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): keep health probes and the Codex attestation out of the proxy log (#4611)
gglucass78736a16bc46
diff --git a/headroom/proxy/server.py b/headroom/proxy/server.pyindex 2e97c8404..9346e2333 100644--- a/headroom/proxy/server.py+++ b/headroom/proxy/server.py@@ -487,6 +487,9 @@ logging.basicConfig()logger = logging.getLogger("headroom.proxy")+# Liveness/readiness probe routes: auth-exempt, and logged at DEBUG.+_HEALTH_PROBE_PATHS = frozenset({"/health", "/livez", "/readyz"})+def _output_reduction_payload(shaper_active: bool, est: Any = None) -> dict[str, Any]:"""/stats payload for the output-shaping layer.@@ -4042,7 +4045,13 @@ def create_app(config: ProxyConfig | None = None) -> FastAPI:safe_headers = redact_for_wire_debug(headers)except Exception:safe_headers = {"redaction_error": True}- logger.info(+ # A desktop or orchestrator polls these every few seconds; at INFO they+ # filled half of an idle proxy log and cut a busy one's history to hours.+ line_level = (+ logging.DEBUG if method == "GET" and path in _HEALTH_PROBE_PATHS else logging.INFO+ )+ logger.log(+ line_level,"event=proxy_inbound_request id=%s method=%s path=%s query=%s client=%s ""content_length=%s headers=%s",inbound_id,@@ -4096,7 +4105,8 @@ def create_app(config: ProxyConfig | None = None) -> FastAPI:proxy.metrics.record_inbound_response(status_code=response.status_code)except Exception:logger.debug("record_inbound_response failed", exc_info=True)- logger.info(+ logger.log(+ line_level,"event=proxy_inbound_response id=%s method=%s path=%s status=%s ""duration_ms=%.2f response_content_length=%s",inbound_id,@@ -4140,7 +4150,7 @@ def create_app(config: ProxyConfig | None = None) -> FastAPI:# Anything else on these paths falls through to the catch-all passthrough# relay (FastAPI's @app.get does not register HEAD), so exempting it would# let an unauthenticated caller relay arbitrary requests upstream.- _AUTH_EXEMPT_PATHS = frozenset({"/health", "/livez", "/readyz"})+ _AUTH_EXEMPT_PATHS = _HEALTH_PROBE_PATHS# A non-loopback bind with no token is the exact shape (``--host 0.0.0.0``# from any launcher) that exposes the unauthenticated /v1/* relay to thediff --git a/headroom/proxy/wire_debug_redaction_policy.py b/headroom/proxy/wire_debug_redaction_policy.pyindex 28fe89533..8387e0281 100644--- a/headroom/proxy/wire_debug_redaction_policy.py+++ b/headroom/proxy/wire_debug_redaction_policy.py@@ -16,6 +16,8 @@ WIRE_DEBUG_SECRET_KEYS = ("x-api-key","openai-api-key","anthropic-api-key",+ # Codex's signed device attestation, sent with every request.+ "x-oai-attestation","access_token","refresh_token","id_token",diff --git a/tests/test_proxy/test_inbound_response_size.py b/tests/test_proxy/test_inbound_response_size.pyindex 9e58e9edd..44c3b2f44 100644--- a/tests/test_proxy/test_inbound_response_size.py+++ b/tests/test_proxy/test_inbound_response_size.py@@ -114,3 +114,22 @@ def test_streaming_response_no_content_length(capture_logs):m = re.search(r"response_content_length=(\S*)", line)assert m, f"expected response_content_length field in: {line}"assert m.group(1) in ("", '""'), f"expected empty for streaming, got: {m.group(1)}"+++def test_health_probes_log_at_debug_and_traffic_at_info(capture_logs):+ """A desktop polls /readyz every few seconds: at INFO it filled half the log."""+ app = create_app(_make_config())+ with TestClient(app) as client:+ client.app.state.proxy.handle_anthropic_messages = _json_handler+ client.get("/readyz")+ client.get("/livez")+ client.post("/v1/messages", json={"model": "glm-5.2", "messages": []})++ levels = {}+ for r in capture_logs:+ m = re.search(r"event=proxy_inbound_(?:request|response) .*?path=(\S+)", r.getMessage())+ if m:+ levels.setdefault(m.group(1), set()).add(r.levelno)+ assert levels["/readyz"] == {logging.DEBUG}+ assert levels["/livez"] == {logging.DEBUG}+ assert levels["/v1/messages"] == {logging.INFO}diff --git a/tests/test_wire_debug_redaction_policy.py b/tests/test_wire_debug_redaction_policy.pyindex 4d6e4ed9c..0a103e36c 100644--- a/tests/test_wire_debug_redaction_policy.py+++ b/tests/test_wire_debug_redaction_policy.py@@ -81,3 +81,13 @@ def test_wire_debug_keeps_token_usage_counters_visible() -> None:"token_count",):assert not should_redact_key(key), key+++def test_wire_debug_redacts_codex_attestation_header() -> None:+ # The inbound request log line runs every header through this policy, so+ # without it each Codex request wrote the ~4 KB signed attestation to proxy.log.+ redacted = redact_for_wire_debug(+ {"x-oai-attestation": '{"v":1,"s":1,"t":"signed"}', "x-codex-window-id": "w:1"}+ )++ assert redacted == {"x-oai-attestation": WIRE_DEBUG_REDACTED, "x-codex-window-id": "w:1"}