fix(context): skip compaction when the window can't fit a summary request

When instructions, the reserved summary output, and any prior checkpoint
already consume the model's context window, the summary-input budget is
zero. Previously a doomed summary request was still submitted (and
rejected). Skip summarization in that case and log, rather than issue a
request guaranteed to overflow.

strix/llm/compaction.py

@@ -297,13 +297,19 @@ async def maybe_compact(
 
     split = _select_split(model, items, context.keep_tokens)
     head, recent = items[:split], items[split:]
-    if not head:
+    previous = _previous_summary(head)
+    input_budget = _summary_input_budget(model, previous)
+    if not head or input_budget <= 0:
+        # Nothing to summarise, or the window can't even fit the summary
+        # instructions plus the reserved output allowance — any request would
+        # be rejected, so skip rather than submit a doomed call.
+        if head:
+            logger.warning(
+                "skipping compaction for %s: no room to summarise within its context window", model
+            )
         return False
 
-    previous = _previous_summary(head)
-    serialized_head = _fit_to_tokens(
-        model, _serialize_items(head), _summary_input_budget(model, previous)
-    )
+    serialized_head = _fit_to_tokens(model, _serialize_items(head), input_budget)
     summary = await _summarize(
         model,
         _build_summary_prompt(serialized_head, previous),

tests/test_compaction.py

@@ -139,11 +139,11 @@ async def test_maybe_compact_noop_when_within_budget(monkeypatch: pytest.MonkeyP
 
 @pytest.mark.asyncio
 async def test_maybe_compact_rewrites_and_keeps_pairs(monkeypatch: pytest.MonkeyPatch) -> None:
-    _patch_budget(monkeypatch, keep_tokens=30, window=50)
+    _patch_budget(monkeypatch, keep_tokens=30, window=4_000)
     _patch_summary(monkeypatch, "SUMMARY BODY")
     session = FakeSession(_turns(12))
 
-    assert await compaction.maybe_compact(session, model="m") is True
+    assert await compaction.maybe_compact(session, model="m", force=True) is True
 
     items = await session.get_items()
     assert items[0]["role"] == "user"
@@ -155,7 +155,8 @@ async def test_maybe_compact_rewrites_and_keeps_pairs(monkeypatch: pytest.Monkey
 
 @pytest.mark.asyncio
 async def test_maybe_compact_updates_previous_summary(monkeypatch: pytest.MonkeyPatch) -> None:
-    _patch_budget(monkeypatch, keep_tokens=30, window=50)
+    # Window large enough to leave real room for the summary instructions.
+    _patch_budget(monkeypatch, keep_tokens=30, window=4_000)
     captured: dict[str, str] = {}
 
     async def fake_acompletion(**kwargs: Any) -> Any:
@@ -167,7 +168,7 @@ async def test_maybe_compact_updates_previous_summary(monkeypatch: pytest.Monkey
     prior = compaction._checkpoint_item("OLD SUMMARY TEXT")
     session = FakeSession([prior, *_turns(12)])
 
-    assert await compaction.maybe_compact(session, model="m") is True
+    assert await compaction.maybe_compact(session, model="m", force=True) is True
     assert "OLD SUMMARY TEXT" in captured["prompt"]
 
 
@@ -227,7 +228,7 @@ async def test_summary_request_fits_when_room_is_below_old_floor(
 
 @pytest.mark.asyncio
 async def test_maybe_compact_skips_when_summary_fails(monkeypatch: pytest.MonkeyPatch) -> None:
-    _patch_budget(monkeypatch, keep_tokens=30, window=50)
+    _patch_budget(monkeypatch, keep_tokens=30, window=4_000)
 
     async def fake_acompletion(**_kwargs: Any) -> Any:
         raise RuntimeError("boom")
@@ -236,5 +237,28 @@ async def test_maybe_compact_skips_when_summary_fails(monkeypatch: pytest.Monkey
     session = FakeSession(_turns(12))
     before = await session.get_items()
 
-    assert await compaction.maybe_compact(session, model="m") is False
+    assert await compaction.maybe_compact(session, model="m", force=True) is False
+    assert await session.get_items() == before
+
+
+@pytest.mark.asyncio
+async def test_maybe_compact_skips_when_no_room_to_summarise(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    # A window too small to fit the summary instructions plus the reserved
+    # output leaves no room for any head, so no (doomed) summary is attempted.
+    _patch_budget(monkeypatch, keep_tokens=30, window=200)
+    called = False
+
+    async def fake_acompletion(**_kwargs: Any) -> Any:
+        nonlocal called
+        called = True
+        return SimpleNamespace(choices=[SimpleNamespace(message=SimpleNamespace(content="S"))])
+
+    monkeypatch.setattr("strix.llm.compaction.litellm.acompletion", fake_acompletion)
+    session = FakeSession(_turns(12))
+    before = await session.get_items()
+
+    assert await compaction.maybe_compact(session, model="m", force=True) is False
+    assert called is False
     assert await session.get_items() == before