fix(context_management): count clear_at_least freed tokens over sent blocks only

litellm/llms/anthropic/messages/context_management/editors/clear_tool_uses.py

@@ -192,10 +192,12 @@ def _cleared_result(block: Mapping[str, object]) -> dict[str, object]:
 
 def _freed_tokens(model: str, result: Mapping[str, object], use: Mapping[str, object], clears_input: bool) -> int:
     """Tokens that clearing this use frees: its result, and its input when ``clear_tool_inputs`` covers it."""
-    before: Final = [result, use] if clears_input else [result]
-    after: Final = [_cleared_result(result), {**use, "input": {}}] if clears_input else [_cleared_result(result)]
+    before: Final[Sequence[Mapping[str, object]]] = (result, use) if clears_input else (result,)
+    after: Final[Sequence[Mapping[str, object]]] = (
+        (_cleared_result(result), {**use, "input": {}}) if clears_input else (_cleared_result(result),)
+    )
     counts: Final = tuple(
-        litellm.token_counter(model=model, messages=[{"role": "user", "content": blocks}]) for blocks in (before, after)
+        _count_tokens(model, [{"role": "user", "content": blocks}], None) for blocks in (before, after)
     )
     return max(counts[0] - counts[1], 0)
 

tests/unit/llms/anthropic/messages/context_management/test_clear_tool_uses.py

@@ -288,6 +288,38 @@ def test_an_input_tokens_trigger_skips_blocks_no_adapter_sends_inside_a_tool_res
     assert _cleared_ids(cleared) == {"toolu_00", "toolu_01"}
 
 
+def _clear_at_least_one_token(
+    messages: list[dict[str, object]],
+) -> tuple[list[dict[str, object]], AppliedEdit | None]:
+    return apply_clear_tool_uses_20250919(
+        model=MODEL,
+        messages=messages,
+        tools=None,
+        system=None,
+        edit_spec={
+            "type": "clear_tool_uses_20250919",
+            "trigger": {"type": "tool_uses", "value": 0},
+            "keep": {"type": "tool_uses", "value": 1},
+            "clear_at_least": {"type": "input_tokens", "value": 1},
+        },
+    )
+
+
+def test_clear_at_least_frees_only_blocks_an_adapter_sends_inside_a_tool_result() -> None:
+    history: Final = _make_history(n_pairs=3)
+    result_text: Final = {"type": "text", "text": "Result 0: " + "x" * 200}
+    search_result: Final = {"type": "web_search_result", "url": "https://example.com", "title": "rain " * 300}
+
+    _, sent_only = _clear_at_least_one_token([*history[:2], _first_result_with([result_text]), *history[3:]])
+    cleared, applied = _clear_at_least_one_token(
+        [*history[:2], _first_result_with([result_text, search_result]), *history[3:]]
+    )
+
+    assert applied is not None
+    assert applied == sent_only
+    assert _cleared_ids(cleared) == {"toolu_00", "toolu_01"}
+
+
 def _blocks_of(messages: list[dict[str, object]], block_type: str) -> tuple[dict[str, object], ...]:
     contents: Final = (message["content"] for message in messages)
     blocks: Final = chain.from_iterable(content for content in contents if isinstance(content, list))