feat(ai): services partages A2 + SSE streaming B4 + confirmations UI B5 (#79)

tests/test_agent_loop.py

@@ -186,6 +186,65 @@ class TestConfirmationAndLimits:
 # ═══════════════════════════════════════════════════════════════════
 
 
+class TestConfirmationResume:
+    @pytest.mark.asyncio
+    async def test_resume_applies_pending_and_continues(self, monkeypatch):
+        seen = {}
+
+        def handler(ctx, params):
+            seen["ctx_confirmed"] = ctx.confirmed
+            return {"done": True}
+
+        _register(monkeypatch, "_write", handler, risk=ToolRisk.WRITE)
+
+        # First run pauses on the mutating tool.
+        llm1 = ScriptedLLM([LLMResponse(tool_calls=[ToolCall(id="1", name="_write", arguments={"x": 1})])])
+        paused = await run_agent([{"role": "user", "content": "write"}], ctx=_ctx(), llm=llm1)
+        assert paused.stopped == STOP_CONFIRMATION_REQUIRED
+        assert paused.pending["error"]["id"] == "1"
+
+        # Resume: the pending call is applied (one-shot confirm), then the loop
+        # continues and produces the final answer.
+        llm2 = ScriptedLLM([LLMResponse(content="applied")])
+        resumed = await run_agent(
+            [{"role": "user", "content": "write"}],
+            ctx=_ctx(),
+            llm=llm2,
+            resume_messages=paused.messages,
+            confirm_pending=paused.pending,
+        )
+        assert resumed.stopped == STOP_DONE
+        assert resumed.content == "applied"
+        assert len(resumed.tool_calls) == 1
+        assert resumed.tool_calls[0].ok is True
+        # The one-shot confirmation must not leak into the context.
+        assert seen["ctx_confirmed"] is False
+        # The tool result is fed back to the model on the resumed turn.
+        tool_msgs = [m for m in llm2.calls[0]["messages"] if m.get("role") == "tool"]
+        assert len(tool_msgs) == 1
+        assert tool_msgs[0]["tool_call_id"] == "1"
+
+    @pytest.mark.asyncio
+    async def test_resume_without_assistant_message_reconstructs_it(self, monkeypatch):
+        _register(monkeypatch, "_write", lambda ctx, params: {"done": True}, risk=ToolRisk.WRITE)
+        llm = ScriptedLLM([LLMResponse(content="ok")])
+        pending = {"error": {"tool": "_write", "arguments": {}, "id": "call_9"}}
+        result = await run_agent(
+            [{"role": "user", "content": "write"}],
+            ctx=_ctx(),
+            llm=llm,
+            resume_messages=[{"role": "user", "content": "write"}],
+            confirm_pending=pending,
+        )
+        assert result.stopped == STOP_DONE
+        assistant_tool_msgs = [
+            m for m in llm.calls[0]["messages"]
+            if m.get("role") == "assistant" and m.get("tool_calls")
+        ]
+        assert len(assistant_tool_msgs) == 1
+        assert assistant_tool_msgs[0]["tool_calls"][0]["id"] == "call_9"
+
+
 class TestAgentPermissions:
     @pytest.mark.asyncio
     async def test_permission_denied_recorded(self, client):