diff --git a/crates/cli/src/gateway/mod.rs b/crates/cli/src/gateway/mod.rs
index 0eefae373..2578ac8c4 100644
--- a/crates/cli/src/gateway/mod.rs
+++ b/crates/cli/src/gateway/mod.rs
@@ -100,7 +100,7 @@ async fn run_managed_gateway(
         let session_id = prep.session_id.clone();
         let session_finish = prep.session_finish;
         let model = prep.model_name.as_deref().unwrap_or("<unknown>");
-        log::warn!(
+        log::debug!(
             target: "nemo_relay.gateway",
             event = "observability_bypassed",
             session_id = session_id.as_str(),
diff --git a/crates/cli/tests/architecture_tests.rs b/crates/cli/tests/architecture_tests.rs
index 45e9a83e2..fe8877b4f 100644
--- a/crates/cli/tests/architecture_tests.rs
+++ b/crates/cli/tests/architecture_tests.rs
@@ -291,10 +291,6 @@ impl<'ast> Visit<'ast> for LogMacroVisitor {
         let macro_name = path.rsplit("::").next().unwrap_or_default();
         if matches!(macro_name, "info" | "warn" | "error" | "debug" | "trace") {
             let tokens = item.tokens.to_string();
-            if matches!(macro_name, "debug" | "trace") {
-                self.failures
-                    .push(format!("production call site uses {path}!"));
-            }
             let target = string_field(&tokens, "target :");
             if target
                 .as_deref()
diff --git a/crates/cli/tests/cli_tests.rs b/crates/cli/tests/cli_tests.rs
index 1575da4f4..0fa1e1dd3 100644
--- a/crates/cli/tests/cli_tests.rs
+++ b/crates/cli/tests/cli_tests.rs
@@ -311,6 +311,86 @@ fn cli_jsonl_logging_records_successful_command_lifecycle_without_leaking_secret
     );
 }
 
+#[test]
+fn cli_claude_startup_probe_bypass_is_debug_only() {
+    let default_stderr = run_claude_startup_probe(None);
+    assert!(
+        !default_stderr.contains(" WARN ") && !default_stderr.contains("observability_bypassed"),
+        "startup probe bypass should not be logged by default: {default_stderr}"
+    );
+
+    let debug_stderr = run_claude_startup_probe(Some("debug"));
+    assert!(
+        debug_stderr.contains(" DEBUG ") && debug_stderr.contains("observability_bypassed"),
+        "startup probe bypass should be logged at debug level: {debug_stderr}"
+    );
+}
+
+fn run_claude_startup_probe(log_level: Option<&str>) -> String {
+    let temp = tempfile::tempdir().unwrap();
+    let probe = TcpListener::bind("127.0.0.1:0").unwrap();
+    let address = probe.local_addr().unwrap();
+    drop(probe);
+    let (upstream_url, received) = spawn_single_request_server(200, r#"{"id":"probe"}"#);
+
+    let mut command = Command::new(gateway_bin());
+    command
+        .args(["--bind", &address.to_string(), "--anthropic-base-url"])
+        .arg(&upstream_url)
+        .env("HOME", temp.path())
+        .env("XDG_CONFIG_HOME", temp.path().join("xdg"))
+        .env_remove("NEMO_RELAY_LOG_STDERR_FORMAT")
+        .env_remove("NEMO_RELAY_LOG_CONFIG_PATH")
+        .stdout(Stdio::null())
+        .stderr(Stdio::piped());
+    match log_level {
+        Some(level) => {
+            command.env("NEMO_RELAY_LOG", level);
+        }
+        None => {
+            command.env_remove("NEMO_RELAY_LOG");
+        }
+    }
+    let child = ChildGuard::new(command.spawn().unwrap());
+
+    let body = r#"{"model":"claude-sonnet-4-5","max_tokens":1,"messages":[{"role":"user","content":"test"}]}"#;
+    let deadline = Instant::now() + Duration::from_secs(5);
+    loop {
+        match TcpStream::connect_timeout(&address, Duration::from_millis(100)) {
+            Ok(mut stream) => {
+                stream
+                    .set_read_timeout(Some(Duration::from_secs(2)))
+                    .unwrap();
+                stream
+                    .write_all(
+                        format!(
+                            "POST /v1/messages HTTP/1.1\r\nHost: {address}\r\ncontent-type: application/json\r\nx-api-key: test\r\nx-claude-code-session-id: probe-session\r\ncontent-length: {}\r\nconnection: close\r\n\r\n{body}",
+                            body.len()
+                        )
+                        .as_bytes(),
+                    )
+                    .unwrap();
+                let mut response = String::new();
+                stream.read_to_string(&mut response).unwrap();
+                assert!(response.starts_with("HTTP/1.1 200"), "{response}");
+                break;
+            }
+            Err(_) if Instant::now() < deadline => thread::sleep(Duration::from_millis(20)),
+            Err(error) => {
+                let output = child.finish();
+                panic!(
+                    "gateway did not accept the startup probe: {error}; stderr: {}",
+                    String::from_utf8_lossy(&output.stderr)
+                );
+            }
+        }
+    }
+
+    let upstream_request = received.recv_timeout(Duration::from_secs(2)).unwrap();
+    assert!(upstream_request.starts_with("POST /v1/messages "));
+    String::from_utf8(child.finish().stderr).unwrap()
+}
+
 #[test]
 fn cli_logs_final_failure_before_shutdown_and_preserves_user_error() {
     let temp = tempfile::tempdir().unwrap();
@@ -1274,6 +1354,29 @@ fn wait_child(child: &mut Child) -> ExitStatus {
     }
 }
 
+struct ChildGuard(Option<Child>);
+
+impl ChildGuard {
+    fn new(child: Child) -> Self {
+        Self(Some(child))
+    }
+
+    fn finish(mut self) -> Output {
+        let mut child = self.0.take().unwrap();
+        let _ = child.kill();
+        wait_child_with_output(child)
+    }
+}
+
+impl Drop for ChildGuard {
+    fn drop(&mut self) {
+        if let Some(child) = self.0.as_mut() {
+            let _ = child.kill();
+            let _ = child.wait();
+        }
+    }
+}
+
 fn wait_child_with_output(mut child: Child) -> Output {
     fn read_pipe(
         pipe: Option<impl Read + Send + 'static>,
diff --git a/docs/nemo-relay-cli/claude-code.mdx b/docs/nemo-relay-cli/claude-code.mdx
index f94d2d59d..e49c7c861 100644
--- a/docs/nemo-relay-cli/claude-code.mdx
+++ b/docs/nemo-relay-cli/claude-code.mdx
@@ -182,8 +182,8 @@ when each span has a real `UserPromptSubmit` payload and assistant output. NeMo
 Relay excludes the known Claude Code startup/preflight probe and late
 uncorrelatable lifecycle hooks from exported user traces so they do not appear
 as synthetic `null`, `user: test`, `idle_timeout`, or lifecycle-only turns.
-Suppressed startup probes are still logged by the gateway as internal pre-turn
-probe bypasses for debugging.
+The gateway records suppressed startup probes as debug-level
+`observability_bypassed` events when debug or trace logging is enabled.
 
 ## Smoke Test
 
diff --git a/integrations/coding-agents/claude-code/README.md b/integrations/coding-agents/claude-code/README.md
index 2ba1e60fa..0ae25300e 100644
--- a/integrations/coding-agents/claude-code/README.md
+++ b/integrations/coding-agents/claude-code/README.md
@@ -41,8 +41,9 @@ root `claude-code-turn` span or ATIF trajectory per user turn. That is expected
 when each turn has a real prompt input and assistant output. Known startup
 probes, uncorrelatable late stop hooks, and other lifecycle-only noise are
 excluded from exported user traces so they do not appear as synthetic `null`,
-`user: test`, or `idle_timeout` turns. Startup probes are still logged by the
-gateway as internal pre-turn probe bypasses for debugging.
+`user: test`, or `idle_timeout` turns.
+The gateway records suppressed startup probes as debug-level
+`observability_bypassed` events when debug or trace logging is enabled.
 
 ## Transparent Setup