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(kompress): prevent loading model when enable_kompress is False (#4667)

NAZIYAa656eb437041
diff --git a/headroom/proxy/handlers/openai.py b/headroom/proxy/handlers/openai.py
index d94233f81..bc77e620b 100644
--- a/headroom/proxy/handlers/openai.py
+++ b/headroom/proxy/handlers/openai.py
@@ -4824,7 +4824,7 @@ class OpenAIHandlerMixin:
_rc_router = find_content_router(self.openai_pipeline)
_rc_kompress = (
(_rc_router._get_remote_kompress() or _rc_router._get_kompress())
- if _rc_router is not None
+ if _rc_router is not None and _rc_router.config.enable_kompress
else None
)
# Sticky drop floor: turns dropped on a prior cold turn must stay
diff --git a/headroom/transforms/cold_prefix.py b/headroom/transforms/cold_prefix.py
index b58ac4d14..8977ebf44 100644
--- a/headroom/transforms/cold_prefix.py
+++ b/headroom/transforms/cold_prefix.py
@@ -216,7 +216,13 @@ def cold_recompact_messages(
# enable_cross_turn_dedup stays on: whether the fold may EMIT pointers
# is decided per-path by the router's recoverability gate below, not
# hardcoded here.
- router = ContentRouter(ContentRouterConfig(lossless=True, enable_cross_turn_dedup=True))
+ router = ContentRouter(
+ ContentRouterConfig(
+ lossless=True,
+ enable_cross_turn_dedup=True,
+ enable_kompress=False,
+ )
+ )
res = router.apply(
list(messages),
tokenizer,
diff --git a/headroom/transforms/content_router.py b/headroom/transforms/content_router.py
index 3f55cf134..c43c85321 100644
--- a/headroom/transforms/content_router.py
+++ b/headroom/transforms/content_router.py
@@ -5854,6 +5854,9 @@ class ContentRouter(Transform):
(model weights are cached at module level in kompress_compressor.py,
so repeated calls with the same model_id are cheap)
"""
+ if not self.config.enable_kompress:
+ return None
+
model_id = getattr(self, "_runtime_kompress_model", None)
# Explicitly disabled — no ML compression
diff --git a/tests/test_proxy_disable_kompress.py b/tests/test_proxy_disable_kompress.py
index c1f3da46e..9f134dc7f 100644
--- a/tests/test_proxy_disable_kompress.py
+++ b/tests/test_proxy_disable_kompress.py
@@ -10,6 +10,12 @@ from headroom.proxy.server import ProxyConfig, create_app
from headroom.transforms import CompressionStrategy, ContentRouter
[email protected](autouse=True)
+def _stub_env(monkeypatch: pytest.MonkeyPatch) -> None:
+ monkeypatch.setenv("HEADROOM_REQUIRE_RUST_CORE", "false")
+ monkeypatch.setenv("HEADROOM_HTTP2", "false")
+
+
def _proxy_router(config: ProxyConfig):
app = create_app(config)
proxy = app.state.proxy
@@ -104,3 +110,155 @@ def test_disable_kompress_fallback_without_disable_kompress_is_noop() -> None:
assert router.config.enable_kompress is True
assert router.config.fallback_strategy == CompressionStrategy.KOMPRESS
+
+
+def test_disable_kompress_prevents_kompress_instantiation_in_content_router(
+ monkeypatch: pytest.MonkeyPatch,
+) -> None:
+ from unittest.mock import MagicMock
+
+ import headroom.transforms.kompress_compressor as kc
+ from headroom.transforms.content_router import ContentRouter, ContentRouterConfig
+
+ mock_avail = MagicMock(return_value=True)
+ mock_compressor = MagicMock()
+ monkeypatch.setattr(kc, "is_kompress_available", mock_avail)
+ monkeypatch.setattr(kc, "KompressCompressor", mock_compressor)
+
+ router = ContentRouter(ContentRouterConfig(enable_kompress=False))
+ result = router._get_kompress()
+ assert result is None
+ mock_avail.assert_not_called()
+ mock_compressor.assert_not_called()
+
+
+def test_disable_kompress_bypasses_remote_kompress(
+ monkeypatch: pytest.MonkeyPatch,
+) -> None:
+ from unittest.mock import MagicMock
+
+ from headroom.transforms.content_router import ContentRouter, ContentRouterConfig
+
+ mock_remote = MagicMock()
+ router = ContentRouter(ContentRouterConfig(enable_kompress=False))
+ monkeypatch.setattr(router, "_get_remote_kompress", mock_remote)
+
+ result = router._get_kompress()
+ assert result is None
+ mock_remote.assert_not_called()
+
+
+def test_disable_kompress_reasoning_compaction_skips_kompress_loader(
+ monkeypatch: pytest.MonkeyPatch,
+) -> None:
+ from unittest.mock import MagicMock
+
+ import headroom.transforms.kompress_compressor as kc
+ from headroom.proxy.server import ProxyConfig, create_app
+ from headroom.transforms.compression_units import find_content_router
+
+ mock_avail = MagicMock(return_value=True)
+ mock_compressor = MagicMock()
+ monkeypatch.setattr(kc, "is_kompress_available", mock_avail)
+ monkeypatch.setattr(kc, "KompressCompressor", mock_compressor)
+
+ config = ProxyConfig(
+ optimize=True,
+ disable_kompress=True,
+ cache_enabled=False,
+ rate_limit_enabled=False,
+ cost_tracking_enabled=False,
+ log_requests=False,
+ )
+ app = create_app(config)
+ proxy = app.state.proxy
+ router = find_content_router(proxy.openai_pipeline)
+ assert router is not None
+ assert router.config.enable_kompress is False
+ assert router._get_kompress() is None
+ mock_avail.assert_not_called()
+ mock_compressor.assert_not_called()
+
+
+def test_disabled_kompress_reasoning_request_never_queries_remote_loader(
+ monkeypatch: pytest.MonkeyPatch,
+) -> None:
+ import json
+ from unittest.mock import MagicMock
+
+ import httpx
+ from fastapi.testclient import TestClient
+
+ from headroom.transforms.compression_units import find_content_router
+
+ monkeypatch.setenv("HEADROOM_SKIP_UPSTREAM_CHECK", "1")
+ monkeypatch.setenv("HEADROOM_THINKING_COMPACT", "1")
+ monkeypatch.setenv("HEADROOM_THINKING_COMPACT_KEEP_LAST", "0")
+ monkeypatch.setenv("HEADROOM_KOMPRESS_ENDPOINT", "http://127.0.0.1:9")
+ app = create_app(
+ ProxyConfig(
+ optimize=True,
+ disable_kompress=True,
+ cache_enabled=False,
+ rate_limit_enabled=False,
+ cost_tracking_enabled=False,
+ log_requests=False,
+ ccr_inject_tool=False,
+ ccr_handle_responses=False,
+ ccr_context_tracking=False,
+ image_optimize=False,
+ )
+ )
+ forwarded = []
+
+ def upstream(request: httpx.Request) -> httpx.Response:
+ forwarded.append(json.loads(request.content))
+ return httpx.Response(
+ 200,
+ json={
+ "id": "fixture",
+ "object": "chat.completion",
+ "model": "kimi-k2.7",
+ "choices": [
+ {
+ "index": 0,
+ "message": {"role": "assistant", "content": "ok"},
+ "finish_reason": "stop",
+ }
+ ],
+ "usage": {"prompt_tokens": 100, "completion_tokens": 1, "total_tokens": 101},
+ },
+ )
+
+ reasoning = " ".join(f"reason{i}" for i in range(100))
+ with TestClient(app, client=("127.0.0.1", 50000)) as client:
+ proxy = app.state.proxy
+
+ async def install_transport() -> None:
+ await proxy.http_client.aclose()
+ if proxy.http_client_h1 is not None:
+ await proxy.http_client_h1.aclose()
+ proxy.http_client = httpx.AsyncClient(transport=httpx.MockTransport(upstream))
+ proxy.http_client_h1 = None
+
+ client.portal.call(install_transport)
+ router = find_content_router(proxy.openai_pipeline)
+ assert router is not None
+ remote_loader = MagicMock(return_value=None)
+ monkeypatch.setattr(router, "_get_remote_kompress", remote_loader)
+ response = client.post(
+ "/v1/chat/completions",
+ headers={"authorization": "Bearer local-fixture"},
+ json={
+ "model": "kimi-k2.7",
+ "messages": [
+ {"role": "user", "content": "question"},
+ {"role": "assistant", "content": "answer", "reasoning_content": reasoning},
+ {"role": "user", "content": "continue"},
+ ],
+ },
+ )
+ assert response.status_code == 200
+ remote_loader.assert_not_called()
+ assert len(forwarded) == 1
+ assert forwarded[0]["messages"][1]["reasoning_content"] == reasoning