fix(compaction): arm verdict after successful boundary

agent/context_compressor.py

@@ -2854,12 +2854,6 @@ This compaction should PRIORITISE preserving all information related to the focu
                 running so a manual ``/compress`` can retry immediately after
                 an auto-compression abort.  Auto-compress callers pass False.
         """
-        # A compaction attempt is running: have should_compress() verify, against
-        # the next real reading, that it actually cleared the threshold. Set on
-        # every path (including the early no-op returns below) so a pass that
-        # changed nothing is still held to account.
-        self._verify_compaction_cleared_threshold = True
-
         # Reset per-call summary failure state — callers inspect these fields
         # after compress() returns to decide whether to surface a warning.
         self._last_summary_dropped_count = 0

agent/conversation_compression.py

@@ -961,6 +961,12 @@ def compress_context(
         agent.context_compressor.last_prompt_tokens = -1
         agent.context_compressor.last_completion_tokens = 0
         agent.context_compressor.awaiting_real_usage_after_compression = True
+        # Arm the effectiveness verdict only after the full compaction boundary
+        # succeeds. Exceptions and aborted/no-op attempts return before here, so
+        # unrelated later usage cannot be charged to an attempt that never took
+        # effect. External engines without the private flag stay untouched.
+        if hasattr(agent.context_compressor, "_verify_compaction_cleared_threshold"):
+            agent.context_compressor._verify_compaction_cleared_threshold = True
 
         # Clear the file-read dedup cache.  After compression the original
         # read content is summarised away — if the model re-reads the same

tests/agent/test_compaction_anti_thrash.py

@@ -70,6 +70,7 @@ def _turn(cc, msgs, real_prompt_tokens):
     if not cc.should_compress(real_prompt_tokens):
         return msgs, False
     msgs = cc.compress(msgs, current_tokens=real_prompt_tokens)
+    cc._verify_compaction_cleared_threshold = True
     cc.update_from_response({"prompt_tokens": real_prompt_tokens})
     return msgs, True
 
@@ -180,6 +181,7 @@ class TestFutilityGuard:
         assert cc.should_compress(real_prompt)
         msgs = cc.compress(msgs, current_tokens=real_prompt)
         real_after = floor + int(skew * estimate_messages_tokens_rough(msgs))
+        cc._verify_compaction_cleared_threshold = True
         cc.update_from_response({"prompt_tokens": real_after})  # provider's real count
 
         assert real_after < cc.threshold_tokens, "compaction really did clear it"
@@ -198,6 +200,7 @@ class TestFutilityGuard:
 
         assert cc.should_compress(33_564)
         cc.compress(msgs, current_tokens=33_564)
+        cc._verify_compaction_cleared_threshold = True
         assert cc._ineffective_compression_count == 0, "no verdict before real usage"
 
         cc.update_from_response({"prompt_tokens": 33_564})  # still over