").strip()
+        reason = str(item.get("reason") or "").strip()
+        if not pattern or not reason:
+            continue
+        record = {"pattern": pattern, "reason": reason}
+        (usable if len(pattern) >= _PREEXISTING_MIN_PATTERN else vague).append(record)
+    return usable, vague
+
+
+def _match_preexisting(
+    entry: dict[str, Any], patterns: list[dict[str, str]]
+) -> dict[str, str] | None:
+    """失败证据里是否确实出现了某个已声明签名。
+
+    只看"声明过"是不够的——那等于声明一次豁免一切。必须是这一条失败的证据里
+    真的带着该签名，声明与现场才对得上。
+    """
+    evidence = entry.get("evidence")
+    haystack = evidence if isinstance(evidence, str) else json.dumps(
+        evidence, ensure_ascii=False, sort_keys=True
+    )
+    for record in patterns:
+        if record["pattern"] in haystack:
+            return record
+    return None
+
+
 def certify_local_candidate(
     change_dir: Path,
     *,
@@ -1029,14 +1085,28 @@ def certify_local_candidate(
         required = ["unitTestFull"]
     required = [str(item).strip() for item in required if str(item).strip()]
 
+    usable_patterns, vague_patterns = _known_preexisting_patterns(project_root)
+
     selected: list[dict[str, Any]] = []
+    exemptions: list[dict[str, Any]] = []
     for name in required:
         entry = validations.get(name)
         if not isinstance(entry, dict):
             raise ValueError(f"required validation missing: {name}")
         missing: list[str] = []
+        exempted = None
         if str(entry.get("status") or "").upper() != "OK":
-            missing.append("status=OK")
+            exempted = _match_preexisting(entry, usable_patterns)
+            if exempted is None:
+                vague = _match_preexisting(entry, vague_patterns)
+                if vague is not None:
+                    raise ValueError(
+                        f"required validation {name} failed and the declared "
+                        f"knownPreexistingErrors pattern {vague['pattern']!r} is too "
+                        "generic to identify it; declare the specific error signature "
+                        f"(at least {_PREEXISTING_MIN_PATTERN} characters)"
+                    )
+                missing.append("status=OK")
         for field in (
             "command",
             "evidence",
@@ -1048,6 +1118,15 @@ def certify_local_candidate(
             raise ValueError(
                 f"required validation {name} is incomplete: {', '.join(missing)}"
             )
+        if exempted is not None:
+            exemptions.append(
+                {
+                    "validation": name,
+                    "status": str(entry.get("status") or "").upper(),
+                    "pattern": exempted["pattern"],
+                    "reason": exempted["reason"],
+                }
+            )
         selected.append(
             {
                 "name": name,
@@ -1160,6 +1239,8 @@ def certify_local_candidate(
                 {str(item["inputsHash"]) for item in selected}
             ),
             "logHashes": sorted({str(item["logHash"]) for item in selected}),
+            # 豁免必须随收据一起留痕，否则事后无法把它与干净通过区分开
+            "preexistingExemptions": exemptions,
         },
     }
     if rebound_from_commit is not None:
@@ -8471,8 +8552,17 @@ def cmd_finalize(
     payload["knowledgeMaintenance"] = _knowledge_maintenance_from_archive_push(
         push_result
     )
-    if push_result.get("archiveStatus") == "durable":
-        payload["archiveDurability"] = "ARCHIVED_REMOTE_DURABLE"
+    remote_durability = remote_durable_archive_durability(
+        payload.get("archiveDurability"), push_result
+    )
+    if remote_durability is not None:
+        payload["archiveDurability"] = remote_durability
+        # 报告必须跟着改口：上传发生在 summary 落盘之后，不回写就会出现
+        # "回执 durable / 报告 local-only" 的自相矛盾
+        if not persist_archive_durability(summary_path, remote_durability):
+            warnings.append(
+                "远端归档已持久化，但 summary-data.json 的 archiveDurability 回写失败"
+            )
     if push_result.get("warning"):
         warnings.append(str(push_result["warning"]))
 
@@ -9040,6 +9130,50 @@ def build_archive_package(
     }
 
 
+def remote_durable_archive_durability(
+    current: Any, push_result: Mapping[str, Any]
+) -> dict[str, Any] | None:
+    """远端上传成功后的耐久性投影；未 durable 时返回 None。
+
+    summary-data.json 在归档流程早期就以 ARCHIVED_LOCAL_ONLY 写盘，而 ZIP 上传在
+    那之后才发生。此前上传成功只改内存里的 payload、还把对象覆盖成裸字符串，落盘
+    报告从没跟着改——于是同一次归档，回执说 durable、报告说 local-only，平台和用户
+    读到的是后者。这里保持对象形状并带上回执标识，供调用方同时更新 payload 与文件。
+    """
+    if str(push_result.get("archiveStatus") or "") != "durable":
+        return None
+    projected = dict(current) if isinstance(current, dict) else {}
+    projected.update({
+        "status": "ARCHIVED_REMOTE_DURABLE",
+        # 已远端持久化：不再声称"只存在于本地、可能随工作区丢失"
+        "risk": None,
+        "archiveId": push_result.get("archiveId"),
+        "uploadStatus": push_result.get("uploadStatus"),
+        "knowledgeStatus": push_result.get("knowledgeStatus"),
+    })
+    return projected
+
+
+def persist_archive_durability(summary_path: Path, durability: dict[str, Any]) -> bool:
+    """把耐久性结论回写到已落盘的 summary，其余字段原样保留。
+
+    summary-data.json 本就被排除在归档校验和覆盖之外（它在覆盖快照之后才写），
+    所以这次回写不会破坏 manifest 对账。
+    """
+    try:
+        summary = json.loads(Path(summary_path).read_text(encoding="utf-8-sig"))
+    except (OSError, json.JSONDecodeError):
+        return False
+    if not isinstance(summary, dict):
+        return False
+    summary["archiveDurability"] = durability
+    try:
+        write_json(Path(summary_path), summary)
+    except OSError:
+        return False
+    return True
+
+
 def auto_push_managed_snapshot(project_root: Path) -> dict[str, Any]:
     """Best-effort upload of rules and Codebase Map through the normal push protocol."""
     project_root = project_root.resolve()
@@ -9095,16 +9229,25 @@ def auto_push_managed_snapshot(project_root: Path) -> dict[str, Any]:
     except (json.JSONDecodeError, TypeError):
         output = None
     if completed.returncode != 0 or not isinstance(output, dict):
-        return {
+        # 失败时丢掉 stdout/stderr 等于把线索一起丢了：plan/spec 与规则、架构地图
+        # 全靠这条推送上平台，用户在平台上只会看到"东西没上去"，却无从查为什么。
+        # 保留有界的诊断片段（push 本身拒绝上传 secret，这里再截断兜底）。
+        detail = (completed.stderr or completed.stdout or "").strip()
+        failure: dict[str, Any] = {
             "ok": False,
             "skipped": False,
             "exitCode": completed.returncode,
             "reasonCode": "MANAGED_SNAPSHOT_UPLOAD_FAILED",
+            "detail": detail[-2_048:],
             "warning": (
-                "项目规则与架构地图未能同步到平台；归档不受影响，"
-                "稍后可运行 npx hunter-harness push --yes --non-interactive 重试。"
+                "项目规则与架构地图未能同步到平台（plan/spec 也随这条推送上传）；"
+                "归档不受影响，稍后可运行 npx hunter-harness push --yes --non-interactive 重试。"
             ),
         }
+        cli_code = output.get("code") or output.get("reason_code") if isinstance(output, dict) else None
+        if isinstance(cli_code, str) and cli_code:
+            failure["cliCode"] = cli_code
+        return failure
     summary = output.get("summary")
     submitted = summary.get("submitted", 0) if isinstance(summary, dict) else 0
     return {

package/harness/bundles/java/claude-code/scripts/harness_change.py

@@ -1081,6 +1081,90 @@ def cmd_resolve(args: argparse.Namespace) -> int:
     )
 
 
+def declare_product_ownership(
+    project_root: Path, change_id: str, *, product_paths: list[str]
+) -> dict[str, Any]:
+    """把 `ownership.productPaths` 写进 change 契约。
+
+    这个字段此前没有任何写入方：plan 的 validate_product_ownership 只校验、缺失时
+    软放行，而归档的 compute_ownership_diff 会把全部改动判成 foreignPaths，
+    filesChanged=0 直接触发 DIFF_ZERO_WITH_NONEMPTY_COMMIT。两端口径不一致，中间
+    没工具能补——只能手改契约。这里给出正规入口。
+
+    规则与 plan 校验一致：只收精确文件或目录前缀，不接受通配。
+    """
+    resolved = resolve_change(project_root, change_id)
+    if not resolved.get("ok"):
+        return resolved
+    change_dir = Path(resolved["changeDir"])
+    normalized = sorted({
+        str(item).strip().replace("\\", "/").removeprefix("./")
+        for item in product_paths
+        if isinstance(item, str) and item.strip()
+    })
+    if not normalized:
+        return {
+            "ok": False,
+            "code": "PLAN_PRODUCT_PATHS_REQUIRED",
+            "message": "至少声明一条 productPaths（精确文件或目录前缀）",
+        }
+    unsupported = sorted(p for p in normalized if any(ch in p for ch in "*?[]"))
+    if unsupported:
+        return {
+            "ok": False,
+            "code": "PLAN_PRODUCT_PATHS_GLOB_UNSUPPORTED",
+            "message": "productPaths 只接受精确文件或目录前缀，不支持通配：" + ", ".join(unsupported),
+            "unsupportedPaths": unsupported,
+        }
+
+    context_path = change_dir / CHANGE_CONTEXT_REL
+    try:
+        context = _read_json(context_path)
+    except (OSError, json.JSONDecodeError) as exc:
+        return {"ok": False, "code": "CHANGE_CONTEXT_INVALID", "message": str(exc)}
+    if not isinstance(context, dict):
+        return {"ok": False, "code": "CHANGE_CONTEXT_INVALID", "message": "change-context.json 不是对象"}
+
+    ownership = context.get("ownership")
+    ownership = dict(ownership) if isinstance(ownership, dict) else {}
+    if ownership.get("productPaths") == normalized:
+        return {
+            "ok": True,
+            "code": "PRODUCT_OWNERSHIP_DECLARED",
+            "idempotent": True,
+            "changeId": change_id,
+            "productPaths": normalized,
+            "path": str(context_path),
+        }
+    ownership["productPaths"] = normalized
+    context["ownership"] = ownership
+    _write_json(context_path, context)
+    return {
+        "ok": True,
+        "code": "PRODUCT_OWNERSHIP_DECLARED",
+        "idempotent": False,
+        "changeId": change_id,
+        "productPaths": normalized,
+        "path": str(context_path),
+    }
+
+
+def cmd_declare_ownership(args: argparse.Namespace) -> int:
+    project = resolve_main_project_root()
+    payload = declare_product_ownership(
+        project, args.change, product_paths=list(args.product_path)
+    )
+    if payload.get("ok"):
+        emit(payload, as_json=bool(args.json))
+        return 0
+    return emit_error(
+        str(payload.get("code", "PRODUCT_OWNERSHIP_FAILED")),
+        str(payload.get("message", "declare ownership failed")),
+        as_json=bool(args.json),
+        extra={k: v for k, v in payload.items() if k not in {"ok", "message"}},
+    )
+
+
 def cmd_migrate(args: argparse.Namespace) -> int:
     project = resolve_main_project_root()
     if not args.change:
@@ -1451,6 +1535,17 @@ def build_parser() -> argparse.ArgumentParser:
     p_migrate.add_argument("--change", required=True)
     p_migrate.set_defaults(func=cmd_migrate)
 
+    p_ownership = sub.add_parser(
+        "declare-ownership", parents=[shared],
+        help="declare ownership.productPaths so archive can project the real diff",
+    )
+    p_ownership.add_argument("--change", required=True)
+    p_ownership.add_argument(
+        "--product-path", required=True, action="append",
+        help="exact file or directory prefix (no globs); repeatable",
+    )
+    p_ownership.set_defaults(func=cmd_declare_ownership)
+
     p_claim = sub.add_parser("claim", parents=[shared])
     p_claim.add_argument("--change", required=True)
     p_claim.add_argument("--phase", required=True)

package/harness/bundles/java/claude-code/scripts/harness_ledger.py

@@ -304,12 +304,23 @@ def _compact_record_payload(payload: dict[str, Any]) -> dict[str, Any]:
 
 
 def _compact_can_reuse_payload(payload: dict[str, Any]) -> dict[str, Any]:
-    """C5: can-reuse compact — ok/reuse/code only."""
-    return {
+    """C5: can-reuse compact — ok/reuse/code；拒绝复用时附带可执行的原因。
+
+    只回 ok/reuse/code 时，调用方拿到裸 `{"reuse": false}` 无从判断该怎么办，
+    只能再跑一次 `--verbose`——原因（profile 缺 verificationInputs、证据不完整等）
+    是单行短文本，值得留在默认输出里。允许复用时保持原样精简。
+    """
+    compact = {
         "ok": payload.get("ok", True),
         "reuse": payload.get("reuse"),
         "code": payload.get("code"),
     }
+    if payload.get("reuse") is not True:
+        for key in ("reason", "executionNeed", "detail"):
+            value = payload.get(key)
+            if value not in (None, "", [], {}):
+                compact[key] = value
+    return compact
 
 
 def emit_compact_or_verbose(
@@ -2581,9 +2592,19 @@ def cmd_record(args: argparse.Namespace) -> int:
                     error_code="PROFILE_INPUT_FILES_CONFLICT",
                 )
         files = profile_files
-    if not files:
+    # NOT_APPLICABLE 的 validation 没有输入文件可言。门禁 close 又要求
+    # requiredValidations 每项都有 entry——两条规则叠加会逼调用方拿无关文件凑数，
+    # ledger 从此声称该 validation 的输入是那些文件，是假证据。
+    not_applicable = str(getattr(args, "applicability", "") or "") == "NOT_APPLICABLE"
+    if not files and not not_applicable:
+        return emit_error(
+            "record requires --files or a non-empty --profile-input file set; "
+            "unless --applicability NOT_APPLICABLE (then --applicability-reason is required)",
+            as_json=as_json,
+        )
+    if not_applicable and not _nonempty_str(getattr(args, "applicability_reason", None)):
         return emit_error(
-            "record requires --files or a non-empty --profile-input file set",
+            "NOT_APPLICABLE requires --applicability-reason",
             as_json=as_json,
         )
 

package/harness/bundles/java/claude-code/scripts/harness_profile.py

@@ -1101,6 +1101,43 @@ def _derive_verification_graph(profile: dict[str, Any]) -> None:
     }
 
 
+def manual_profile_template(
+    excluded: tuple[str, ...] | list[str], components: list[str]
+) -> dict[str, Any]:
+    """歧义时交给调用方的可填骨架。
+
+    骨架本身就是 `empty_profile_skeleton`——defaultsFingerprint、excludedRoots
+    默认集、commands/verificationInputs/verificationGraph 的键都已就位；这里只是
+    补上 commands 的条目形状与 requiredCoverage 取值，省掉调用方去源码里翻。
+    占位值用 `<...>` 包裹，便于替换后 grep 自检。
+    """
+    template = empty_profile_skeleton(excluded)
+    template["projectType"] = (
+        f"<从 detectedComponents 选定：{'|'.join(components)}>" if components else "<项目类型>"
+    )
+    template["commands"] = {
+        "unitTest": {
+            "command": "<在项目根可直接执行的增量测试命令>",
+            "scope": "incremental",
+            "inputs": ["<相对项目根的 glob，如 <component>/test/**>"],
+            "coverage": "unitTest",
+            "requiredCoverage": "incremental",
+            "dependsOn": [],
+            "source": "user",
+        },
+        "unitTestFull": {
+            "command": "<在项目根可直接执行的全量测试命令>",
+            "scope": "full",
+            "inputs": ["<相对项目根的 glob，如 <component>/**>"],
+            "coverage": "unitTestFull",
+            "requiredCoverage": "module",
+            "dependsOn": [],
+            "source": "user",
+        },
+    }
+    return template
+
+
 def detect(project: Path) -> dict[str, Any]:
     project = project.resolve()
     existing = load_profile(project)
@@ -1120,6 +1157,14 @@ def detect(project: Path) -> dict[str, Any]:
             "applied": False,
             "detectedComponents": components,
             "message": "multiple or nested project components require an explicit profile",
+            # 拒绝生成可以，但不能连结构都不给：调用方否则只能反读本文件源码去凑
+            # v3 形状。骨架不落盘——一份占位 profile 看起来"已配好"比没有更糟。
+            "profileTemplate": manual_profile_template(excluded, components),
+            "hint": (
+                "手工声明一份 profile：以 profileTemplate 为骨架，按 detectedComponents "
+                "选定产品包，commands 里的 inputs 用相对项目根的 glob（如 "
+                f"'<component>/test/**'），填好后写入 {project / PROFILE_REL}"
+            ),
         }
     if project_type == "unknown" and existing:
         return {

package/harness/bundles/java/claude-code/scripts/harness_runtime.py

@@ -157,6 +157,35 @@ def _private_evidence_root(explicit: Path | None = None) -> Path:
             Path.home() / ".harness" / "private-evidence").expanduser().resolve()
 
 
+def private_evidence_root_for(source: Path, explicit: Path | None = None) -> Path:
+    """私有隔离根：必须与源同一驱动器。
+
+    `os.replace` 跨驱动器会直接失败（Windows 上是 WinError 17）。默认根
+    `~/.harness/private-evidence` 在 Windows 落 C 盘，项目在别的盘时隔离必然失败
+    ——而隔离是归档的硬前置，等于整条归档路被堵死。同盘时沿用配置/默认值；不同盘
+    时退到源所在驱动器根下的 `.harness-private-evidence`（天然在任何项目根之外）。
+    POSIX 上 drive 恒为空串，行为不变。
+    """
+    source = source.expanduser().resolve()
+    preferred = _private_evidence_root(explicit)
+    if preferred.drive.lower() == source.drive.lower():
+        return preferred
+    return (Path(source.anchor) / ".harness-private-evidence").resolve()
+
+
+def _infer_project_root(change_root: Path) -> Path | None:
+    """change 目录形如 `<project>/.harness/changes/<name>`：取路径里 `.harness` 段的父目录。
+
+    不做"向上找存在 .harness 的目录"式搜索——用户 HOME 下通常也有 `.harness`，
+    那样会把 HOME 误判成项目根，把合法的临时隔离位置一并拒掉。
+    """
+    parts = change_root.parts
+    for index in range(len(parts) - 1, 0, -1):
+        if parts[index] == ".harness":
+            return Path(*parts[:index])
+    return None
+
+
 def _secure_private_path(path: Path, *, directory: bool) -> str:
     """Best-effort owner-only permissions, with an explicit ACL result on Windows."""
     mode = 0o700 if directory else 0o600
@@ -288,11 +317,12 @@ def quarantine_sensitive_evidence(
     change_root: Path,
     reason: str = "sensitive evidence",
     private_root: Path | None = None,
+    project_root: Path | None = None,
 ) -> dict[str, Any]:
     """Atomically move legacy plaintext evidence outside any publishable tree."""
     source = source.expanduser().resolve()
     change_root = change_root.expanduser().resolve()
-    private = _private_evidence_root(private_root)
+    private = private_evidence_root_for(source, private_root)
     if not source.is_file():
         return {"ok": False, "reasonCode": "SENSITIVE_EVIDENCE_SOURCE_MISSING", "sourcePath": str(source)}
     if private == change_root or private.is_relative_to(change_root):
@@ -301,6 +331,25 @@ def quarantine_sensitive_evidence(
             "reasonCode": "SENSITIVE_EVIDENCE_QUARANTINE_FAILED",
             "error": "private quarantine root must be outside the publishable change root",
         }
+    # 归档的密钥扫描门禁按**项目根**判定（SECRET_SCAN_PRIVATE_PATH_IN_COPY_ROOT）。
+    # 这里此前只看 change_root，比门禁宽——于是隔离"成功"了，门禁再拒一次，调用方
+    # 得多试一轮才找到合法位置。对齐到同一条边界，错就错在当场。
+    resolved_project = (
+        project_root.expanduser().resolve()
+        if project_root is not None
+        else _infer_project_root(change_root)
+    )
+    if resolved_project is not None and (
+        private == resolved_project or private.is_relative_to(resolved_project)
+    ):
+        return {
+            "ok": False,
+            "reasonCode": "SENSITIVE_EVIDENCE_QUARANTINE_FAILED",
+            "error": (
+                "private quarantine root must be outside the project root "
+                f"({resolved_project}); archive's secret scan rejects paths inside it"
+            ),
+        }
     try:
         raw_digest = "sha256:" + hashlib.sha256(source.read_bytes()).hexdigest()
         quarantine_dir = private / uuid.uuid4().hex
@@ -1850,6 +1899,37 @@ def doctor(project: Path, change_dir: Path, *, agent: str) -> dict[str, Any]:
     return {"ok": True, "action": "doctor", **payload}
 
 
+def cmd_quarantine_evidence(args: argparse.Namespace) -> dict[str, Any]:
+    """隔离命令：把明文敏感证据原子移出可发布树，逐条返回结果。
+
+    单条失败不吞掉其余条目——归档阻断项常常一次报多个文件，逐条报告才能一轮修完。
+    """
+    project = Path(args.project).expanduser().resolve()
+    change_root = Path(args.change_dir).expanduser().resolve()
+    private_root = Path(args.private_root) if args.private_root else None
+    entries: list[dict[str, Any]] = []
+    for raw in args.file:
+        candidate = Path(raw)
+        source = candidate if candidate.is_absolute() else change_root / candidate
+        outcome = quarantine_sensitive_evidence(
+            source,
+            change_root=change_root,
+            reason=args.reason,
+            private_root=private_root,
+            project_root=project,
+        )
+        entries.append({"file": raw, **outcome})
+    failed = [item for item in entries if not item.get("ok")]
+    return {
+        "ok": not failed,
+        "action": "quarantine-evidence",
+        "code": "SENSITIVE_EVIDENCE_QUARANTINED" if not failed else "SENSITIVE_EVIDENCE_QUARANTINE_FAILED",
+        "quarantined": len(entries) - len(failed),
+        "failed": len(failed),
+        "entries": entries,
+    }
+
+
 def build_parser() -> argparse.ArgumentParser:
     parser = argparse.ArgumentParser(prog="harness_runtime.py")
     sub = parser.add_subparsers(dest="command", required=True)
@@ -1858,6 +1938,24 @@ def build_parser() -> argparse.ArgumentParser:
     p_doctor.add_argument("--change-dir", required=True)
     p_doctor.add_argument("--agent", choices=sorted(_ADAPTERS), required=True)
     p_doctor.add_argument("--json", action="store_true")
+    # 归档的 SENSITIVE_EVIDENCE_UNQUARANTINED 是硬阻断，此前却没有任何命令行入口
+    # ——调用方只能写 python -c + sys.path hack 去调内部函数。
+    p_quarantine = sub.add_parser(
+        "quarantine-evidence",
+        help="move plaintext sensitive evidence outside the publishable tree",
+    )
+    p_quarantine.add_argument("--project", required=True)
+    p_quarantine.add_argument("--change-dir", required=True)
+    p_quarantine.add_argument(
+        "--file", required=True, action="append",
+        help="path to quarantine, relative to --change-dir or absolute; repeatable",
+    )
+    p_quarantine.add_argument("--reason", default="sensitive evidence")
+    p_quarantine.add_argument(
+        "--private-root", default=None,
+        help="override the private root; must be outside the project root and on its drive",
+    )
+    p_quarantine.add_argument("--json", action="store_true")
     p_adapter = sub.add_parser("adapter")
     p_adapter.add_argument("--agent", choices=sorted(_ADAPTERS), required=True)
     p_adapter.add_argument("--change", required=True)
@@ -1899,6 +1997,8 @@ def main(argv: list[str] | None = None) -> int:
     try:
         if args.command == "doctor":
             result = doctor(Path(args.project), Path(args.change_dir), agent=args.agent)
+        elif args.command == "quarantine-evidence":
+            result = cmd_quarantine_evidence(args)
         elif args.command == "adapter":
             result = {"ok": True, "action": "adapter", **adapter_worktree(args.agent, args.change)}
         elif args.command == "_worker":

package/harness/bundles/java/codebuddy/.harness-build.json

@@ -2,5 +2,5 @@
   "schemaVersion": 1,
   "overlay": "java",
   "agent": "codebuddy",
-  "coreHash": "984aee382a36595c"
+  "coreHash": "4f3a290efefe6a68"
 }