refactor(context): detect overflow via LiteLLM's typed error only

Drop the brittle substring matching for context-overflow detection. LiteLLM
already normalises every provider's overflow error to
ContextWindowExceededError and maintains the provider-specific matching
upstream, so is_context_overflow just checks that type.

strix/llm/compaction.py

@@ -13,6 +13,7 @@ import logging
 from typing import TYPE_CHECKING, Any
 
 import litellm
+from litellm.exceptions import ContextWindowExceededError
 
 from strix.config import load_settings
 from strix.core.sessions import replace_session_items, session_write_lock
@@ -30,30 +31,16 @@ _TOOL_OUTPUT_MAX_CHARS = 2_000
 _MIN_ITEMS_TO_COMPACT = 6
 _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.
-# Deliberately excludes rate-limit/throttle wording, which must not trigger
-# compaction.
-_OVERFLOW_MARKERS = (
-    "context length",
-    "context window",
-    "maximum context",
-    "context_length_exceeded",
-    "too many tokens",
-    "reduce the length",
-    "input is too long",
-    "prompt is too long",
-    "exceeds the maximum",
-    "string too long",
-)
-
 
 def is_context_overflow(exc: BaseException) -> bool:
-    """Whether ``exc`` looks like a model context-window-overflow error."""
-    overflow_error = getattr(litellm, "ContextWindowExceededError", None)
-    if overflow_error is not None and isinstance(exc, overflow_error):
-        return True
-    message = str(exc).lower()
-    return any(marker in message for marker in _OVERFLOW_MARKERS)
+    """Whether ``exc`` is a model context-window-overflow error.
+
+    LiteLLM normalises every provider's overflow error to
+    ``ContextWindowExceededError`` (a ``BadRequestError`` subclass) and keeps
+    the provider-specific detection upstream, so we rely on that type rather
+    than matching error message strings ourselves.
+    """
+    return isinstance(exc, ContextWindowExceededError)
 
 
 _SUMMARY_INSTRUCTIONS = """\

tests/test_compaction.py

@@ -6,6 +6,7 @@ from types import SimpleNamespace
 from typing import TYPE_CHECKING, Any
 
 import pytest
+from litellm.exceptions import ContextWindowExceededError, RateLimitError
 
 from strix.config import ContextSettings
 from strix.llm import compaction
@@ -70,12 +71,17 @@ def _has_orphan_tool_output(items: list[Any]) -> bool:
     return any(i["call_id"] not in call_ids for i in items if compaction._is_tool_output(i))
 
 
-def test_is_context_overflow_matches_and_excludes() -> None:
-    assert compaction.is_context_overflow(
-        RuntimeError("This model's maximum context length is 8192")
+def test_is_context_overflow_uses_litellm_typed_error() -> None:
+    # LiteLLM maps every provider's overflow error to this type; anything else
+    # (including a rate-limit error) must not trigger compaction.
+    overflow = ContextWindowExceededError(
+        message="context length exceeded", model="m", llm_provider="openai"
     )
-    assert compaction.is_context_overflow(ValueError("input is too long for the model"))
-    assert not compaction.is_context_overflow(RuntimeError("rate limit exceeded, retry later"))
+    assert compaction.is_context_overflow(overflow)
+    assert not compaction.is_context_overflow(
+        RateLimitError(message="slow down", model="m", llm_provider="openai")
+    )
+    assert not compaction.is_context_overflow(RuntimeError("maximum context length is 8192"))
 
 
 def test_select_split_never_orphans_tool_output(monkeypatch: pytest.MonkeyPatch) -> None: