feat: require runway before prune-only compaction

Make prune-first compression cache-aware by only accepting prune-only
compaction when it gets comfortably below threshold. If pruning merely
dips under threshold, fall through to the existing summary compaction
so we avoid frequent near-threshold recompressions.

Tests cover both the conservative fallback and the prune-only fast path.

agent/context_compressor.py

@@ -74,6 +74,8 @@ class ContextCompressor:
         self.summary_model = summary_model_override or ""
         self._prune_protect_tokens = _adaptive_prune_protect(self.context_length)
         self._prune_minimum_tokens = _adaptive_prune_minimum(self.context_length)
+        self._prune_runway_tokens = max(self._prune_minimum_tokens, int(self.threshold_tokens * 0.15))
+        self._prune_target_tokens = max(0, self.threshold_tokens - self._prune_runway_tokens)
 
     def update_from_response(self, usage: Dict[str, Any]):
         """Update tracked token usage from API response."""
@@ -354,7 +356,7 @@ Write only the summary, starting with "[CONTEXT SUMMARY]:" prefix."""
                     f"   ✂️  Phase 1 (prune): removed {chars_saved:,} chars of old tool outputs "
                     f"(~{tokens_saved_phase1:,} tokens saved)"
                 )
-            if pruned_tokens < self.threshold_tokens:
+            if pruned_tokens <= self._prune_target_tokens:
                 self.compression_count += 1
                 pruned_messages = self._sanitize_tool_pairs(pruned_messages)
                 if not self.quiet_mode:
@@ -364,6 +366,11 @@ Write only the summary, starting with "[CONTEXT SUMMARY]:" prefix."""
                     )
                     print(f"   💡 Compression #{self.compression_count} complete (prune only — no LLM call needed)")
                 return pruned_messages
+            if not self.quiet_mode and pruned_tokens < self.threshold_tokens:
+                print(
+                    f"   ↪️  Phase 1 recovered tokens but not enough runway "
+                    f"({pruned_tokens:,} > target {self._prune_target_tokens:,}); continuing to compaction"
+                )
             messages = pruned_messages
             n_messages = len(messages)
             compress_start = self.protect_first_n

tests/agent/test_context_compressor.py

@@ -400,3 +400,57 @@ class TestPruneToolOutputs:
 
         assert pruned[-2]["content"] == huge_content
         assert pruned[-1]["content"] == "latest"
+
+
+class TestPruneAcceptancePolicy:
+    def _make_compressor(self, *, context_length=128000):
+        with patch("agent.context_compressor.get_model_context_length", return_value=context_length):
+            return ContextCompressor(
+                model="test/model",
+                threshold_percent=0.50,
+                protect_first_n=2,
+                protect_last_n=1,
+                quiet_mode=True,
+            )
+
+    def test_prune_near_threshold_still_falls_back_to_summary(self):
+        c = self._make_compressor()
+        huge_content = "x" * 180000
+        messages = [
+            {"role": "system", "content": "sys"},
+            {"role": "user", "content": "task"},
+            {"role": "assistant", "content": "older"},
+            {"role": "tool", "content": huge_content, "name": "terminal"},
+            {"role": "assistant", "content": "newer"},
+            {"role": "tool", "content": huge_content, "name": "terminal"},
+            {"role": "assistant", "content": "tail"},
+        ]
+        mock_response = MagicMock()
+        mock_response.choices = [MagicMock()]
+        mock_response.choices[0].message.content = "[CONTEXT SUMMARY]: compacted"
+
+        with patch("agent.context_compressor.estimate_messages_tokens_rough", return_value=62000), \
+             patch("agent.context_compressor.call_llm", return_value=mock_response):
+            result = c.compress(messages, current_tokens=68000)
+
+        assert any("CONTEXT SUMMARY" in (msg.get("content") or "") for msg in result)
+
+    def test_prune_only_is_allowed_when_it_buys_real_runway(self):
+        c = self._make_compressor()
+        huge_content = "x" * 180000
+        messages = [
+            {"role": "system", "content": "sys"},
+            {"role": "user", "content": "task"},
+            {"role": "assistant", "content": "older"},
+            {"role": "tool", "content": huge_content, "name": "terminal"},
+            {"role": "assistant", "content": "newer"},
+            {"role": "tool", "content": huge_content, "name": "terminal"},
+            {"role": "assistant", "content": "tail"},
+        ]
+
+        with patch("agent.context_compressor.estimate_messages_tokens_rough", return_value=48000), \
+             patch.object(ContextCompressor, "_generate_summary", side_effect=AssertionError("summary should not be called")):
+            result = c.compress(messages, current_tokens=68000)
+
+        assert result[3]["content"].startswith("[Tool output pruned")
+        assert result[5]["content"] == huge_content