Skip to content

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

fix(transforms/adaptive-sizer): honor max_k on small-input fast path (#2319)

Abhay Singh8a905232091d
diff --git a/headroom/transforms/adaptive_sizer.py b/headroom/transforms/adaptive_sizer.py
index 9a867d60a..5f73aabd8 100644
--- a/headroom/transforms/adaptive_sizer.py
+++ b/headroom/transforms/adaptive_sizer.py
@@ -51,7 +51,12 @@ def compute_optimal_k(
# Tier 1: Fast path
if n <= 8:
- return n
+ # Still honor the caller's hard cap: ``max_k`` is documented as "never
+ # return more than this". Returning the raw ``n`` let a small ``max_k``
+ # (e.g. a tight search/log budget) be exceeded on tiny inputs, so the
+ # caller kept more items than it asked for. The near-duplicate branch
+ # below already clamps to ``effective_max``; do the same here.
+ return min(n, effective_max)
# Check for near-total redundancy
unique_count = count_unique_simhash(items)
diff --git a/tests/test_adaptive_sizer.py b/tests/test_adaptive_sizer.py
index f20746d56..f9ae55c6d 100644
--- a/tests/test_adaptive_sizer.py
+++ b/tests/test_adaptive_sizer.py
@@ -107,6 +107,18 @@ class TestSmallArrays:
items = _make_unique_items(8)
assert compute_optimal_k(items) == 8
+ def test_small_array_respects_max_k(self):
+ """A small array (n <= 8) must still honor a tight ``max_k`` cap.
+
+ ``max_k`` is documented as "never return more than this"; the fast path
+ used to return the raw ``n`` and blow past a small cap.
+ """
+ items = _make_unique_items(8)
+ assert compute_optimal_k(items, max_k=5) == 5
+ assert compute_optimal_k(items, max_k=3) == 3
+ # A cap >= n leaves the array unchanged.
+ assert compute_optimal_k(items, max_k=20) == 8
+
class TestNearTotalRedundancy:
def test_identical_items_returns_min(self):