diff --git a/.github/workflows/ubuntu-focal.yml b/.github/workflows/ubuntu-focal.yml
index 069330e208..3c1336191c 100644
--- a/.github/workflows/ubuntu-focal.yml
+++ b/.github/workflows/ubuntu-focal.yml
@@ -589,6 +589,8 @@ jobs:
           -DOMNITRACE_BUILD_CI=OFF
           -DOMNITRACE_BUILD_TESTING=ON
           -DOMNITRACE_BUILD_DYNINST=OFF
+          -DOMNITRACE_BUILD_DEBUG=ON
+          -DOMNITRACE_BUILD_HIDDEN_VISIBILITY=OFF
           -DOMNITRACE_USE_MPI=ON
           -DOMNITRACE_USE_PYTHON=ON
           -DOMNITRACE_USE_OMPT=ON
diff --git a/cmake/BuildSettings.cmake b/cmake/BuildSettings.cmake
index 2ed03f4203..f108ebd1b0 100644
--- a/cmake/BuildSettings.cmake
+++ b/cmake/BuildSettings.cmake
@@ -19,7 +19,6 @@ omnitrace_add_option(OMNITRACE_BUILD_EXTRA_OPTIMIZATIONS "Extra optimization fla
 omnitrace_add_option(OMNITRACE_BUILD_LTO "Build with link-time optimization" OFF)
 omnitrace_add_option(OMNITRACE_USE_COMPILE_TIMING
                      "Build with timing metrics for compilation" OFF)
-omnitrace_add_option(OMNITRACE_USE_COVERAGE "Build with code-coverage flags" OFF)
 omnitrace_add_option(OMNITRACE_USE_SANITIZER
                      "Build with -fsanitze=\${OMNITRACE_SANITIZER_TYPE}" OFF)
 omnitrace_add_option(OMNITRACE_BUILD_STATIC_LIBGCC
@@ -145,13 +144,15 @@ endif()
 # non-debug optimizations
 #
 omnitrace_add_interface_library(omnitrace-compile-extra "Extra optimization flags")
-if(NOT OMNITRACE_USE_COVERAGE AND OMNITRACE_BUILD_EXTRA_OPTIMIZATIONS)
+if(NOT OMNITRACE_BUILD_CODECOV AND OMNITRACE_BUILD_EXTRA_OPTIMIZATIONS)
     add_target_flag_if_avail(
         omnitrace-compile-extra "-finline-functions" "-funroll-loops" "-ftree-vectorize"
         "-ftree-loop-optimize" "-ftree-loop-vectorize")
 endif()
 
-if(NOT "${CMAKE_BUILD_TYPE}" STREQUAL "Debug" AND OMNITRACE_BUILD_EXTRA_OPTIMIZATIONS)
+if(NOT "${CMAKE_BUILD_TYPE}" STREQUAL "Debug"
+   AND OMNITRACE_BUILD_EXTRA_OPTIMIZATIONS
+   AND NOT OMNITRACE_BUILD_CODECOV)
     target_link_libraries(omnitrace-compile-options
                           INTERFACE $<BUILD_INTERFACE:omnitrace-compile-extra>)
     add_flag_if_avail(
@@ -166,29 +167,32 @@ endif()
 #
 add_cxx_flag_if_avail("-faligned-new")
 
-omnitrace_save_variables(FLTO VARIABLES CMAKE_CXX_FLAGS)
-set(_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
-set(CMAKE_CXX_FLAGS "-flto=thin ${_CXX_FLAGS}")
-
 omnitrace_add_interface_library(omnitrace-lto "Adds link-time-optimization flags")
-add_target_flag_if_avail(omnitrace-lto "-flto=thin")
-if(NOT cxx_omnitrace_lto_flto_thin)
-    set(CMAKE_CXX_FLAGS "-flto ${_CXX_FLAGS}")
-    add_target_flag_if_avail(omnitrace-lto "-flto")
-    if(NOT cxx_omnitrace_lto_flto)
-        set(OMNITRACE_BUILD_LTO OFF)
-    else()
-        target_link_options(omnitrace-lto INTERFACE -flto)
-    endif()
-    add_target_flag_if_avail(omnitrace-lto "-fno-fat-lto-objects")
-    if(cxx_omnitrace_lto_fno_fat_lto_objects)
-        target_link_options(omnitrace-lto INTERFACE -fno-fat-lto-objects)
-    endif()
-else()
-    target_link_options(omnitrace-lto INTERFACE -flto=thin)
-endif()
 
-omnitrace_restore_variables(FLTO VARIABLES CMAKE_CXX_FLAGS)
+if(NOT OMNITRACE_BUILD_CODECOV)
+    omnitrace_save_variables(FLTO VARIABLES CMAKE_CXX_FLAGS)
+    set(_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+    set(CMAKE_CXX_FLAGS "-flto=thin ${_CXX_FLAGS}")
+
+    add_target_flag_if_avail(omnitrace-lto "-flto=thin")
+    if(NOT cxx_omnitrace_lto_flto_thin)
+        set(CMAKE_CXX_FLAGS "-flto ${_CXX_FLAGS}")
+        add_target_flag_if_avail(omnitrace-lto "-flto")
+        if(NOT cxx_omnitrace_lto_flto)
+            set(OMNITRACE_BUILD_LTO OFF)
+        else()
+            target_link_options(omnitrace-lto INTERFACE -flto)
+        endif()
+        add_target_flag_if_avail(omnitrace-lto "-fno-fat-lto-objects")
+        if(cxx_omnitrace_lto_fno_fat_lto_objects)
+            target_link_options(omnitrace-lto INTERFACE -fno-fat-lto-objects)
+        endif()
+    else()
+        target_link_options(omnitrace-lto INTERFACE -flto=thin)
+    endif()
+
+    omnitrace_restore_variables(FLTO VARIABLES CMAKE_CXX_FLAGS)
+endif()
 
 # ----------------------------------------------------------------------------------------#
 # print compilation timing reports (Clang compiler)
diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index 07a26b3749..8c8451e305 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -54,3 +54,4 @@ add_subdirectory(rccl)
 add_subdirectory(rewrite-caller)
 add_subdirectory(causal)
 add_subdirectory(trace-time-window)
+add_subdirectory(fork)
diff --git a/examples/fork/CMakeLists.txt b/examples/fork/CMakeLists.txt
new file mode 100644
index 0000000000..57dab6947c
--- /dev/null
+++ b/examples/fork/CMakeLists.txt
@@ -0,0 +1,18 @@
+cmake_minimum_required(VERSION 3.15 FATAL_ERROR)
+
+project(omnitrace-fork LANGUAGES CXX)
+
+set(CMAKE_BUILD_TYPE "RelWithDebInfo")
+string(REPLACE " " ";" _FLAGS "${CMAKE_CXX_FLAGS_DEBUG}")
+
+find_package(Threads REQUIRED)
+add_executable(fork-example fork.cpp)
+target_link_libraries(fork-example PRIVATE Threads::Threads)
+target_compile_options(fork-example PRIVATE ${_FLAGS})
+
+if(OMNITRACE_INSTALL_EXAMPLES)
+    install(
+        TARGETS fork-example
+        DESTINATION bin
+        COMPONENT omnitrace-examples)
+endif()
diff --git a/examples/fork/fork.cpp b/examples/fork/fork.cpp
new file mode 100644
index 0000000000..fd53670024
--- /dev/null
+++ b/examples/fork/fork.cpp
@@ -0,0 +1,78 @@
+
+#include <chrono>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <string>
+#include <sys/wait.h>
+#include <thread>
+#include <unistd.h>
+
+void
+print_info(const char* _name)
+{
+    printf("[%s] pid = %i, ppid = %i\n", _name, getpid(), getppid());
+}
+
+int
+run(const char* _name, int nchildren)
+{
+    for(int i = 0; i < nchildren; ++i)
+    {
+        auto _run = [i, _name]() {
+            pid_t _pid = fork();
+            if(_pid == 0)
+            {
+                // child code
+                print_info(_name);
+                auto _sleep = [=]() {
+                    std::this_thread::sleep_for(std::chrono::seconds{ i + 1 });
+                };
+                std::thread{ _sleep }.join();
+                exit(EXIT_SUCCESS);
+            }
+        };
+        std::thread{ _run }.join();
+        //_run();
+    }
+
+    int   _status   = 0;
+    pid_t _wait_pid = 0;
+    // parent waits for all the child processes
+    while((_wait_pid = wait(&_status)) > 0)
+    {
+        printf("[%s][%i] returned from wait with pid = %i :: ", _name, getpid(),
+               _wait_pid);
+        if(WIFEXITED(_status))
+        {
+            printf("exited, status=%d\n", WEXITSTATUS(_status));
+        }
+        else if(WIFSIGNALED(_status))
+        {
+            printf("killed by signal %d\n", WTERMSIG(_status));
+        }
+        else if(WIFSTOPPED(_status))
+        {
+            printf("stopped by signal %d\n", WSTOPSIG(_status));
+        }
+        else if(WIFCONTINUED(_status))
+        {
+            printf("continued\n");
+        }
+        else
+        {
+            printf("unknown\n");
+        }
+    }
+    return _status;
+}
+
+int
+main(int argc, char** argv)
+{
+    int _n = 4;
+    if(argc > 1) _n = std::stoi(argv[1]);
+
+    print_info(argv[0]);
+    return run(argv[0], _n);
+}
diff --git a/external/timemory b/external/timemory
index b5e815c8a4..146b91cb87 160000
--- a/external/timemory
+++ b/external/timemory
@@ -1 +1 @@
-Subproject commit b5e815c8a4f5c3c1e468c83e7790c3c011194a7b
+Subproject commit 146b91cb87f30d66044a418b549c3448c4c79099
diff --git a/scripts/run-ci.sh b/scripts/run-ci.sh
index 03f27b11cf..9ac5c72561 100755
--- a/scripts/run-ci.sh
+++ b/scripts/run-ci.sh
@@ -200,51 +200,62 @@ set(CTEST_BINARY_DIRECTORY ${BINARY_DIR})
 
 set(CTEST_UPDATE_COMMAND ${GIT_CMD})
 set(CTEST_CONFIGURE_COMMAND "${CMAKE_CMD} -B ${BINARY_DIR} ${SOURCE_DIR} -DOMNITRACE_BUILD_CI=ON ${CMAKE_ARGS}")
-set(CTEST_BUILD_COMMAND "${CMAKE_CMD} --build ${BINARY_DIR} --target all")
+set(CTEST_BUILD_COMMAND "${CMAKE_CMD} --build ${BINARY_DIR} --target all --parallel ${CMAKE_BUILD_PARALLEL_LEVEL}")
 set(CTEST_COVERAGE_COMMAND ${GCOV_CMD})
 EOF
 
 verbose-run cd ${BINARY_DIR}
 
 cat << EOF > dashboard.cmake
+cmake_minimum_required(VERSION 3.16 FATAL_ERROR)
 
 include("\${CMAKE_CURRENT_LIST_DIR}/CTestCustom.cmake")
 
+set(_STAGES ${DASHBOARD_STAGES})
+
+macro(handle_submit)
+    if("Submit" IN_LIST _STAGES)
+        ctest_submit(
+            ${ARGN}
+            CAPTURE_CMAKE_ERROR _submit_err)
+        if(NOT \${_submit_err} EQUAL 0)
+            message(WARNING "Submission failed: ctest_submit(\${ARGN})")
+        endif()
+    endif()
+endmacro()
+
 macro(handle_error _message _ret)
     if(NOT \${\${_ret}} EQUAL 0)
-        ctest_submit(PARTS Done RETURN_VALUE _submit_ret)
+        handle_submit(PARTS Done RETURN_VALUE _submit_ret)
         message(FATAL_ERROR "\${_message} failed: \${\${_ret}}")
     endif()
 endmacro()
 
 ctest_start(${DASHBOARD_MODE})
 ctest_update(SOURCE "${SOURCE_DIR}")
-ctest_submit(PARTS Start Update RETURN_VALUE _submit_ret)
-
 ctest_configure(BUILD "${BINARY_DIR}" RETURN_VALUE _configure_ret)
-ctest_submit(PARTS Configure RETURN_VALUE _submit_ret)
 
+handle_submit(PARTS Start Update Configure RETURN_VALUE _submit_ret)
 handle_error("Configure" _configure_ret)
 
 ctest_build(BUILD "${BINARY_DIR}" RETURN_VALUE _build_ret)
-ctest_submit(PARTS Build RETURN_VALUE _submit_ret)
-
+handle_submit(PARTS Build RETURN_VALUE _submit_ret)
 handle_error("Build" _build_ret)
 
 ctest_test(BUILD "${BINARY_DIR}" RETURN_VALUE _test_ret)
-ctest_submit(PARTS Test RETURN_VALUE _submit_ret)
+handle_submit(PARTS Test RETURN_VALUE _submit_ret)
 
 if("${CODECOV}" GREATER 0)
     ctest_coverage(
         BUILD "${BINARY_DIR}"
         RETURN_VALUE _coverage_ret
         CAPTURE_CMAKE_ERROR _coverage_err)
-    ctest_submit(PARTS Coverage RETURN_VALUE _submit_ret)
+    handle_submit(PARTS Coverage RETURN_VALUE _submit_ret)
 endif()
 
 handle_error("Testing" _test_ret)
 
-ctest_submit(PARTS Done RETURN_VALUE _submit_ret)
+handle_submit(PARTS Done RETURN_VALUE _submit_ret)
 EOF
 
 verbose-run cat CTestCustom.cmake
diff --git a/source/bin/omnitrace-avail/avail.cpp b/source/bin/omnitrace-avail/avail.cpp
index a29f55d925..395f2b05c7 100644
--- a/source/bin/omnitrace-avail/avail.cpp
+++ b/source/bin/omnitrace-avail/avail.cpp
@@ -29,6 +29,7 @@
 #include "get_availability.hpp"
 #include "info_type.hpp"
 
+#include "api.hpp"
 #include "core/config.hpp"
 #include "core/gpu.hpp"
 #include "library/rocprofiler.hpp"
diff --git a/source/lib/common/invoke.hpp b/source/lib/common/invoke.hpp
index 43afef233e..9de2645c80 100644
--- a/source/lib/common/invoke.hpp
+++ b/source/lib/common/invoke.hpp
@@ -29,6 +29,7 @@
 #include <cstring>
 #include <functional>
 #include <string>
+#include <unistd.h>
 
 #if !defined(OMNITRACE_COMMON_LIBRARY_NAME)
 #    define OMNITRACE_COMMON_LIBRARY_NAME "common"
@@ -77,9 +78,9 @@ ignore(const char* _name, int _verbose, int _value, const char* _reason, Args...
         fflush(stderr);
         fprintf(stderr,
                 "[omnitrace][" OMNITRACE_COMMON_LIBRARY_NAME
-                "][%li] %s(%s) was ignored :: %s\n",
-                get_thread_index(), _name, join(QuoteStrings{}, ", ", _args...).c_str(),
-                _reason);
+                "][%i][%li] %s(%s) was ignored :: %s\n",
+                getpid(), get_thread_index(), _name,
+                join(QuoteStrings{}, ", ", _args...).c_str(), _reason);
         fflush(stderr);
     }
 }
@@ -109,8 +110,8 @@ invoke(const char* _name, int _verbose, bool& _toggle, FuncT&& _func, Args... _a
                 OMNITRACE_COMMON_LIBRARY_LOG_START
                 fprintf(stderr,
                         "[omnitrace][" OMNITRACE_COMMON_LIBRARY_NAME
-                        "][%li][%i] %s(%s)\n",
-                        get_thread_index(), _lk, _name,
+                        "][%i][%li][%i] %s(%s)\n",
+                        getpid(), get_thread_index(), _lk, _name,
                         join(QuoteStrings{}, ", ", _args...).c_str());
                 OMNITRACE_COMMON_LIBRARY_LOG_END
                 fflush(stderr);
@@ -123,8 +124,8 @@ invoke(const char* _name, int _verbose, bool& _toggle, FuncT&& _func, Args... _a
             OMNITRACE_COMMON_LIBRARY_LOG_START
             fprintf(stderr,
                     "[omnitrace][" OMNITRACE_COMMON_LIBRARY_NAME
-                    "][%li] %s(%s) was guarded :: value = %i\n",
-                    get_thread_index(), _name,
+                    "][%i][%li] %s(%s) was guarded :: value = %i\n",
+                    getpid(), get_thread_index(), _name,
                     join(QuoteStrings{}, ", ", _args...).c_str(), _lk);
             OMNITRACE_COMMON_LIBRARY_LOG_END
             fflush(stderr);
@@ -135,8 +136,9 @@ invoke(const char* _name, int _verbose, bool& _toggle, FuncT&& _func, Args... _a
         OMNITRACE_COMMON_LIBRARY_LOG_START
         fprintf(stderr,
                 "[omnitrace][" OMNITRACE_COMMON_LIBRARY_NAME
-                "][%li] %s(%s) ignored :: null function pointer\n",
-                get_thread_index(), _name, join(QuoteStrings{}, ", ", _args...).c_str());
+                "][%i][%li] %s(%s) ignored :: null function pointer\n",
+                getpid(), get_thread_index(), _name,
+                join(QuoteStrings{}, ", ", _args...).c_str());
         OMNITRACE_COMMON_LIBRARY_LOG_END
     }
 
diff --git a/source/lib/core/config.cpp b/source/lib/core/config.cpp
index 8653f4cb5f..12197648c1 100644
--- a/source/lib/core/config.cpp
+++ b/source/lib/core/config.cpp
@@ -227,20 +227,18 @@ finalize()
 bool
 settings_are_configured()
 {
-    volatile bool _v = _settings_are_configured();
-    return _v;
+    return _settings_are_configured();
 }
 
 void
 configure_settings(bool _init)
 {
-    volatile bool _v = _settings_are_configured();
-    if(_v) return;
-
     static bool _once = false;
     if(_once) return;
     _once = true;
 
+    if(settings_are_configured()) return;
+
     if(get_is_continuous_integration() && get_state() < State::Init)
     {
         timemory_print_demangled_backtrace<64>();
@@ -2192,9 +2190,11 @@ get_perfetto_output_filename()
         _ext = _val.substr(_pos_ext + 1);
         _val = _val.substr(0, _pos_ext);
     }
-    _val = settings::compose_output_filename(_val, _ext, settings::use_output_suffix(),
-                                             settings::default_process_suffix(), false,
-                                             _dir);
+
+    auto _cfg = settings::compose_filename_config{ settings::use_output_suffix(),
+                                                   settings::default_process_suffix(),
+                                                   false, _dir };
+    _val      = settings::compose_output_filename(_val, _ext, _cfg);
     if(!_val.empty() && _val.at(0) != '/')
         return settings::format(JOIN('/', "%env{PWD}%", _val), get_config()->get_tag());
     return _val;
diff --git a/source/lib/core/config.hpp b/source/lib/core/config.hpp
index c74e7f6fc6..2af2ffe12e 100644
--- a/source/lib/core/config.hpp
+++ b/source/lib/core/config.hpp
@@ -100,11 +100,16 @@ template <typename Tp>
 bool
 set_setting_value(const std::string& _name, Tp&& _v)
 {
+    auto _user_upd = tim::settings::update_type::user;
     auto _instance = tim::settings::shared_instance();
     auto _setting  = _instance->find(_name);
     if(_setting == _instance->end()) return false;
     if(!_setting->second) return false;
-    return _setting->second->set(std::forward<Tp>(_v));
+    auto& itr      = _setting->second;
+    auto  _upd     = itr->set_user_updated();
+    auto  _success = itr->set(std::forward<Tp>(_v), _user_upd);
+    if(!_success) itr->set_updated(_upd);
+    return _success;
 }
 
 template <typename Tp>
diff --git a/source/lib/core/debug.cpp b/source/lib/core/debug.cpp
index 1076d34e5c..0dfe182355 100644
--- a/source/lib/core/debug.cpp
+++ b/source/lib/core/debug.cpp
@@ -25,6 +25,7 @@
 #include "state.hpp"
 
 #include <timemory/log/color.hpp>
+#include <timemory/process/threading.hpp>
 #include <timemory/utility/filepath.hpp>
 
 #include <iomanip>
@@ -43,6 +44,29 @@ struct source_location_history
     size_t                          size = 0;
 };
 
+const std::string&
+get_file_name()
+{
+    static auto _fname = tim::get_env<std::string>("OMNITRACE_LOG_FILE", "");
+    return _fname;
+}
+
+std::atomic<FILE*>&
+get_file_pointer()
+{
+    static auto _v = std::atomic<FILE*>{ []() {
+        const auto&_fname= get_file_name();
+        if(!_fname.empty()) tim::log::monochrome() = true;
+        return (_fname.empty())
+                   ? stderr
+                   : filepath::fopen(
+                         settings::format(_fname, filepath::basename(filepath::realpath(
+                                                      "/proc/self/exe", nullptr, false))),
+                         "w");
+    }() };
+    return _v;
+}
+
 auto&
 get_source_location_history()
 {
@@ -88,11 +112,28 @@ lock::~lock()
 FILE*
 get_file()
 {
-    static FILE* _v = []() {
-        auto&& _fname = tim::get_env<std::string>("OMNITRACE_LOG_FILE", "");
-        if(!_fname.empty()) tim::log::monochrome() = true;
-        return (_fname.empty()) ? stderr : tim::filepath::fopen(_fname, "w");
-    }();
+    return get_file_pointer();
+}
+
+void
+close_file()
+{
+    if(get_file() != stderr)
+    {
+        auto* _file = get_file_pointer().load();
+        get_file_pointer().store(stderr);
+        fclose(_file);
+        // Write the trace into a file.
+        if(get_verbose() >= 0)
+            operation::file_output_message<tim::project::omnitrace>{}(
+                get_file_name(), std::string{ "debug" });
+    }
+}
+
+int64_t
+get_tid()
+{
+    static thread_local auto _v = threading::get_id();
     return _v;
 }
 }  // namespace debug
diff --git a/source/lib/core/debug.hpp b/source/lib/core/debug.hpp
index 63902c3678..777d8593e3 100644
--- a/source/lib/core/debug.hpp
+++ b/source/lib/core/debug.hpp
@@ -86,6 +86,12 @@ set_source_location(source_location&&);
 FILE*
 get_file();
 //
+void
+close_file();
+//
+int64_t
+get_tid();
+//
 inline void
 flush()
 {
@@ -168,7 +174,7 @@ as_hex<void*>(void*, size_t);
 #endif
 
 #if !defined(OMNITRACE_DEBUG_THREAD_IDENTIFIER)
-#    define OMNITRACE_DEBUG_THREAD_IDENTIFIER ::tim::threading::get_id()
+#    define OMNITRACE_DEBUG_THREAD_IDENTIFIER ::omnitrace::debug::get_tid()
 #endif
 
 #if !defined(OMNITRACE_SOURCE_LOCATION)
@@ -394,7 +400,6 @@ as_hex<void*>(void*, size_t);
         fprintf(::omnitrace::debug::get_file(), __VA_ARGS__);                            \
         ::omnitrace::debug::flush();                                                     \
         ::omnitrace::set_state(::omnitrace::State::Finalized);                           \
-        ::tim::signals::disable_signal_detection();                                      \
         timemory_print_demangled_backtrace<64>();                                        \
         METHOD;                                                                          \
     }
@@ -410,7 +415,6 @@ as_hex<void*>(void*, size_t);
         fprintf(::omnitrace::debug::get_file(), __VA_ARGS__);                            \
         ::omnitrace::debug::flush();                                                     \
         ::omnitrace::set_state(::omnitrace::State::Finalized);                           \
-        ::tim::signals::disable_signal_detection();                                      \
         timemory_print_demangled_backtrace<64>();                                        \
         METHOD;                                                                          \
     }
@@ -427,7 +431,6 @@ as_hex<void*>(void*, size_t);
         fprintf(::omnitrace::debug::get_file(), __VA_ARGS__);                            \
         ::omnitrace::debug::flush();                                                     \
         ::omnitrace::set_state(::omnitrace::State::Finalized);                           \
-        ::tim::signals::disable_signal_detection();                                      \
         timemory_print_demangled_backtrace<64>();                                        \
         METHOD;                                                                          \
     }
@@ -443,7 +446,6 @@ as_hex<void*>(void*, size_t);
         fprintf(::omnitrace::debug::get_file(), __VA_ARGS__);                            \
         ::omnitrace::debug::flush();                                                     \
         ::omnitrace::set_state(::omnitrace::State::Finalized);                           \
-        ::tim::signals::disable_signal_detection();                                      \
         timemory_print_demangled_backtrace<64>();                                        \
         METHOD;                                                                          \
     }
diff --git a/source/lib/omnitrace-dl/dl.cpp b/source/lib/omnitrace-dl/dl.cpp
index 251eefe337..82b97d2b24 100644
--- a/source/lib/omnitrace-dl/dl.cpp
+++ b/source/lib/omnitrace-dl/dl.cpp
@@ -128,6 +128,16 @@ reset_omnitrace_preload()
     }
 }
 
+inline pid_t
+get_omnitrace_root_pid()
+{
+    auto _pid = getpid();
+    setenv("OMNITRACE_ROOT_PROCESS", std::to_string(_pid).c_str(), 0);
+    return get_env("OMNITRACE_ROOT_PROCESS", _pid);
+}
+
+pid_t _omnitrace_root_pid = get_omnitrace_root_pid();
+
 // environment priority:
 //  - OMNITRACE_DL_DEBUG
 //  - OMNITRACE_DL_VERBOSE
@@ -538,7 +548,9 @@ bool _omnitrace_dl_fini = (std::atexit([]() {
     {                                                                                    \
         fflush(stderr);                                                                  \
         OMNITRACE_COMMON_LIBRARY_LOG_START                                               \
-        fprintf(stderr, "[omnitrace][" OMNITRACE_COMMON_LIBRARY_NAME "] " __VA_ARGS__);  \
+        fprintf(stderr, "[omnitrace][" OMNITRACE_COMMON_LIBRARY_NAME "][%i] ",           \
+                getpid());                                                               \
+        fprintf(stderr, __VA_ARGS__);                                                    \
         OMNITRACE_COMMON_LIBRARY_LOG_END                                                 \
         fflush(stderr);                                                                  \
     }
@@ -1040,7 +1052,6 @@ extern "C"
     ompt_start_tool_result_t* ompt_start_tool(unsigned int omp_version,
                                               const char*  runtime_version)
     {
-        if(!omnitrace::common::get_env("OMNITRACE_USE_OMPT", true)) return nullptr;
         return OMNITRACE_DL_INVOKE(get_indirect().ompt_start_tool_f, omp_version,
                                    runtime_version);
     }
@@ -1060,6 +1071,7 @@ bool
 omnitrace_preload()
 {
     auto _preload = get_omnitrace_preload() && get_env("OMNITRACE_ENABLED", true);
+    auto _use_mpi = get_env("OMNITRACE_USE_MPI", get_env("OMNITRACE_USE_MPIP", false));
 
     static bool _once = false;
     if(_once) return _preload;
@@ -1067,7 +1079,7 @@ omnitrace_preload()
 
     if(_preload)
     {
-        // reset_omnitrace_preload();
+        reset_omnitrace_preload();
         omnitrace_preinit_library();
         auto _causal = get_env("OMNITRACE_USE_CAUSAL", false);
         auto _mode   = get_env("OMNITRACE_MODE", (_causal) ? "causal" : "sampling");
@@ -1075,6 +1087,16 @@ omnitrace_preload()
                          ::omnitrace::join(::omnitrace::QuoteStrings{}, ", ", _mode,
                                            false, "omnitrace")
                              .c_str());
+        if(_use_mpi && !(_causal && _mode == "causal"))
+        {
+            // only make this call if true bc otherwise, if
+            // false, it will disable the MPIP component and
+            // we may intercept the MPI init call later.
+            // If _use_mpi defaults to true above, calling this
+            // will override can current env or config value for
+            // OMNITRACE_USE_PID.
+            omnitrace_set_mpi(_use_mpi, false);
+        }
         omnitrace_init(_mode.c_str(), false, nullptr);
         omnitrace_init_tooling();
     }
diff --git a/source/lib/omnitrace/library.cpp b/source/lib/omnitrace/library.cpp
index c6bf76d7ba..de183119fe 100644
--- a/source/lib/omnitrace/library.cpp
+++ b/source/lib/omnitrace/library.cpp
@@ -68,6 +68,7 @@
 #include <timemory/signals/signal_mask.hpp>
 #include <timemory/signals/types.hpp>
 #include <timemory/utility/backtrace.hpp>
+#include <timemory/utility/join.hpp>
 #include <timemory/utility/procfs/maps.hpp>
 
 #include <atomic>
@@ -679,6 +680,8 @@ omnitrace_finalize_hidden(void)
     // disable initialization callback
     threading::remove_callback(&ensure_initialization);
 
+    bool _is_child = is_child_process();
+
     set_thread_state(ThreadState::Completed);
 
     // return if not active
@@ -688,6 +691,12 @@ omnitrace_finalize_hidden(void)
                                 std::to_string(get_state()).c_str());
         return;
     }
+    else if(_is_child)
+    {
+        set_state(State::Finalized);
+        std::quick_exit(EXIT_SUCCESS);
+        return;
+    }
 
     if(get_verbose() >= 0 || get_debug()) fprintf(stderr, "\n");
     OMNITRACE_VERBOSE_F(0, "finalizing...\n");
@@ -1104,9 +1113,14 @@ omnitrace_finalize_hidden(void)
                       _push_count, "vs. popped:", _pop_count)
             .c_str());
 
+    debug::close_file();
     config::finalize();
 
     OMNITRACE_VERBOSE_F(0, "Finalized: %s\n", _finalization.as_string().c_str());
+
+    tim::signals::enable_signal_detection(
+        { tim::signals::sys_signal::SegFault, tim::signals::sys_signal::Stop },
+        [](int) {});
 }
 
 //======================================================================================//
diff --git a/source/lib/omnitrace/library/components/backtrace.cpp b/source/lib/omnitrace/library/components/backtrace.cpp
index 5d9d44a6d7..c9d65dfe8c 100644
--- a/source/lib/omnitrace/library/components/backtrace.cpp
+++ b/source/lib/omnitrace/library/components/backtrace.cpp
@@ -127,6 +127,7 @@ backtrace::filter_and_patch(const std::vector<entry_type>& _data)
         if(_lbl.find("rocprofiler_") != _npos) return -1;
         if(_lbl.find("roctracer_") != _npos) return -1;
         if(_lbl.find("perfetto::") != _npos) return -1;
+        if(_lbl.find("protozero::") == 0) return -1;
         return 1;
     };
 
diff --git a/source/lib/omnitrace/library/components/backtrace.hpp b/source/lib/omnitrace/library/components/backtrace.hpp
index e077d4c3f2..9652b67468 100644
--- a/source/lib/omnitrace/library/components/backtrace.hpp
+++ b/source/lib/omnitrace/library/components/backtrace.hpp
@@ -28,11 +28,11 @@
 #include "core/timemory.hpp"
 #include "library/thread_data.hpp"
 
-#include <timemory/components/base.hpp>
-#include <timemory/macros/language.hpp>
+#include <timemory/components/base/declaration.hpp>
 #include <timemory/mpl/concepts.hpp>
-#include <timemory/utility/unwind.hpp>
-#include <timemory/variadic/types.hpp>
+#include <timemory/unwind/cache.hpp>
+#include <timemory/unwind/processed_entry.hpp>
+#include <timemory/unwind/stack.hpp>
 
 #include <array>
 #include <chrono>
diff --git a/source/lib/omnitrace/library/components/backtrace_metrics.cpp b/source/lib/omnitrace/library/components/backtrace_metrics.cpp
index 8c2527ca69..386b39f553 100644
--- a/source/lib/omnitrace/library/components/backtrace_metrics.cpp
+++ b/source/lib/omnitrace/library/components/backtrace_metrics.cpp
@@ -28,7 +28,6 @@
 #include "library/components/ensure_storage.hpp"
 #include "library/ptl.hpp"
 #include "library/runtime.hpp"
-#include "library/sampling.hpp"
 #include "library/thread_info.hpp"
 #include "library/tracing.hpp"
 
diff --git a/source/lib/omnitrace/library/components/category_region.hpp b/source/lib/omnitrace/library/components/category_region.hpp
index ec89f747c0..15351fa64b 100644
--- a/source/lib/omnitrace/library/components/category_region.hpp
+++ b/source/lib/omnitrace/library/components/category_region.hpp
@@ -137,6 +137,8 @@ category_region<CategoryT>::start(std::string_view name, Args&&... args)
     if(get_thread_state() == ThreadState::Disabled) return;
     if(get_state() >= State::Finalized) return;
 
+    if(name.empty()) return;
+
     OMNITRACE_SCOPED_THREAD_STATE(ThreadState::Internal);
 
     // the expectation here is that if the state is not active then the call
@@ -186,7 +188,7 @@ category_region<CategoryT>::start(std::string_view name, Args&&... args)
     {
         if(get_use_timemory())
         {
-            tracing::push_timemory(CategoryT{}, name.data(), std::forward<Args>(args)...);
+            tracing::push_timemory(CategoryT{}, name, std::forward<Args>(args)...);
         }
     }
 
@@ -248,8 +250,7 @@ category_region<CategoryT>::stop(std::string_view name, Args&&... args)
         {
             if(get_use_timemory())
             {
-                tracing::pop_timemory(CategoryT{}, name.data(),
-                                      std::forward<Args>(args)...);
+                tracing::pop_timemory(CategoryT{}, name, std::forward<Args>(args)...);
             }
         }
 
diff --git a/source/lib/omnitrace/library/components/exit_gotcha.cpp b/source/lib/omnitrace/library/components/exit_gotcha.cpp
index ce955833da..ce171f2814 100644
--- a/source/lib/omnitrace/library/components/exit_gotcha.cpp
+++ b/source/lib/omnitrace/library/components/exit_gotcha.cpp
@@ -60,28 +60,33 @@ invoke_exit_gotcha(const exit_gotcha::gotcha_data& _data, FuncT _func, Args... _
 {
     threading::clear_callbacks();
 
-    if(config::settings_are_configured())
+    if(get_state() < State::Finalized)
     {
-        OMNITRACE_VERBOSE(0, "%s called %s(%s)...\n", get_exe_name().c_str(),
-                          _data.tool_id.c_str(), JOIN(", ", _args...).c_str());
-    }
-    else
-    {
-        OMNITRACE_BASIC_VERBOSE(0, "%s called %s(%s)...\n", get_exe_name().c_str(),
-                                _data.tool_id.c_str(), JOIN(", ", _args...).c_str());
-    }
+        if(config::settings_are_configured())
+        {
+            OMNITRACE_VERBOSE(0, "finalizing %s before calling %s(%s)...\n",
+                              get_exe_name().c_str(), _data.tool_id.c_str(),
+                              JOIN(", ", _args...).c_str());
+        }
+        else
+        {
+            OMNITRACE_BASIC_VERBOSE(0, "finalizing %s before calling %s(%s)...\n",
+                                    get_exe_name().c_str(), _data.tool_id.c_str(),
+                                    JOIN(", ", _args...).c_str());
+        }
 
-    if(get_state() != State::Finalized) omnitrace_finalize_hidden();
+        omnitrace_finalize();
+    }
 
     if(config::settings_are_configured())
     {
-        OMNITRACE_VERBOSE(0, "%s called %s(%s)...\n", get_exe_name().c_str(),
-                          _data.tool_id.c_str(), JOIN(", ", _args...).c_str());
+        OMNITRACE_VERBOSE(0, "calling %s(%s) in %s...\n", _data.tool_id.c_str(),
+                          JOIN(", ", _args...).c_str(), get_exe_name().c_str());
     }
     else
     {
-        OMNITRACE_BASIC_VERBOSE(0, "%s called %s(%s)...\n", get_exe_name().c_str(),
-                                _data.tool_id.c_str(), JOIN(", ", _args...).c_str());
+        OMNITRACE_BASIC_VERBOSE(0, "calling %s(%s) in %s...\n", _data.tool_id.c_str(),
+                                JOIN(", ", _args...).c_str(), get_exe_name().c_str());
     }
 
     (*_func)(_args...);
diff --git a/source/lib/omnitrace/library/components/fork_gotcha.cpp b/source/lib/omnitrace/library/components/fork_gotcha.cpp
index cf7f120c26..fabdd85b18 100644
--- a/source/lib/omnitrace/library/components/fork_gotcha.cpp
+++ b/source/lib/omnitrace/library/components/fork_gotcha.cpp
@@ -20,10 +20,15 @@
 // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 // SOFTWARE.
 
-#include "library/components/fork_gotcha.hpp"
+#include "api.hpp"
+
 #include "core/config.hpp"
 #include "core/debug.hpp"
+#include "core/perfetto.hpp"
 #include "core/state.hpp"
+#include "library/components/fork_gotcha.hpp"
+#include "library/runtime.hpp"
+#include "library/sampling.hpp"
 
 #include <timemory/backends/process.hpp>
 #include <timemory/backends/threading.hpp>
@@ -46,8 +51,10 @@ void
 fork_gotcha::audit(const gotcha_data_t&, audit::incoming)
 {
     tim::set_env("OMNITRACE_PRELOAD", "0", 1);
-    OMNITRACE_VERBOSE(1, "fork() called on PID %i (rank: %i), TID %li\n",
-                      process::get_id(), dmp::rank(), threading::get_id());
+    tim::set_env("OMNITRACE_ROOT_PROCESS", process::get_id(), 0);
+    omnitrace_reset_preload_hidden();
+    OMNITRACE_BASIC_VERBOSE(0, "fork() called on PID %i (rank: %i), TID %li\n",
+                            process::get_id(), dmp::rank(), threading::get_id());
     OMNITRACE_BASIC_DEBUG(
         "Warning! Calling fork() within an OpenMPI application using libfabric "
         "may result is segmentation fault\n");
@@ -59,9 +66,31 @@ fork_gotcha::audit(const gotcha_data_t&, audit::outgoing, pid_t _pid)
 {
     if(_pid != 0)
     {
-        OMNITRACE_VERBOSE(1, "fork() called on PID %i created PID %i\n", getppid(), _pid);
+        OMNITRACE_BASIC_VERBOSE(0, "fork() called on PID %i created PID %i\n", getppid(),
+                                _pid);
+    }
+    else
+    {
+        OMNITRACE_REQUIRE(is_child_process())
+            << "Error! child process " << process::get_id()
+            << " believes it is the root process " << get_root_process_id() << "\n";
+        settings::enabled() = false;
+        settings::verbose() = -127;
+        settings::debug()   = false;
+        omnitrace::sampling::shutdown();
+        omnitrace::categories::shutdown();
+        omnitrace::get_perfetto_session().release();
+        set_thread_state(::omnitrace::ThreadState::Disabled);
+    }
+
+    if(!settings::use_output_suffix())
+    {
         settings::use_output_suffix()      = true;
         settings::default_process_suffix() = process::get_id();
+
+        OMNITRACE_BASIC_VERBOSE(
+            0, "call to fork() enables using an output suffix. PID %i will use %i\n",
+            process::get_id(), process::get_id());
     }
 }
 }  // namespace component
diff --git a/source/lib/omnitrace/library/components/mpi_gotcha.cpp b/source/lib/omnitrace/library/components/mpi_gotcha.cpp
index bc34ca42c6..ed48ee9053 100644
--- a/source/lib/omnitrace/library/components/mpi_gotcha.cpp
+++ b/source/lib/omnitrace/library/components/mpi_gotcha.cpp
@@ -102,29 +102,34 @@ auto     mpi_comm_records  = std::map<uintptr_t, comm_rank_data>{};
 using tim::auto_lock_t;
 using tim::type_mutex;
 
+#if defined(TIMEMORY_USE_MPI)
+int
+omnitrace_mpi_copy(MPI_Comm, int, void*, void*, void*, int*)
+{
+    return MPI_SUCCESS;
+}
+
+int
+omnitrace_mpi_fini(MPI_Comm, int, void*, void*)
+{
+    OMNITRACE_DEBUG("MPI Comm attribute finalize\n");
+    auto _blocked = get_sampling_signals();
+    if(!_blocked.empty())
+        tim::signals::block_signals(_blocked, tim::signals::sigmask_scope::process);
+    if(mpip_index != std::numeric_limits<uint64_t>::max())
+        comp::deactivate_mpip<mpip_bundle_t, project::omnitrace>(mpip_index);
+    if(is_root_process()) omnitrace_finalize_hidden();
+    return MPI_SUCCESS;
+}
+#endif
+
 // this ensures omnitrace_finalize is called before MPI_Finalize
 void
 omnitrace_mpi_set_attr()
 {
 #if defined(TIMEMORY_USE_MPI)
-    static auto _mpi_copy = [](MPI_Comm, int, void*, void*, void*, int*) {
-        return MPI_SUCCESS;
-    };
-    static auto _mpi_fini = [](MPI_Comm, int, void*, void*) {
-        OMNITRACE_DEBUG("MPI Comm attribute finalize\n");
-        auto _blocked = get_sampling_signals();
-        if(!_blocked.empty())
-            tim::signals::block_signals(_blocked, tim::signals::sigmask_scope::process);
-        if(mpip_index != std::numeric_limits<uint64_t>::max())
-            comp::deactivate_mpip<mpip_bundle_t, project::omnitrace>(mpip_index);
-        omnitrace_finalize_hidden();
-        return MPI_SUCCESS;
-    };
-    using copy_func_t = int (*)(MPI_Comm, int, void*, void*, void*, int*);
-    using fini_func_t = int (*)(MPI_Comm, int, void*, void*);
-    int _comm_key     = -1;
-    if(PMPI_Comm_create_keyval(static_cast<copy_func_t>(_mpi_copy),
-                               static_cast<fini_func_t>(_mpi_fini), &_comm_key,
+    int _comm_key = -1;
+    if(PMPI_Comm_create_keyval(&omnitrace_mpi_copy, &omnitrace_mpi_fini, &_comm_key,
                                nullptr) == MPI_SUCCESS)
         PMPI_Comm_set_attr(MPI_COMM_SELF, _comm_key, nullptr);
 #endif
@@ -258,7 +263,7 @@ mpi_gotcha::audit(const gotcha_data_t& _data, audit::incoming)
     tim::mpi::is_initialized_callback() = []() { return false; };
     tim::mpi::is_finalized()            = true;
 #else
-    omnitrace_finalize_hidden();
+    if(is_root_process()) omnitrace_finalize_hidden();
 #endif
 }
 
diff --git a/source/lib/omnitrace/library/components/pthread_gotcha.cpp b/source/lib/omnitrace/library/components/pthread_gotcha.cpp
index ba014bc369..72de8334ea 100644
--- a/source/lib/omnitrace/library/components/pthread_gotcha.cpp
+++ b/source/lib/omnitrace/library/components/pthread_gotcha.cpp
@@ -27,7 +27,6 @@
 #include "library/components/pthread_create_gotcha.hpp"
 #include "library/components/pthread_mutex_gotcha.hpp"
 #include "library/runtime.hpp"
-#include "library/sampling.hpp"
 #include "library/thread_data.hpp"
 
 #include <timemory/backends/threading.hpp>
diff --git a/source/lib/omnitrace/library/components/pthread_mutex_gotcha.cpp b/source/lib/omnitrace/library/components/pthread_mutex_gotcha.cpp
index e4732e4915..ada298833a 100644
--- a/source/lib/omnitrace/library/components/pthread_mutex_gotcha.cpp
+++ b/source/lib/omnitrace/library/components/pthread_mutex_gotcha.cpp
@@ -27,7 +27,6 @@
 #include "library/components/category_region.hpp"
 #include "library/critical_trace.hpp"
 #include "library/runtime.hpp"
-#include "library/sampling.hpp"
 #include "library/thread_info.hpp"
 
 #include <timemory/backends/threading.hpp>
diff --git a/source/lib/omnitrace/library/components/roctracer.cpp b/source/lib/omnitrace/library/components/roctracer.cpp
index c169eccf19..d72aebd165 100644
--- a/source/lib/omnitrace/library/components/roctracer.cpp
+++ b/source/lib/omnitrace/library/components/roctracer.cpp
@@ -29,7 +29,6 @@
 #include "core/redirect.hpp"
 #include "library/roctracer.hpp"
 #include "library/runtime.hpp"
-#include "library/sampling.hpp"
 #include "library/thread_data.hpp"
 
 namespace omnitrace
diff --git a/source/lib/omnitrace/library/coverage.cpp b/source/lib/omnitrace/library/coverage.cpp
index 70f46fc607..bb42c8b683 100644
--- a/source/lib/omnitrace/library/coverage.cpp
+++ b/source/lib/omnitrace/library/coverage.cpp
@@ -337,7 +337,7 @@ omnitrace_register_coverage_hidden(const char* file, const char* func, size_t ad
     if(omnitrace::get_state() < omnitrace::State::Active &&
        !omnitrace_init_tooling_hidden())
         return;
-    else if(omnitrace::get_state() == omnitrace::State::Finalized)
+    else if(omnitrace::get_state() >= omnitrace::State::Finalized)
         return;
 
     OMNITRACE_BASIC_VERBOSE_F(3, "[0x%x] %-20s :: %20s\n", (unsigned int) address, func,
diff --git a/source/lib/omnitrace/library/ompt.cpp b/source/lib/omnitrace/library/ompt.cpp
index d8e3eb6b48..26db750c77 100644
--- a/source/lib/omnitrace/library/ompt.cpp
+++ b/source/lib/omnitrace/library/ompt.cpp
@@ -35,6 +35,7 @@
 
 #    include <timemory/components/ompt.hpp>
 #    include <timemory/components/ompt/extern.hpp>
+#    include <timemory/mpl/type_traits.hpp>
 #    include <timemory/timemory.hpp>
 
 #    include <memory>
@@ -85,6 +86,7 @@ shutdown()
     _protect = true;
     if(f_bundle)
     {
+        if(tim::manager::instance()) tim::manager::instance()->cleanup("omnitrace-ompt");
         f_bundle->stop();
         ompt_context_t::cleanup();
         trait::runtime_enabled<ompt_toolset_t>::set(false);
diff --git a/source/lib/omnitrace/library/process_sampler.cpp b/source/lib/omnitrace/library/process_sampler.cpp
index 38032526a1..29bd4efcfd 100644
--- a/source/lib/omnitrace/library/process_sampler.cpp
+++ b/source/lib/omnitrace/library/process_sampler.cpp
@@ -26,7 +26,6 @@
 #include "library/cpu_freq.hpp"
 #include "library/rocm_smi.hpp"
 #include "library/runtime.hpp"
-#include "library/sampling.hpp"
 
 #include <memory>
 #include <vector>
@@ -95,11 +94,11 @@ sampler::poll(std::atomic<State>* _state, nsec_t _interval, promise_t* _ready)
     auto _now = std::chrono::steady_clock::now();
     auto _end =
         _now + std::chrono::nanoseconds{ static_cast<uint64_t>(_duration * units::sec) };
-    while(_state && _state->load() != State::Finalized && get_state() != State::Finalized)
+    while(_state && _state->load() < State::Finalized && get_state() < State::Finalized)
     {
         std::this_thread::sleep_until(_now);
         if(_state->load() != State::Active) continue;
-        if(get_state() == State::Finalized) break;
+        if(get_state() >= State::Finalized) break;
         if(get_state() != State::Active) continue;
         get_sampler_is_sampling().store(true);
         for(auto& itr : instances)
@@ -112,7 +111,7 @@ sampler::poll(std::atomic<State>* _state, nsec_t _interval, promise_t* _ready)
     // ensure this is always false
     get_sampler_is_sampling().store(false);
 
-    if(_has_duration && _now >= _end && get_state() != State::Finalized)
+    if(_has_duration && _now >= _end && get_state() < State::Finalized)
     {
         OMNITRACE_VERBOSE(
             1,
diff --git a/source/lib/omnitrace/library/ptl.cpp b/source/lib/omnitrace/library/ptl.cpp
index 44683cfd31..46a864e190 100644
--- a/source/lib/omnitrace/library/ptl.cpp
+++ b/source/lib/omnitrace/library/ptl.cpp
@@ -96,9 +96,9 @@ get_thread_pool_state()
 PTL::ThreadPool&
 get_thread_pool()
 {
-    static auto _v =
-        (get_thread_pool_state() = State::Active, PTL::ThreadPool{ _thread_pool_cfg() });
-    return _v;
+    static auto* _v = (get_thread_pool_state() = State::Active,
+                       new PTL::ThreadPool{ _thread_pool_cfg() });
+    return *_v;
 }
 }  // namespace
 
diff --git a/source/lib/omnitrace/library/rocm.cpp b/source/lib/omnitrace/library/rocm.cpp
index ae82a97b01..3e6b9683cc 100644
--- a/source/lib/omnitrace/library/rocm.cpp
+++ b/source/lib/omnitrace/library/rocm.cpp
@@ -32,7 +32,6 @@
 #include "library/rocprofiler.hpp"
 #include "library/roctracer.hpp"
 #include "library/runtime.hpp"
-#include "library/sampling.hpp"
 #include "library/thread_data.hpp"
 #include "library/tracing.hpp"
 
diff --git a/source/lib/omnitrace/library/rocm_smi.cpp b/source/lib/omnitrace/library/rocm_smi.cpp
index f851af6403..44418a679c 100644
--- a/source/lib/omnitrace/library/rocm_smi.cpp
+++ b/source/lib/omnitrace/library/rocm_smi.cpp
@@ -232,8 +232,6 @@ data::shutdown()
 void
 data::post_process(uint32_t _dev_id)
 {
-    OMNITRACE_VERBOSE(1, "Post-processing rocm-smi data for device %u\n", _dev_id);
-
     using component::sampling_gpu_busy;
     using component::sampling_gpu_memory;
     using component::sampling_gpu_power;
@@ -245,6 +243,9 @@ data::post_process(uint32_t _dev_id)
     auto        _rocm_smi   = (_rocm_smi_v) ? *_rocm_smi_v : std::deque<rocm_smi::data>{};
     const auto& _thread_info = thread_info::get(0, InternalTID);
 
+    OMNITRACE_VERBOSE(1, "Post-processing %zu rocm-smi samples from device %u\n",
+                      _rocm_smi.size(), _dev_id);
+
     OMNITRACE_CI_THROW(!_thread_info, "Missing thread info for thread 0");
     if(!_thread_info) return;
 
diff --git a/source/lib/omnitrace/library/runtime.cpp b/source/lib/omnitrace/library/runtime.cpp
index ac0bc793fe..9ef575977a 100644
--- a/source/lib/omnitrace/library/runtime.cpp
+++ b/source/lib/omnitrace/library/runtime.cpp
@@ -35,6 +35,7 @@
 #include <timemory/backends/threading.hpp>
 #include <timemory/components/rusage/backends.hpp>
 #include <timemory/environment.hpp>
+#include <timemory/process/process.hpp>
 #include <timemory/sampling/allocator.hpp>
 #include <timemory/settings.hpp>
 #include <timemory/settings/types.hpp>
@@ -55,6 +56,8 @@ namespace omnitrace
 {
 namespace
 {
+auto root_process_id = get_env<pid_t>("OMNITRACE_ROOT_PROCESS", process::get_id(), false);
+
 auto&
 get_sampling_on_child_threads_history(int64_t _idx = utility::get_thread_index())
 {
@@ -259,4 +262,22 @@ set_sampling_on_all_future_threads(bool _v)
     for(size_t i = 0; i < max_supported_threads; ++i)
         get_sampling_on_child_threads_history(i).emplace_back(_v);
 }
+
+pid_t
+get_root_process_id()
+{
+    return root_process_id;
+}
+
+bool
+is_root_process()
+{
+    return (root_process_id == process::get_id());
+}
+
+bool
+is_child_process()
+{
+    return (root_process_id != process::get_id());
+}
 }  // namespace omnitrace
diff --git a/source/lib/omnitrace/library/runtime.hpp b/source/lib/omnitrace/library/runtime.hpp
index 0677e1b4da..ec3bf9c304 100644
--- a/source/lib/omnitrace/library/runtime.hpp
+++ b/source/lib/omnitrace/library/runtime.hpp
@@ -121,6 +121,15 @@ struct scoped_child_sampling
     scoped_child_sampling(bool _v) { push_enable_sampling_on_child_threads(_v); }
     ~scoped_child_sampling() { pop_enable_sampling_on_child_threads(); }
 };
+
+pid_t
+get_root_process_id();
+
+bool
+is_root_process();
+
+bool
+is_child_process();
 }  // namespace omnitrace
 
 #define OMNITRACE_SCOPED_SAMPLING_ON_CHILD_THREADS(VALUE)                                \
diff --git a/source/lib/omnitrace/library/sampling.cpp b/source/lib/omnitrace/library/sampling.cpp
index 449e506cc3..5ad997b9c1 100644
--- a/source/lib/omnitrace/library/sampling.cpp
+++ b/source/lib/omnitrace/library/sampling.cpp
@@ -333,7 +333,7 @@ start_duration_thread()
                 if(!_lk.owns_lock()) _lk.lock();
                 get_duration_cv().wait_until(_lk, _end);
                 auto _premature = (std::chrono::steady_clock::now() < _end);
-                auto _finalized = (get_state() == State::Finalized);
+                auto _finalized = (get_state() >= State::Finalized);
                 if(_premature && !_finalized)
                 {
                     // protect against spurious wakeups
@@ -480,6 +480,7 @@ configure(bool _setup, int64_t _tid)
                                 "causal profiling is enabled");
 
     OMNITRACE_SCOPED_SAMPLING_ON_CHILD_THREADS(false);
+    OMNITRACE_SCOPED_THREAD_STATE(ThreadState::Internal);
 
     auto&& _cpu_tids  = get_sampling_cpu_tids();
     auto&& _real_tids = get_sampling_real_tids();
@@ -673,6 +674,13 @@ setup()
 std::set<int>
 shutdown()
 {
+    if(is_child_process())
+    {
+        for(auto& itr : sampler_instances::instances())
+            itr.release();
+        return std::set<int>{};
+    }
+
     auto _v = configure(false);
     if(utility::get_thread_index() == 0) stop_duration_thread();
     return _v;
diff --git a/source/lib/omnitrace/library/thread_data.hpp b/source/lib/omnitrace/library/thread_data.hpp
index ecedc69b45..420269c313 100644
--- a/source/lib/omnitrace/library/thread_data.hpp
+++ b/source/lib/omnitrace/library/thread_data.hpp
@@ -22,7 +22,6 @@
 
 #pragma once
 
-#include "api.hpp"
 #include "core/common.hpp"
 #include "core/concepts.hpp"
 #include "core/config.hpp"
diff --git a/source/lib/omnitrace/library/thread_deleter.cpp b/source/lib/omnitrace/library/thread_deleter.cpp
index 7ecbdba74a..d90849fcd8 100644
--- a/source/lib/omnitrace/library/thread_deleter.cpp
+++ b/source/lib/omnitrace/library/thread_deleter.cpp
@@ -21,6 +21,7 @@
 // SOFTWARE.
 
 #include "library/thread_deleter.hpp"
+#include "api.hpp"
 #include "core/utility.hpp"
 #include "library/components/pthread_create_gotcha.hpp"
 #include "library/thread_info.hpp"
@@ -35,10 +36,22 @@ template struct component_bundle_cache<instrumentation_bundle_t>;
 void
 thread_deleter<void>::operator()() const
 {
-    component::pthread_create_gotcha::shutdown(threading::get_id());
-    set_thread_state(ThreadState::Completed);
-    if(get_state() != State::Finalized && threading::get_id() == 0)
-        omnitrace_finalize_hidden();
+    // called after thread info is deleted
+    if(!thread_info::exists()) return;
+
+    const auto& _info = thread_info::get();
+    if(_info && _info->index_data)
+    {
+        auto _tid = _info->index_data->sequent_value;
+
+        component::pthread_create_gotcha::shutdown(_tid);
+        set_thread_state(ThreadState::Completed);
+        if(get_state() < State::Finalized && _tid == 0) omnitrace_finalize_hidden();
+    }
+    else
+    {
+        set_thread_state(ThreadState::Completed);
+    }
 }
 
 template struct thread_deleter<void>;
diff --git a/source/lib/omnitrace/library/thread_info.cpp b/source/lib/omnitrace/library/thread_info.cpp
index 7841171488..6512d94a01 100644
--- a/source/lib/omnitrace/library/thread_info.cpp
+++ b/source/lib/omnitrace/library/thread_info.cpp
@@ -78,7 +78,7 @@ init_index_data(int64_t _tid, bool _offset = false)
         itr       = thread_index_data{};
         int _verb = 2;
         // if thread created using finalization, bump up the minimum verbosity level
-        if(get_state() == State::Finalized && _offset) _verb += 2;
+        if(get_state() >= State::Finalized && _offset) _verb += 2;
         if(!config::settings_are_configured())
         {
             OMNITRACE_BASIC_VERBOSE_F(
@@ -145,6 +145,12 @@ grow_data(int64_t _tid)
     return _max_threads;
 }
 
+bool
+thread_info::exists()
+{
+    return (get_info_data() != nullptr);
+}
+
 const std::optional<thread_info>&
 thread_info::init(bool _offset)
 {
@@ -177,6 +183,11 @@ thread_info::init(bool _offset)
 const std::optional<thread_info>&
 thread_info::get()
 {
+    if(!exists())
+    {
+        static thread_local auto _v = std::optional<thread_info>{};
+        return _v;
+    }
     return get_info_data(utility::get_thread_index());
 }
 
diff --git a/source/lib/omnitrace/library/thread_info.hpp b/source/lib/omnitrace/library/thread_info.hpp
index 8469df505e..e9c0a3f326 100644
--- a/source/lib/omnitrace/library/thread_info.hpp
+++ b/source/lib/omnitrace/library/thread_info.hpp
@@ -95,6 +95,7 @@ struct thread_info
 
     std::string as_string() const;
 
+    static bool                              exists();
     static const std::optional<thread_info>& init(bool _offset = false);
     static const std::optional<thread_info>& get();
     static const std::optional<thread_info>& get(int64_t _tid, ThreadIdType _type);
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 7dff069cf5..715e639bb7 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -810,6 +810,13 @@ omnitrace_add_validation_test(
          0
          -p)
 
+omnitrace_add_test(
+    NAME fork
+    TARGET fork-example
+    REWRITE_ARGS -e -v 2 --print-instrumented modules -i 16
+    RUNTIME_ARGS -e -v 1 --label file -i 16
+    ENVIRONMENT "${_base_environment};OMNITRACE_CRITICAL_TRACE=ON")
+
 # -------------------------------------------------------------------------------------- #
 #
 # critical-trace tests
@@ -846,7 +853,6 @@ set(_parallel_overhead_critical_trace_environ
     "OMNITRACE_OUTPUT_PREFIX=parallel-overhead-critical-trace/"
     "OMNITRACE_CRITICAL_TRACE_DEBUG=ON"
     "OMNITRACE_VERBOSE=4"
-    "OMNITRACE_DEBUG=ON"
     "OMNITRACE_USE_PID=OFF"
     "OMNITRACE_TIME_OUTPUT=OFF")