From 7e417c8f8b3cf37bf598ef07aa526e92d9623409 Mon Sep 17 00:00:00 2001
From: JiahuiGu <jiahuigu@sjtu.edu.cn>
Date: Wed, 4 Nov 2026 20:34:41 +0800
Subject: [PATCH 1/6] fix(samples): use latest task in LangGraph HITL
 conversations

---
 .../07-human-in-the-loop/test-spec.yml        | 32 +++++++++++++++++++
 .../responses/07-human-in-the-loop/README.md  | 16 +++++++---
 .../main.py                                   | 20 ++++++++++--
 3 files changed, 60 insertions(+), 8 deletions(-)
 create mode 100644 .azure-pipelines/hosted-agent-tests/python/langgraph/responses/07-human-in-the-loop/test-spec.yml

diff --git a/.azure-pipelines/hosted-agent-tests/python/langgraph/responses/07-human-in-the-loop/test-spec.yml b/.azure-pipelines/hosted-agent-tests/python/langgraph/responses/07-human-in-the-loop/test-spec.yml
new file mode 100644
index 000000000..277c07a27
--- /dev/null
+++ b/.azure-pipelines/hosted-agent-tests/python/langgraph/responses/07-human-in-the-loop/test-spec.yml
@@ -0,0 +1,32 @@
+version: 1
+
+sample:
+  owner: jiahuigu
+  experiences: [azd, vscode]
+
+tests:
+  - name: approved-tasks-use-latest-user-message
+    turns:
+      - input: Draft a one-sentence launch email containing the exact marker SEATTLE-LAUNCH-071. Do not include any other city or event marker.
+        approvals:
+          mcp:
+            - server_label: langgraph
+              name: __hosted_agent_adapter_interrupt__
+        assertions:
+          - source: assistant_text
+            type: contains
+            value: SEATTLE-LAUNCH-071
+      - input: This is a new task. Draft a one-sentence invitation containing the exact marker PARIS-EVENT-072. Do not mention the previous task or its marker.
+        approvals:
+          mcp:
+            - server_label: langgraph
+              name: __hosted_agent_adapter_interrupt__
+        assertions:
+          - source: assistant_text
+            type: contains
+            value: PARIS-EVENT-072
+          - source: assistant_text
+            type: contains
+            value: SEATTLE-LAUNCH-071
+            min_matches: 0
+            max_matches: 0
diff --git a/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md b/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md
index 6aa494564..d4ecfa74e 100644
--- a/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md
+++ b/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md
@@ -7,11 +7,11 @@ A [LangGraph](https://langchain-ai.github.io/langgraph/) **human-in-the-loop** a
 ### Graph shape
 
 ```
-START → draft → await_approval ──approve──► END (emit final draft)
-                       │
-                       └──revise(feedback)──► draft (loop)
-                       │
-                       └──reject──► host returns response.failed (interrupt_rejected)
+START → start_task → draft → await_approval ──approve──► END (emit final draft)
+                                 │
+                                 └──revise(feedback)──► draft (loop)
+                                 │
+                                 └──reject──► host returns response.failed (interrupt_rejected)
 ```
 
 The graph state declares three channels:
@@ -20,6 +20,12 @@ The graph state declares three channels:
 - **`draft`** — the current proposal text under review.
 - **`revision_history`** — list of `{draft, feedback}` pairs from prior revision rounds; each round seeds the next `draft` call with the rejected draft and the reviewer's feedback.
 
+After completing a task, a new user message in the same conversation starts a new
+task. `start_task` clears the previous draft and revision history, while `draft`
+selects the latest user message rather than the first conversation message.
+Revision feedback resumes the existing task directly and retains its revision
+history; it does not pass through `start_task`.
+
 Routing out of `await_approval` is done by returning a [`Command(goto=...)`](https://langchain-ai.github.io/langgraph/how-tos/command/) instead of static edges, so the same node can finalize, loop, or stay paused depending on the resume value.
 
 State is persisted by `FoundryCheckpointSaver` and keyed by the `conversation.id` from the Responses request, so follow-up requests continue the paused run even after a process restart. The saver uses Foundry State Store when hosted and a local file-backed store during local development. See [main.py](src/langgraph-human-in-the-loop-responses/main.py) for the full implementation.
diff --git a/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/src/langgraph-human-in-the-loop-responses/main.py b/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/src/langgraph-human-in-the-loop-responses/main.py
index a7391bd15..22fb88690 100644
--- a/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/src/langgraph-human-in-the-loop-responses/main.py
+++ b/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/src/langgraph-human-in-the-loop-responses/main.py
@@ -92,11 +92,23 @@ class State(TypedDict):
 
 
 def _build_graph(model: ChatOpenAI):
+    def start_task(state: State) -> dict:
+        return {"draft": "", "revision_history": []}
+
     async def draft(state: State) -> dict:
-        task = state["messages"][0].content
+        task = next(
+            (
+                message
+                for message in reversed(state["messages"])
+                if isinstance(message, HumanMessage)
+            ),
+            None,
+        )
+        if task is None:
+            raise ValueError("A user message is required to draft a proposal")
         msgs: list[BaseMessage] = [
             SystemMessage(content=_DRAFT_PROMPT),
-            HumanMessage(content=f"Task: {task}"),
+            HumanMessage(content=task.content),
         ]
         for rev in state.get("revision_history", []):
             msgs.append(AIMessage(content=rev["draft"]))
@@ -125,9 +137,11 @@ def await_approval(state: State) -> Command:
         )
 
     builder = StateGraph(State)
+    builder.add_node("start_task", start_task)
     builder.add_node("draft", draft)
     builder.add_node("await_approval", await_approval)
-    builder.add_edge(START, "draft")
+    builder.add_edge(START, "start_task")
+    builder.add_edge("start_task", "draft")
     builder.add_edge("draft", "await_approval")
     checkpointer = FoundryCheckpointSaver(user_isolation=True)
     return builder.compile(checkpointer=checkpointer)

From 5b80ac603fe17d591d8b0eef16d5dc1301e83f30 Mon Sep 17 00:00:00 2001
From: JiahuiGu <jiahuigu@sjtu.edu.cn>
Date: Thu, 5 Nov 2026 00:29:19 +0800
Subject: [PATCH 2/6] test(samples): cover HITL revision lifecycle over
 Responses

---
 .azure-pipelines/hosted-agent-tests/README.md |  20 ++
 .../tests/test_langgraph_sample_lifecycle.py  | 198 ++++++++++++++++++
 2 files changed, 218 insertions(+)
 create mode 100644 .azure-pipelines/scripts/hosted-agents/tests/test_langgraph_sample_lifecycle.py

diff --git a/.azure-pipelines/hosted-agent-tests/README.md b/.azure-pipelines/hosted-agent-tests/README.md
index e4d5aa100..cca8a6523 100644
--- a/.azure-pipelines/hosted-agent-tests/README.md
+++ b/.azure-pipelines/hosted-agent-tests/README.md
@@ -147,6 +147,26 @@ as rejecting MCP approval policies for Invocations. The PR policy applies both
 checks to new sample contracts. Cloud runners must validate and plan the contracts
 they consume before executing their turns.
 
+### Local sample lifecycle regression checks
+
+The shared tests directory also contains deterministic lifecycle checks against
+the actual sample graph and local Responses HTTP host. Run the HITL check from
+the repository root using the sample runtime dependencies:
+
+```bash
+uv run --frozen --project samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/src/langgraph-human-in-the-loop-responses python .azure-pipelines/scripts/hosted-agents/tests/test_langgraph_sample_lifecycle.py
+```
+
+This standard-library `unittest` check sends revision feedback through the rich
+`function_call_output` resume channel, approves the revised draft, then submits
+a new task in the same conversation. It asserts the exact model inputs: the
+revision includes its draft and feedback, while the new task includes neither.
+Mutation checks ensure that removing history reset or feedback application is
+detected at the corresponding assertion. The model is scripted and the graph
+checkpointer is in-memory; no Azure credentials or model calls are required.
+These are local regression checks, not cloud or UI acceptance. They require the
+sample runtime environment and are separate from the policy-only test suite.
+
 ## Complete document shape
 
 ```yaml
diff --git a/.azure-pipelines/scripts/hosted-agents/tests/test_langgraph_sample_lifecycle.py b/.azure-pipelines/scripts/hosted-agents/tests/test_langgraph_sample_lifecycle.py
new file mode 100644
index 000000000..c20b3e66b
--- /dev/null
+++ b/.azure-pipelines/scripts/hosted-agents/tests/test_langgraph_sample_lifecycle.py
@@ -0,0 +1,198 @@
+"""Local sample lifecycle regressions using runtime dependencies, not Azure access.
+
+Run from the repository root in the HITL sample's Python environment:
+    python .azure-pipelines/scripts/hosted-agents/tests/test_langgraph_sample_lifecycle.py
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import json
+import os
+import sys
+import tempfile
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+from langchain_azure_ai.agents.hosting import ResponsesHostServer
+from langchain_core.messages import AIMessage, HumanMessage
+from langgraph.checkpoint.memory import InMemorySaver
+from starlette.testclient import TestClient
+
+ROOT = Path(__file__).resolve().parents[4]
+SAMPLE = ROOT / (
+    "samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/"
+    "src/langgraph-human-in-the-loop-responses/main.py"
+)
+
+
+class LangGraphSampleLifecycleTests(unittest.TestCase):
+    def exercise_lifecycle(self, source: str) -> None:
+        state_root = self.enterContext(tempfile.TemporaryDirectory())
+        source_path = Path(state_root) / "main.py"
+        source_path.write_text(source, encoding="utf-8")
+        spec = importlib.util.spec_from_file_location(
+            "hitl_sample_lifecycle", source_path
+        )
+        self.assertIsNotNone(spec)
+        self.assertIsNotNone(spec.loader)
+        module = importlib.util.module_from_spec(spec)
+        model = types.SimpleNamespace(
+            ainvoke=AsyncMock(
+                side_effect=[
+                    AIMessage(content="SEATTLE-DRAFT-ORIGINAL"),
+                    AIMessage(content="SEATTLE-DRAFT-REVISED"),
+                    AIMessage(content="PARIS-DRAFT"),
+                ]
+            )
+        )
+        with (
+            patch.dict(sys.modules, {module.__name__: module}),
+            patch.dict(
+                os.environ,
+                {"AGENTSERVER_STATE_ROOT": state_root, "OTEL_SDK_DISABLED": "true"},
+            ),
+            patch("dotenv.load_dotenv"),
+        ):
+            spec.loader.exec_module(module)
+            with patch.object(
+                module, "FoundryCheckpointSaver", return_value=InMemorySaver()
+            ):
+                graph = module._build_graph(model)
+            with TestClient(ResponsesHostServer(graph).app) as client:
+                conversation = {"id": "hitl-lifecycle"}
+
+                def post(input_items: object) -> dict:
+                    response = client.post(
+                        "/responses",
+                        json={"conversation": conversation, "input": input_items},
+                    )
+                    self.assertEqual(response.status_code, 200, response.text)
+                    payload = response.json()
+                    self.assertEqual(payload["status"], "completed", payload)
+                    return payload
+
+                def pending(payload: dict) -> dict:
+                    requests = [
+                        item
+                        for item in payload["output"]
+                        if item.get("type") == "mcp_approval_request"
+                    ]
+                    self.assertEqual(len(requests), 1, payload)
+                    return requests[0]
+
+                def approve(payload: dict) -> dict:
+                    return post(
+                        [
+                            {
+                                "type": "mcp_approval_response",
+                                "approval_request_id": pending(payload)["id"],
+                                "approve": True,
+                            }
+                        ]
+                    )
+
+                def assistant_text(payload: dict) -> str:
+                    return "".join(
+                        part.get("text", "")
+                        for item in payload["output"]
+                        if item.get("type") == "message"
+                        for part in item.get("content", [])
+                    )
+
+                initial = post("Draft a Seattle launch email.")
+                pending(initial)
+                original_call = model.ainvoke.await_args_list[0]
+                self.assertEqual(
+                    original_call.args[0][1].content, "Draft a Seattle launch email."
+                )
+                sentinels = [
+                    item
+                    for item in initial["output"]
+                    if item.get("type") == "function_call"
+                    and item.get("name") == "__hosted_agent_adapter_interrupt__"
+                ]
+                self.assertEqual(len(sentinels), 1, initial)
+                feedback = "Keep the Seattle launch date, but shorten the email."
+                revised = post(
+                    [
+                        {
+                            "type": "function_call_output",
+                            "call_id": sentinels[0]["call_id"],
+                            "output": json.dumps({"resume": {"feedback": feedback}}),
+                        }
+                    ]
+                )
+                self.assertEqual(
+                    json.loads(pending(revised)["arguments"])["value"]["draft"],
+                    "SEATTLE-DRAFT-REVISED",
+                )
+                self.assertEqual(model.ainvoke.await_count, 2)
+                revision_messages = model.ainvoke.await_args_list[1].args[0]
+                self.assertEqual(
+                    [message.content for message in revision_messages[1:]],
+                    [
+                        "Draft a Seattle launch email.",
+                        "SEATTLE-DRAFT-ORIGINAL",
+                        f"Revision feedback: {feedback}",
+                    ],
+                    "revision must apply the current draft and feedback",
+                )
+                self.assertIsInstance(revision_messages[2], AIMessage)
+                self.assertIsInstance(revision_messages[3], HumanMessage)
+                approved = approve(revised)
+                self.assertFalse(
+                    any(
+                        item.get("type") == "mcp_approval_request"
+                        for item in approved["output"]
+                    ),
+                    approved,
+                )
+                self.assertEqual(model.ainvoke.await_count, 2)
+                self.assertEqual(assistant_text(approved), "SEATTLE-DRAFT-REVISED")
+
+                following = post("Draft a Paris event invitation.")
+                self.assertEqual(model.ainvoke.await_count, 3)
+                following_messages = model.ainvoke.await_args_list[2].args[0]
+                self.assertEqual(
+                    [message.content for message in following_messages[1:]],
+                    ["Draft a Paris event invitation."],
+                    "new task must exclude the previous draft and revision feedback",
+                )
+                self.assertEqual(
+                    json.loads(pending(following)["arguments"])["value"]["draft"],
+                    "PARIS-DRAFT",
+                )
+                final = approve(following)
+                self.assertEqual(model.ainvoke.await_count, 3)
+                self.assertEqual(assistant_text(final), "PARIS-DRAFT")
+                self.assertFalse(
+                    any(
+                        item.get("type") == "mcp_approval_request"
+                        for item in final["output"]
+                    ),
+                    final,
+                )
+
+    def test_revision_then_new_task_isolates_model_context(self) -> None:
+        self.exercise_lifecycle(SAMPLE.read_text(encoding="utf-8"))
+
+    def test_detects_missing_history_reset(self) -> None:
+        source = SAMPLE.read_text(encoding="utf-8")
+        original = 'return {"draft": "", "revision_history": []}'
+        self.assertIn(original, source)
+        with self.assertRaisesRegex(AssertionError, "new task must exclude"):
+            self.exercise_lifecycle(source.replace(original, 'return {"draft": ""}', 1))
+
+    def test_detects_missing_feedback_application(self) -> None:
+        source = SAMPLE.read_text(encoding="utf-8")
+        original = 'for rev in state.get("revision_history", []):'
+        self.assertIn(original, source)
+        with self.assertRaisesRegex(AssertionError, "revision must apply"):
+            self.exercise_lifecycle(source.replace(original, "for rev in []:", 1))
+
+
+if __name__ == "__main__":
+    unittest.main()

From 5e17122a1c8502b5fe4456e9235a026e6df133c2 Mon Sep 17 00:00:00 2001
From: JiahuiGu <jiahuigu@sjtu.edu.cn>
Date: Thu, 5 Nov 2026 01:03:11 +0800
Subject: [PATCH 3/6] fix(ci): isolate and run frozen HITL runtime regressions

---
 .azure-pipelines/hosted-agent-tests/README.md |  24 ++--
 .azure-pipelines/hosted-agents-samples-ci.yml |   6 +
 .../hosted-agents/run_runtime_tests.py        |  42 ++++++
 .../test_langgraph_sample_lifecycle.py        |  29 +---
 .../tests/test_runtime_test_entrypoint.py     | 134 ++++++++++++++++++
 .github/workflows/scripts-selftest.yml        |  16 +++
 6 files changed, 217 insertions(+), 34 deletions(-)
 create mode 100644 .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
 rename .azure-pipelines/scripts/hosted-agents/{tests => runtime-tests}/test_langgraph_sample_lifecycle.py (84%)
 create mode 100644 .azure-pipelines/scripts/hosted-agents/tests/test_runtime_test_entrypoint.py

diff --git a/.azure-pipelines/hosted-agent-tests/README.md b/.azure-pipelines/hosted-agent-tests/README.md
index cca8a6523..b8135b137 100644
--- a/.azure-pipelines/hosted-agent-tests/README.md
+++ b/.azure-pipelines/hosted-agent-tests/README.md
@@ -149,23 +149,29 @@ they consume before executing their turns.
 
 ### Local sample lifecycle regression checks
 
-The shared tests directory also contains deterministic lifecycle checks against
-the actual sample graph and local Responses HTTP host. Run the HITL check from
-the repository root using the sample runtime dependencies:
+Runtime lifecycle checks live in the separate `runtime-tests` directory, outside
+the lightweight `tests/test_*.py` policy discovery path. They exercise the actual
+sample graph and local Responses HTTP host. Run them from the repository root:
 
 ```bash
-uv run --frozen --project samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/src/langgraph-human-in-the-loop-responses python .azure-pipelines/scripts/hosted-agents/tests/test_langgraph_sample_lifecycle.py
+python -m pip install "uv==0.11.7"
+python .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
 ```
 
 This standard-library `unittest` check sends revision feedback through the rich
 `function_call_output` resume channel, approves the revised draft, then submits
 a new task in the same conversation. It asserts the exact model inputs: the
 revision includes its draft and feedback, while the new task includes neither.
-Mutation checks ensure that removing history reset or feedback application is
-detected at the corresponding assertion. The model is scripted and the graph
-checkpointer is in-memory; no Azure credentials or model calls are required.
-These are local regression checks, not cloud or UI acceptance. They require the
-sample runtime environment and are separate from the policy-only test suite.
+The model is scripted and the graph checkpointer is in-memory; no Azure
+credentials or model calls are required. No test rewrites or matches the sample's
+source implementation. Mutation checks are development evidence only.
+
+The entrypoint first runs `uv sync --frozen --python 3.13` against the sample's
+committed runtime project, then executes runtime test discovery in that isolated
+environment. Dependency or test failures propagate; missing dependencies are not
+skipped. GitHub scripts self-test runs it in a dedicated credential-free job,
+and ADO runs it in a separate step after lightweight policy tests. These are
+local runtime regressions, not cloud or UI acceptance.
 
 ## Complete document shape
 
diff --git a/.azure-pipelines/hosted-agents-samples-ci.yml b/.azure-pipelines/hosted-agents-samples-ci.yml
index 1b302a497..7a7fe291d 100644
--- a/.azure-pipelines/hosted-agents-samples-ci.yml
+++ b/.azure-pipelines/hosted-agents-samples-ci.yml
@@ -186,6 +186,12 @@ stages:
               fi
             displayName: Test and validate hosted-agent test specs
 
+          - bash: |
+              set -euo pipefail
+              python3 -m pip install --disable-pip-version-check "uv==0.11.7"
+              python3 .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
+            displayName: Run frozen hosted-agent runtime regressions
+
           - bash: .azure-pipelines/scripts/hosted-agents/tests/test-hosted-agent-ci-toolboxes.sh
             displayName: Test hosted-agent CI toolbox lifecycle scripts
 
diff --git a/.azure-pipelines/scripts/hosted-agents/run_runtime_tests.py b/.azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
new file mode 100644
index 000000000..1d9243950
--- /dev/null
+++ b/.azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
@@ -0,0 +1,42 @@
+"""Run hosted-agent runtime regressions outside the lightweight policy suite."""
+
+from __future__ import annotations
+
+import subprocess
+from pathlib import Path
+
+
+def main() -> None:
+    root = Path(__file__).resolve().parents[3]
+    project = root / (
+        "samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/"
+        "src/langgraph-human-in-the-loop-responses"
+    )
+    subprocess.run(
+        ["uv", "sync", "--frozen", "--python", "3.13", "--project", str(project)],
+        cwd=root,
+        check=True,
+    )
+    subprocess.run(
+        [
+            "uv",
+            "run",
+            "--no-sync",
+            "--project",
+            str(project),
+            "python",
+            "-m",
+            "unittest",
+            "discover",
+            "-s",
+            ".azure-pipelines/scripts/hosted-agents/runtime-tests",
+            "-p",
+            "test_*.py",
+        ],
+        cwd=root,
+        check=True,
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/.azure-pipelines/scripts/hosted-agents/tests/test_langgraph_sample_lifecycle.py b/.azure-pipelines/scripts/hosted-agents/runtime-tests/test_langgraph_sample_lifecycle.py
similarity index 84%
rename from .azure-pipelines/scripts/hosted-agents/tests/test_langgraph_sample_lifecycle.py
rename to .azure-pipelines/scripts/hosted-agents/runtime-tests/test_langgraph_sample_lifecycle.py
index c20b3e66b..2c27a1bbe 100644
--- a/.azure-pipelines/scripts/hosted-agents/tests/test_langgraph_sample_lifecycle.py
+++ b/.azure-pipelines/scripts/hosted-agents/runtime-tests/test_langgraph_sample_lifecycle.py
@@ -1,7 +1,7 @@
 """Local sample lifecycle regressions using runtime dependencies, not Azure access.
 
-Run from the repository root in the HITL sample's Python environment:
-    python .azure-pipelines/scripts/hosted-agents/tests/test_langgraph_sample_lifecycle.py
+Run with the sample's frozen runtime dependencies:
+    python .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
 """
 
 from __future__ import annotations
@@ -29,13 +29,9 @@
 
 
 class LangGraphSampleLifecycleTests(unittest.TestCase):
-    def exercise_lifecycle(self, source: str) -> None:
+    def test_revision_then_new_task_isolates_model_context(self) -> None:
         state_root = self.enterContext(tempfile.TemporaryDirectory())
-        source_path = Path(state_root) / "main.py"
-        source_path.write_text(source, encoding="utf-8")
-        spec = importlib.util.spec_from_file_location(
-            "hitl_sample_lifecycle", source_path
-        )
+        spec = importlib.util.spec_from_file_location("hitl_sample_lifecycle", SAMPLE)
         self.assertIsNotNone(spec)
         self.assertIsNotNone(spec.loader)
         module = importlib.util.module_from_spec(spec)
@@ -176,23 +172,6 @@ def assistant_text(payload: dict) -> str:
                     final,
                 )
 
-    def test_revision_then_new_task_isolates_model_context(self) -> None:
-        self.exercise_lifecycle(SAMPLE.read_text(encoding="utf-8"))
-
-    def test_detects_missing_history_reset(self) -> None:
-        source = SAMPLE.read_text(encoding="utf-8")
-        original = 'return {"draft": "", "revision_history": []}'
-        self.assertIn(original, source)
-        with self.assertRaisesRegex(AssertionError, "new task must exclude"):
-            self.exercise_lifecycle(source.replace(original, 'return {"draft": ""}', 1))
-
-    def test_detects_missing_feedback_application(self) -> None:
-        source = SAMPLE.read_text(encoding="utf-8")
-        original = 'for rev in state.get("revision_history", []):'
-        self.assertIn(original, source)
-        with self.assertRaisesRegex(AssertionError, "revision must apply"):
-            self.exercise_lifecycle(source.replace(original, "for rev in []:", 1))
-
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/.azure-pipelines/scripts/hosted-agents/tests/test_runtime_test_entrypoint.py b/.azure-pipelines/scripts/hosted-agents/tests/test_runtime_test_entrypoint.py
new file mode 100644
index 000000000..cdd383848
--- /dev/null
+++ b/.azure-pipelines/scripts/hosted-agents/tests/test_runtime_test_entrypoint.py
@@ -0,0 +1,134 @@
+"""Keep runtime regression execution isolated, frozen, and wired into both CIs."""
+
+from __future__ import annotations
+
+import importlib.util
+import subprocess
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+import yaml
+
+ROOT = Path(__file__).resolve().parents[4]
+SCRIPTS = ROOT / ".azure-pipelines/scripts/hosted-agents"
+SPEC = importlib.util.spec_from_file_location(
+    "hosted_agent_runtime_tests", SCRIPTS / "run_runtime_tests.py"
+)
+assert SPEC and SPEC.loader
+runner = importlib.util.module_from_spec(SPEC)
+SPEC.loader.exec_module(runner)
+
+
+class RuntimeTestEntrypointTests(unittest.TestCase):
+    def test_syncs_frozen_sample_before_running_separate_discovery(self):
+        project = str(
+            ROOT
+            / "samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop"
+            / "src/langgraph-human-in-the-loop-responses"
+        )
+        with patch.object(runner.subprocess, "run") as run:
+            runner.main()
+        self.assertEqual(run.call_count, 2)
+        self.assertEqual(
+            run.call_args_list[0].args[0],
+            ["uv", "sync", "--frozen", "--python", "3.13", "--project", project],
+        )
+        self.assertEqual(
+            run.call_args_list[1].args[0],
+            [
+                "uv",
+                "run",
+                "--no-sync",
+                "--project",
+                project,
+                "python",
+                "-m",
+                "unittest",
+                "discover",
+                "-s",
+                ".azure-pipelines/scripts/hosted-agents/runtime-tests",
+                "-p",
+                "test_*.py",
+            ],
+        )
+        for call in run.call_args_list:
+            self.assertEqual(call.kwargs, {"cwd": ROOT, "check": True})
+        self.assertFalse(
+            (SCRIPTS / "tests/test_langgraph_sample_lifecycle.py").exists()
+        )
+        self.assertTrue(
+            (SCRIPTS / "runtime-tests/test_langgraph_sample_lifecycle.py").is_file()
+        )
+
+    def test_sync_failure_prevents_test_execution(self):
+        with (
+            patch.object(
+                runner.subprocess,
+                "run",
+                side_effect=subprocess.CalledProcessError(1, "uv"),
+            ) as run,
+            self.assertRaises(subprocess.CalledProcessError),
+        ):
+            runner.main()
+        self.assertEqual(run.call_count, 1)
+
+    def test_runtime_failure_propagates(self):
+        with (
+            patch.object(
+                runner.subprocess,
+                "run",
+                side_effect=[None, subprocess.CalledProcessError(1, "unittest")],
+            ),
+            self.assertRaises(subprocess.CalledProcessError),
+        ):
+            runner.main()
+
+    def test_github_and_ado_run_the_entrypoint_without_skipping(self):
+        workflow = yaml.load(
+            (ROOT / ".github/workflows/scripts-selftest.yml").read_text(
+                encoding="utf-8"
+            ),
+            Loader=yaml.BaseLoader,
+        )
+        job = workflow["jobs"]["hosted-agent-runtime-regressions"]
+        self.assertEqual(workflow["permissions"], {"contents": "read"})
+        self.assertIn(
+            "samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/**",
+            workflow["on"]["pull_request"]["paths"],
+        )
+        self.assertEqual(job["steps"][0]["with"]["persist-credentials"], "false")
+        for step in job["steps"]:
+            self.assertNotIn("if", step)
+            self.assertNotIn("continue-on-error", step)
+        self.assertIn('"uv==0.11.7"', job["steps"][-2]["run"])
+        self.assertEqual(
+            job["steps"][-1]["run"],
+            "python .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py",
+        )
+        pipeline = yaml.safe_load(
+            (ROOT / ".azure-pipelines/hosted-agents-samples-ci.yml").read_text(
+                encoding="utf-8"
+            )
+        )
+        discover = next(
+            stage for stage in pipeline["stages"] if stage.get("stage") == "Discover"
+        )
+        steps = discover["jobs"][0]["steps"]
+        runtime = next(
+            step
+            for step in steps
+            if step.get("displayName") == "Run frozen hosted-agent runtime regressions"
+        )
+        self.assertIn('"uv==0.11.7"', runtime["bash"])
+        self.assertIn("set -euo pipefail", runtime["bash"])
+        self.assertIn(
+            "python3 .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py",
+            runtime["bash"],
+        )
+        self.assertNotIn("condition", runtime)
+        self.assertNotIn("continueOnError", runtime)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/.github/workflows/scripts-selftest.yml b/.github/workflows/scripts-selftest.yml
index 1c76e6468..580154b68 100644
--- a/.github/workflows/scripts-selftest.yml
+++ b/.github/workflows/scripts-selftest.yml
@@ -27,11 +27,27 @@ on:
       - '.github/workflows/validation-pilot.yml'
       - '.github/scripts/discover-validation-samples.py'
       - 'samples/*/prompt-agents/**'
+      - 'samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/**'
 
 permissions:
   contents: read
 
 jobs:
+  hosted-agent-runtime-regressions:
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
+        with:
+          python-version: '3.13'
+      - name: Install runtime test resolver
+        run: python -m pip install --disable-pip-version-check "uv==0.11.7"
+      - name: Run frozen hosted-agent runtime regressions
+        run: python .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
+
   ado-pipelines-harness:
     runs-on: ubuntu-latest
     timeout-minutes: 15

From 58042aea57a25d54d8af7084248c590ee09cfb81 Mon Sep 17 00:00:00 2001
From: JiahuiGu <jiahuigu@sjtu.edu.cn>
Date: Tue, 10 Nov 2026 16:14:13 +0800
Subject: [PATCH 4/6] test(samples): scope HITL regressions without changing
 cloud CI

---
 .azure-pipelines/hosted-agent-tests/README.md | 26 -----------
 .azure-pipelines/hosted-agents-samples-ci.yml |  6 ---
 .../scripts/run_hosted_agent_runtime_tests.py |  4 +-
 .../test_langgraph_sample_lifecycle.py        |  4 +-
 .../test_hosted_agent_runtime_entrypoint.py   | 44 ++++++-------------
 .github/workflows/scripts-selftest.yml        |  3 +-
 .../responses/07-human-in-the-loop/README.md  | 20 +++++++++
 7 files changed, 40 insertions(+), 67 deletions(-)
 rename .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py => .github/scripts/run_hosted_agent_runtime_tests.py (88%)
 rename {.azure-pipelines/scripts/hosted-agents => .github/scripts}/runtime-tests/test_langgraph_sample_lifecycle.py (98%)
 rename .azure-pipelines/scripts/hosted-agents/tests/test_runtime_test_entrypoint.py => .github/scripts/test/test_hosted_agent_runtime_entrypoint.py (70%)

diff --git a/.azure-pipelines/hosted-agent-tests/README.md b/.azure-pipelines/hosted-agent-tests/README.md
index 94e558fa4..5d96c1e5f 100644
--- a/.azure-pipelines/hosted-agent-tests/README.md
+++ b/.azure-pipelines/hosted-agent-tests/README.md
@@ -147,32 +147,6 @@ as rejecting MCP approval policies for Invocations. The PR policy applies both
 checks to new sample contracts. Cloud runners must validate and plan the contracts
 they consume before executing their turns.
 
-### Local sample lifecycle regression checks
-
-Runtime lifecycle checks live in the separate `runtime-tests` directory, outside
-the lightweight `tests/test_*.py` policy discovery path. They exercise the actual
-sample graph and local Responses HTTP host. Run them from the repository root:
-
-```bash
-python -m pip install "uv==0.11.7"
-python .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
-```
-
-This standard-library `unittest` check sends revision feedback through the rich
-`function_call_output` resume channel, approves the revised draft, then submits
-a new task in the same conversation. It asserts the exact model inputs: the
-revision includes its draft and feedback, while the new task includes neither.
-The model is scripted and the graph checkpointer is in-memory; no Azure
-credentials or model calls are required. No test rewrites or matches the sample's
-source implementation. Mutation checks are development evidence only.
-
-The entrypoint first runs `uv sync --frozen --python 3.13` against the sample's
-committed runtime project, then executes runtime test discovery in that isolated
-environment. Dependency or test failures propagate; missing dependencies are not
-skipped. GitHub scripts self-test runs it in a dedicated credential-free job,
-and ADO runs it in a separate step after lightweight policy tests. These are
-local runtime regressions, not cloud or UI acceptance.
-
 ## Complete document shape
 
 ```yaml
diff --git a/.azure-pipelines/hosted-agents-samples-ci.yml b/.azure-pipelines/hosted-agents-samples-ci.yml
index 7a7fe291d..1b302a497 100644
--- a/.azure-pipelines/hosted-agents-samples-ci.yml
+++ b/.azure-pipelines/hosted-agents-samples-ci.yml
@@ -186,12 +186,6 @@ stages:
               fi
             displayName: Test and validate hosted-agent test specs
 
-          - bash: |
-              set -euo pipefail
-              python3 -m pip install --disable-pip-version-check "uv==0.11.7"
-              python3 .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
-            displayName: Run frozen hosted-agent runtime regressions
-
           - bash: .azure-pipelines/scripts/hosted-agents/tests/test-hosted-agent-ci-toolboxes.sh
             displayName: Test hosted-agent CI toolbox lifecycle scripts
 
diff --git a/.azure-pipelines/scripts/hosted-agents/run_runtime_tests.py b/.github/scripts/run_hosted_agent_runtime_tests.py
similarity index 88%
rename from .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
rename to .github/scripts/run_hosted_agent_runtime_tests.py
index 1d9243950..10449078c 100644
--- a/.azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
+++ b/.github/scripts/run_hosted_agent_runtime_tests.py
@@ -7,7 +7,7 @@
 
 
 def main() -> None:
-    root = Path(__file__).resolve().parents[3]
+    root = Path(__file__).resolve().parents[2]
     project = root / (
         "samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/"
         "src/langgraph-human-in-the-loop-responses"
@@ -29,7 +29,7 @@ def main() -> None:
             "unittest",
             "discover",
             "-s",
-            ".azure-pipelines/scripts/hosted-agents/runtime-tests",
+            ".github/scripts/runtime-tests",
             "-p",
             "test_*.py",
         ],
diff --git a/.azure-pipelines/scripts/hosted-agents/runtime-tests/test_langgraph_sample_lifecycle.py b/.github/scripts/runtime-tests/test_langgraph_sample_lifecycle.py
similarity index 98%
rename from .azure-pipelines/scripts/hosted-agents/runtime-tests/test_langgraph_sample_lifecycle.py
rename to .github/scripts/runtime-tests/test_langgraph_sample_lifecycle.py
index 2c27a1bbe..7a5d9280d 100644
--- a/.azure-pipelines/scripts/hosted-agents/runtime-tests/test_langgraph_sample_lifecycle.py
+++ b/.github/scripts/runtime-tests/test_langgraph_sample_lifecycle.py
@@ -1,7 +1,7 @@
 """Local sample lifecycle regressions using runtime dependencies, not Azure access.
 
 Run with the sample's frozen runtime dependencies:
-    python .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
+    python .github/scripts/run_hosted_agent_runtime_tests.py
 """
 
 from __future__ import annotations
@@ -21,7 +21,7 @@
 from langgraph.checkpoint.memory import InMemorySaver
 from starlette.testclient import TestClient
 
-ROOT = Path(__file__).resolve().parents[4]
+ROOT = Path(__file__).resolve().parents[3]
 SAMPLE = ROOT / (
     "samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/"
     "src/langgraph-human-in-the-loop-responses/main.py"
diff --git a/.azure-pipelines/scripts/hosted-agents/tests/test_runtime_test_entrypoint.py b/.github/scripts/test/test_hosted_agent_runtime_entrypoint.py
similarity index 70%
rename from .azure-pipelines/scripts/hosted-agents/tests/test_runtime_test_entrypoint.py
rename to .github/scripts/test/test_hosted_agent_runtime_entrypoint.py
index cdd383848..73b10373c 100644
--- a/.azure-pipelines/scripts/hosted-agents/tests/test_runtime_test_entrypoint.py
+++ b/.github/scripts/test/test_hosted_agent_runtime_entrypoint.py
@@ -1,4 +1,4 @@
-"""Keep runtime regression execution isolated, frozen, and wired into both CIs."""
+"""Keep runtime regression execution isolated, frozen, and wired into GitHub CI."""
 
 from __future__ import annotations
 
@@ -10,10 +10,10 @@
 
 import yaml
 
-ROOT = Path(__file__).resolve().parents[4]
-SCRIPTS = ROOT / ".azure-pipelines/scripts/hosted-agents"
+ROOT = Path(__file__).resolve().parents[3]
+SCRIPTS = ROOT / ".github/scripts"
 SPEC = importlib.util.spec_from_file_location(
-    "hosted_agent_runtime_tests", SCRIPTS / "run_runtime_tests.py"
+    "hosted_agent_runtime_tests", SCRIPTS / "run_hosted_agent_runtime_tests.py"
 )
 assert SPEC and SPEC.loader
 runner = importlib.util.module_from_spec(SPEC)
@@ -47,16 +47,14 @@ def test_syncs_frozen_sample_before_running_separate_discovery(self):
                 "unittest",
                 "discover",
                 "-s",
-                ".azure-pipelines/scripts/hosted-agents/runtime-tests",
+                ".github/scripts/runtime-tests",
                 "-p",
                 "test_*.py",
             ],
         )
         for call in run.call_args_list:
             self.assertEqual(call.kwargs, {"cwd": ROOT, "check": True})
-        self.assertFalse(
-            (SCRIPTS / "tests/test_langgraph_sample_lifecycle.py").exists()
-        )
+        self.assertFalse((SCRIPTS / "test/test_langgraph_sample_lifecycle.py").exists())
         self.assertTrue(
             (SCRIPTS / "runtime-tests/test_langgraph_sample_lifecycle.py").is_file()
         )
@@ -84,7 +82,7 @@ def test_runtime_failure_propagates(self):
         ):
             runner.main()
 
-    def test_github_and_ado_run_the_entrypoint_without_skipping(self):
+    def test_github_runs_the_entrypoint_without_skipping(self):
         workflow = yaml.load(
             (ROOT / ".github/workflows/scripts-selftest.yml").read_text(
                 encoding="utf-8"
@@ -104,30 +102,16 @@ def test_github_and_ado_run_the_entrypoint_without_skipping(self):
         self.assertIn('"uv==0.11.7"', job["steps"][-2]["run"])
         self.assertEqual(
             job["steps"][-1]["run"],
-            "python .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py",
-        )
-        pipeline = yaml.safe_load(
-            (ROOT / ".azure-pipelines/hosted-agents-samples-ci.yml").read_text(
-                encoding="utf-8"
-            )
+            "python .github/scripts/run_hosted_agent_runtime_tests.py",
         )
-        discover = next(
-            stage for stage in pipeline["stages"] if stage.get("stage") == "Discover"
-        )
-        steps = discover["jobs"][0]["steps"]
-        runtime = next(
-            step
-            for step in steps
-            if step.get("displayName") == "Run frozen hosted-agent runtime regressions"
-        )
-        self.assertIn('"uv==0.11.7"', runtime["bash"])
-        self.assertIn("set -euo pipefail", runtime["bash"])
         self.assertIn(
-            "python3 .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py",
-            runtime["bash"],
+            "python .github/scripts/test/test_hosted_agent_runtime_entrypoint.py",
+            next(
+                step["run"]
+                for step in workflow["jobs"]["ado-pipelines-harness"]["steps"]
+                if step.get("name") == "Test hosted-agent helpers and fixture contracts"
+            ),
         )
-        self.assertNotIn("condition", runtime)
-        self.assertNotIn("continueOnError", runtime)
 
 
 if __name__ == "__main__":
diff --git a/.github/workflows/scripts-selftest.yml b/.github/workflows/scripts-selftest.yml
index 580154b68..1ec22b43b 100644
--- a/.github/workflows/scripts-selftest.yml
+++ b/.github/workflows/scripts-selftest.yml
@@ -46,7 +46,7 @@ jobs:
       - name: Install runtime test resolver
         run: python -m pip install --disable-pip-version-check "uv==0.11.7"
       - name: Run frozen hosted-agent runtime regressions
-        run: python .azure-pipelines/scripts/hosted-agents/run_runtime_tests.py
+        run: python .github/scripts/run_hosted_agent_runtime_tests.py
 
   ado-pipelines-harness:
     runs-on: ubuntu-latest
@@ -66,6 +66,7 @@ jobs:
       - name: Test hosted-agent helpers and fixture contracts
         run: |
           python -m unittest discover -s .azure-pipelines/scripts/hosted-agents/tests -p 'test_*.py'
+          python .github/scripts/test/test_hosted_agent_runtime_entrypoint.py
           bash .azure-pipelines/scripts/hosted-agents/tests/test-hosted-agent-ci-toolboxes.sh
           bash .azure-pipelines/scripts/hosted-agents/tests/test-hosted-agent-session-quota.sh
           python .github/scripts/test/test_ado_pipeline_fixtures.py
diff --git a/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md b/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md
index d4ecfa74e..f2463d218 100644
--- a/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md
+++ b/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md
@@ -47,6 +47,26 @@ The Responses host emits two paired output items for each pause, both keyed by t
 
 The graph factory is registered in `langgraph.json` and loaded by `langchain_azure_ai.agents.hosting.run`, which exposes the OpenAI-compatible Responses endpoint at `/responses` and handles conversation history, interrupt serialization, and streaming lifecycle events automatically.
 
+### Lifecycle regression check
+
+From the repository root, run the shared runtime regression entrypoint:
+
+```bash
+python -m pip install "uv==0.11.7"
+python .github/scripts/run_hosted_agent_runtime_tests.py
+```
+
+It installs this sample's committed dependencies with `uv sync --frozen` before
+running the actual graph through a local Responses HTTP host. A scripted model
+lets the test assert that rich `function_call_output` revision feedback reaches
+the model, and that a new task after approval receives neither the prior draft
+nor its revision feedback. No Azure credentials or model calls are needed.
+
+GitHub scripts self-test runs this in a dedicated runtime job, separate from
+lightweight policy tests. Missing dependencies and test failures are not skipped.
+The cloud contract separately tests latest-task selection in Code and Container
+deployments; this local regression is not cloud or UI acceptance.
+
 ## Option 1: Azure Developer CLI (`azd`)
 
 ### Prerequisites

From efb4ec2aefb9ffc5bc5cedc7f251e2be22bf8e79 Mon Sep 17 00:00:00 2001
From: JiahuiGu <jiahuigu@sjtu.edu.cn>
Date: Fri, 13 Nov 2026 14:53:25 +0800
Subject: [PATCH 5/6] docs(samples): remove internal regression details from
 HITL README

---
 .../responses/07-human-in-the-loop/README.md  | 20 -------------------
 1 file changed, 20 deletions(-)

diff --git a/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md b/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md
index f2463d218..d4ecfa74e 100644
--- a/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md
+++ b/samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/README.md
@@ -47,26 +47,6 @@ The Responses host emits two paired output items for each pause, both keyed by t
 
 The graph factory is registered in `langgraph.json` and loaded by `langchain_azure_ai.agents.hosting.run`, which exposes the OpenAI-compatible Responses endpoint at `/responses` and handles conversation history, interrupt serialization, and streaming lifecycle events automatically.
 
-### Lifecycle regression check
-
-From the repository root, run the shared runtime regression entrypoint:
-
-```bash
-python -m pip install "uv==0.11.7"
-python .github/scripts/run_hosted_agent_runtime_tests.py
-```
-
-It installs this sample's committed dependencies with `uv sync --frozen` before
-running the actual graph through a local Responses HTTP host. A scripted model
-lets the test assert that rich `function_call_output` revision feedback reaches
-the model, and that a new task after approval receives neither the prior draft
-nor its revision feedback. No Azure credentials or model calls are needed.
-
-GitHub scripts self-test runs this in a dedicated runtime job, separate from
-lightweight policy tests. Missing dependencies and test failures are not skipped.
-The cloud contract separately tests latest-task selection in Code and Container
-deployments; this local regression is not cloud or UI acceptance.
-
 ## Option 1: Azure Developer CLI (`azd`)
 
 ### Prerequisites

From 06f2a9c5f7d6c1f41520e7ce3e27d8c698b8ef35 Mon Sep 17 00:00:00 2001
From: JiahuiGu <jiahuigu@sjtu.edu.cn>
Date: Fri, 13 Nov 2026 15:12:18 +0800
Subject: [PATCH 6/6] test(samples): simplify and isolate HITL lifecycle CI

---
 .../scripts/run_hosted_agent_runtime_tests.py |  42 -------
 .../test_langgraph_sample_lifecycle.py        |   4 +-
 .../test_hosted_agent_runtime_entrypoint.py   | 118 ------------------
 .../workflows/hosted-agent-runtime-tests.yml  |  32 +++++
 .github/workflows/scripts-selftest.yml        |  17 ---
 5 files changed, 34 insertions(+), 179 deletions(-)
 delete mode 100644 .github/scripts/run_hosted_agent_runtime_tests.py
 delete mode 100644 .github/scripts/test/test_hosted_agent_runtime_entrypoint.py
 create mode 100644 .github/workflows/hosted-agent-runtime-tests.yml

diff --git a/.github/scripts/run_hosted_agent_runtime_tests.py b/.github/scripts/run_hosted_agent_runtime_tests.py
deleted file mode 100644
index 10449078c..000000000
--- a/.github/scripts/run_hosted_agent_runtime_tests.py
+++ /dev/null
@@ -1,42 +0,0 @@
-"""Run hosted-agent runtime regressions outside the lightweight policy suite."""
-
-from __future__ import annotations
-
-import subprocess
-from pathlib import Path
-
-
-def main() -> None:
-    root = Path(__file__).resolve().parents[2]
-    project = root / (
-        "samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/"
-        "src/langgraph-human-in-the-loop-responses"
-    )
-    subprocess.run(
-        ["uv", "sync", "--frozen", "--python", "3.13", "--project", str(project)],
-        cwd=root,
-        check=True,
-    )
-    subprocess.run(
-        [
-            "uv",
-            "run",
-            "--no-sync",
-            "--project",
-            str(project),
-            "python",
-            "-m",
-            "unittest",
-            "discover",
-            "-s",
-            ".github/scripts/runtime-tests",
-            "-p",
-            "test_*.py",
-        ],
-        cwd=root,
-        check=True,
-    )
-
-
-if __name__ == "__main__":
-    main()
diff --git a/.github/scripts/runtime-tests/test_langgraph_sample_lifecycle.py b/.github/scripts/runtime-tests/test_langgraph_sample_lifecycle.py
index 7a5d9280d..6884f7745 100644
--- a/.github/scripts/runtime-tests/test_langgraph_sample_lifecycle.py
+++ b/.github/scripts/runtime-tests/test_langgraph_sample_lifecycle.py
@@ -1,7 +1,7 @@
 """Local sample lifecycle regressions using runtime dependencies, not Azure access.
 
-Run with the sample's frozen runtime dependencies:
-    python .github/scripts/run_hosted_agent_runtime_tests.py
+Run from the repository root in the sample's frozen runtime environment:
+    python .github/scripts/runtime-tests/test_langgraph_sample_lifecycle.py
 """
 
 from __future__ import annotations
diff --git a/.github/scripts/test/test_hosted_agent_runtime_entrypoint.py b/.github/scripts/test/test_hosted_agent_runtime_entrypoint.py
deleted file mode 100644
index 73b10373c..000000000
--- a/.github/scripts/test/test_hosted_agent_runtime_entrypoint.py
+++ /dev/null
@@ -1,118 +0,0 @@
-"""Keep runtime regression execution isolated, frozen, and wired into GitHub CI."""
-
-from __future__ import annotations
-
-import importlib.util
-import subprocess
-import unittest
-from pathlib import Path
-from unittest.mock import patch
-
-import yaml
-
-ROOT = Path(__file__).resolve().parents[3]
-SCRIPTS = ROOT / ".github/scripts"
-SPEC = importlib.util.spec_from_file_location(
-    "hosted_agent_runtime_tests", SCRIPTS / "run_hosted_agent_runtime_tests.py"
-)
-assert SPEC and SPEC.loader
-runner = importlib.util.module_from_spec(SPEC)
-SPEC.loader.exec_module(runner)
-
-
-class RuntimeTestEntrypointTests(unittest.TestCase):
-    def test_syncs_frozen_sample_before_running_separate_discovery(self):
-        project = str(
-            ROOT
-            / "samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop"
-            / "src/langgraph-human-in-the-loop-responses"
-        )
-        with patch.object(runner.subprocess, "run") as run:
-            runner.main()
-        self.assertEqual(run.call_count, 2)
-        self.assertEqual(
-            run.call_args_list[0].args[0],
-            ["uv", "sync", "--frozen", "--python", "3.13", "--project", project],
-        )
-        self.assertEqual(
-            run.call_args_list[1].args[0],
-            [
-                "uv",
-                "run",
-                "--no-sync",
-                "--project",
-                project,
-                "python",
-                "-m",
-                "unittest",
-                "discover",
-                "-s",
-                ".github/scripts/runtime-tests",
-                "-p",
-                "test_*.py",
-            ],
-        )
-        for call in run.call_args_list:
-            self.assertEqual(call.kwargs, {"cwd": ROOT, "check": True})
-        self.assertFalse((SCRIPTS / "test/test_langgraph_sample_lifecycle.py").exists())
-        self.assertTrue(
-            (SCRIPTS / "runtime-tests/test_langgraph_sample_lifecycle.py").is_file()
-        )
-
-    def test_sync_failure_prevents_test_execution(self):
-        with (
-            patch.object(
-                runner.subprocess,
-                "run",
-                side_effect=subprocess.CalledProcessError(1, "uv"),
-            ) as run,
-            self.assertRaises(subprocess.CalledProcessError),
-        ):
-            runner.main()
-        self.assertEqual(run.call_count, 1)
-
-    def test_runtime_failure_propagates(self):
-        with (
-            patch.object(
-                runner.subprocess,
-                "run",
-                side_effect=[None, subprocess.CalledProcessError(1, "unittest")],
-            ),
-            self.assertRaises(subprocess.CalledProcessError),
-        ):
-            runner.main()
-
-    def test_github_runs_the_entrypoint_without_skipping(self):
-        workflow = yaml.load(
-            (ROOT / ".github/workflows/scripts-selftest.yml").read_text(
-                encoding="utf-8"
-            ),
-            Loader=yaml.BaseLoader,
-        )
-        job = workflow["jobs"]["hosted-agent-runtime-regressions"]
-        self.assertEqual(workflow["permissions"], {"contents": "read"})
-        self.assertIn(
-            "samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/**",
-            workflow["on"]["pull_request"]["paths"],
-        )
-        self.assertEqual(job["steps"][0]["with"]["persist-credentials"], "false")
-        for step in job["steps"]:
-            self.assertNotIn("if", step)
-            self.assertNotIn("continue-on-error", step)
-        self.assertIn('"uv==0.11.7"', job["steps"][-2]["run"])
-        self.assertEqual(
-            job["steps"][-1]["run"],
-            "python .github/scripts/run_hosted_agent_runtime_tests.py",
-        )
-        self.assertIn(
-            "python .github/scripts/test/test_hosted_agent_runtime_entrypoint.py",
-            next(
-                step["run"]
-                for step in workflow["jobs"]["ado-pipelines-harness"]["steps"]
-                if step.get("name") == "Test hosted-agent helpers and fixture contracts"
-            ),
-        )
-
-
-if __name__ == "__main__":
-    unittest.main()
diff --git a/.github/workflows/hosted-agent-runtime-tests.yml b/.github/workflows/hosted-agent-runtime-tests.yml
new file mode 100644
index 000000000..9199abb12
--- /dev/null
+++ b/.github/workflows/hosted-agent-runtime-tests.yml
@@ -0,0 +1,32 @@
+name: Hosted-agent runtime regressions
+
+on:
+  workflow_dispatch:
+  pull_request:
+    paths:
+      - 'samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/src/**'
+      - '.github/scripts/runtime-tests/test_langgraph_sample_lifecycle.py'
+      - '.github/workflows/hosted-agent-runtime-tests.yml'
+
+permissions:
+  contents: read
+
+jobs:
+  hitl-lifecycle:
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    env:
+      RUNTIME_PROJECT: samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/src/langgraph-human-in-the-loop-responses
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4.4.0
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
+        with:
+          python-version: '3.13'
+      - name: Install runtime test resolver
+        run: python -m pip install --disable-pip-version-check "uv==0.11.7"
+      - name: Install frozen sample dependencies
+        run: uv sync --frozen --python 3.13 --project "$RUNTIME_PROJECT"
+      - name: Verify HITL revision lifecycle
+        run: uv run --no-sync --project "$RUNTIME_PROJECT" python .github/scripts/runtime-tests/test_langgraph_sample_lifecycle.py
diff --git a/.github/workflows/scripts-selftest.yml b/.github/workflows/scripts-selftest.yml
index 1ec22b43b..1c76e6468 100644
--- a/.github/workflows/scripts-selftest.yml
+++ b/.github/workflows/scripts-selftest.yml
@@ -27,27 +27,11 @@ on:
       - '.github/workflows/validation-pilot.yml'
       - '.github/scripts/discover-validation-samples.py'
       - 'samples/*/prompt-agents/**'
-      - 'samples/python/hosted-agents/langgraph/responses/07-human-in-the-loop/**'
 
 permissions:
   contents: read
 
 jobs:
-  hosted-agent-runtime-regressions:
-    runs-on: ubuntu-latest
-    timeout-minutes: 15
-    steps:
-      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4.4.0
-        with:
-          persist-credentials: false
-      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
-        with:
-          python-version: '3.13'
-      - name: Install runtime test resolver
-        run: python -m pip install --disable-pip-version-check "uv==0.11.7"
-      - name: Run frozen hosted-agent runtime regressions
-        run: python .github/scripts/run_hosted_agent_runtime_tests.py
-
   ado-pipelines-harness:
     runs-on: ubuntu-latest
     timeout-minutes: 15
@@ -66,7 +50,6 @@ jobs:
       - name: Test hosted-agent helpers and fixture contracts
         run: |
           python -m unittest discover -s .azure-pipelines/scripts/hosted-agents/tests -p 'test_*.py'
-          python .github/scripts/test/test_hosted_agent_runtime_entrypoint.py
           bash .azure-pipelines/scripts/hosted-agents/tests/test-hosted-agent-ci-toolboxes.sh
           bash .azure-pipelines/scripts/hosted-agents/tests/test-hosted-agent-session-quota.sh
           python .github/scripts/test/test_ado_pipeline_fixtures.py