[codex] Improve proxy Anthropic token counting

CONTRIBUTING.md

@@ -59,7 +59,8 @@ python -m venv .venv
 source .venv/bin/activate  # or `.venv\Scripts\activate` on Windows
 
 # Install in development mode with all dependencies
-pip install -e ".[dev,relevance,proxy]"
+# Include langchain so LangChain integration tests import cleanly (full pytest).
+pip install -e ".[dev,relevance,proxy,langchain]"
 
 # Run tests
 pytest

headroom/proxy/server.py

@@ -139,6 +139,26 @@ def _get_image_compressor():
 )
 logger = logging.getLogger("headroom.proxy")
 
+
+def _optional_anthropic_client_for_token_count() -> Any | None:
+    """Anthropic SDK client for accurate pipeline token counts (count_tokens API).
+
+    Returns None when the package is missing or ANTHROPIC_API_KEY is unset (OpenAI-only
+    proxy use). In those cases AnthropicProvider falls back to tiktoken approximation.
+    """
+    if not os.environ.get("ANTHROPIC_API_KEY"):
+        return None
+    try:
+        from anthropic import Anthropic
+
+        return Anthropic()
+    except ImportError:
+        logger.debug(
+            "anthropic package not installed; pipeline uses tiktoken approximation for Claude models"
+        )
+        return None
+
+
 # Always-on file logging to ~/.headroom/logs/ for `headroom perf` analysis
 _HEADROOM_LOG_DIR = Path.home() / ".headroom" / "logs"
 
@@ -1597,8 +1617,10 @@ def __init__(self, config: ProxyConfig):
             gurl = config.gemini_api_url.rstrip("/")
             HeadroomProxy.GEMINI_API_URL = gurl
 
-        # Initialize providers
-        self.anthropic_provider = AnthropicProvider()
+        # Initialize providers (SDK client enables accurate Anthropic token counting)
+        self.anthropic_provider = AnthropicProvider(
+            client=_optional_anthropic_client_for_token_count()
+        )
         self.openai_provider = OpenAIProvider()
 
         # Initialize transforms based on routing mode
@@ -2412,11 +2434,11 @@ async def handle_anthropic_messages(
         # Hook: pre_compress — let hooks modify messages before compression
         if self.config.hooks:
             from headroom.hooks import CompressContext
-            from headroom.transforms.query_echo import extract_user_query
+            from headroom.transforms import query_echo as _query_echo
 
             _hook_ctx = CompressContext(
                 model=model,
-                user_query=extract_user_query(messages),
+                user_query=_query_echo.extract_user_query(messages),
                 provider="anthropic",
             )
             try:

pyproject.toml

@@ -47,7 +47,7 @@ dependencies = [
     # Core: lightweight compression (SmartCrusher, ContentRouter, CCR, TOIN)
     "tiktoken>=0.5.0",            # Tokenizer for all compressors
     "pydantic>=2.0.0",            # Config and data models
-    "litellm==1.82.3",            # Model registry, pricing, and provider support
+    "litellm>=1.82.6,<2",         # Model registry, pricing, and provider support
     "click>=8.1.0",               # CLI framework
     "rich>=13.0.0",               # Rich terminal output
 ]
@@ -58,6 +58,7 @@ proxy = [
     "fastapi>=0.100.0",
     "uvicorn>=0.23.0",
     "httpx[http2]>=0.24.0",
+    "anthropic>=0.18.0",          # Accurate Anthropic token counting in proxy pipeline
     "openai>=2.14.0",             # OpenAI API format support
     "mcp>=1.0.0",                 # MCP server (headroom_compress, retrieve, stats)
     "magika>=0.6.0",              # ML content detection for ContentRouter
@@ -163,7 +164,7 @@ dev = [
     "pre-commit>=3.0.0",
     "openai>=1.0.0",
     "anthropic>=0.18.0",
-    "litellm==1.82.3",
+    "litellm>=1.82.6,<2",
     "fastapi>=0.100.0",
     "uvicorn>=0.23.0",
     "httpx[http2]>=0.24.0",

tests/test_cli/test_mcp.py

@@ -53,6 +53,9 @@ def mock_claude_config_path(temp_claude_dir):
     def which_no_claude(cmd):
         if cmd == "claude":
             return None
+        if cmd == "headroom":
+            # Deterministic: install path uses `headroom` when on PATH; avoid env-specific python -m fallback.
+            return "/opt/headroom/bin/headroom"
         return _real_which(cmd)
 
     with patch("headroom.cli.mcp.MCP_CONFIG_PATH", config_path):