feat: add remote workspaces and verified deliverables

tests/test_agent_loop.py

@@ -6,7 +6,7 @@ from typing import Any
 
 from agent_platform.auth import UserIdentity
 from agent_platform.models import get_model_spec
-from agent_platform.runtime.loop import AgentLoop
+from agent_platform.runtime.loop import AgentLoop, tool_event_details
 from agent_platform.runtime.tools import TOOL_METADATA
 from agent_platform.store import RuntimeStore
 
@@ -119,11 +119,23 @@ async def test_mutating_tool_calls_are_serialized(settings) -> None:
             }
         ]
     }
+    verify = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "content": None,
+                    "tool_calls": [tool_call("3", "read_file", {"path": "a"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
     final = {"choices": [{"message": {"role": "assistant", "content": "done"}, "finish_reason": "stop"}]}
     store = RuntimeStore(settings.database_url)
     await store.initialize()
     registry = SerialRegistry()
-    loop = AgentLoop(ScriptedProvider([first, final]), registry, store, max_tool_output_chars=10_000)
+    loop = AgentLoop(ScriptedProvider([first, verify, final]), registry, store, max_tool_output_chars=10_000)
 
     async def callback(event_type, payload):
         return None
@@ -140,3 +152,123 @@ async def test_mutating_tool_calls_are_serialized(settings) -> None:
         assert registry.max_active == 1
     finally:
         await store.close()
+
+
+class RecordingRegistry:
+    def __init__(self) -> None:
+        self.calls = []
+
+    def specs(self, *, read_only=False, allow_delegate=True):
+        return [metadata.openai_spec() for metadata in TOOL_METADATA.values()]
+
+    async def execute(self, name, arguments, context):
+        self.calls.append((name, arguments))
+        return {"ok": True, "output": f"{name} complete"}
+
+
+async def test_artifact_completion_is_rejected_until_written_and_verified(settings) -> None:
+    premature = {"choices": [{"message": {"role": "assistant", "content": "报告已经完成"}, "finish_reason": "stop"}]}
+    write = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [tool_call("1", "write_file", {"path": "report.md", "content": "ok"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    verify = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [tool_call("2", "read_file", {"path": "report.md"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    final = {"choices": [{"message": {"role": "assistant", "content": "已完成 report.md"}, "finish_reason": "stop"}]}
+    provider = ScriptedProvider([premature, write, verify, final])
+    registry = RecordingRegistry()
+    store = RuntimeStore(settings.database_url)
+    await store.initialize()
+    events = []
+
+    async def callback(event_type, payload):
+        events.append(event_type)
+
+    try:
+        answer = await AgentLoop(provider, registry, store, max_tool_output_chars=10_000).run(
+            spec=get_model_spec("work-light"),
+            messages=[{"role": "user", "content": "写一个脚本并生成报告"}],
+            identity=UserIdentity("u1", "", "", "user"),
+            raw_user_jwt="jwt",
+            chat_id="c1",
+            callback=callback,
+        )
+        assert answer == "已完成 report.md"
+        assert [name for name, _ in registry.calls] == ["write_file", "read_file"]
+        assert "completion.rejected" in events
+    finally:
+        await store.close()
+
+
+async def test_bare_interactive_exec_is_blocked_before_registry(settings) -> None:
+    first = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [tool_call("1", "exec", {"command": "python3"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    final = {"choices": [{"message": {"role": "assistant", "content": "无法执行"}, "finish_reason": "stop"}]}
+    registry = RecordingRegistry()
+    store = RuntimeStore(settings.database_url)
+    await store.initialize()
+
+    async def callback(event_type, payload):
+        return None
+
+    try:
+        answer = await AgentLoop(
+            ScriptedProvider([first, final]),
+            registry,
+            store,
+            max_tool_output_chars=10_000,
+        ).run(
+            spec=get_model_spec("work-light"),
+            messages=[{"role": "user", "content": "你好"}],
+            identity=UserIdentity("u1", "", "", "user"),
+            raw_user_jwt="jwt",
+            chat_id="c1",
+            callback=callback,
+        )
+        assert answer == "无法执行"
+        assert registry.calls == []
+    finally:
+        await store.close()
+
+
+def test_tool_events_render_one_friendly_completed_card() -> None:
+    assert tool_event_details("tool.started", {"name": "exec"}) is None
+    rendered = tool_event_details(
+        "tool.completed",
+        {
+            "call_id": "call-1",
+            "name": "exec",
+            "arguments": {"command": "pytest -q"},
+            "ok": True,
+            "summary": "2 passed",
+        },
+    )
+    assert rendered is not None
+    assert 'name="执行命令"' in rendered
+    assert "2 passed" in rendered
+    assert 'done="false"' not in rendered