fix(context): never floor summary input above the model's remaining room

The 1,000-token floor on summary-input room could exceed the space
actually left after instructions, the reserved summary output, and any
existing checkpoint. On a small window the summary request then overflowed
and returned nothing, leaving the oversized session uncompacted. Clamp to
the real room instead so the request always fits.

strix/llm/compaction.py

@@ -28,8 +28,6 @@ logger = logging.getLogger(__name__)
 _CHECKPOINT_TAG = "<conversation-checkpoint>"
 _TOOL_OUTPUT_MAX_CHARS = 2_000
 _MIN_ITEMS_TO_COMPACT = 6
-# Floor for how much of the head we still try to summarise even on a tiny model.
-_MIN_SUMMARY_INPUT_TOKENS = 1_000
 _HEAD_TRUNCATED_MARKER = "\n\n[... older conversation omitted to fit the summary request ...]\n\n"
 
 # Substrings that identify a context-window-overflow error across providers.
@@ -214,9 +212,11 @@ def _summary_input_budget(model: str, previous: str | None) -> int:
     if previous:
         overhead += count_tokens(model, previous)
     # Leave slack for the prompt's wrapper text ("Conversation to summarise:",
-    # the update instructions, etc.) that is not part of ``overhead``.
+    # the update instructions, etc.) that is not part of ``overhead``. Never
+    # floor above the actual room: doing so would let the summary request
+    # itself overflow a small window (and then compaction silently fails).
     room = context_window(model) - context.summary_max_tokens - overhead - 256
-    return max(_MIN_SUMMARY_INPUT_TOKENS, room)
+    return max(0, room)
 
 
 def _build_summary_prompt(serialized_head: str, previous: str | None) -> str:

tests/test_compaction.py

@@ -202,6 +202,29 @@ async def test_maybe_compact_bounds_summary_prompt(monkeypatch: pytest.MonkeyPat
     assert len(captured["prompt"]) <= 4_000
 
 
+@pytest.mark.asyncio
+async def test_summary_request_fits_when_room_is_below_old_floor(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    # When the window leaves less head-input room than the old fixed floor, the
+    # budget must shrink to the real room so the request still fits the window
+    # (a fixed floor above the room would overflow and silently fail).
+    instructions = len(compaction._SUMMARY_INSTRUCTIONS)
+    window = instructions + 64 + 256 + 300  # summary_max(64)+slack(256)+room(300)
+    _patch_budget(monkeypatch, keep_tokens=30, window=window)
+    captured: dict[str, str] = {}
+
+    async def fake_acompletion(**kwargs: Any) -> Any:
+        captured["prompt"] = kwargs["messages"][0]["content"]
+        return SimpleNamespace(choices=[SimpleNamespace(message=SimpleNamespace(content="S"))])
+
+    monkeypatch.setattr("strix.llm.compaction.litellm.acompletion", fake_acompletion)
+    session = FakeSession([{"role": "user", "content": "y" * 5_000} for _ in range(20)])
+
+    assert await compaction.maybe_compact(session, model="m") is True
+    assert len(captured["prompt"]) <= window
+
+
 @pytest.mark.asyncio
 async def test_maybe_compact_skips_when_summary_fails(monkeypatch: pytest.MonkeyPatch) -> None:
     _patch_budget(monkeypatch, keep_tokens=30, window=50)