test(fase2): assert BudgetMiddleware short-circuits a real chat call + the debate

tests/test_budget.py

@@ -33,9 +33,7 @@ async def _noop() -> None:
 
 def _resp(total: int | None) -> ChatResponse:
     usage = UsageDetails(total_token_count=total) if total is not None else None
-    return ChatResponse(
-        messages=[Message(role="assistant", contents=["x"])], usage_details=usage
-    )
+    return ChatResponse(messages=[Message(role="assistant", contents=["x"])], usage_details=usage)
 
 
 async def test_meter_reads_total_token_count_and_accumulates() -> None:
@@ -69,6 +67,24 @@ async def test_strict_usage_none_hard_fails() -> None:
         await mw.process(_Ctx(_resp(None)), _noop)  # type: ignore[arg-type]
 
 
+async def test_budget_middleware_fires_on_real_agent_chat(make_client_factory) -> None:
+    """F8 (real-client half): registering BudgetMiddleware on a real Agent (layered chat client
+    that carries ChatMiddlewareLayer) and running an actual chat call short-circuits with
+    BudgetExceeded — the middleware<->client integration the prior suite never exercised (the
+    minimal-base stand-in silently no-ops the middleware)."""
+    from agent_framework import Agent
+
+    client = make_client_factory("ok", tokens=8)("proposer")  # 8 tokens/reply > cap 5
+    meter = TokenMeter(Budget(max_tokens=5, max_rounds=10))
+    agent = Agent(
+        client, "propose a measure", name="proposer", middleware=[BudgetMiddleware(meter)]
+    )
+    with pytest.raises(BudgetExceeded) as exc:
+        await agent.run("hi")
+    assert exc.value.kind == "tokens"
+    assert meter.tokens == 8  # charged from the synthetic UsageDetails via the middleware
+
+
 def test_no_word_count_token_proxy_in_src() -> None:
     # The meter is fed from real UsageDetails, never a len(text.split()) word-count proxy
     # (research 03 Rec 3 — the Fase 1 _word_tokens anti-pattern is retired). NOTE: a bare