feat(tuning): release V0.8.2 Agent 界面重构

training_server/tests/test_tuning_manager.py

@@ -148,6 +148,8 @@ class TuningManagerTest(unittest.TestCase):
             self.fail("session did not wait for approval")
         proposal = detail["proposals"][-1]
         patch = {"weights": {"pose": 1.1}, "params": {}}
+        with self.assertRaisesRegex(Exception, "feedback"):
+            self.manager.approve(session["id"], proposal["id"], {"feedback": {}})
         approved = self.manager.approve(
             session["id"], proposal["id"], {"feedback": "ok", "patch": patch}
         )
@@ -155,6 +157,34 @@ class TuningManagerTest(unittest.TestCase):
         self.manager.cancel(session["id"])
         self.assertEqual(self.wait_terminal(session["id"])["state"], "cancelled")
 
+    def test_runtime_mode_switch_auto_approves_pending_proposal(self):
+        session = self.manager.create(self.payload("approval"))
+        deadline = time.monotonic() + 3
+        while time.monotonic() < deadline:
+            detail = self.manager.detail(session["id"])
+            if detail["state"] == "awaiting_approval":
+                break
+            time.sleep(0.01)
+        else:
+            self.fail("session did not wait for approval")
+        changed = self.manager.set_mode(session["id"], {"mode": "automatic"})
+        self.assertEqual(changed["mode"], "automatic")
+        self.assertEqual(changed["proposals"][-1]["state"], "approved")
+        completed = self.wait_terminal(session["id"])
+        self.assertEqual(completed["state"], "succeeded", completed["message"])
+
+    def test_trial_count_is_user_configurable(self):
+        payload = self.payload()
+        payload["trialCount"] = 1
+        _, config, _, _ = self.manager.parse_create(payload)
+        self.assertEqual(config["trialCount"], 1)
+        payload["trialCount"] = 100
+        _, config, _, _ = self.manager.parse_create(payload)
+        self.assertEqual(config["trialCount"], 100)
+        payload["trialCount"] = 101
+        with self.assertRaisesRegex(Exception, "trialCount"):
+            self.manager.parse_create(payload)
+
     def test_resume_discards_only_interrupted_trial_and_continues(self):
         mode, config, objective, fallback = self.manager.parse_create(self.payload())
         session = self.manager.storage.create_session(mode, config, objective, fallback)
@@ -178,8 +208,10 @@ class TuningManagerTest(unittest.TestCase):
             "trial-001-rung-0",
         )
         self.manager.storage.update_trial(interrupted["id"], state="interrupted")
+        self.manager.storage.reset_step_gate(session["id"])
         self.manager.storage.update_session(session["id"], state="interrupted")
         self.manager.resume(session["id"])
+        self.assertEqual(self.manager.storage.get_control(session["id"])["runPolicy"], "continuous")
         completed = self.wait_terminal(session["id"])
         self.assertEqual(completed["state"], "succeeded", completed["message"])
         self.assertNotIn(interrupted["id"], [trial["id"] for trial in completed["trials"]])
@@ -190,6 +222,129 @@ class TuningManagerTest(unittest.TestCase):
             self.manager.parse_create({"taskId": "Other"})
         self.assertEqual(self.manager.test_agent()["model"], "fake")
 
+    def test_pause_closes_persistent_dispatch_gate_at_trial_boundary(self):
+        mode, config, objective, fallback = self.manager.parse_create(self.payload())
+        session = self.manager.storage.create_session(mode, config, objective, fallback)
+        self.manager.storage.update_session(session["id"], state="running")
+        self.manager.storage.grant_dispatch_token(session["id"])
+
+        paused = self.manager.pause(session["id"])
+
+        self.assertEqual(paused["state"], "paused")
+        self.assertEqual(paused["control"]["runPolicy"], "step")
+        self.assertEqual(paused["control"]["dispatchTokens"], 0)
+
+    def test_step_token_executes_exactly_one_trial_then_pauses(self):
+        session = self.manager.create(self.payload("approval"))
+        deadline = time.monotonic() + 3
+        while time.monotonic() < deadline:
+            detail = self.manager.detail(session["id"])
+            if detail["state"] == "awaiting_approval":
+                break
+            time.sleep(0.01)
+        else:
+            self.fail("session did not wait for approval")
+        baseline_count = len([trial for trial in detail["trials"] if trial["state"] == "completed"])
+        stepped = self.manager.step(session["id"], {"count": 1})
+        self.assertEqual(stepped["control"]["runPolicy"], "step")
+        self.assertEqual(stepped["control"]["dispatchTokens"], 1)
+        proposal = stepped["proposals"][-1]
+        self.manager.approve(session["id"], proposal["id"], {})
+        deadline = time.monotonic() + 3
+        while time.monotonic() < deadline:
+            detail = self.manager.detail(session["id"])
+            completed_count = len(
+                [trial for trial in detail["trials"] if trial["state"] == "completed"]
+            )
+            if detail["state"] == "paused" and completed_count == baseline_count + 1:
+                break
+            time.sleep(0.01)
+        else:
+            self.fail("single-step trial did not pause at the next boundary")
+        time.sleep(0.05)
+        self.assertEqual(
+            len(
+                [
+                    trial
+                    for trial in self.manager.detail(session["id"])["trials"]
+                    if trial["state"] == "completed"
+                ]
+            ),
+            baseline_count + 1,
+        )
+        self.manager.cancel(session["id"])
+        self.assertEqual(self.wait_terminal(session["id"])["state"], "cancelled")
+
+    def test_constraints_are_revisioned_and_enforced_during_approval(self):
+        session = self.manager.create(self.payload("approval"))
+        deadline = time.monotonic() + 3
+        while time.monotonic() < deadline:
+            detail = self.manager.detail(session["id"])
+            if detail["state"] == "awaiting_approval":
+                break
+            time.sleep(0.01)
+        else:
+            self.fail("session did not wait for approval")
+        proposal = detail["proposals"][-1]
+        constrained = self.manager.set_constraints(
+            session["id"],
+            {
+                "revision": 0,
+                "constraints": {"weights.track_linear_velocity": {"kind": "fixed", "value": 1.0}},
+            },
+        )
+        self.assertEqual(constrained["control"]["constraintsRevision"], 1)
+        with self.assertRaisesRegex(Exception, "已固定"):
+            self.manager.approve(
+                session["id"],
+                proposal["id"],
+                {"patch": {"weights": {"track_linear_velocity": 1.1}, "params": {}}},
+            )
+        with self.assertRaisesRegex(Exception, "revision"):
+            self.manager.set_constraints(session["id"], {"revision": 0, "constraints": {}})
+        self.manager.approve(session["id"], proposal["id"], {})
+        self.manager.cancel(session["id"])
+        self.assertEqual(self.wait_terminal(session["id"])["state"], "cancelled")
+
+    def test_rollback_uses_safe_completed_trial_as_next_proposal_base(self):
+        session = self.manager.create(self.payload("approval"))
+        deadline = time.monotonic() + 3
+        while time.monotonic() < deadline:
+            detail = self.manager.detail(session["id"])
+            if detail["state"] == "awaiting_approval":
+                break
+            time.sleep(0.01)
+        else:
+            self.fail("session did not wait for approval")
+        baseline = detail["trials"][0]
+        old_proposal = detail["proposals"][-1]
+        rolled_back = self.manager.rollback(
+            session["id"], {"trialId": baseline["id"], "checkpoint": True}
+        )
+        self.assertEqual(rolled_back["state"], "paused")
+        self.assertEqual(rolled_back["control"]["activeBaseTrialId"], baseline["id"])
+        self.assertEqual(
+            next(item for item in rolled_back["proposals"] if item["id"] == old_proposal["id"])[
+                "state"
+            ],
+            "rejected",
+        )
+        self.manager.step(session["id"], {"count": 1})
+        deadline = time.monotonic() + 3
+        while time.monotonic() < deadline:
+            detail = self.manager.detail(session["id"])
+            if (
+                detail["state"] == "awaiting_approval"
+                and detail["proposals"][-1]["id"] != old_proposal["id"]
+            ):
+                break
+            time.sleep(0.01)
+        else:
+            self.fail("rollback base did not produce a replacement proposal")
+        self.assertEqual(detail["proposals"][-1]["baseTrialId"], baseline["id"])
+        self.manager.cancel(session["id"])
+        self.assertEqual(self.wait_terminal(session["id"])["state"], "cancelled")
+
 
 if __name__ == "__main__":
     unittest.main()