mirror of
https://github.com/headroomlabs-ai/headroom.git
synced 2026-08-27 14:17:10 -04:00
fix(proxy): compress Anthropic user text blocks when enabled (#1875)
## Summary - make Anthropic content-block user text honor compress_user_messages - keep cache_control text blocks protected even when user-message compression is enabled - add regression coverage for user text blocks, default protection, and cache_control protection ## Why The string-message path already compresses role=user content when compress_user_messages=True, but the Anthropic content-block path always fell through to the unknown-role protection branch for role=user. In proxy mode this produced router:noop for large user text blocks even with HEADROOM_COMPRESS_USER_MESSAGES=1 and HEADROOM_FORCE_KOMPRESS_ALL=1. ## Tests - python -m pytest tests/test_content_router_user_blocks.py tests/test_compression_safety_rails.py tests/test_force_kompress_all.py -q Note: running tests/test_agent_savings.py locally still hits an unrelated Rust binding mismatch: SmartCrusherConfig.__new__() got an unexpected keyword argument 'lossless_only'. Co-authored-by: Julien Guarino <julien.guarino@fashiondata.io>
This commit is contained in:
parent
739f654bbd
commit
e36439a941
2 changed files with 117 additions and 7 deletions
|
|
@ -4360,14 +4360,16 @@ class ContentRouter(Transform):
|
|||
# Role-based gate for `text` blocks. Tool/function roles are tool
|
||||
# outputs and compress freely; assistant defaults to skip (cache
|
||||
# safety) with explicit opt-in; unknown roles default to skip.
|
||||
if (skip_user and role == "user") or (skip_system and role in {"system", "developer"}):
|
||||
protect_text_blocks = True
|
||||
elif role == "assistant" and not compress_assistant_text_blocks:
|
||||
protect_text_blocks = True
|
||||
elif role not in ("assistant", "tool", "function"):
|
||||
protect_text_blocks = True
|
||||
else:
|
||||
if role == "user":
|
||||
protect_text_blocks = skip_user
|
||||
elif role in {"system", "developer"}:
|
||||
protect_text_blocks = skip_system
|
||||
elif role == "assistant":
|
||||
protect_text_blocks = not compress_assistant_text_blocks
|
||||
elif role in ("tool", "function"):
|
||||
protect_text_blocks = False
|
||||
else:
|
||||
protect_text_blocks = True
|
||||
|
||||
for block in content_blocks:
|
||||
if not isinstance(block, dict):
|
||||
|
|
|
|||
108
tests/test_content_router_user_blocks.py
Normal file
108
tests/test_content_router_user_blocks.py
Normal file
|
|
@ -0,0 +1,108 @@
|
|||
"""Content-router coverage for Anthropic user text blocks."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any
|
||||
|
||||
from headroom.transforms.content_router import ContentRouter, ContentRouterConfig
|
||||
|
||||
|
||||
class _Tokenizer:
|
||||
def count_text(self, text: str) -> int:
|
||||
return max(1, len(str(text)) // 4)
|
||||
|
||||
|
||||
def _long_text() -> str:
|
||||
return "This user message should be compressible when opted in. " * 80
|
||||
|
||||
|
||||
def test_user_text_block_compresses_when_user_messages_are_enabled() -> None:
|
||||
router = ContentRouter(ContentRouterConfig(force_kompress_all=True))
|
||||
calls: list[dict[str, Any]] = []
|
||||
|
||||
def fake_compress_block_content(**kwargs: Any) -> tuple[str | None, bool]:
|
||||
calls.append(kwargs)
|
||||
kwargs["transforms_applied"].append("router:text_block:fake")
|
||||
return "COMPRESSED", True
|
||||
|
||||
router._compress_block_content = fake_compress_block_content # type: ignore[method-assign]
|
||||
|
||||
result = router.apply(
|
||||
[
|
||||
{
|
||||
"role": "user",
|
||||
"content": [{"type": "text", "text": _long_text()}],
|
||||
}
|
||||
],
|
||||
_Tokenizer(),
|
||||
compress_user_messages=True,
|
||||
force_kompress=True,
|
||||
target_ratio=0.10,
|
||||
)
|
||||
|
||||
assert len(calls) == 1
|
||||
assert result.messages[0]["content"][0]["text"] == "COMPRESSED"
|
||||
assert result.transforms_applied == ["router:text_block:fake"]
|
||||
|
||||
|
||||
def test_user_text_block_stays_protected_by_default() -> None:
|
||||
router = ContentRouter(ContentRouterConfig(force_kompress_all=True))
|
||||
calls: list[dict[str, Any]] = []
|
||||
text = _long_text()
|
||||
|
||||
def fake_compress_block_content(**kwargs: Any) -> tuple[str | None, bool]:
|
||||
calls.append(kwargs)
|
||||
return "COMPRESSED", True
|
||||
|
||||
router._compress_block_content = fake_compress_block_content # type: ignore[method-assign]
|
||||
|
||||
result = router.apply(
|
||||
[
|
||||
{
|
||||
"role": "user",
|
||||
"content": [{"type": "text", "text": text}],
|
||||
}
|
||||
],
|
||||
_Tokenizer(),
|
||||
force_kompress=True,
|
||||
target_ratio=0.10,
|
||||
)
|
||||
|
||||
assert calls == []
|
||||
assert result.messages[0]["content"][0]["text"] == text
|
||||
assert result.transforms_applied == ["router:noop"]
|
||||
|
||||
|
||||
def test_user_cache_control_text_block_stays_protected_when_enabled() -> None:
|
||||
router = ContentRouter(ContentRouterConfig(force_kompress_all=True))
|
||||
calls: list[dict[str, Any]] = []
|
||||
text = _long_text()
|
||||
|
||||
def fake_compress_block_content(**kwargs: Any) -> tuple[str | None, bool]:
|
||||
calls.append(kwargs)
|
||||
return "COMPRESSED", True
|
||||
|
||||
router._compress_block_content = fake_compress_block_content # type: ignore[method-assign]
|
||||
|
||||
result = router.apply(
|
||||
[
|
||||
{
|
||||
"role": "user",
|
||||
"content": [
|
||||
{
|
||||
"type": "text",
|
||||
"text": text,
|
||||
"cache_control": {"type": "ephemeral"},
|
||||
}
|
||||
],
|
||||
}
|
||||
],
|
||||
_Tokenizer(),
|
||||
compress_user_messages=True,
|
||||
force_kompress=True,
|
||||
target_ratio=0.10,
|
||||
)
|
||||
|
||||
assert calls == []
|
||||
assert result.messages[0]["content"][0]["text"] == text
|
||||
assert result.transforms_applied == ["router:noop"]
|
||||
Loading…
Add table
Add a link
Reference in a new issue