The agent that grows with you
Python aiai-agentai-agentsanthropicchatgptclaudeclaude-codecodexhermeshermes-agentllmnous-researchopenai
"""Regression tests for #89886: part-level cache_control on role:tool messages
must be suppressed on LiteLLM-style envelope routes.
LiteLLM's OpenAI→Anthropic translation copies tool-message content parts
verbatim, so a part-level ``cache_control`` lands at ``tool_result.content[0]``
— a placement the Anthropic Messages schema rejects with a non-retryable
HTTP 400 that kills the whole turn. OpenRouter relocates the marker correctly,
so the part-level form stays on for it.
"""
import copy
from agent.prompt_caching import (
apply_anthropic_cache_control,
build_prompt_cache_plan,
envelope_tool_part_cache_markers_supported,
)
def _tool_history():
return [
{"role": "system", "content": "sys"},
{"role": "user", "content": "run the tool"},
{
"role": "assistant",
"content": "",
"tool_calls": [
{
"id": "call_1",
"type": "function",
"function": {"name": "t", "arguments": "{}"},
}
],
},
{"role": "tool", "tool_call_id": "call_1", "content": "tool output"},
]
def _tool_messages(msgs):
return [m for m in msgs if m.get("role") == "tool"]
def _has_part_marker(msg):
content = msg.get("content")
if isinstance(content, list):
return any(
isinstance(p, dict) and "cache_control" in p for p in content
)
return False
class TestEnvelopeToolPartMarkerSuppression:
def test_litellm_route_detected_as_unsupported(self):
assert not envelope_tool_part_cache_markers_supported(
"litellm", "https://litellm.internal.example/v1"
)
assert not envelope_tool_part_cache_markers_supported(
"custom:litellm", "https://gw.example.com/v1"
)
assert not envelope_tool_part_cache_markers_supported(
"custom", "https://litellm-proxy.example.com/v1"
)
def test_openrouter_route_stays_supported(self):
assert envelope_tool_part_cache_markers_supported(
"openrouter", "https://openrouter.ai/api/v1"
)
# Lookalike names must not be swept in.
assert envelope_tool_part_cache_markers_supported(
"custom:notlitellm", "https://notlitellm.example.com/v1"
)
def test_tool_message_not_marked_when_part_markers_off(self):
decorated = apply_anthropic_cache_control(
copy.deepcopy(_tool_history()),
native_anthropic=False,
tool_part_markers=False,
)
for tool_msg in _tool_messages(decorated):
assert "cache_control" not in tool_msg
assert not _has_part_marker(tool_msg)
# Content must stay a plain string — no decoration-produced
# part list that LiteLLM would forward into tool_result.content.
assert isinstance(tool_msg.get("content"), str)
def test_breakpoint_reallocates_to_non_tool_message(self):
decorated = apply_anthropic_cache_control(
copy.deepcopy(_tool_history()),
native_anthropic=False,
tool_part_markers=False,
)
non_tool_marked = [
m
for m in decorated
if m.get("role") != "tool"
and ("cache_control" in m or _has_part_marker(m))
]
assert non_tool_marked, "budget must reallocate, not vanish"
def test_openrouter_default_keeps_tool_part_marker(self):
decorated = apply_anthropic_cache_control(
copy.deepcopy(_tool_history()), native_anthropic=False
)
assert any(_has_part_marker(m) for m in _tool_messages(decorated))
def test_build_prompt_cache_plan_threads_flag(self):
plan = build_prompt_cache_plan(
_tool_history(),
[],
native_anthropic=False,
tool_part_markers=False,
)
for tool_msg in _tool_messages(plan.messages):
assert "cache_control" not in tool_msg
assert not _has_part_marker(tool_msg)
def test_native_anthropic_layout_unaffected(self):
# Native layout relies on the adapter to relocate a top-level marker
# onto the tool_result block; the flag must not interfere.
decorated = apply_anthropic_cache_control(
copy.deepcopy(_tool_history()),
native_anthropic=True,
tool_part_markers=False,
)
assert any(
"cache_control" in m for m in _tool_messages(decorated)
)