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.
Python agentaianthropicclaude-codecompressioncontext-engineeringcontext-windowcursorfastapilangchainllmmcpopenaiprompt-engineeringproxypythonragtoken-optimizationtokenstypescript
fix(proxy/anthropic): None-guard usage token counts on the direct buffered path (#2434)
Abhay Singh2b5ee7cde809
diff --git a/headroom/proxy/handlers/anthropic.py b/headroom/proxy/handlers/anthropic.pyindex 8a17311a5..d19413ac0 100644--- a/headroom/proxy/handlers/anthropic.py+++ b/headroom/proxy/handlers/anthropic.py@@ -3576,13 +3576,13 @@ class AnthropicHandlerMixin:uncached_input_tokens = 0if resp_json:usage = resp_json.get("usage", {})- output_tokens = usage.get("output_tokens", 0)- cr_tokens = usage.get("cache_read_input_tokens", 0)- cw_tokens = usage.get("cache_creation_input_tokens", 0)+ output_tokens = int(usage.get("output_tokens", 0) or 0)+ cr_tokens = int(usage.get("cache_read_input_tokens", 0) or 0)+ cw_tokens = int(usage.get("cache_creation_input_tokens", 0) or 0)cw_5m_tokens, cw_1h_tokens = self._extract_anthropic_cache_ttl_metrics(usage)- uncached_input_tokens = usage.get("input_tokens", 0)+ uncached_input_tokens = int(usage.get("input_tokens", 0) or 0)# Track cache bust: tokens that lost their cache discount due to compression.# If we had X tokens cached last turn and only Y hit cache this turn,diff --git a/tests/test_proxy/test_anthropic_buffered_timeout.py b/tests/test_proxy/test_anthropic_buffered_timeout.pyindex 69e44e7f5..6e0c753bb 100644--- a/tests/test_proxy/test_anthropic_buffered_timeout.py+++ b/tests/test_proxy/test_anthropic_buffered_timeout.py@@ -110,6 +110,59 @@ def _anthropic_list_response() -> httpx.Response:)+def _anthropic_null_usage_response() -> dict[str, object]:+ """An Anthropic Messages response whose usage counts are present but null.++ An Anthropic-compatible gateway (custom ANTHROPIC_TARGET_API_URL) can emit+ this shape on a stopped or empty turn — the same class that crashed the+ Gemini path in #2347.+ """+ return {+ "id": "msg_test_null",+ "type": "message",+ "role": "assistant",+ "content": [{"type": "text", "text": "ok"}],+ "usage": {+ "input_tokens": None,+ "output_tokens": None,+ "cache_read_input_tokens": None,+ "cache_creation_input_tokens": None,+ },+ }+++def test_anthropic_messages_buffered_survives_null_usage_counts():+ config = _make_config()+ app = create_app(config)+ with TestClient(app) as client:+ proxy = client.app.state.proxy+ # Use the real prefix tracker (its `_cached_token_count` is read on this+ # path); only pin the session id so the tracker resolves deterministically.+ proxy.session_tracker_store.compute_session_id = lambda request, model, messages: "s1"++ async def _fake_retry(method, url, headers, body, stream=False, **kwargs): # noqa: ANN001+ return httpx.Response(200, json=_anthropic_null_usage_response())++ proxy._retry_request = _fake_retry # type: ignore[assignment]++ response = client.post(+ "/v1/messages",+ headers={+ "x-api-key": "test-key",+ "anthropic-version": "2023-06-01",+ "content-type": "application/json",+ },+ json={+ "model": "claude-sonnet-4-6",+ "max_tokens": 64,+ "messages": [{"role": "user", "content": "hello"}],+ },+ )++ # Null usage counts must not crash outcome recording / cache-bust math.+ assert response.status_code == 200, response.text++def test_anthropic_messages_buffered_timeout_override_reaches_retry_request():config = _make_config()app = create_app(config)