@cannbot-plugin/cannbot-harness 0.0.0-stage → 0.1.4

package/assets/ops-direct-invoke/skills/workflow-orchestrator/tests/test_orchestrator.py

@@ -0,0 +1,1383 @@
+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+
+"""Blackbox tests for the workflow-orchestrator skill.
+
+These tests exercise ``orchestrator.py`` only through its documented CLI in
+``--dry-run`` mode (simulated agents), asserting on exit codes, stdout, and the
+persisted artifacts (``status.json`` / ``log.jsonl``). ``OrchestratorTest``
+drives the orchestrator only through this CLI — it stays a blackbox, its source
+intentionally not read here. ``BuildPromptTest`` is the sole exception: a
+whitebox unit test of the pure function ``get_task.build_prompt``.
+
+Run from the skill root (or anywhere)::
+
+    python3 -m unittest discover -s harness/workflow-orchestrator/tests
+
+or directly::
+
+    python3 harness/workflow-orchestrator/tests/test_orchestrator.py
+"""
+
+import json
+import os
+import shutil
+import subprocess
+import sys
+import tempfile
+import unittest
+from dataclasses import dataclass
+from unittest.mock import patch
+from pathlib import Path
+
+import yaml
+
+SKILL_ROOT = Path(__file__).resolve().parents[1]
+ORCHESTRATOR = SKILL_ROOT / "scripts" / "orchestrator.py"
+
+sys.path.insert(0, str(SKILL_ROOT / "scripts"))
+import get_task  # noqa: E402  白盒：借用纯函数 build_prompt / validate_workflow
+from verdict_common import verdict_file_path  # noqa: E402
+
+
+def _node(node_id, **overrides):
+    """A valid normal task node; overrides let a test break one field."""
+    node = {
+        "id": node_id,
+        "task_type": "normal",
+        "title": f"Task {node_id}",
+        "goal": ["finish the task"],
+        "approach": ["do the work"],
+        "acceptance": ["true"],
+        "out_of_scope": ["nothing"],
+        "depends_on": [],
+        "executor": "general-executor",
+        "verifier": "general-verifier",
+        "max_retries": 0,
+        "on_exhaust": "exit",
+    }
+    node.update(overrides)
+    return node
+
+
+def _workflow_yaml(nodes, max_parallel=1, max_rollbacks=None):
+    wf = {"workflow": "test", "max_parallel": max_parallel, "nodes": nodes}
+    if max_rollbacks is not None:
+        wf["max_rollbacks"] = max_rollbacks
+    return yaml.safe_dump(wf, sort_keys=False)
+
+
+@dataclass(frozen=True)
+class WorkflowOptions:
+    max_parallel: int = 1
+    max_rollbacks: int | None = None
+
+
+def _rollback_nodes():
+    """Two-node workflow in which b rolls back to a."""
+    return [
+        _node("a"),
+        _node("b", depends_on=["a"], on_exhaust="rollback", rollback_to="a"),
+    ]
+
+
+def run_orchestrator(
+    nodes=None,
+    yaml_text=None,
+    dry_replies=None,
+    options=None,
+    pre_files=None,
+):
+    """Run orchestrator.py with --dry-run; return (exit_code, output, work_dir).
+
+    pre_files: {相对路径: 内容}，运行前写入 work_dir（如子图 yaml、手工 status.json）。
+    """
+    options = options or WorkflowOptions()
+    work_dir = tempfile.mkdtemp(prefix="wo-test-")
+    wf_path = Path(work_dir) / "workflow.yaml"
+    wf_path.write_text(
+        yaml_text
+        if yaml_text is not None
+        else _workflow_yaml(nodes, options.max_parallel, options.max_rollbacks)
+    )
+    for rel, content in (pre_files or {}).items():
+        p = Path(work_dir) / rel
+        p.parent.mkdir(parents=True, exist_ok=True)
+        p.write_text(content)
+    if dry_replies is not None:
+        wr = Path(work_dir) / ".workflow"
+        wr.mkdir(parents=True, exist_ok=True)
+        (wr / "dry_replies.json").write_text(json.dumps(dry_replies))
+    proc = subprocess.run(
+        [
+            sys.executable,
+            str(ORCHESTRATOR),
+            "--yaml",
+            str(wf_path),
+            "--work-dir",
+            work_dir,
+            "--dry-run",
+            "--prompt",
+            "test",
+        ],
+        capture_output=True,
+        text=True,
+    )
+    return proc.returncode, proc.stdout + proc.stderr, work_dir
+
+
+def read_status(work_dir):
+    return json.loads((Path(work_dir) / ".workflow" / "status.json").read_text())
+
+
+def read_log(work_dir):
+    lines = (Path(work_dir) / ".workflow" / "log.jsonl").read_text().splitlines()
+    return [json.loads(line) for line in lines if line.strip()]
+
+
+def _run_orchestrator_dry(wf, work_dir):
+    """以 --dry-run 运行编排器（work_dir 已就绪）；返回 (returncode, out)。"""
+    proc = subprocess.run(
+        [
+            sys.executable,
+            str(ORCHESTRATOR),
+            "--yaml",
+            str(wf),
+            "--work-dir",
+            work_dir,
+            "--dry-run",
+        ],
+        capture_output=True,
+        text=True,
+    )
+    return proc.returncode, proc.stdout + proc.stderr
+
+
+class OrchestratorTest(unittest.TestCase):
+    def setUp(self):
+        self._work_dirs = []
+
+    def tearDown(self):
+        for d in self._work_dirs:
+            shutil.rmtree(d, ignore_errors=True)
+
+    def _run(self, **kwargs):
+        code, out, work_dir = run_orchestrator(**kwargs)
+        self._work_dirs.append(work_dir)
+        return code, out, work_dir
+
+    def _legacy_work_dir(self, nodes, tasks, seq=1, dry_replies=None):
+        """构造「旧账本」work_dir：写 workflow.yaml + 手工 status.json，返回 (work_dir, wf, wr)。"""
+        work_dir = tempfile.mkdtemp(prefix="wo-test-")
+        self._work_dirs.append(work_dir)
+        wf = Path(work_dir) / "workflow.yaml"
+        wf.write_text(_workflow_yaml(nodes))
+        wr = Path(work_dir) / ".workflow"
+        wr.mkdir()
+        (wr / "user_prompt.md").write_text("test\n")
+        (wr / "status.json").write_text(
+            json.dumps(
+                {
+                    "workflow": str(wf),
+                    "work_dir": work_dir,
+                    "user_prompt": str(wr / "user_prompt.md"),
+                    "provider": "dry",
+                    "seq": seq,
+                    "rollbacks_used": 0,
+                    "tasks": tasks,
+                }
+            )
+        )
+        (wr / "dry_replies.json").write_text(json.dumps(dry_replies or {}))
+        return work_dir, wf, wr
+
+    def _pending_advice_fixture(self):
+        nodes = [
+            _node("a"),
+            _node("b", depends_on=["a"], on_exhaust="rollback", rollback_to="a"),
+        ]
+        wd, wf, wr = self._legacy_work_dir(
+            nodes,
+            {
+                "a": {
+                    "status": "pass",
+                    "retries": 0,
+                    "pass_seq": 1,
+                    "checkpoint_seq": 0,
+                },
+                "b": {"status": "fail", "retries": 1},
+            },
+        )
+        ckpt = wr / "checkpoints" / "a"
+        ckpt.mkdir(parents=True)
+        (ckpt / "workflow.yaml").write_text(wf.read_text())
+        (ckpt / "marker.txt").write_text("baseline")
+        (Path(wd) / "marker.txt").write_text("failed attempt")
+        return wd, wf, wr
+
+    def _poll_advice(self, wd):
+        proc = subprocess.run(
+            [sys.executable, str(SKILL_ROOT / "scripts/get_task.py"), wd],
+            capture_output=True,
+            text=True,
+        )
+        self.assertIn(proc.returncode, (0, 1), proc.stderr)
+        return json.loads(proc.stdout)
+
+    def test_advice_two_polls_archive_once_and_inline_content(self):
+        wd, wf, wr = self._pending_advice_fixture()
+        first = self._poll_advice(wd)
+        self.assertEqual(first[0].get("phase"), "advice")
+        self.assertEqual(set(read_status(wd)["tasks"]), {"a", "b"})
+        self.assertIn("advice_pending", read_status(wd))
+        history = wr / "history" / "b.to.a.1"
+        self.assertEqual((history / "marker.txt").read_text(), "failed attempt")
+        self.assertEqual((Path(wd) / "marker.txt").read_text(), "baseline")
+        self.assertEqual(first, self._poll_advice(wd))
+        self.assertEqual(len(list((wr / "history").iterdir())), 1)
+        self.assertEqual(read_status(wd)["rollbacks_used"], 1)
+        advice = wr / "advice" / "b.to.a.1.md"
+        self.assertFalse(advice.exists())
+        advice.write_text("Unique recovery instruction: preserve the seed.")
+        second = self._poll_advice(wd)
+        self.assertEqual(second[0]["task_id"], "a")
+        self.assertIn("Unique recovery instruction", second[0]["prompt"])
+        self.assertNotIn("advice_pending", read_status(wd))
+        status = read_status(wd)
+        status["tasks"]["a"].update(status="pass", pass_seq=2)
+        (wr / "status.json").write_text(json.dumps(status))
+        downstream = self._poll_advice(wd)
+        self.assertEqual(downstream[0]["task_id"], "b")
+        self.assertNotIn("Unique recovery instruction", downstream[0]["prompt"])
+
+    def test_empty_advice_does_not_dispatch_redo(self):
+        wd, wf, wr = self._pending_advice_fixture()
+        first = self._poll_advice(wd)
+        (wr / "advice" / "b.to.a.1.md").write_text("   ")
+        self.assertEqual(first, self._poll_advice(wd))
+        self.assertIn("advice_pending", read_status(wd))
+
+    def test_orchestrator_stops_when_advice_output_missing(self):
+        wd, wf, wr = self._pending_advice_fixture()
+        with patch.object(
+            get_task.orch, "run_agent", return_value=("advice-written", False)
+        ) as run:
+            result = get_task.orch.loop(wd, None, None)
+        self.assertEqual(result, 1)
+        self.assertEqual(run.call_count, 1)
+        self.assertEqual(run.call_args.args[0]["phase"], "advice")
+        self.assertIn("error", read_status(wd)["advice_pending"])
+        self.assertEqual(read_status(wd)["rollbacks_used"], 1)
+
+    def test_orchestrator_stops_after_advice_crash(self):
+        wd, wf, wr = self._pending_advice_fixture()
+        with patch.object(get_task.orch, "run_agent", return_value=(None, True)) as run:
+            result = get_task.orch.loop(wd, None, None)
+        self.assertEqual(result, 1)
+        self.assertEqual(run.call_count, 1)
+        self.assertIn("error", read_status(wd)["advice_pending"])
+
+    def test_init_without_provider(self):
+        wd = tempfile.mkdtemp(prefix="wo-test-")
+        self._work_dirs.append(wd)
+        wf = Path(wd) / "workflow.yaml"
+        wf.write_text(_workflow_yaml([_node("a")]))
+        proc = subprocess.run(
+            [
+                sys.executable,
+                str(SKILL_ROOT / "scripts/init_status.py"),
+                "--yaml",
+                str(wf),
+                "--work-dir",
+                wd,
+                "--prompt",
+                "test",
+            ],
+            capture_output=True,
+            text=True,
+        )
+        self.assertEqual(proc.returncode, 0, proc.stderr)
+        self.assertNotIn("provider", read_status(wd))
+
+    def test_happy_path_single_node(self):
+        code, out, wd = self._run(nodes=[_node("setup")])
+        self.assertEqual(code, 0, out)
+        self.assertEqual(read_status(wd)["tasks"]["setup"]["status"], "pass")
+        events = [e["event"] for e in read_log(wd)]
+        self.assertIn("init", events)
+        self.assertIn("finish", events)
+
+    def test_status_ledger_fields(self):
+        code, out, wd = self._run(nodes=[_node("setup")])
+        self.assertEqual(code, 0, out)
+        status = read_status(wd)
+        self.assertNotIn("provider", status)
+        self.assertEqual(status.get("rollbacks_used"), 0)
+        self.assertGreaterEqual(status.get("seq", 0), 1)
+        task = status["tasks"]["setup"]
+        self.assertEqual(task.get("pass_seq"), 1)
+        self.assertNotIn("advice", task)
+
+    def test_checkpoint_snapshot_taken(self):
+        nodes = [
+            _node("setup"),
+            _node(
+                "rollback",
+                depends_on=["setup"],
+                on_exhaust="rollback",
+                rollback_to="setup",
+            ),
+            _node("downstream", depends_on=["rollback"]),
+        ]
+        code, out, wd = self._run(nodes=nodes, pre_files={"hello.txt": "hi\n"})
+        self.assertEqual(code, 0, out)
+        ckpt = Path(wd) / ".workflow" / "checkpoints" / "setup"
+        self.assertTrue((ckpt / "hello.txt").is_file())
+        self.assertEqual((ckpt / "hello.txt").read_text(), "hi\n")
+        self.assertFalse((ckpt / ".workflow").exists())  # .workflow 不进快照
+        # 首次派发时还没有任何 pass，seq=0
+        status = read_status(wd)
+        self.assertEqual(status["tasks"]["setup"].get("checkpoint_seq"), 0)
+        self.assertNotIn("checkpoint_seq", status["tasks"]["rollback"])
+        self.assertNotIn("checkpoint_seq", status["tasks"]["downstream"])
+        self.assertFalse((Path(wd) / ".workflow" / "checkpoints" / "rollback").exists())
+        self.assertFalse(
+            (Path(wd) / ".workflow" / "checkpoints" / "downstream").exists()
+        )
+
+    def test_dependency_chain_completes(self):
+        nodes = [_node("a"), _node("b", depends_on=["a"])]
+        code, out, wd = self._run(nodes=nodes)
+        self.assertEqual(code, 0, out)
+        tasks = read_status(wd)["tasks"]
+        self.assertEqual(tasks["a"]["status"], "pass")
+        self.assertEqual(tasks["b"]["status"], "pass")
+
+    def test_independent_nodes_run_in_parallel(self):
+        nodes = [_node("a"), _node("b")]
+        code, out, wd = self._run(nodes=nodes, options=WorkflowOptions(max_parallel=2))
+        self.assertEqual(code, 0, out)
+        tasks = read_status(wd)["tasks"]
+        self.assertEqual(tasks["a"]["status"], "pass")
+        self.assertEqual(tasks["b"]["status"], "pass")
+
+    def test_empty_nodes_rejected(self):
+        code, out, _ = self._run(
+            yaml_text="workflow: test\nmax_parallel: 1\nnodes: []\n"
+        )
+        self.assertEqual(code, 1, out)
+        self.assertIn("没有任务节点", out)
+
+    def test_missing_node_key_rejected(self):
+        node = _node("t1")
+        del node["approach"]
+        code, out, _ = self._run(nodes=[node])
+        self.assertEqual(code, 2, out)
+        self.assertIn("approach", out)
+
+    def test_invalid_task_type_rejected(self):
+        code, out, _ = self._run(nodes=[_node("t1", task_type="bogus")])
+        self.assertEqual(code, 2, out)
+        self.assertIn("normal/subgraph", out)
+
+    def test_retry_then_pass(self):
+        node = _node("t1", max_retries=1)
+        code, out, wd = self._run(
+            nodes=[node],
+            dry_replies={
+                "t1": ["executed", "$VERDICT:fail", "executed", "$VERDICT:pass"]
+            },
+        )
+        self.assertEqual(code, 0, out)
+        task = read_status(wd)["tasks"]["t1"]
+        self.assertEqual(task["status"], "pass")
+        self.assertEqual(task["retries"], 1)
+
+    def test_fail_exhausts_budget(self):
+        node = _node("t1", max_retries=1)
+        code, out, wd = self._run(
+            nodes=[node],
+            dry_replies={
+                "t1": ["executed", "$VERDICT:fail", "executed", "$VERDICT:fail"]
+            },
+        )
+        self.assertEqual(code, 1, out)
+        task = read_status(wd)["tasks"]["t1"]
+        self.assertEqual(task["status"], "fail")
+        self.assertTrue(task["exhausted"])
+        self.assertGreaterEqual(task["retries"], 1)
+
+    def test_max_rollbacks_invalid_rejected(self):
+        node = _node("t1")
+        for bad in (-1, 1.5, "2", True):
+            code, out, _ = self._run(
+                yaml_text=_workflow_yaml([node], max_rollbacks=bad)
+            )
+            self.assertEqual(code, 2, out)
+            self.assertIn("max_rollbacks", out)
+
+    def test_rollback_without_rollback_to_rejected(self):
+        code, out, _ = self._run(nodes=[_node("t1", on_exhaust="rollback")])
+        self.assertEqual(code, 2, out)
+        self.assertIn("rollback_to", out)
+
+    def test_rollback_to_ignored_when_on_exhaust_not_rollback(self):
+        # on_exhaust!=rollback：rollback_to 完全忽略，即使指向不存在的节点也照常跑完
+        code, out, wd = self._run(
+            nodes=[_node("t1", on_exhaust="exit", rollback_to="ghost")]
+        )
+        self.assertEqual(code, 0, out)
+        self.assertEqual(read_status(wd)["tasks"]["t1"]["status"], "pass")
+
+    def test_rollback_to_unknown_rejected(self):
+        code, out, _ = self._run(
+            nodes=[_node("t1", on_exhaust="rollback", rollback_to="ghost")]
+        )
+        self.assertEqual(code, 2, out)
+        self.assertIn("rollback_to", out)
+
+    def test_rollback_to_non_ancestor_rejected(self):
+        nodes = [
+            _node("a"),
+            _node("b", depends_on=["a"]),
+            _node("c", on_exhaust="rollback", rollback_to="b"),
+        ]  # c 不依赖 b
+        code, out, _ = self._run(nodes=nodes)
+        self.assertEqual(code, 2, out)
+        self.assertIn("上游", out)
+
+    def test_rollback_to_subgraph_node_rejected(self):
+        nodes = [
+            _node("a"),
+            {
+                "id": "sg",
+                "task_type": "subgraph",
+                "file": "sub.yaml",
+                "depends_on": ["a"],
+            },
+            _node("b", depends_on=["a"], on_exhaust="rollback", rollback_to="sg"),
+        ]
+        code, out, _ = self._run(nodes=nodes)
+        self.assertEqual(code, 2, out)
+        self.assertIn("subgraph", out)
+
+    def test_rollback_budget_exhausted_stuck(self):
+        nodes = _rollback_nodes()
+        code, out, wd = self._run(
+            nodes=nodes,
+            options=WorkflowOptions(max_rollbacks=0),
+            dry_replies={
+                "a": ["executed", "$VERDICT:pass"],
+                "b": ["executed", "$VERDICT:fail"],
+            },
+        )
+        self.assertEqual(code, 1, out)
+        self.assertIn("回滚预算耗尽", out)
+        self.assertIn("工作流卡住", out)
+        self.assertIn("rollbacks_used=0 >= max_rollbacks=0", out)
+        self.assertNotIn("rollback_pending", read_status(wd))  # 预算耗尽路径不置标记
+
+    def test_subgraph_child_rollback_to_outside_rejected(self):
+        # 子图子节点的 rollback_to 在子图命名空间解析：引用父级节点 = 引用不存在
+        sub = {"nodes": [_node("x", on_exhaust="rollback", rollback_to="a")]}
+        nodes = [
+            _node("a"),
+            {
+                "id": "sg",
+                "task_type": "subgraph",
+                "file": "sub.yaml",
+                "depends_on": ["a"],
+            },
+        ]
+        code, out, _ = self._run(
+            nodes=nodes, pre_files={"sub.yaml": yaml.safe_dump(sub)}
+        )
+        self.assertEqual(code, 2, out)
+        self.assertIn("rollback_to", out)
+
+    def test_rollback_full_cycle(self):
+        nodes = [
+            _node("a"),
+            _node("b", depends_on=["a"]),
+            _node(
+                "c",
+                depends_on=["b"],
+                max_retries=0,
+                on_exhaust="rollback",
+                rollback_to="a",
+            ),
+        ]
+        replies = {
+            "a": ["executed", "$VERDICT:pass", "executed", "$VERDICT:pass"],
+            "b": ["executed", "$VERDICT:pass", "executed", "$VERDICT:pass"],
+            "c": ["executed", "$VERDICT:fail", "executed", "$VERDICT:pass"],
+        }
+        code, out, wd = self._run(
+            nodes=nodes, dry_replies=replies, pre_files={"seed.txt": "v1\n"}
+        )
+        self.assertEqual(code, 0, out)
+        status = read_status(wd)
+        self.assertEqual(status.get("rollbacks_used"), 1)
+        self.assertNotIn("rollback_pending", status)
+        for tid in ("a", "b", "c"):
+            self.assertEqual(status["tasks"][tid]["status"], "pass")
+            self.assertEqual(status["tasks"][tid]["retries"], 0)
+        # advice 占位文件（dry）；Y=a 的 advice 字段 pass 后已清除
+        self.assertTrue((Path(wd) / ".workflow" / "advice" / "c.to.a.1.md").is_file())
+        self.assertNotIn("advice", status["tasks"]["a"])
+        # 失败时间线产物已归档；seed.txt 在 checkpoint(a) 拍摄前已存在，恢复后仍在
+        self.assertTrue(
+            (Path(wd) / ".workflow" / "history" / "c.to.a.1" / "seed.txt").is_file()
+        )
+        self.assertTrue((Path(wd) / "seed.txt").is_file())
+        # 重置时删旧 checkpoint，重跑时重拍
+        self.assertTrue(
+            (Path(wd) / ".workflow" / "checkpoints" / "a" / "seed.txt").is_file()
+        )
+        # 日志含 rollback 事件（带 history/advice 字段）
+        rollbacks = [e for e in read_log(wd) if e["event"] == "rollback"]
+        self.assertEqual(len(rollbacks), 1)
+        self.assertEqual(rollbacks[0]["from"], "c")
+        self.assertEqual(rollbacks[0]["to"], "a")
+        self.assertIn("history", rollbacks[0])
+        self.assertIn("advice", rollbacks[0])
+
+    def test_rollback_drains_inflight(self):
+        # b 快速失败触发回滚时 d 仍在执行 → 输出 [] 自然排干，d 收尾后才执行回滚
+        nodes = [
+            _node("a"),
+            _node(
+                "b",
+                depends_on=["a"],
+                max_retries=0,
+                on_exhaust="rollback",
+                rollback_to="a",
+            ),
+            _node("d", depends_on=["a"]),
+        ]
+        replies = {
+            "a": ["executed", "$VERDICT:pass", "executed", "$VERDICT:pass"],
+            "b": ["executed", "$VERDICT:fail", "executed", "$VERDICT:pass"],
+            "d": [
+                "$SLEEP:2",
+                "executed",
+                "$VERDICT:pass",
+            ],  # 首轮 execute 慢；executed 态被重置后重做
+        }
+        code, out, wd = self._run(
+            nodes=nodes, dry_replies=replies, options=WorkflowOptions(max_parallel=2)
+        )
+        self.assertEqual(code, 0, out)
+        status = read_status(wd)
+        self.assertEqual(status.get("rollbacks_used"), 1)
+        for tid in ("a", "b", "d"):
+            self.assertEqual(status["tasks"][tid]["status"], "pass")
+        # d 首轮从未 pass（executed 态被重置）；d 不是 rollback 目标，不创建 checkpoint
+        self.assertNotIn("checkpoint_seq", status["tasks"]["d"])
+
+    def test_rollback_without_checkpoint_stuck(self):
+        # 旧账本：a 无 checkpoint_seq → 回滚无法执行 → STUCK，rollback_pending 留账
+        nodes = _rollback_nodes()
+        tasks = {
+            "a": {"status": "pass", "retries": 0},  # 旧账本：无 pass_seq/checkpoint_seq
+            "b": {"status": "fail", "retries": 1, "exhausted": True},
+        }
+        work_dir, wf, _ = self._legacy_work_dir(nodes, tasks)
+        code, out = _run_orchestrator_dry(wf, work_dir)
+        self.assertEqual(code, 1, out)
+        self.assertIn("checkpoint", out)
+        self.assertIn("rollback_pending", read_status(work_dir))
+
+    def test_subgraph_child_rollback(self):
+        sub = {
+            "nodes": [
+                _node("x"),
+                _node(
+                    "y",
+                    depends_on=["x"],
+                    max_retries=0,
+                    on_exhaust="rollback",
+                    rollback_to="x",
+                ),
+            ]
+        }
+        nodes = [
+            _node("pre"),
+            {
+                "id": "sg",
+                "task_type": "subgraph",
+                "file": "sub.yaml",
+                "depends_on": ["pre"],
+            },
+        ]
+        replies = {
+            "pre": ["executed", "$VERDICT:pass"],
+            "sg/x": ["executed", "$VERDICT:pass", "executed", "$VERDICT:pass"],
+            "sg/y": ["executed", "$VERDICT:fail", "executed", "$VERDICT:pass"],
+        }
+        code, out, wd = self._run(
+            nodes=nodes,
+            dry_replies=replies,
+            pre_files={"sub.yaml": yaml.safe_dump(sub)},
+        )
+        self.assertEqual(code, 0, out)
+        status = read_status(wd)
+        self.assertEqual(status.get("rollbacks_used"), 1)
+        for tid in ("pre", "sg/x", "sg/y"):
+            self.assertEqual(status["tasks"][tid]["status"], "pass")
+        # 子图子任务的 checkpoint 按 sg/child 嵌套；advice/history 文件名 / 转写为 __
+        self.assertTrue((Path(wd) / ".workflow" / "checkpoints" / "sg" / "x").is_dir())
+        self.assertTrue(
+            (Path(wd) / ".workflow" / "advice" / "sg__y.to.sg__x.1.md").is_file()
+        )
+        self.assertTrue(
+            (Path(wd) / ".workflow" / "history" / "sg__y.to.sg__x.1").is_dir()
+        )
+        self.assertEqual(status.get("checkpoint_targets"), ["sg/x"])
+        self.assertNotIn("checkpoint_seq", status["tasks"]["sg/y"])
+        self.assertFalse((Path(wd) / ".workflow" / "checkpoints" / "sg" / "y").exists())
+        # sg/x 的 checkpoint 是稳定基线，回滚后重跑不重拍
+        self.assertEqual(status["tasks"]["sg/x"].get("checkpoint_seq"), 1)
+
+    def test_node_id_with_slash_rejected(self):
+        code, out, _ = self._run(nodes=[_node("a/b")])
+        self.assertEqual(code, 2, out)
+        self.assertIn("'/'", out)
+
+    def test_rollback_resets_later_parallel_pass(self):
+        # c 与 b 平行（都依赖 a）；c 在 checkpoint(a) 之后 pass → 回滚时一并重置重跑
+        nodes = [
+            _node("a"),
+            _node(
+                "b",
+                depends_on=["a"],
+                max_retries=0,
+                on_exhaust="rollback",
+                rollback_to="a",
+            ),
+            _node("c", depends_on=["a"]),
+        ]
+        replies = {
+            "a": ["executed", "$VERDICT:pass", "executed", "$VERDICT:pass"],
+            "b": [
+                "$SLEEP:2",
+                "$VERDICT:fail",
+                "executed",
+                "$VERDICT:pass",
+            ],  # 首轮 execute 慢，让 c 先 pass
+            "c": ["executed", "$VERDICT:pass", "executed", "$VERDICT:pass"],
+        }
+        code, out, wd = self._run(
+            nodes=nodes, dry_replies=replies, options=WorkflowOptions(max_parallel=2)
+        )
+        self.assertEqual(code, 0, out)
+        status = read_status(wd)
+        self.assertEqual(status.get("rollbacks_used"), 1)
+        for tid in ("a", "b", "c"):
+            self.assertEqual(status["tasks"][tid]["status"], "pass")
+        # c 被重置后第二轮重跑时仍不创建 checkpoint：它不是 rollback 目标
+        self.assertNotIn("checkpoint_seq", status["tasks"]["c"])
+
+    def test_rollback_resets_cross_subgraph_continue(self):
+        # x2(continue 耗尽) 经子图节点 sg 才是 a 的传递下游 → 豁免失效，一并重置重跑
+        sub = {"nodes": [_node("k")]}
+        nodes = [
+            _node("a"),
+            {
+                "id": "sg",
+                "task_type": "subgraph",
+                "file": "sub.yaml",
+                "depends_on": ["a"],
+            },
+            _node("x2", depends_on=["sg"], max_retries=0, on_exhaust="continue"),
+            _node(
+                "x",
+                depends_on=["a"],
+                max_retries=0,
+                on_exhaust="rollback",
+                rollback_to="a",
+            ),
+        ]
+        replies = {
+            "a": ["executed", "$VERDICT:pass", "executed", "$VERDICT:pass"],
+            "sg/k": ["executed", "$VERDICT:pass", "executed", "$VERDICT:pass"],
+            "x": [
+                "$SLEEP:2",
+                "$VERDICT:fail",
+                "executed",
+                "$VERDICT:pass",
+            ],  # 首轮 execute 慢，让 x2 先耗尽
+            "x2": ["executed", "$VERDICT:fail", "executed", "$VERDICT:pass"],
+        }
+        code, out, wd = self._run(
+            nodes=nodes,
+            dry_replies=replies,
+            options=WorkflowOptions(max_parallel=2),
+            pre_files={"sub.yaml": yaml.safe_dump(sub)},
+        )
+        self.assertEqual(code, 0, out)
+        status = read_status(wd)
+        self.assertEqual(status.get("rollbacks_used"), 1)
+        for tid in ("a", "sg/k", "x2", "x"):
+            self.assertEqual(status["tasks"][tid]["status"], "pass")
+
+    def test_rollback_pending_exit_exhausted_stuck_first(self):
+        # rollback_pending 待执行时另一节点 exit 耗尽 → STUCK 优先，回滚不执行、标记留账
+        nodes = [
+            _node("a"),
+            _node(
+                "b",
+                depends_on=["a"],
+                max_retries=0,
+                on_exhaust="rollback",
+                rollback_to="a",
+            ),
+            _node("e", depends_on=["a"], max_retries=0, on_exhaust="exit"),
+        ]
+        tasks = {
+            "a": {"status": "pass", "retries": 0},
+            "b": {"status": "fail", "retries": 1, "exhausted": True},
+            "e": {"status": "fail", "retries": 1, "exhausted": True},
+        }
+        work_dir, wf, _ = self._legacy_work_dir(nodes, tasks)
+        code, out = _run_orchestrator_dry(wf, work_dir)
+        self.assertEqual(code, 1, out)
+        self.assertIn("工作流卡住", out)
+        # 回滚未执行：rollbacks_used 仍为 0、b 保持 fail、标记留账供排查
+        status = read_status(work_dir)
+        self.assertEqual(status.get("rollbacks_used"), 0)
+        self.assertIn("rollback_pending", status)
+        self.assertEqual(status["tasks"]["b"]["status"], "fail")
+
+    def test_second_rollback_trigger_free_redo(self):
+        # 同轮第二个 rollback 触发不置标记不吃预算：回滚执行时一并重置（免费重做）
+        nodes = [
+            _node("a"),
+            _node(
+                "b",
+                depends_on=["a"],
+                max_retries=0,
+                on_exhaust="rollback",
+                rollback_to="a",
+            ),
+            _node(
+                "c",
+                depends_on=["a"],
+                max_retries=0,
+                on_exhaust="rollback",
+                rollback_to="a",
+            ),
+        ]
+        replies = {
+            "a": ["executed", "$VERDICT:pass", "executed", "$VERDICT:pass"],
+            "b": ["executed", "$VERDICT:fail", "executed", "$VERDICT:pass"],
+            "c": ["executed", "$VERDICT:fail", "executed", "$VERDICT:pass"],
+        }
+        code, out, wd = self._run(
+            nodes=nodes, dry_replies=replies, options=WorkflowOptions(max_parallel=2)
+        )
+        self.assertEqual(code, 0, out)
+        status = read_status(wd)
+        self.assertEqual(status.get("rollbacks_used"), 1)
+        for tid in ("a", "b", "c"):
+            self.assertEqual(status["tasks"][tid]["status"], "pass")
+
+    def test_multiple_rollbacks_sequence(self):
+        # max_rollbacks=2：同一节点两次耗尽触发两次回滚，history/advice 按序号区分
+        nodes = _rollback_nodes()
+        replies = {
+            "a": [
+                "executed",
+                "$VERDICT:pass",
+                "executed",
+                "$VERDICT:pass",
+                "executed",
+                "$VERDICT:pass",
+            ],
+            "b": [
+                "executed",
+                "$VERDICT:fail",
+                "executed",
+                "$VERDICT:fail",
+                "executed",
+                "$VERDICT:pass",
+            ],
+        }
+        code, out, wd = self._run(
+            nodes=nodes, dry_replies=replies, options=WorkflowOptions(max_rollbacks=2)
+        )
+        self.assertEqual(code, 0, out)
+        status = read_status(wd)
+        self.assertEqual(status.get("rollbacks_used"), 2)
+        for n in (1, 2):
+            self.assertTrue(
+                (Path(wd) / ".workflow" / "advice" / ("b.to.a.%d.md" % n)).is_file()
+            )
+            self.assertTrue(
+                (Path(wd) / ".workflow" / "history" / ("b.to.a.%d" % n)).is_dir()
+            )
+        self.assertEqual(status["tasks"]["b"]["status"], "pass")
+
+    def test_rollback_preserves_legacy_pass(self):
+        # 旧账本 pass 任务（无 pass_seq）视为 pass_seq=0：≤ checkpoint_seq(回滚目标) 保持不动
+        nodes = _rollback_nodes()
+        tasks = {
+            "a": {"status": "pass", "retries": 0, "checkpoint_seq": 0},  # 无 pass_seq
+            "b": {"status": "fail", "retries": 1, "exhausted": True},
+        }
+        work_dir, wf, wr = self._legacy_work_dir(
+            nodes, tasks, seq=0, dry_replies={"b": ["executed", "$VERDICT:pass"]}
+        )
+        # execute_rollback 要恢复 checkpoint(a)：真实快照含 work_dir 顶层条目（恢复后
+        # clear_tree 删掉的 workflow.yaml 随之还原），这里手工对齐，只需 workflow.yaml
+        ckpt_a = wr / "checkpoints" / "a"
+        ckpt_a.mkdir(parents=True)
+        (ckpt_a / "workflow.yaml").write_text(wf.read_text())
+        code, out = _run_orchestrator_dry(wf, work_dir)
+        self.assertEqual(code, 0, out)
+        status = read_status(work_dir)
+        self.assertEqual(status.get("rollbacks_used"), 1)
+        # a 保持 pass 未被重做（若被重置重跑，a 的 pass 会让 seq 变成 2）
+        self.assertEqual(status.get("seq"), 1)
+        self.assertEqual(status["tasks"]["a"]["status"], "pass")
+        self.assertEqual(status["tasks"]["b"]["status"], "pass")
+
+    def test_repeated_rollback_keeps_same_target_checkpoint(self):
+        """A second rollback to the same target must consume the original baseline."""
+        work_dir = tempfile.mkdtemp(prefix="wo-test-")
+        self._work_dirs.append(work_dir)
+        root = Path(work_dir)
+        wr = root / ".workflow"
+        wr.mkdir()
+        marker = root / "marker.txt"
+        marker.write_text("baseline-a\n")
+        ckpt = wr / "checkpoints" / "a"
+        ckpt.mkdir(parents=True)
+        (ckpt / "marker.txt").write_text("baseline-a\n")
+        a = _node("a")
+        b = _node("b", depends_on=["a"], on_exhaust="rollback", rollback_to="a")
+        tasks = {
+            "a": {"status": "pass", "retries": 0, "pass_seq": 1, "checkpoint_seq": 0},
+            "b": {"status": "fail", "retries": 1, "exhausted": True},
+        }
+        status = {
+            "provider": "dry",
+            "seq": 1,
+            "rollbacks_used": 0,
+            "tasks": tasks,
+            "rollback_pending": {"from": "b", "to": "a"},
+        }
+        status_path = wr / "status.json"
+        status_path.write_text(json.dumps(status))
+        graph = ([(a, "a", []), (b, "b", ["a"])], {"a": a, "b": b}, {})
+
+        self.assertEqual(
+            get_task.execute_rollback(work_dir, status, str(status_path), graph, tasks),
+            0,
+        )
+        self.assertTrue((ckpt / "marker.txt").is_file())
+
+        # Simulate the target's successful redo, then trigger the same rollback again.
+        tasks["a"].update(status="pass", pass_seq=2)
+        tasks["b"].update(status="fail", retries=1, exhausted=True)
+        status["rollback_pending"] = {"from": "b", "to": "a"}
+        marker.write_text("changed-after-first-rollback\n")
+        self.assertEqual(
+            get_task.execute_rollback(work_dir, status, str(status_path), graph, tasks),
+            0,
+        )
+        self.assertEqual(marker.read_text(), "baseline-a\n")
+
+    def test_distinct_rollback_targets_keep_independent_checkpoints(self):
+        """Resetting for target a must retain target b's independent baseline."""
+        work_dir = tempfile.mkdtemp(prefix="wo-test-")
+        self._work_dirs.append(work_dir)
+        root = Path(work_dir)
+        wr = root / ".workflow"
+        wr.mkdir()
+        marker = root / "marker.txt"
+        marker.write_text("baseline-a\n")
+        for target, value in (("a", "baseline-a\n"), ("b", "baseline-b\n")):
+            d = wr / "checkpoints" / target
+            d.mkdir(parents=True)
+            (d / "marker.txt").write_text(value)
+        a = _node("a")
+        b = _node("b", depends_on=["a"])
+        c = _node("c", depends_on=["b"], on_exhaust="rollback", rollback_to="b")
+        tasks = {
+            "a": {"status": "pass", "retries": 0, "pass_seq": 1, "checkpoint_seq": 0},
+            "b": {"status": "pass", "retries": 0, "pass_seq": 2, "checkpoint_seq": 1},
+            "c": {"status": "fail", "retries": 1, "exhausted": True},
+        }
+        status = {
+            "provider": "dry",
+            "seq": 2,
+            "rollbacks_used": 0,
+            "tasks": tasks,
+            "rollback_pending": {"from": "c", "to": "a"},
+        }
+        status_path = wr / "status.json"
+        status_path.write_text(json.dumps(status))
+        graph = (
+            [(a, "a", []), (b, "b", ["a"]), (c, "c", ["b"])],
+            {"a": a, "b": b, "c": c},
+            {},
+        )
+
+        self.assertEqual(
+            get_task.execute_rollback(work_dir, status, str(status_path), graph, tasks),
+            0,
+        )
+        self.assertTrue((wr / "checkpoints" / "b" / "marker.txt").is_file())
+
+        tasks["a"].update(status="pass", pass_seq=3)
+        tasks["b"].update(status="pass", pass_seq=4)
+        tasks["c"].update(status="fail", retries=1, exhausted=True)
+        status["rollback_pending"] = {"from": "c", "to": "b"}
+        marker.write_text("changed-before-second-target\n")
+        self.assertEqual(
+            get_task.execute_rollback(work_dir, status, str(status_path), graph, tasks),
+            0,
+        )
+        self.assertEqual(marker.read_text(), "baseline-b\n")
+
+    def test_rollback_preserves_preexisting_non_target_checkpoint(self):
+        """Rollback must retain checkpoints and metadata for non-target tasks."""
+        work_dir = tempfile.mkdtemp(prefix="wo-test-")
+        self._work_dirs.append(work_dir)
+        root = Path(work_dir)
+        wr = root / ".workflow"
+        wr.mkdir()
+        for target, value in (("a", "baseline-a\n"), ("b", "baseline-b\n")):
+            d = wr / "checkpoints" / target
+            d.mkdir(parents=True)
+            (d / "marker.txt").write_text(value)
+        a = _node("a")
+        b = _node("b", depends_on=["a"])
+        c = _node("c", depends_on=["b"], on_exhaust="rollback", rollback_to="a")
+        tasks = {
+            "a": {"status": "pass", "retries": 0, "pass_seq": 1, "checkpoint_seq": 0},
+            "b": {"status": "pass", "retries": 0, "pass_seq": 2, "checkpoint_seq": 1},
+            "c": {"status": "fail", "retries": 1, "exhausted": True},
+        }
+        status = {
+            "provider": "dry",
+            "seq": 2,
+            "rollbacks_used": 0,
+            "tasks": tasks,
+            "rollback_pending": {"from": "c", "to": "a"},
+        }
+        status_path = wr / "status.json"
+        status_path.write_text(json.dumps(status))
+        graph = (
+            [(a, "a", []), (b, "b", ["a"]), (c, "c", ["b"])],
+            {"a": a, "b": b, "c": c},
+            {},
+        )
+
+        self.assertEqual(
+            get_task.execute_rollback(work_dir, status, str(status_path), graph, tasks),
+            0,
+        )
+        self.assertEqual(tasks["b"].get("checkpoint_seq"), 1)
+        self.assertTrue((wr / "checkpoints" / "b" / "marker.txt").is_file())
+
+    def test_child_rollback_target_uses_namespaced_checkpoint(self):
+        """A bare child rollback_to resolves to and restores sg/x's checkpoint."""
+        work_dir = tempfile.mkdtemp(prefix="wo-test-")
+        self._work_dirs.append(work_dir)
+        root = Path(work_dir)
+        wr = root / ".workflow"
+        wr.mkdir()
+        marker = root / "marker.txt"
+        marker.write_text("child-baseline\n")
+        ckpt = wr / "checkpoints" / "sg" / "x"
+        ckpt.mkdir(parents=True)
+        (ckpt / "marker.txt").write_text("child-baseline\n")
+        x = _node("x")
+        y = _node("y", depends_on=["x"], on_exhaust="rollback", rollback_to="x")
+        tasks = {
+            "sg/x": {
+                "status": "pass",
+                "retries": 0,
+                "pass_seq": 1,
+                "checkpoint_seq": 0,
+            },
+            "sg/y": {"status": "fail", "retries": 1, "exhausted": True},
+        }
+        status = {
+            "provider": "dry",
+            "seq": 1,
+            "rollbacks_used": 0,
+            "tasks": tasks,
+            "rollback_pending": {"from": "sg/y", "to": "sg/x"},
+        }
+        status_path = wr / "status.json"
+        status_path.write_text(json.dumps(status))
+        graph = (
+            [(x, "sg/x", []), (y, "sg/y", ["sg/x"])],
+            {"sg": {"id": "sg", "task_type": "subgraph", "depends_on": []}},
+            {"sg": [x, y]},
+        )
+
+        self.assertEqual(
+            get_task.execute_rollback(work_dir, status, str(status_path), graph, tasks),
+            0,
+        )
+        self.assertTrue(ckpt.is_dir())
+        marker.write_text("changed-child\n")
+        tasks["sg/x"].update(status="pass", pass_seq=2)
+        tasks["sg/y"].update(status="fail", retries=1, exhausted=True)
+        status["rollback_pending"] = {"from": "sg/y", "to": "sg/x"}
+        self.assertEqual(
+            get_task.execute_rollback(work_dir, status, str(status_path), graph, tasks),
+            0,
+        )
+        self.assertEqual(marker.read_text(), "child-baseline\n")
+
+    def test_checkpoint_targets_ignore_non_rollback_nodes(self):
+        nodes = [_node("a", on_exhaust="exit", rollback_to="ghost")]
+        self.assertEqual(get_task.collect_checkpoint_targets(nodes, {}), set())
+
+    def test_checkpoint_creation_is_idempotent(self):
+        """Repeated task polls preserve the original rollback baseline."""
+        with tempfile.TemporaryDirectory() as wd:
+            root = Path(wd)
+            workflow = root / ".workflow"
+            workflow.mkdir()
+            marker = root / "marker.txt"
+            marker.write_text("baseline\n")
+            status = {"seq": 4, "tasks": {"a": {"status": "pending", "retries": 0}}}
+            status_path = workflow / "status.json"
+            status_path.write_text(json.dumps(status))
+
+            get_task.create_checkpoints(wd, status, str(status_path), ["a"], {"a"})
+            marker.write_text("changed\n")
+            get_task.create_checkpoints(wd, status, str(status_path), ["a"], {"a"})
+
+            self.assertEqual(
+                (workflow / "checkpoints" / "a" / "marker.txt").read_text(),
+                "baseline\n",
+            )
+            self.assertEqual(status["tasks"]["a"]["checkpoint_seq"], 4)
+
+    def test_get_task_creates_checkpoint_without_orchestrator(self):
+        nodes = [
+            _node("a"),
+            _node("b", depends_on=["a"], on_exhaust="rollback", rollback_to="a"),
+        ]
+        wd, wf, wr = self._legacy_work_dir(
+            nodes,
+            {
+                "a": {"status": "pending", "retries": 0},
+                "b": {"status": "pending", "retries": 0},
+            },
+            seq=0,
+        )
+        marker = Path(wd) / "marker.txt"
+        marker.write_text("original")
+        first = self._poll_advice(wd)
+        self.assertEqual(first[0]["task_id"], "a")
+        checkpoint = wr / "checkpoints" / "a" / "marker.txt"
+        self.assertEqual(checkpoint.read_text(), "original")
+        self.assertEqual(read_status(wd)["tasks"]["a"]["checkpoint_seq"], 0)
+        self.assertEqual(read_status(wd)["tasks"]["a"]["status"], "pending")
+        marker.write_text("changed after polling")
+        self.assertEqual(self._poll_advice(wd), first)
+        self.assertEqual(checkpoint.read_text(), "original")
+        self.assertFalse((wr / "checkpoints" / "b").exists())
+
+
+class BuildPromptTest(unittest.TestCase):
+    """get_task.build_prompt 的白盒单测（纯函数；orchestrator 本体仍是黑盒）。"""
+
+    NODE = {
+        "id": "a",
+        "title": "t",
+        "goal": ["g"],
+        "approach": ["ap"],
+        "acceptance": ["ac"],
+        "out_of_scope": ["o"],
+    }
+
+    def test_execute_prompt_includes_advice(self):
+        with tempfile.TemporaryDirectory() as wd:
+            advice = Path(wd) / "advice.md"
+            advice.write_text("Investigate the original failure before retrying.")
+            p = get_task.build_prompt(
+                self.NODE,
+                "execute",
+                wd,
+                "/up",
+                get_task.PromptContext(advice=str(advice)),
+            )
+            self.assertIn("Rollback-Advice:", p)
+            self.assertIn(advice.read_text(), p)
+
+    def test_execute_prompt_includes_latest_verdict_failure_reason(self):
+        with tempfile.TemporaryDirectory() as wd:
+            verdict = Path(verdict_file_path(wd, "a"))
+            verdict.parent.mkdir(parents=True)
+            verdict.write_text(json.dumps({"verdict": "fail", "reason": "报告未生成"}))
+            reason = get_task.read_verdict_failure_reason(wd, "a")
+            p = get_task.build_prompt(
+                self.NODE,
+                "execute",
+                wd,
+                "/up",
+                get_task.PromptContext(failure_reason=reason),
+            )
+            self.assertIn("Previous-Verification-Failure:", p)
+            self.assertIn("报告未生成", p)
+
+    def test_verify_prompt_omits_verdict_failure_reason(self):
+        p = get_task.build_prompt(
+            self.NODE,
+            "verify",
+            "/wd",
+            "/up",
+            get_task.PromptContext(failure_reason="报告未生成"),
+        )
+        self.assertNotIn("Previous-Verification-Failure", p)
+
+    def test_missing_or_non_fail_verdict_has_no_failure_reason(self):
+        with tempfile.TemporaryDirectory() as wd:
+            self.assertIsNone(get_task.read_verdict_failure_reason(wd, "a"))
+            verdict = Path(verdict_file_path(wd, "a"))
+            verdict.parent.mkdir(parents=True)
+            verdict.write_text(json.dumps({"verdict": "pass"}))
+            self.assertIsNone(get_task.read_verdict_failure_reason(wd, "a"))
+            verdict.write_text("not json")
+            self.assertIsNone(get_task.read_verdict_failure_reason(wd, "a"))
+
+    def test_verify_prompt_omits_advice(self):
+        p = get_task.build_prompt(
+            self.NODE,
+            "verify",
+            "/wd",
+            "/up",
+            get_task.PromptContext(advice="/wd/.workflow/advice/x.md"),
+        )
+        self.assertNotIn("Rollback-Advice", p)
+
+    def test_prompt_without_advice_unchanged(self):
+        p = get_task.build_prompt(self.NODE, "execute", "/wd", "/up")
+        self.assertNotIn("Rollback-Advice", p)
+
+    def test_execute_prompt_includes_system_prompt(self):
+        p = get_task.build_prompt(
+            self.NODE,
+            "execute",
+            "/wd",
+            "/up",
+            get_task.PromptContext(system_prompt="全局指令"),
+        )
+        self.assertIn("$SYSTEM_PROMPT=全局指令", p)
+        self.assertIn("$WORK_DIR=/wd", p)
+
+    def test_verify_prompt_includes_system_prompt(self):
+        p = get_task.build_prompt(
+            self.NODE,
+            "verify",
+            "/wd",
+            "/up",
+            get_task.PromptContext(system_prompt="全局指令"),
+        )
+        self.assertIn("$SYSTEM_PROMPT=全局指令", p)
+
+    def test_verify_prompt_embeds_verdict_commands(self):
+        p = get_task.build_prompt(self.NODE, "verify", "/wd", "/up")
+        scripts = get_task.orch.SCRIPTS_DIR
+        self.assertIn(
+            'python3 "%s" --work-dir "/wd" --task-id "a"'
+            % (scripts + "/verdict_pass.py"),
+            p,
+        )
+        self.assertIn(
+            'python3 "%s" --work-dir "/wd" --task-id "a"'
+            ' --reason "<why acceptance fails>"' % (scripts + "/verdict_fail.py"),
+            p,
+        )
+        self.assertIn("The verdict file is required; a text reply is not a verdict.", p)
+        self.assertNotIn("After verifying, reply with only", p)
+
+    def test_verify_prompt_uses_namespaced_task_id(self):
+        p = get_task.build_prompt(
+            self.NODE, "verify", "/wd", "/up", get_task.PromptContext(task_id="sub/a")
+        )
+        self.assertIn('--task-id "sub/a"', p)
+
+    def test_execute_prompt_has_no_verdict_commands(self):
+        p = get_task.build_prompt(self.NODE, "execute", "/wd", "/up")
+        self.assertNotIn("verdict_", p)
+        self.assertNotIn("reply with only", p)
+
+    def test_prompt_without_system_prompt_unchanged(self):
+        p = get_task.build_prompt(self.NODE, "execute", "/wd", "/up")
+        self.assertNotIn("$SYSTEM_PROMPT=", p)
+        self.assertTrue(p.startswith("$WORK_DIR=/wd"))
+
+
+class ValidateSystemPromptTest(unittest.TestCase):
+    """validate_workflow 白盒校验顶层可选键 system_prompt。"""
+
+    BASE = {
+        "workflow": "wf",
+        "max_parallel": 1,
+        "nodes": [
+            {
+                "id": "a",
+                "task_type": "normal",
+                "title": "t",
+                "goal": ["g"],
+                "approach": ["ap"],
+                "acceptance": ["ac"],
+                "out_of_scope": ["o"],
+                "depends_on": [],
+                "executor": "e",
+                "verifier": "v",
+                "max_retries": 0,
+                "on_exhaust": "exit",
+            }
+        ],
+    }
+
+    def test_absent_ok(self):
+        self.assertIsNone(get_task.validate_workflow(self.BASE))
+
+    def test_non_empty_str_ok(self):
+        self.assertIsNone(
+            get_task.validate_workflow(dict(self.BASE, system_prompt="全局指令"))
+        )
+
+    def test_invalid_rejected(self):
+        for bad in (123, ["x"], "", "   "):
+            err = get_task.validate_workflow(dict(self.BASE, system_prompt=bad))
+            self.assertIsNotNone(err, "system_prompt=%r 应校验失败" % (bad,))
+
+
+class VerdictFileTest(unittest.TestCase):
+    """verdict 裁决文件消费（黑盒 dry-run）。
+
+    $VERDICT 令牌模拟 verifier 调脚本写裁决文件；编排器以裁决文件为权威，不使用文本兜底。
+    """
+
+    @staticmethod
+    def _events(wd, name):
+        return [e for e in read_log(wd) if e.get("event") == name]
+
+    def test_verdict_file_pass_with_non_compliant_text(self):
+        code, out, wd = self._run(
+            nodes=[_node("t1")], dry_replies={"t1": ["executed", "$VERDICT:pass"]}
+        )
+        self.assertEqual(code, 0, out)
+        self.assertEqual(read_status(wd)["tasks"]["t1"]["status"], "pass")
+        verdicts = self._events(wd, "verdict")
+        self.assertEqual(len(verdicts), 1)
+        self.assertEqual(verdicts[0]["verdict"], "pass")
+        self.assertFalse(self._events(wd, "reset"))  # 一次验证即过，无重验证
+        self.assertTrue(Path(verdict_file_path(wd, "t1")).exists())  # 保留裁决文件
+
+    def test_verdict_file_fail_with_reason(self):
+        code, out, wd = self._run(
+            nodes=[_node("t1")],  # max_retries=0 → fail 即耗尽 → on_exhaust=exit
+            dry_replies={"t1": ["executed", "$VERDICT:fail:报告未生成"]},
+        )
+        self.assertEqual(code, 1, out)
+        task = read_status(wd)["tasks"]["t1"]
+        self.assertEqual(task["status"], "fail")
+        self.assertTrue(task["exhausted"])
+        verdicts = self._events(wd, "verdict")
+        self.assertEqual(len(verdicts), 1)
+        self.assertEqual(verdicts[0]["verdict"], "fail")
+        self.assertEqual(verdicts[0]["reason"], "报告未生成")
+
+    def test_missing_verdict_triggers_reverification(self):
+        code, out, wd = self._run(
+            nodes=[_node("t1")], dry_replies={"t1": ["executed", "通过"]}
+        )
+        # 缺失裁决 → 滞留 verifying → reset 自愈重验证（dry 默认写 pass 裁决）。
+        self.assertEqual(code, 0, out)
+        self.assertEqual(read_status(wd)["tasks"]["t1"]["status"], "pass")
+        self.assertTrue(self._events(wd, "reset"))
+        self.assertEqual(len(self._events(wd, "verdict")), 1)
+
+    def test_stale_verdict_file_cleared_on_dispatch(self):
+        code, out, wd = self._run(
+            nodes=[_node("t1")],
+            pre_files={
+                os.path.relpath(
+                    verdict_file_path("/", "t1"), "/"
+                ): '{"verdict": "fail"}\n'
+            },
+        )
+        # 陈旧裁决文件先归档；dry 默认生成新的 pass 裁决。
+        self.assertEqual(code, 0, out)
+        self.assertEqual(read_status(wd)["tasks"]["t1"]["status"], "pass")
+        self.assertEqual(
+            json.loads(Path(verdict_file_path(wd, "t1")).read_text())["verdict"], "pass"
+        )
+        history = Path(wd) / ".workflow" / "verdicts" / "history"
+        archived = list(history.glob("t1.*.json"))
+        self.assertEqual(len(archived), 1)
+        self.assertRegex(
+            archived[0].name, r"^t1\.20[0-9]{6}T[0-9]{6}\.[0-9]{6}Z\.json$"
+        )
+        self.assertEqual(len(self._events(wd, "verdict")), 1)
+
+    def test_verdict_archive_failure_stops_dispatch(self):
+        code, out, wd = self._run(
+            nodes=[_node("t1")],
+            pre_files={
+                os.path.relpath(
+                    verdict_file_path("/", "t1"), "/"
+                ): '{"verdict": "fail"}\n',
+                ".workflow/verdicts/history": "not a directory\n",
+            },
+        )
+        self.assertEqual(code, 2, out)
+        self.assertIn("裁决文件归档失败", out)
+        self.assertEqual(read_status(wd)["tasks"]["t1"]["status"], "executed")
+
+    def test_invalid_verdict_token_fails_fast(self):
+        code, out, _ = self._run(
+            nodes=[_node("t1")], dry_replies={"t1": ["executed", "$VERDICT:bogus"]}
+        )
+        # 非法 verdict 令牌属测试数据 bug，fail fast（与非法 $SLEEP 令牌同款语义）
+        self.assertEqual(code, 2, out)
+        self.assertIn("非法 verdict", out)
+
+    def test_verdict_file_wins_over_contradictory_text(self):
+        code, out, wd = self._run(
+            nodes=[_node("t1")],
+            dry_replies={"t1": ["executed", "$VERDICT:fail:报告缺失|pass"]},
+        )
+        # 文本谎称 pass，裁决文件为 fail：忽略文本，只消费裁决文件。
+        self.assertEqual(code, 1, out)
+        self.assertEqual(read_status(wd)["tasks"]["t1"]["status"], "fail")
+        verdicts = self._events(wd, "verdict")
+        self.assertEqual(len(verdicts), 1)
+        self.assertEqual(verdicts[0]["verdict"], "fail")
+        self.assertEqual(verdicts[0]["reason"], "报告缺失")
+        self.assertNotIn("warning", verdicts[0])
+
+    def _run(self, **kwargs):
+        code, out, wd = run_orchestrator(**kwargs)
+        self._work_dirs.append(wd)
+        return code, out, wd
+
+    def setUp(self):
+        self._work_dirs = []
+
+    def tearDown(self):
+        for d in self._work_dirs:
+            shutil.rmtree(d, ignore_errors=True)
+
+
+if __name__ == "__main__":
+    unittest.main()