fix(anthropic): canonicalize system prompts with the shared stripper

.github/workflows/ci.yml

@@ -82,6 +82,7 @@ jobs:
             tests/test_request.py \
             tests/test_anthropic_models.py \
             tests/test_anthropic_adapter.py \
+            tests/test_anthropic_adapter_canonicalization.py \
             tests/test_dependency_floors.py \
             tests/test_ssd_cache_shutdown.py \
             tests/test_chat_template_safety.py \

tests/test_anthropic_adapter_canonicalization.py

@@ -0,0 +1,112 @@
+# SPDX-License-Identifier: Apache-2.0
+"""The Anthropic adapter must canonicalize system prompts the shared way.
+
+`anthropic_to_openai()` used to carry its own unanchored `re.sub` for the
+billing header. It runs *before* `canonicalize_system_messages()`, so anything
+it removed was gone by the time the shared, line-anchored stripper saw the
+messages — including the middle of a user's own sentence. These tests pin both
+stages: what the adapter alone does, and what the prepared-message path
+produces end to end.
+"""
+
+import pytest
+
+from vllm_mlx.api.anthropic_adapter import anthropic_to_openai
+from vllm_mlx.api.anthropic_models import AnthropicRequest
+from vllm_mlx.api.prompt_canonicalize import canonicalize_system_messages
+
+HEADER = "x-anthropic-billing-header: cch=9f2a1b;"
+BODY = "You are helpful.\nBe concise."
+
+
+def _system_text(system) -> str:
+    request = AnthropicRequest(
+        model="test-model",
+        max_tokens=16,
+        system=system,
+        messages=[{"role": "user", "content": "hi"}],
+    )
+    converted = anthropic_to_openai(request)
+    systems = [m for m in converted.messages if m.role == "system"]
+    assert len(systems) == 1, f"expected one system message, got {len(systems)}"
+    return systems[0].content
+
+
+def _two_stage(system) -> str:
+    """Adapter output after the shared pass the server applies next."""
+    text = _system_text(system)
+    canonicalized = canonicalize_system_messages([{"role": "system", "content": text}])
+    return canonicalized[0]["content"]
+
+
+class TestStandaloneHeaderIsRemoved:
+    @pytest.mark.parametrize(
+        "header",
+        [
+            "x-anthropic-billing-header: cch=9f2a1b;",
+            "X-Anthropic-Billing-Header: cch=9f2a1b;",
+            "X-ANTHROPIC-BILLING-HEADER: cch=9f2a1b;",
+        ],
+        ids=["lower", "mixed", "upper"],
+    )
+    def test_case_insensitive_on_its_own_line(self, header):
+        text = _system_text(f"You are helpful.\n{header}\nBe concise.")
+        assert "cch=" not in text, f"volatile hash survived: {text!r}"
+        assert "You are helpful." in text
+        assert "Be concise." in text
+
+    def test_header_as_the_entire_prompt(self):
+        assert _system_text(HEADER).strip() == ""
+
+    def test_trailing_header_without_newline(self):
+        text = _system_text(f"{BODY}\n{HEADER}")
+        assert "cch=" not in text
+        assert "Be concise." in text
+
+
+class TestUserTextIsPreserved:
+    """The regression this replaces: an unanchored strip ate real prompt text."""
+
+    def test_mid_line_mention_survives(self):
+        prompt = "Explain what x-anthropic-billing-header: means in HTTP terms."
+        assert _system_text(prompt) == prompt
+
+    def test_mid_line_mention_survives_the_second_pass_too(self):
+        prompt = "Explain what x-anthropic-billing-header: means in HTTP terms."
+        assert _two_stage(prompt) == prompt
+
+    @pytest.mark.parametrize(
+        "prompt",
+        [
+            BODY,
+            "Timestamps look like 2026-05-10T13:42:18.123Z in our logs.",
+            "Use the header: Authorization: Bearer <token> when calling the API.",
+            "Discuss anthropic billing headers in general.",
+        ],
+        ids=["plain", "timestamp", "other-header", "prose"],
+    )
+    def test_unrelated_content_is_untouched(self, prompt):
+        assert _system_text(prompt) == prompt
+        assert _two_stage(prompt) == prompt
+
+
+class TestBothStagesAgree:
+    """Adapter and shared pass must not disagree about what to remove."""
+
+    @pytest.mark.parametrize(
+        "prompt",
+        [
+            f"You are helpful.\n{HEADER}\nBe concise.",
+            "You are helpful.\nX-Anthropic-Billing-Header: cch=abc;\nBe concise.",
+            "Explain what x-anthropic-billing-header: means in HTTP terms.",
+            BODY,
+        ],
+        ids=["lower", "mixed", "mid-line", "clean"],
+    )
+    def test_second_pass_is_a_no_op(self, prompt):
+        """Whatever the adapter returns is already canonical."""
+        once = _system_text(prompt)
+        twice = canonicalize_system_messages([{"role": "system", "content": once}])[0][
+            "content"
+        ]
+        assert twice == once, f"shared pass still changed {once!r} -> {twice!r}"

vllm_mlx/api/anthropic_adapter.py

@@ -9,7 +9,6 @@
 """
 
 import json
-import re
 import uuid
 
 from .anthropic_models import (
@@ -26,6 +25,7 @@
     Message,
     ToolDefinition,
 )
+from .prompt_canonicalize import canonicalize_system_prompt
 
 
 def anthropic_to_openai(request: AnthropicRequest) -> ChatCompletionRequest:
@@ -64,7 +64,13 @@ def anthropic_to_openai(request: AnthropicRequest) -> ChatCompletionRequest:
         # Strip per-request billing/tracking headers injected by some
         # clients (e.g. Claude Code).  These contain a per-request hash
         # that prevents prefix-cache reuse across turn boundaries.
-        system_text = re.sub(r"x-anthropic-billing-header:[^\n]*\n?", "", system_text)
+        #
+        # Use the shared canonicalizer rather than an inline pattern. The
+        # local one was unanchored, so it deleted the header text wherever
+        # it appeared -- including mid-sentence in a user's own prompt --
+        # and this pass runs before canonicalize_system_messages(), which
+        # cannot restore what was already removed.
+        system_text = canonicalize_system_prompt(system_text) or ""
         messages.append(Message(role="system", content=system_text))
 
     # Convert each message