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(router): keep tagged blocks out of the lossy size-gate fallback (#4543)

HetCreepf333edf20f0d
diff --git a/headroom/transforms/content_router.py b/headroom/transforms/content_router.py
index 87e06c4cc..e5905f0fb 100644
--- a/headroom/transforms/content_router.py
+++ b/headroom/transforms/content_router.py
@@ -5193,6 +5193,13 @@ class ContentRouter(Transform):
self._kompress_max_tokens,
self._kompress_gate_fires,
)
+ if protected:
+ # The fallback compressors take no ``ccr_original`` and would
+ # store the ``{{HEADROOM_TAG_N}}`` placeholder text in CCR; any
+ # placeholder line they drop is also discarded by
+ # ``restore_tags``, losing the tagged block outright. Pass the
+ # original through unchanged instead (as the budget branch above).
+ return content, _estimate_tokens(content)
out = text_to_compress
crusher = self._get_text_crusher()
if crusher is not None:
@@ -5213,8 +5220,6 @@ class ContentRouter(Transform):
"Kompress size-gate -> LogCompressor failed (%s); passing through", e
)
out = text_to_compress
- if protected:
- out = restore_tags(out, protected)
return out, _estimate_tokens(out)
# Reached only when the gate is enabled and this eligible block is
diff --git a/tests/test_transforms/test_kompress_size_gate.py b/tests/test_transforms/test_kompress_size_gate.py
index 3c718401d..eb2344678 100644
--- a/tests/test_transforms/test_kompress_size_gate.py
+++ b/tests/test_transforms/test_kompress_size_gate.py
@@ -74,3 +74,28 @@ def test_gate_fires_through_strategy_dispatch(monkeypatch, strategy):
router._apply_strategy_to_content(big, strategy, "")
assert router._kompress_gate_fires == 1
+
+
+def test_gate_keeps_tagged_blocks_out_of_lossy_fallback(monkeypatch):
+ # Over the gate, with custom tags protected: the TextCrusher/LogCompressor
+ # fallback must not see the {{HEADROOM_TAG_N}} placeholder text. It would
+ # store the placeholders in CCR and ``restore_tags`` discards any wrap whose
+ # placeholder line it dropped, so the tagged content would vanish from both
+ # the output and the retrieval path.
+ router = ContentRouter()
+ router._kompress_max_tokens = 50 # tiny ceiling (x4 = 200 chars)
+
+ def _boom():
+ raise AssertionError("kompress must not be invoked for oversized (gated) input")
+
+ monkeypatch.setattr(router, "_get_kompress", _boom)
+
+ lines = [f"<item>row {i} value</item> note text here for item number {i}" for i in range(200)]
+ text = "Results from the fetch tool:\n\n" + "\n".join(lines)
+
+ out, _ = router._try_ml_compressor(text, "")
+
+ assert router._kompress_gate_fires == 1
+ assert out == text
+ assert out.count("<item>") == 200
+ assert "{{HEADROOM_TAG" not in out