diff --git a/studio/backend/core/inference/context_compaction.py b/studio/backend/core/inference/context_compaction.py
index 18c36bb07c..f4a52c675a 100644
--- a/studio/backend/core/inference/context_compaction.py
+++ b/studio/backend/core/inference/context_compaction.py
@@ -195,17 +195,22 @@ class SlidingWindowCompact(CompactStrategy):
                 group_id[ti] = g
 
         # The "recent window" is the last ``keep_recent`` distinct
-        # groups encountered scanning from the end.
+        # groups encountered scanning from the end. ``keep_recent == 0``
+        # must collect ZERO groups so the caller can drop everything
+        # outside the anchor set (system + first user + multimodal).
+        # The pre-fix loop tested the limit AFTER appending and so
+        # always preserved at least one group even when keep_recent
+        # was 0; flip the check to BEFORE appending so the bound holds.
         recent_groups: list[int] = []
         seen_groups: set[int] = set()
         for i in range(len(messages) - 1, -1, -1):
+            if len(recent_groups) >= self.keep_recent:
+                break
             g = group_id[i]
             if g in seen_groups:
                 continue
             seen_groups.add(g)
             recent_groups.append(g)
-            if len(recent_groups) >= self.keep_recent:
-                break
         recent_groups_set = set(recent_groups)
 
         # Decide drop set: every index whose group is NOT in the recent
@@ -230,13 +235,23 @@ class SlidingWindowCompact(CompactStrategy):
         # When dropping an assistant-with-tool-calls message we must
         # also drop the matching tool-role messages (and vice versa)
         # so the chat template stays valid. Iterate pair_map once.
+        # Anchor indices stay regardless: dragging an anchored
+        # multimodal assistant or first-user message into the drop
+        # set just because its tool-pair partner was dropped would
+        # violate the structural invariant the anchor set exists to
+        # enforce, and llama-server would 400 on the resulting
+        # template (a tool message whose tool_call_id has no
+        # surviving assistant tool_calls entry).
         for asst_idx, tool_idxs in pair_map.items():
             if asst_idx in dropped:
-                dropped.update(tool_idxs)
+                dropped.update(t for t in tool_idxs if t not in anchor_idx)
             elif tool_idxs and tool_idxs <= dropped:
                 # All matching tool messages were dropped: drop the
-                # assistant tool-call message too.
-                dropped.add(asst_idx)
+                # assistant tool-call message too -- unless it's
+                # anchored, in which case we'd rather leak an
+                # orphan tool_call shape than violate the invariant.
+                if asst_idx not in anchor_idx:
+                    dropped.add(asst_idx)
 
         return [m for i, m in enumerate(messages) if i not in dropped]
 
diff --git a/studio/backend/tests/test_context_compaction.py b/studio/backend/tests/test_context_compaction.py
index c42b39d9ad..aa629093d7 100644
--- a/studio/backend/tests/test_context_compaction.py
+++ b/studio/backend/tests/test_context_compaction.py
@@ -277,6 +277,88 @@ class TestStrategyRegistry:
         assert isinstance(get_strategy("totally-made-up"), NoCompact)
 
 
+class TestKeepRecentZero:
+    """Regression: keep_recent=0 must collect ZERO recent groups, not 1.
+
+    Pre-fix, the loop tested the limit AFTER appending, so the first
+    iteration always added one group and broke. Trim with keep_recent=0
+    should drop everything outside the anchor set (system + first-user
+    + multimodal). This matters when a caller intentionally wants only
+    the anchors to survive (extreme-pressure regime).
+    """
+
+    def test_keep_recent_zero_keeps_only_anchors(self):
+        msgs = [
+            _msg("system", "sys"),
+            _msg("user", "task"),
+            _long("assistant", 4000),
+            _long("user", 4000),
+            _long("assistant", 4000),
+            _long("user", 4000),
+        ]
+        out = SlidingWindowCompact(keep_recent = 0).compact(
+            msgs, budget_tokens = 50
+        )
+        roles = [m["role"] for m in out]
+        # System + first-user only.
+        assert roles == ["system", "user"]
+        assert out[1]["content"].startswith("task")
+
+
+class TestAnchoredMultimodalPairCleanup:
+    """Regression: an anchored multimodal assistant whose paired tool
+    messages all get dropped must survive pair-cleanup. Without this
+    guard the pair_map's "all tool messages dropped -> drop the
+    assistant too" rule would yank a multimodal anchor out from under
+    the structural invariant that says multimodal turns are never
+    dropped, and llama-server would 400 the resulting template.
+    """
+
+    def test_anchored_multimodal_assistant_survives_pair_cleanup(self):
+        # Multimodal assistant carrying tool_calls (rare but valid:
+        # vision models can call tools while emitting image parts).
+        multimodal_asst = {
+            "role": "assistant",
+            "content": [
+                {"type": "text", "text": "describe and call tool"},
+                {"type": "image_url", "image_url": {"url": "x"}},
+            ],
+            "tool_calls": [
+                {
+                    "id": "call_mm",
+                    "type": "function",
+                    "function": {
+                        "name": "web_search",
+                        "arguments": '{"q":"y"}',
+                    },
+                }
+            ],
+        }
+        msgs = [
+            _msg("system", "sys"),
+            _msg("user", "task"),
+            multimodal_asst,
+            _msg(
+                "tool",
+                content = "result for y" * 200,
+                tool_call_id = "call_mm",
+                name = "web_search",
+            ),
+            _long("assistant", 4000),
+            _long("user", 4000),
+            _long("assistant", 4000),
+            _long("user", 4000),
+        ]
+        out = SlidingWindowCompact(keep_recent = 1).compact(
+            msgs, budget_tokens = 50
+        )
+        # Anchored multimodal assistant must still be there.
+        assert any(
+            m.get("role") == "assistant" and isinstance(m.get("content"), list)
+            for m in out
+        ), "multimodal assistant got dropped by pair_map cleanup"
+
+
 class TestConstructorValidation:
     def test_negative_keep_recent_raises(self):
         import pytest