diff --git a/raven/agent/loop/main.py b/raven/agent/loop/main.py
index f6d7661..375e2ba 100644
--- a/raven/agent/loop/main.py
+++ b/raven/agent/loop/main.py
@@ -2012,7 +2012,12 @@ async def _run_agent_loop(
                 usage_sink["cost_usd"] = usage_snapshot.estimated_cost_usd
                 usage_sink["context_max"] = context_max
                 usage_sink["context_used"] = context_used
-                usage_sink["context_percent"] = round(100 * context_used / context_max) if context_max else 0
+                # context_max is a real provider window when resolve_context_window()
+                # succeeds, but otherwise a guessed default (self.context_window_tokens)
+                # that can be smaller than the model's true window. Clamp so the gauge
+                # never reports past full: an unresolved model must not render as
+                # "over budget" while the request is still well within its real limit.
+                usage_sink["context_percent"] = min(100, round(100 * context_used / context_max)) if context_max else 0
 
             # Context-window overflow recovery: the structured classifier flags
             # should_compress (a smaller window won't help, but eliding the bulk
diff --git a/tests/test_agent_loop_usage_sink.py b/tests/test_agent_loop_usage_sink.py
index 7963e59..d2723da 100644
--- a/tests/test_agent_loop_usage_sink.py
+++ b/tests/test_agent_loop_usage_sink.py
@@ -336,3 +336,31 @@ def test_refresh_context_window_on_an_openrouter_model_never_touches_the_network
 
     assert agent.context_window_tokens == rates.DEFAULT_CONTEXT_WINDOW_TOKENS
     assert counter["calls"] == 0
+
+
+@pytest.mark.asyncio
+async def test_usage_sink_context_percent_clamps_when_window_is_a_guess(workspace):
+    """Usage past the configured fallback window must not render past 100%.
+
+    ``stub`` never resolves via resolve_context_window(), so context_max here is
+    the configured (guessed) default. Real usage can legitimately exceed that
+    guess while staying under the model's true, unresolved window, so the
+    reported percent must clamp rather than imply an impossible over-budget bar.
+    """
+    provider = UsageProvider("stub", prompt_tokens=6000, completion_tokens=2000)
+    agent = _make_agent(workspace, provider, model="stub", window=1000)
+    sink: dict = {}
+
+    await agent._process_message(
+        TurnRequest(
+            origin=Origin.USER,
+            source=Source(channel="test", chat_id="c1", sender_id="user", chat_type=ChatType.DM),
+            text="hi",
+        ),
+        session_key="s1",
+        usage_sink=sink,
+    )
+
+    assert sink["context_max"] == 1000
+    assert sink["context_used"] == 8000
+    assert sink["context_percent"] == 100