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(ccr): skip compact summaries for proactive expansion (#2242)

Vinay Gupta3f241e472b07
diff --git a/headroom/cache/compression_store.py b/headroom/cache/compression_store.py
index f78fedcbb..b06108a58 100644
--- a/headroom/cache/compression_store.py
+++ b/headroom/cache/compression_store.py
@@ -480,6 +480,7 @@ class CompressionStore:
"compressed_item_count": entry.compressed_item_count,
"query_context": entry.query_context,
"compressed_content": entry.compressed_content,
+ "original_content_preview": entry.original_content[:2000],
"created_at": entry.created_at,
"ttl": entry.ttl,
}
diff --git a/headroom/ccr/context_tracker.py b/headroom/ccr/context_tracker.py
index aa98dcebe..885966407 100644
--- a/headroom/ccr/context_tracker.py
+++ b/headroom/ccr/context_tracker.py
@@ -31,6 +31,38 @@ from ..cache.compression_store import get_compression_store
logger = logging.getLogger(__name__)
+def looks_like_claude_code_compact_summary(*texts: str | None) -> bool:
+ """Return true for Claude Code `/compact` continuation summaries.
+
+ Claude Code can carry a previous session forward by injecting a compact
+ conversation summary into a fresh session. Those summaries are already
+ context; tracking them for CCR proactive expansion makes Headroom re-add
+ stale session state to later turns. Keep this detector deliberately narrow
+ so ordinary tool output that happens to mention "summary" remains eligible.
+ """
+ combined = " ".join(text.strip() for text in texts if text and text.strip())
+ if not combined:
+ return False
+
+ normalized = " ".join(combined.lower().split())
+ has_summary = "summary" in normalized or "summarized" in normalized
+
+ if "this session is being continued from a previous conversation" in normalized and has_summary:
+ return True
+
+ if "conversation is summarized below" in normalized and (
+ "ran out of context" in normalized or "previous conversation" in normalized
+ ):
+ return True
+
+ return (
+ "/compact" in normalized
+ and "claude" in normalized
+ and has_summary
+ and ("conversation" in normalized or "session" in normalized)
+ )
+
+
@dataclass
class CompressedContext:
"""Represents a piece of compressed context from the conversation.
@@ -157,6 +189,13 @@ class ContextTracker:
if not self.config.enabled:
return
+ if looks_like_claude_code_compact_summary(query_context, sample_content):
+ logger.debug(
+ "CCR Tracker: skipped Claude Code compact summary %s for proactive expansion",
+ hash_key,
+ )
+ return
+
context = CompressedContext(
hash_key=hash_key,
turn_number=turn_number,
diff --git a/headroom/proxy/handlers/anthropic.py b/headroom/proxy/handlers/anthropic.py
index 427ee6413..bb9be767b 100644
--- a/headroom/proxy/handlers/anthropic.py
+++ b/headroom/proxy/handlers/anthropic.py
@@ -24,6 +24,7 @@ if TYPE_CHECKING:
import httpx
from headroom.agent_savings import proxy_pipeline_kwargs
+from headroom.ccr.context_tracker import looks_like_claude_code_compact_summary
from headroom.copilot_auth import build_copilot_upstream_url
from headroom.pipeline import PipelineStage, summarize_routing_markers
from headroom.proxy.auth_mode import classify_auth_mode, classify_client
@@ -1934,6 +1935,16 @@ class AnthropicHandlerMixin:
store = get_compression_store()
entry = store.get_metadata(hash_key)
if entry:
+ if looks_like_claude_code_compact_summary(
+ entry.get("query_context"),
+ entry.get("compressed_content"),
+ entry.get("original_content_preview"),
+ ):
+ logger.info(
+ f"[{request_id}] CCR: skipping proactive "
+ f"tracking for Claude Code compact summary {hash_key}"
+ )
+ continue
self.ccr_context_tracker.track_compression(
hash_key=hash_key,
turn_number=self._turn_counter,
diff --git a/tests/test_ccr_context_tracker.py b/tests/test_ccr_context_tracker.py
index ac8427cbc..fe45b3d69 100644
--- a/tests/test_ccr_context_tracker.py
+++ b/tests/test_ccr_context_tracker.py
@@ -23,6 +23,7 @@ from headroom.ccr.context_tracker import (
ContextTrackerConfig,
ExpansionRecommendation,
get_context_tracker,
+ looks_like_claude_code_compact_summary,
reset_context_tracker,
)
@@ -58,6 +59,37 @@ class TestContextTrackerBasics:
stats = tracker.get_stats()
assert stats["tracked_contexts"] == 1
+ def test_claude_code_compact_summary_not_tracked(self):
+ """Claude Code `/compact` summaries are not proactively re-expanded."""
+ tracker = ContextTracker()
+
+ tracker.track_compression(
+ hash_key="compact_summary",
+ turn_number=1,
+ tool_name="Claude Code",
+ original_count=1,
+ compressed_count=1,
+ query_context="Claude Code /compact summary",
+ sample_content=(
+ "This session is being continued from a previous conversation "
+ "that ran out of context. The conversation is summarized below: "
+ "We changed the auth middleware and still need to run tests."
+ ),
+ workspace_key="ws-test",
+ )
+
+ assert tracker.get_tracked_hashes() == []
+
+ def test_claude_code_compact_summary_detector_is_narrow(self):
+ """The compact-summary detector does not reject ordinary summaries."""
+ assert looks_like_claude_code_compact_summary(
+ "This session is being continued from a previous conversation. "
+ "The conversation is summarized below."
+ )
+ assert not looks_like_claude_code_compact_summary(
+ "search results summary for auth middleware files"
+ )
+
def test_track_multiple_compressions(self):
"""Track multiple compression events."""
tracker = ContextTracker()