fix(compression,llm): keep the summary across mid-execution compression, ignore empty saved points, commit the head hash on success

Three review findings on the bounded-chain change.

Mid-execution compression rebuilt the conversation from the in-flight
messages, which after a turn-start reuse hold only the recent turns: the
summary living in the system prompt never reached the compressor, so the
new summary replaced the old one, and the persisted point's query_index
was relative to that shortened list. The summary the agent is running
under now rides into the synthetic conversation as its latest point
(query_index -1, so every in-flight query is new), for both the database
and the in-memory path, and the database path persists the index of the
saved conversation's last row.

Saved points with an empty summary, which earlier versions wrote, were
treated as reusable: get_compressed_context sliced the raw history away
and the effective token count made the conversation look small. Point
selection everywhere now takes the latest usable point (non-blank
summary, positive token count) and falls back to the raw history when
there is none.

The chained system-head hash was committed while building the request,
so a transport failure followed by the same-primary retry omitted a
changed system message. The hash is now staged per request and committed
only when the provider records the response.

tests/api/answer/services/compression/test_compression_reuse.py

@@ -338,3 +338,164 @@ class TestIncrementalTailExcludesSummaryRows:
         assert result.success and not result.compression_performed
         assert result.compressed_summary == "S"
         svc.compress_and_save.assert_not_called()
+
+
+@pytest.mark.unit
+class TestAbsolutePersistIndex:
+    def test_compress_conversation_persists_the_given_absolute_index(self):
+        llm = MagicMock()
+        llm.gen.return_value = "<summary>new</summary>"
+        svc = CompressionService(llm=llm, model_id="m")
+        svc.prompt_builder = MagicMock(version="v1.0")
+        svc.prompt_builder.build_prompt.return_value = [{"role": "user", "content": "p"}]
+        conv = {"queries": [{"prompt": "q18", "response": "r"}, {"prompt": "q19", "response": ""}]}
+
+        metadata = svc.compress_conversation(conv, compress_up_to_index=1, persist_query_index=19)
+
+        assert metadata.query_index == 19
+
+    @patch(
+        "application.api.answer.services.compression.orchestrator.get_provider_from_model_id",
+        return_value="openai",
+    )
+    @patch(
+        "application.api.answer.services.compression.orchestrator.get_api_key_for_provider",
+        return_value="sk",
+    )
+    @patch("application.api.answer.services.compression.orchestrator.LLMCreator")
+    @patch("application.api.answer.services.compression.orchestrator.CompressionService")
+    @patch("application.api.answer.services.compression.orchestrator.settings")
+    def test_mid_execution_builds_on_the_carried_summary_and_persists_the_absolute_index(
+        self, mock_settings, MockCompressionService, MockLLMCreator, _key, _provider,
+        orchestrator, conversation_service,
+    ):
+        mock_settings.COMPRESSION_MODEL_OVERRIDE = None
+        MockLLMCreator.create_llm.return_value = MagicMock()
+        metadata = MagicMock()
+        metadata.compression_ratio = 3.0
+        metadata.original_token_count = 30
+        metadata.compressed_token_count = 10
+        metadata.timestamp = "2026-09-03T10:00:00+00:00"
+        svc = MagicMock()
+        svc.compress_and_save.return_value = metadata
+        svc.get_compressed_context.return_value = ("new", [])
+        MockCompressionService.return_value = svc
+        conversation_service.get_conversation.return_value = {"queries": [], "compression_metadata": {}}
+        synthetic = {
+            "queries": [{"prompt": "q18", "response": "r"}, {"prompt": "q19", "response": ""}],
+            "compression_metadata": {
+                "is_compressed": True,
+                "compression_points": [{"query_index": -1, "compressed_summary": "prior",
+                                        "compressed_token_count": 5, "original_token_count": 5}],
+            },
+        }
+
+        result = orchestrator.compress_mid_execution(
+            "conv1", "user1", "m", {"sub": "user1"}, current_conversation=synthetic,
+            persist_query_index=19,
+        )
+
+        assert result.success and result.compression_performed
+        args, kwargs = svc.compress_and_save.call_args
+        assert kwargs["start_index"] == 0          # every synthetic query is newer than the summary
+        assert kwargs["persist_query_index"] == 19  # indexed against the database conversation
+
+
+@pytest.mark.unit
+class TestUnusableSavedPoints:
+    """A saved point with an empty summary (older versions wrote them) must
+    never make a turn drop history."""
+
+    def _conversation_with_empty_point(self):
+        return {
+            "queries": [
+                {"prompt": "q0", "response": "first saved turn"},
+                {"prompt": "q1", "response": "second saved turn"},
+            ],
+            "compression_metadata": {
+                "is_compressed": True,
+                "last_compression_at": EPOCH,
+                "compression_points": [{
+                    "timestamp": EPOCH, "query_index": 1, "compressed_summary": "",
+                    "original_token_count": 493541, "compressed_token_count": 0,
+                }],
+            },
+            "agent_id": "agent-1",
+        }
+
+    def test_reuse_ignores_an_empty_point_and_keeps_history(
+        self, orchestrator, conversation_service, threshold_checker
+    ):
+        conversation_service.get_conversation.return_value = self._conversation_with_empty_point()
+        threshold_checker.should_compress.return_value = False
+
+        result = orchestrator.compress_if_needed("conv1", "user1", "m", {"sub": "user1"})
+
+        assert result.success is True
+        assert result.compressed_summary is None
+        assert [q["prompt"] for q in result.recent_queries] == ["q0", "q1"]
+        assert [h["prompt"] for h in result.as_history()] == ["q0", "q1"]
+
+    def test_effective_count_ignores_an_empty_point(self):
+        conv = self._conversation_with_empty_point()
+        assert TokenCounter.count_effective_conversation_tokens(conv) == (
+            TokenCounter.count_conversation_tokens(conv)
+        )
+
+    def test_get_compressed_context_ignores_an_empty_point(self):
+        summary, recent = CompressionService(llm=None, model_id="m").get_compressed_context(
+            self._conversation_with_empty_point()
+        )
+        assert summary is None
+        assert [q["prompt"] for q in recent] == ["q0", "q1"]
+
+    def test_get_compressed_context_falls_back_to_the_latest_usable_point(self):
+        conv = _compressed_conversation()
+        conv["queries"].append({"prompt": "q4", "response": "r4"})
+        conv["compression_metadata"]["compression_points"].append(
+            {"timestamp": "2026-09-04T09:00:00+00:00", "query_index": 3,
+             "compressed_summary": "   ", "compressed_token_count": 0}
+        )
+        summary, recent = CompressionService(llm=None, model_id="m").get_compressed_context(conv)
+        assert summary == "S"
+        assert [q["prompt"] for q in recent] == ["q2", "q3", "q4"]
+
+    @patch(
+        "application.api.answer.services.compression.orchestrator.get_provider_from_model_id",
+        return_value="openai",
+    )
+    @patch(
+        "application.api.answer.services.compression.orchestrator.get_api_key_for_provider",
+        return_value="sk",
+    )
+    @patch("application.api.answer.services.compression.orchestrator.LLMCreator")
+    @patch("application.api.answer.services.compression.orchestrator.CompressionService")
+    @patch("application.api.answer.services.compression.orchestrator.settings")
+    def test_recompression_starts_after_the_latest_usable_point(
+        self, mock_settings, MockCompressionService, MockLLMCreator, _key, _provider,
+        orchestrator, conversation_service, threshold_checker,
+    ):
+        mock_settings.COMPRESSION_MODEL_OVERRIDE = None
+        conv = _compressed_conversation()
+        conv["queries"].append({"prompt": "q4", "response": "r4"})
+        conv["compression_metadata"]["compression_points"].append(
+            {"timestamp": "2026-09-04T09:00:00+00:00", "query_index": 3,
+             "compressed_summary": "", "compressed_token_count": 0}
+        )
+        conversation_service.get_conversation.return_value = conv
+        threshold_checker.should_compress.return_value = True
+        MockLLMCreator.create_llm.return_value = MagicMock()
+        metadata = MagicMock()
+        metadata.compression_ratio = 3.0
+        metadata.original_token_count = 30
+        metadata.compressed_token_count = 10
+        metadata.timestamp = "2026-09-05T10:00:00+00:00"
+        svc = MagicMock()
+        svc.compress_and_save.return_value = metadata
+        svc.get_compressed_context.return_value = ("S2", [])
+        MockCompressionService.return_value = svc
+
+        result = orchestrator.compress_if_needed("conv1", "user1", "m", {"sub": "user1"})
+
+        assert result.compression_performed
+        assert svc.compress_and_save.call_args.kwargs["start_index"] == 2

tests/api/answer/services/compression/test_orchestrator.py

@@ -373,7 +373,7 @@ class TestCompressMidExecution:
                 "gpt-4",
                 decoded_token,
                 user_id="user1",
-                model_user_id=None,
+                model_user_id=None, persist_query_index=None,
             )
 
     def test_loads_conversation_when_not_provided(

tests/llm/handlers/test_mid_execution_chain.py

@@ -107,3 +107,112 @@ def test_empty_summary_is_a_failure_not_a_success():
     assert success is False and messages is None
     conv_service.append_compression_message.assert_not_called()
     agent.llm.start_responses_turn.assert_not_called()
+
+
+# ── the summary already in play must survive a mid-execution compression ────
+
+
+PRIOR = "Original delivery location: Warehouse Seven."
+
+
+def _run_with(handler, agent, result, db_conversation, synthetic):
+    conv_service = Mock()
+    conv_service.get_conversation.return_value = db_conversation
+    orchestrator = Mock()
+    orchestrator.compress_mid_execution.return_value = result
+    with patch(
+        "application.api.answer.services.compression.CompressionOrchestrator",
+        return_value=orchestrator,
+    ), patch(
+        "application.api.answer.services.conversation_service.ConversationService",
+        return_value=conv_service,
+    ), patch.object(
+        handler, "_build_conversation_from_messages", return_value=synthetic
+    ), patch.object(
+        handler,
+        "_rebuild_messages_after_compression",
+        return_value=[{"role": "system", "content": "rebuilt"}],
+    ), patch.object(handler, "_prune_messages_minimal", return_value=None):
+        outcome = handler._perform_mid_execution_compression(
+            agent, [{"role": "user", "content": "hi"}]
+        )
+    return outcome, orchestrator
+
+
+@pytest.mark.unit
+def test_db_path_carries_the_current_summary_and_persists_the_absolute_index():
+    agent = _agent()
+    agent.compressed_summary = PRIOR
+    agent.last_compression_at = "2026-09-03T09:00:00+00:00"
+    db_conversation = {
+        "queries": [{"prompt": f"q{i}", "response": "r"} for i in range(20)],
+        "compression_metadata": {
+            "is_compressed": True,
+            "compression_points": [{"query_index": 17, "compressed_summary": PRIOR,
+                                    "compressed_token_count": 12}],
+        },
+    }
+    # What the turn actually replayed: the two turns after the saved point
+    # plus the in-flight one. Their indexes are 0-2 here, 18-19 in the DB.
+    synthetic = {"queries": [{"prompt": "q18", "response": "r"}, {"prompt": "q19", "response": ""}],
+                 "compression_metadata": {"is_compressed": False, "compression_points": []}}
+
+    (success, _), orchestrator = _run_with(_Handler(), agent, _result(), db_conversation, synthetic)
+
+    assert success is True
+    kwargs = orchestrator.compress_mid_execution.call_args.kwargs
+    sent = kwargs["current_conversation"]
+    points = sent["compression_metadata"]["compression_points"]
+    assert sent["compression_metadata"]["is_compressed"] is True
+    # The summary the system prompt carries is what the compressor builds on...
+    assert points[-1]["compressed_summary"] == PRIOR
+    # ...and it predates every synthetic query, so all of them are "new".
+    assert points[-1]["query_index"] == -1
+    # The persisted point must index the DATABASE conversation, not the
+    # shortened synthetic list.
+    assert kwargs["persist_query_index"] == 19
+
+
+@pytest.mark.unit
+def test_db_path_without_a_summary_in_play_sends_no_carried_point():
+    agent = _agent()
+    agent.compressed_summary = None
+    db_conversation = {"queries": [{"prompt": "q0", "response": "r"}]}
+    synthetic = {"queries": [{"prompt": "q0", "response": ""}],
+                 "compression_metadata": {"is_compressed": False, "compression_points": []}}
+
+    _, orchestrator = _run_with(_Handler(), agent, _result(), db_conversation, synthetic)
+
+    sent = orchestrator.compress_mid_execution.call_args.kwargs["current_conversation"]
+    assert sent["compression_metadata"]["compression_points"] == []
+
+
+@pytest.mark.unit
+def test_in_memory_path_carries_the_current_summary():
+    agent = _agent()
+    agent.compressed_summary = PRIOR
+    agent.decoded_token = {"sub": "user1"}
+    metadata = Mock()
+    metadata.compressed_token_count = 10
+    metadata.original_token_count = 100
+    metadata.compression_ratio = 10.0
+    metadata.timestamp = "2026-09-03T10:00:00+00:00"
+    metadata.to_dict.return_value = {"query_index": 1, "compressed_summary": "new", "compressed_token_count": 10}
+    svc = Mock()
+    svc.compress_conversation.return_value = metadata
+    svc.get_compressed_context.return_value = ("new", [])
+    handler = _Handler()
+    with patch("application.api.answer.services.compression.service.CompressionService", return_value=svc), patch(
+        "application.llm.llm_creator.LLMCreator"
+    ), patch("application.core.model_utils.get_provider_from_model_id", return_value="openai"), patch(
+        "application.core.model_utils.get_api_key_for_provider", return_value="sk"
+    ), patch.object(
+        handler, "_build_conversation_from_messages",
+        return_value={"queries": [{"prompt": "q", "response": "r"}, {"prompt": "q2", "response": ""}],
+                      "compression_metadata": {"is_compressed": False, "compression_points": []}},
+    ), patch.object(handler, "_rebuild_messages_after_compression", return_value=[{"role": "system", "content": "x"}]):
+        success, _ = handler._perform_in_memory_compression(agent, [{"role": "user", "content": "hi"}])
+
+    assert success is True
+    sent = svc.compress_conversation.call_args[0][0]
+    assert sent["compression_metadata"]["compression_points"][-1]["compressed_summary"] == PRIOR

tests/llm/test_responses_chain_budget.py

@@ -58,6 +58,11 @@ def _roles(items):
     return [i.get("role") for i in items if isinstance(i, dict) and i.get("role")]
 
 
+def _accepted(llm, rid="resp_1"):
+    """The provider accepted the request just built: record its response."""
+    llm._record_responses_metadata(types.SimpleNamespace(id=rid, output=[], usage=None))
+
+
 # ── chained system head ──────────────────────────────────────────────────────
 
 
@@ -69,6 +74,7 @@ def test_chained_input_omits_unchanged_system_head(monkeypatch):
     first, prev = llm._build_responses_input(_messages(), None)
     assert prev is None
     assert "system" in _roles(first)
+    _accepted(llm)
 
     chained, prev = llm._build_responses_input(_messages(), "resp_1")
     assert prev == "resp_1"
@@ -79,9 +85,11 @@ def test_chained_input_omits_unchanged_system_head(monkeypatch):
 def test_chained_input_resends_changed_system_head(monkeypatch):
     llm = _make_llm(monkeypatch)
     llm._build_responses_input(_messages("sys v1"), None)
+    _accepted(llm)
 
     chained, _ = llm._build_responses_input(_messages("sys v2"), "resp_1")
     assert _roles(chained) == ["system", "user"]
+    _accepted(llm, "resp_2")
 
     # ...and the new head becomes the one the chain holds.
     again, _ = llm._build_responses_input(_messages("sys v2"), "resp_2")
@@ -92,6 +100,7 @@ def test_chained_input_resends_changed_system_head(monkeypatch):
 def test_chained_system_head_hash_roundtrips_through_state(monkeypatch):
     llm = _make_llm(monkeypatch)
     llm._build_responses_input(_messages(), None)
+    _accepted(llm)
     state = llm.export_responses_state()
     assert state.get("system_hash")
 
@@ -105,6 +114,7 @@ def test_chained_system_head_hash_roundtrips_through_state(monkeypatch):
 def test_start_responses_turn_forgets_system_head(monkeypatch):
     llm = _make_llm(monkeypatch)
     llm._build_responses_input(_messages(), None)
+    _accepted(llm)
     llm.start_responses_turn()
     chained, _ = llm._build_responses_input(_messages(), "resp_1")
     # A fresh chain has no head on the server yet, so the head is sent.
@@ -312,3 +322,35 @@ def test_cache_key_for_user_is_opaque_and_stable():
     assert key == _cache_key_for_user("user_2Vhzgd63RSgixvvbF8Z2nhtqnE9")
     assert key != _cache_key_for_user("someone-else")
     assert _cache_key_for_user(None) is None
+
+
+# ── the head hash commits only once the provider accepted the request ───────
+
+
+@pytest.mark.unit
+def test_system_hash_commits_only_when_the_response_is_recorded(monkeypatch):
+    llm = _make_llm(monkeypatch)
+    llm._build_responses_input(_messages(), None)
+    # Not recorded (the request failed before the provider stored it): a
+    # chained follow-up must still carry the head.
+    chained, _ = llm._build_responses_input(_messages(), "resp_1")
+    assert _roles(chained) == ["system", "user"]
+    _accepted(llm)
+    chained, _ = llm._build_responses_input(_messages(), "resp_1")
+    assert _roles(chained) == ["user"]
+
+
+@pytest.mark.unit
+def test_failed_chained_request_resends_changed_head_on_retry(monkeypatch):
+    llm = _make_llm(monkeypatch)
+    llm._build_responses_input(_messages("sys v1"), None)
+    _accepted(llm)
+    first, _ = llm._build_responses_input(_messages("sys v2"), "resp_1")
+    assert _roles(first) == ["system", "user"]
+    # Transport error before any chunk: nothing recorded. The same-primary
+    # retry chains onto resp_1 again, whose stored head is still "sys v1".
+    retry, _ = llm._build_responses_input(_messages("sys v2"), "resp_1")
+    assert _roles(retry) == ["system", "user"]
+    _accepted(llm, "resp_2")
+    after, _ = llm._build_responses_input(_messages("sys v2"), "resp_2")
+    assert _roles(after) == ["user"]