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(integrations): pin MCP server + LangChain evals to lossy+CCR path

chopratejas168800329b7c
diff --git a/headroom/integrations/mcp/server.py b/headroom/integrations/mcp/server.py
index 9fe0b2626..e002a0001 100644
--- a/headroom/integrations/mcp/server.py
+++ b/headroom/integrations/mcp/server.py
@@ -245,13 +245,23 @@ class HeadroomMCPCompressor:
items_after = 0
errors_preserved = 0
- # Create SmartCrusher with profile settings
+ # Create SmartCrusher with profile settings.
+ #
+ # The MCP integration emits JSON-shaped output that downstream
+ # tool consumers parse and iterate. The PR4 lossless path
+ # substitutes a CSV+schema STRING in place of arrays — great
+ # for LLM prompts but wire-format-incompatible with consumers
+ # that expect JSON arrays. So we keep the runtime MCP wrapper
+ # on the lossy + CCR-Dropped path: the LLM sees row-level
+ # subsets inline, with the full payload retrievable via CCR
+ # cache. (Same retention semantics as Python's pre-PR4
+ # SmartCrusher behavior.)
smart_config = SmartCrusherConfig(
enabled=True,
min_tokens_to_crush=profile.min_tokens_to_compress,
max_items_after_crush=profile.max_items,
)
- crusher = SmartCrusher(config=smart_config) # type: ignore[arg-type]
+ crusher = SmartCrusher(config=smart_config, with_compaction=False) # type: ignore[arg-type]
# Build messages for SmartCrusher (it expects conversation format)
messages = [
diff --git a/tests/test_integrations/langchain/test_evals.py b/tests/test_integrations/langchain/test_evals.py
index 8da5862bf..c18b8b9f3 100644
--- a/tests/test_integrations/langchain/test_evals.py
+++ b/tests/test_integrations/langchain/test_evals.py
@@ -30,13 +30,21 @@ def tokenizer():
@pytest.fixture
def smart_crusher():
- """Create SmartCrusher with default config."""
+ """Create SmartCrusher with default config.
+
+ These eval tests assert row-level retention semantics (errors
+ preserved, anomalies preserved, schema unchanged in JSON shape).
+ Those properties belong to the lossy + CCR-Dropped path, not
+ the lossless path which substitutes a CSV+schema string.
+ `with_compaction=False` keeps these tests on the legacy lossy
+ path — same as the retention tests in `test_quality_retention.py`.
+ """
config = SmartCrusherConfig(
enabled=True,
min_tokens_to_crush=200,
max_items_after_crush=20,
)
- return SmartCrusher(config=config)
+ return SmartCrusher(config=config, with_compaction=False)
def generate_log_entries(count: int, error_rate: float = 0.15) -> list[dict]: