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: use all user text blocks for relevance scoring (#4628)

cuteNBD12dca26aa11d
diff --git a/headroom/utils.py b/headroom/utils.py
index cbd085895..e0ba74e5d 100644
--- a/headroom/utils.py
+++ b/headroom/utils.py
@@ -15,6 +15,9 @@ from typing import Any
MARKER_PREFIX = "<headroom:"
MARKER_SUFFIX = ">"
+_REMINDER_OPEN = re.compile(r"<system-reminder(?=[\s>])[^<>]*>", re.IGNORECASE)
+_REMINDER_CLOSE = re.compile(r"</system-reminder\s*>", re.IGNORECASE)
+
def generate_request_id() -> str:
"""Generate a unique request ID."""
@@ -61,6 +64,24 @@ def fast_hash(data: str | bytes, length: int = 16) -> str:
return hashlib.md5(data).hexdigest()[:length] # nosec B324
+def _strip_system_reminders(text: str) -> str:
+ """Remove complete harness reminders while keeping surrounding user text."""
+ parts: list[str] = []
+ cursor = 0
+ while opening := _REMINDER_OPEN.search(text, cursor):
+ closing = _REMINDER_CLOSE.search(text, opening.end())
+ if closing is None:
+ break
+ parts.append(text[cursor : opening.start()])
+ end = closing.end()
+ if opening.start() and end < len(text):
+ if not text[opening.start() - 1].isspace() and not text[end].isspace():
+ parts.append(" ")
+ cursor = end
+ parts.append(text[cursor:])
+ return "".join(parts).strip()
+
+
def extract_user_query(
messages: list[dict[str, Any]], *, latest_user_turn_only: bool = False
) -> str:
@@ -82,14 +103,20 @@ def extract_user_query(
for msg in reversed(messages):
if msg.get("role") == "user":
content = msg.get("content", "")
- if isinstance(content, str) and content.strip():
- return content.strip()
+ if isinstance(content, str):
+ user_text = _strip_system_reminders(content)
+ if user_text:
+ return user_text
if isinstance(content, list):
- for block in content:
- if isinstance(block, dict) and block.get("type") == "text":
- text = str(block.get("text", "")).strip()
- if text:
- return text
+ user_text = "\n".join(
+ text
+ for block in content
+ if isinstance(block, dict) and block.get("type") == "text"
+ for text in (_strip_system_reminders(str(block.get("text", ""))),)
+ if text
+ )
+ if user_text:
+ return user_text
if latest_user_turn_only:
return ""
return ""
diff --git a/tests/test_relevance_split.py b/tests/test_relevance_split.py
index 3e2e57132..6f6e610f3 100644
--- a/tests/test_relevance_split.py
+++ b/tests/test_relevance_split.py
@@ -125,6 +125,31 @@ def test_split_keeps_relevant_drops_irrelevant():
assert "".join(t for _, t in runs) == content
+def test_split_uses_all_user_text_blocks():
+ from headroom.relevance.bm25 import BM25Scorer
+ from headroom.utils import extract_user_query
+
+ messages = [
+ {
+ "role": "user",
+ "content": [
+ {"type": "text", "text": "Review production incidents"},
+ {"type": "text", "text": "Focus on Kafka replication"},
+ ],
+ }
+ ]
+ content = (
+ "Production incident: payment gateway high latency\n\n"
+ "Kafka replication lag on broker 17 exceeded 30 minutes\n\n"
+ "Disk full on backup server\n\n"
+ "Production incident: expired TLS certificate\n"
+ )
+ runs = plan_relevance_split(content, extract_user_query(messages), BM25Scorer(), threshold=0.25)
+ kept = "".join(text for keep, text in runs if keep)
+ assert "Kafka replication lag" in kept
+ assert "Disk full on backup server" not in kept
+
+
def test_empty_query_yields_no_split():
assert plan_relevance_split("x\ny\n", "", KeywordScorer(), threshold=0.5) == [(True, "x\ny\n")]
diff --git a/tests/test_utils.py b/tests/test_utils.py
index fafae8ead..3b749b9d7 100644
--- a/tests/test_utils.py
+++ b/tests/test_utils.py
@@ -48,6 +48,85 @@ def test_extract_user_query_and_message_hashes() -> None:
assert utils.extract_user_query([{"role": "assistant", "content": "skip"}]) == ""
[email protected]("latest_only", [False, True])
+def test_extract_user_query_combines_text_blocks(latest_only: bool) -> None:
+ messages = [
+ {
+ "role": "user",
+ "content": [
+ {"type": "text", "text": " Review incidents "},
+ {"type": "image", "source": "example"},
+ {"type": "text", "text": " "},
+ {"type": "text", "text": " Focus on Kafka replication "},
+ ],
+ }
+ ]
+ assert utils.extract_user_query(messages, latest_user_turn_only=latest_only) == (
+ "Review incidents\nFocus on Kafka replication"
+ )
+
+
+def test_extract_user_query_skips_system_reminder_blocks() -> None:
+ messages = [
+ {"role": "user", "content": "older question"},
+ {
+ "role": "user",
+ "content": [
+ {
+ "type": "text",
+ "text": "<system-reminder>injected instructions</system-reminder>",
+ },
+ {"type": "text", "text": "real question"},
+ {"type": "text", "text": "<system-reminder>more instructions</system-reminder>"},
+ ],
+ },
+ ]
+ assert utils.extract_user_query(messages) == "real question"
+ assert utils.extract_user_query(messages, latest_user_turn_only=True) == "real question"
+
+ messages[-1]["content"] = [
+ {"type": "text", "text": "<system-reminder>injected</system-reminder>"}
+ ]
+ assert utils.extract_user_query(messages) == "older question"
+ assert utils.extract_user_query(messages, latest_user_turn_only=True) == ""
+
+
[email protected]("as_blocks", [False, True])
+ ("content", "expected"),
+ [
+ ("<system-reminder>context</system-reminder>\nFind Kafka lag", "Find Kafka lag"),
+ ("Find <system-reminder>context</system-reminder> Kafka lag", "Find Kafka lag"),
+ (
+ "<system-reminder id='1'>a</system-reminder>Find <system-reminder>b</system-reminder>Kafka",
+ "Find Kafka",
+ ),
+ ("<System-Reminder>context</System-Reminder>Find Kafka", "Find Kafka"),
+ ("İ<system-reminder>x</system-reminder>Kafka", "İ Kafka"),
+ ("Find <system-reminder>unfinished", "Find <system-reminder>unfinished"),
+ (
+ "Find <system-reminder-extra>example</system-reminder> Kafka",
+ "Find <system-reminder-extra>example</system-reminder> Kafka",
+ ),
+ ],
+)
+def test_extract_user_query_keeps_text_around_reminders(
+ content: str, expected: str, as_blocks: bool
+) -> None:
+ body = [{"type": "text", "text": content}] if as_blocks else content
+ messages = [{"role": "user", "content": body}]
+ assert utils.extract_user_query(messages, latest_user_turn_only=True) == expected
+
+
[email protected]("as_blocks", [False, True])
+def test_reminder_only_turn_keeps_user_query_fallback(as_blocks: bool) -> None:
+ reminder = "<system-reminder>context</system-reminder>"
+ content = [{"type": "text", "text": reminder}] if as_blocks else reminder
+ messages = [{"role": "user", "content": "older question"}, {"role": "user", "content": content}]
+ assert utils.extract_user_query(messages) == "older question"
+ assert utils.extract_user_query(messages, latest_user_turn_only=True) == ""
+
+
def test_extract_user_query_latest_user_turn_only() -> None:
"""The flag stops at the newest user turn instead of resurrecting an older one."""
# A tool_result continuation: the newest user turn carries no text at all.