diff --git a/opencollab/application/compaction_summary.py b/opencollab/application/compaction_summary.py
index 088b2ed4..a3913d03 100644
--- a/opencollab/application/compaction_summary.py
+++ b/opencollab/application/compaction_summary.py
@@ -153,6 +153,8 @@ async def asummarize(self, segment: list[dict[str, Any]]) -> str:
 
     async def _complete(self, request: list[dict[str, Any]]) -> str:
         response = await self._acomplete(request)
+        if getattr(response, "finish_reason", None) in {"length", "max_tokens", "model_context_window_exceeded"}:
+            raise ValueError("The model returned a truncated conversation summary")
         return format_compact_summary(getattr(response, "content", None) or "")
 
     def _fits(self, request: list[dict[str, Any]], *, allow_budget_stop: bool = False) -> bool:
diff --git a/tests/runtime/test_bounded_summary.py b/tests/runtime/test_bounded_summary.py
index 4e27ac24..47febca9 100644
--- a/tests/runtime/test_bounded_summary.py
+++ b/tests/runtime/test_bounded_summary.py
@@ -113,7 +113,10 @@ async def complete(request):
     assert requests == [build_summary_request(segment)]
     assert summarizer.last_call_cacheable
 
-@pytest.mark.parametrize("failure", [RuntimeError("provider failed"), "<summary> </summary>"])
+@pytest.mark.parametrize("failure", [
+    RuntimeError("provider failed"), "<summary> </summary>",
+    SimpleNamespace(content="<summary>Truncated earlier changes.</summary>", finish_reason="max_tokens"),
+])
 async def test_chunk_failure_stops_before_the_tail_and_never_returns_a_partial_summary(failure):
     requests = []
 
@@ -123,7 +126,7 @@ async def complete(request):
             return SimpleNamespace(content="<summary>earlier user changes</summary>")
         if isinstance(failure, Exception):
             raise failure
-        return SimpleNamespace(content=failure)
+        return SimpleNamespace(content=failure) if isinstance(failure, str) else failure
 
     summarizer = ReadTimeSummarizer(
         complete, context_window=WINDOW, max_output_tokens=OUTPUT, transcript_path="session.json",
diff --git a/tests/runtime/test_compaction_summary.py b/tests/runtime/test_compaction_summary.py
index 0dcaf249..4c650536 100644
--- a/tests/runtime/test_compaction_summary.py
+++ b/tests/runtime/test_compaction_summary.py
@@ -5,6 +5,8 @@
 import asyncio
 from types import SimpleNamespace
 
+import pytest
+
 import opencollab.application.compaction_summary as compaction_summary
 from opencollab.application.compaction_summary import ReadTimeSummarizer, run_coro_blocking
 
@@ -68,6 +70,24 @@ def test_summarizer_falls_back_on_llm_error():
     assert "Summary:" not in out
     assert s.last_call_cacheable is False
 
+@pytest.mark.parametrize("finish_reason", ["length", "max_tokens", "model_context_window_exceeded"])
+async def test_truncation_rejects_even_a_well_formed_summary(finish_reason):
+    async def complete(_request):
+        return SimpleNamespace(content="<summary>Partial context.</summary>", finish_reason=finish_reason)
+
+    summarizer = ReadTimeSummarizer(complete)
+    assert await summarizer.asummarize(SEGMENT) == "[user]: build the thing\n[assistant]: working on it"
+    assert not summarizer.last_call_cacheable
+
+@pytest.mark.parametrize("finish_reason", [None, "stop", "end_turn", "stop_sequence"])
+async def test_complete_summary_accepts_supported_stop_reasons(finish_reason):
+    async def complete(_request):
+        return SimpleNamespace(content="<summary>Complete context.</summary>", finish_reason=finish_reason)
+
+    summarizer = ReadTimeSummarizer(complete)
+    assert await summarizer.asummarize(SEGMENT) == "Summary:\nComplete context."
+    assert summarizer.last_call_cacheable
+
 def test_summarizer_falls_back_when_no_summary_block():
     s = _summarizer("<analysis>only thinking, forgot the summary</analysis>")
     out = s(SEGMENT)
diff --git a/tests/runtime/test_summary_completion_http.py b/tests/runtime/test_summary_completion_http.py
new file mode 100644
index 00000000..20667243
--- /dev/null
+++ b/tests/runtime/test_summary_completion_http.py
@@ -0,0 +1,116 @@
+"""Provider truncation retains history until a complete summary can be cached."""
+
+from __future__ import annotations
+
+import copy
+import json
+
+import pytest
+
+from opencollab.adapters.env import LocalEnvironment
+from opencollab.adapters.llm.client import LLMClient
+from opencollab.bootstrap import build_session
+from opencollab.domain.agent import Agent
+from tests.support.provider_sdk_http import http_response, install_sdk_transport
+from tests.support.responses_provider_test_support import message_item
+
+REQUIREMENT = "Preserve the SQLite backup before replacing the database."
+
+
+@pytest.mark.parametrize("configuration", [
+    {"provider": "openai", "wire_protocol": "chat_completions", "model": "gpt-4o"},
+    {"provider": "openai", "wire_protocol": "responses", "model": "gpt-4o"},
+    {"provider": "anthropic", "wire_protocol": "chat_completions", "model": "claude-sonnet-4-6"},
+], ids=["chat", "responses", "anthropic"])
+@pytest.mark.parametrize("truncated", [False, True], ids=["complete", "truncated"])
+async def test_native_summary_completion_preserves_history_usage_and_cache(
+    tmp_path, monkeypatch, configuration, truncated,
+):
+    requests = []
+    monkeypatch.setenv("OPENCOLLAB_API_USAGE_LOG", str(tmp_path / "usage.jsonl"))
+    monkeypatch.setenv("OPENCOLLAB_BUDGET_NUDGE_MODE", "off")
+    monkeypatch.setenv("OPENCOLLAB_WRITE_NUDGE_MODE", "off")
+
+    async def handler(request):
+        body = json.loads(request.content)
+        requests.append(body)
+        summary = "create a detailed summary of the conversation" in json.dumps(body)
+        partial = summary and truncated and len(requests) == 1
+        text = (
+            "<summary>Earlier setup notes." if partial else f"<summary>{REQUIREMENT}</summary>"
+        ) if summary else ("Backup preserved." if REQUIREMENT in json.dumps(body) else "Backup missing.")
+        output_tokens = 1_024 if partial else 10
+        if configuration["provider"] == "anthropic":
+            return http_response(request, 200, json={
+                "id": "msg_summary", "type": "message", "role": "assistant", "model": body["model"],
+                "content": [{"type": "text", "text": text}],
+                "stop_reason": "max_tokens" if partial else "end_turn", "stop_sequence": None,
+                "usage": {"input_tokens": 100, "output_tokens": output_tokens},
+            })
+        if configuration["wire_protocol"] == "responses":
+            item = message_item(text)
+            item["status"] = "incomplete" if partial else "completed"
+            terminal = {
+                "id": "resp_summary", "object": "response", "model": body["model"],
+                "status": "incomplete" if partial else "completed", "error": None,
+                "incomplete_details": {"reason": "max_output_tokens"} if partial else None,
+                "output": [item],
+                "usage": {"input_tokens": 100, "output_tokens": output_tokens, "total_tokens": 100 + output_tokens},
+            }
+            events = [
+                {"type": "response.output_item.done", "output_index": 0, "item": item},
+                {"type": "response.incomplete" if partial else "response.completed", "response": terminal},
+            ]
+            wire = "".join(f"data: {json.dumps(event)}\n\n" for event in events)
+            return http_response(request, 200, headers={"content-type": "text/event-stream"}, text=wire)
+        return http_response(request, 200, json={
+            "id": "chat_summary", "object": "chat.completion", "created": 1, "model": body["model"],
+            "choices": [{"index": 0, "message": {"role": "assistant", "content": text},
+                         "finish_reason": "length" if partial else "stop"}],
+            "usage": {"prompt_tokens": 100, "completion_tokens": output_tokens, "total_tokens": 100 + output_tokens},
+        })
+
+    install_sdk_transport(monkeypatch, configuration["provider"], handler)
+    async with LLMClient(
+        **configuration, api_key="controlled-test",  # pragma: allowlist secret
+        base_url="https://controlled.invalid/v1", max_retries=0, context_window=40_000,
+    ) as client:
+        session = build_session(
+            agent=Agent(name="notes", system_prompt="Complete the user task.",
+                        **configuration, tools=[], max_tokens_per_step=1_024),
+            llm=client, env=LocalEnvironment(str(tmp_path)), max_budget_tokens=None,
+        )
+        original = [
+            {"role": "system", "content": "Complete the user task."},
+            {"role": "user", "content": "Earlier setup notes. " * 1_800 + REQUIREMENT},
+            {"role": "assistant", "content": "Earlier result."},
+            {"role": "user", "content": "The next work item is ready."},
+            {"role": "assistant", "content": "Another result."},
+            {"role": "user", "content": "Now finish the task."},
+        ]
+        session.messages = copy.deepcopy(original)
+        try:
+            assert await session.run_loop() == "Backup preserved."
+            assert len(requests) == 2
+            assert REQUIREMENT in json.dumps(requests[0])
+            assert REQUIREMENT in json.dumps(requests[1])
+            assert session.messages[:len(original)] == original
+            assert session.used_tokens == (1_234 if truncated else 220)
+
+            compact = session.runner.shaper._shapers[-1]
+            assert compact.summarizer.last_call_cacheable is (not truncated)
+            assert len(compact._summary_cache) == (0 if truncated else 1)
+            if truncated:
+                assert original[1]["content"] in json.dumps(requests[1])
+
+            projection = await session.runner.shaper.ashape(original)
+            assert REQUIREMENT in json.dumps(projection)
+            assert len(requests) == (3 if truncated else 2)
+            assert compact.summarizer.last_call_cacheable
+            assert len(compact._summary_cache) == 1
+            assert await session.runner.shaper.ashape(original) == projection
+            assert len(requests) == (3 if truncated else 2)
+            assert session.used_tokens == (1_344 if truncated else 220)
+        finally:
+            await session.aclose()
+            await session.env.cleanup()