fix(agent): re-anchor after post-tool compression

agent/turn_preflight.py

@@ -22,7 +22,7 @@ from agent.conversation_compression import (
 from agent.turn_context import _review_fork_first_request_pending
 from agent.turn_context_compaction import (
     _apply_grown_window, _blocked_compress_reason, _clear_overflow_warn, _refund_api_call,
-    _reset_retry_state_after_compaction,
+    _reanchor, _reset_retry_state_after_compaction,
 )
 
 logger = logging.getLogger("agent.conversation_loop")
@@ -238,13 +238,14 @@ class PostToolCompressionVerdict:
     compression_attempts: int
     final_response: Any
     turn_exit_reason: Any
+    current_turn_user_idx: int
 
 
 def compress_after_tool_results(
     agent: Any, *, messages: List[Dict[str, Any]], system_message: Any, user_message: Any,
     active_system_prompt: Any, conversation_history: Any, compression_attempts: int,
     max_compression_attempts: int, effective_task_id: Any, final_response: Any,
-    turn_exit_reason: Any,
+    turn_exit_reason: Any, current_turn_user_idx: int,
 ) -> PostToolCompressionVerdict:
     """Post-tool-call compression decision. Pressure comes from API-reported
     ``prompt_tokens`` (a tight lower bound; thinking models inflate completion tokens),
@@ -263,6 +264,7 @@ def compress_after_tool_results(
             end_turn=end_turn, messages=messages, active_system_prompt=active_system_prompt,
             conversation_history=conversation_history, compression_attempts=compression_attempts,
             final_response=final_response, turn_exit_reason=turn_exit_reason,
+            current_turn_user_idx=current_turn_user_idx,
         )
 
     _compressor = agent.context_compressor
@@ -355,6 +357,7 @@ def compress_after_tool_results(
                     final_response = _HANDOFF_SKIP_FINAL_RESPONSE
                 turn_exit_reason = "compaction_handoff_not_actionable"
                 return _verdict(True)
+            current_turn_user_idx = _reanchor(agent, messages, user_message)
     elif agent.compression_enabled:
         # Over threshold but compression blocked (cooldown/anti-thrash): deduped
         # warning so context can't silently overflow. ``attempts_spent`` names the

agent/turn_tool_round.py

@@ -39,6 +39,7 @@ class ToolRoundVerdict:
     failed: Any
     _turn_exit_reason: Any
     truncated_tool_call_retries: Any
+    current_turn_user_idx: Any
     result: Optional[Dict[str, Any]] = None
 
 
@@ -47,7 +48,7 @@ def run_tool_round(
     conversation_history: Any, api_call_count: Any, effective_task_id: Any, user_message: Any,
     system_message: Any, active_system_prompt: Any, compression_attempts: Any,
     max_compression_attempts: Any, final_response: Any, failed: Any, _turn_exit_reason: Any,
-    truncated_tool_call_retries: Any,
+    truncated_tool_call_retries: Any, current_turn_user_idx: Any,
 ) -> ToolRoundVerdict:
     """Execute one tool round in the exact original order. Persist-before-execute is a
     durability invariant: resume must see the executed block if a destructive tool restarts
@@ -60,7 +61,8 @@ def run_tool_round(
             action=action, messages=messages, conversation_history=conversation_history,
             active_system_prompt=active_system_prompt, compression_attempts=compression_attempts,
             final_response=final_response, failed=failed, _turn_exit_reason=_turn_exit_reason,
-            truncated_tool_call_retries=truncated_tool_call_retries, result=result,
+            truncated_tool_call_retries=truncated_tool_call_retries,
+            current_turn_user_idx=current_turn_user_idx, result=result,
         )
 
     if not agent.quiet_mode:
@@ -191,6 +193,7 @@ def run_tool_round(
         compression_attempts=compression_attempts,
         max_compression_attempts=max_compression_attempts, effective_task_id=effective_task_id,
         final_response=final_response, turn_exit_reason=_turn_exit_reason,
+        current_turn_user_idx=current_turn_user_idx,
     )
     messages = _ptc.messages
     active_system_prompt = _ptc.active_system_prompt
@@ -198,6 +201,7 @@ def run_tool_round(
     compression_attempts = _ptc.compression_attempts
     final_response = _ptc.final_response
     _turn_exit_reason = _ptc.turn_exit_reason
+    current_turn_user_idx = _ptc.current_turn_user_idx
     if _ptc.end_turn:
         return _verdict("break")
 

tests/agent/test_post_tool_compression_turn_boundary.py

@@ -0,0 +1,56 @@
+from types import SimpleNamespace
+
+from agent.turn_preflight import compress_after_tool_results
+
+
+def test_post_tool_compression_reanchors_the_active_user_boundary(monkeypatch):
+    compressed = [
+        {"role": "user", "content": "compressed history"},
+        {"role": "user", "content": "current ask"},
+        {"role": "assistant", "tool_calls": [{"id": "2"}]},
+        {"role": "tool", "content": "fresh result", "tool_call_id": "2"},
+    ]
+
+    class Compressor:
+        last_prompt_tokens = 100
+        threshold_tokens = 50
+
+        @staticmethod
+        def should_compress(_tokens):
+            return True
+
+    agent = SimpleNamespace(
+        context_compressor=Compressor(),
+        compression_enabled=True,
+        _clear_context_overflow_warn=lambda: None,
+        _safe_print=lambda *_args: None,
+        _compress_context=lambda *_args, **_kwargs: (compressed, "system"),
+        _persist_user_message_idx=4,
+    )
+    monkeypatch.setattr(
+        "agent.turn_preflight.conversation_history_after_compression",
+        lambda _agent, _messages, _history: [],
+    )
+    monkeypatch.setattr(
+        "agent.conversation_loop._should_skip_model_call_for_reference_handoff",
+        lambda _messages, _user_message: False,
+    )
+
+    verdict = compress_after_tool_results(
+        agent,
+        messages=[{"role": "user", "content": "current ask"}],
+        system_message="system",
+        user_message="current ask",
+        active_system_prompt="system",
+        conversation_history=[],
+        compression_attempts=0,
+        max_compression_attempts=1,
+        effective_task_id="task",
+        final_response="",
+        turn_exit_reason=None,
+        current_turn_user_idx=0,
+    )
+
+    assert verdict.messages is compressed
+    assert verdict.current_turn_user_idx == 1
+    assert agent._persist_user_message_idx == 1