diff --git a/.github/workflows/inshpect.yml b/.github/workflows/inshpect.yml
new file mode 100644
index 0000000000..9d4387f4b8
--- /dev/null
+++ b/.github/workflows/inshpect.yml
@@ -0,0 +1,30 @@
+name: Inshpect
+
+on:
+  pull_request:
+
+jobs:
+  inshpect:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Install tools
+        run: |
+          sudo apt update
+          sudo apt install -y fd-find ripgrep dasel
+
+          pushd /tmp
+          curl --output inshpect \
+            --location https://raw.githubusercontent.com/msimberg/inshpect/d31ba8ed7e636c57cb9085ed3a8450d998e6b8a5/inshpect
+          echo "4b891a178fa99ac800b9b92acc0596d72f640ca1ae6de7f35ac4b3785dcd7af9 inshpect" | sha256sum --check
+          chmod +x inshpect
+          mv inshpect /usr/local/bin/inshpect
+          popd
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Inshpect
+        if: always()
+        run: |
+          export INSHPECT_NUMTHREADS=4
+          export INSHPECT_FD=fdfind
+          inshpect . .inshpect.toml
diff --git a/.inshpect.toml b/.inshpect.toml
new file mode 100644
index 0000000000..a29b864060
--- /dev/null
+++ b/.inshpect.toml
@@ -0,0 +1,176 @@
+#
+# Distributed Linear Algebra with Future (DLAF)
+#
+# Copyright (c) 2018-2023, ETH Zurich
+# All rights reserved.
+#
+# Please, refer to the LICENSE file in the root directory.
+# SPDX-License-Identifier: BSD-3-Clause
+#
+
+extensions = ["c", "cpp", "cu", "h", "h.in"]
+
+[includes]
+enable = true
+patterns = [
+    { pattern = '\bstd::make_shared\b', include = 'memory' },
+    { pattern = '\bstd::map\b', include = 'map' },
+    { pattern = '\bstd::set\b', include = 'set' },
+    { pattern = '\bstd::multi_map\b', include = 'multi_map' },
+    { pattern = '\bstd::multi_set\b', include = 'multi_set' },
+    { pattern = '\bstd::unordered_map\b', include = 'unordered_map' },
+    { pattern = '\bstd::unordered_set\b', include = 'unordered_set' },
+    { pattern = '\bstd::unordered_multi_map\b', include = 'unordered_map' },
+    { pattern = '\bstd::unordered_multi_set\b', include = 'unordered_set' },
+    { pattern = '\bstd::((shared|unique)_ptr)\b', include = 'memory' },
+    { pattern = '\bstd::addressof\b', include = 'memory' },
+    { pattern = '\bstd::list\b', include = 'list' },
+    { pattern = '\bstd::string\b', include = 'string' },
+    { pattern = '\bstd::vector\b', include = 'vector' },
+    { pattern = '\bstd::array\b', include = 'array' },
+    { pattern = '\bstd::reference_wrapper\b', include = 'functional' },
+    { pattern = '\bstd::(c?ref)\b', include = 'functional' },
+    { pattern = '\bstd::(u?int[0-9]+_t)\b', include = 'cstdint' },
+    { pattern = '\bstd::((i|o)?stringstream)\b', include = 'sstream' },
+    { pattern = '\bstd::((i|o)?fstream)\b', include = 'fstream' },
+    { pattern = '\bstd::(cin|cout|cerr|clog)\b', include = 'iostream' },
+    { pattern = '\bstd::setw\b', include = 'iomanip' },
+    { pattern = '\bstd::setprecision\b', include = 'iomanip' },
+    { pattern = '\bstd::size_t\b', include = 'cstddef' },
+    { pattern = '\bstd::ptrdiff_t\b', include = 'cstddef' },
+    { pattern = '\bstd::nullptr_t\b', include = 'cstddef' },
+    { pattern = '\bstd::(true|false)_type\b', include = 'type_traits' },
+    { pattern = '\bstd::(bool|integral)_constant\b', include = 'type_traits' },
+    { pattern = '\bstd::invoke_result\b', include = 'type_traits' },
+    { pattern = '\bstd::invoke_result_t\b', include = 'type_traits' },
+    { pattern = '\bstd::is_void\b', include = 'type_traits' },
+    { pattern = '\bstd::is_void_t\b', include = 'type_traits' },
+    { pattern = '\bstd::is_null_pointer\b', include = 'type_traits' },
+    { pattern = '\bstd::is_integral\b', include = 'type_traits' },
+    { pattern = '\bstd::is_floating_point\b', include = 'type_traits' },
+    { pattern = '\bstd::is_array\b', include = 'type_traits' },
+    { pattern = '\bstd::is_enum\b', include = 'type_traits' },
+    { pattern = '\bstd::is_union\b', include = 'type_traits' },
+    { pattern = '\bstd::is_class\b', include = 'type_traits' },
+    { pattern = '\bstd::is_function\b', include = 'type_traits' },
+    { pattern = '\bstd::is_pointer\b', include = 'type_traits' },
+    { pattern = '\bstd::is_lvalue_reference\b', include = 'type_traits' },
+    { pattern = '\bstd::is_rvalue_reference\b', include = 'type_traits' },
+    { pattern = '\bstd::is_member_object_pointer\b', include = 'type_traits' },
+    { pattern = '\bstd::is_member_function_pointer\b', include = 'type_traits' },
+    { pattern = '\bstd::is_fundamental\b', include = 'type_traits' },
+    { pattern = '\bstd::is_arithmetic\b', include = 'type_traits' },
+    { pattern = '\bstd::is_scalar\b', include = 'type_traits' },
+    { pattern = '\bstd::is_object\b', include = 'type_traits' },
+    { pattern = '\bstd::is_compound\b', include = 'type_traits' },
+    { pattern = '\bstd::is_reference\b', include = 'type_traits' },
+    { pattern = '\bstd::is_member_pointer\b', include = 'type_traits' },
+    { pattern = '\bstd::is_const\b', include = 'type_traits' },
+    { pattern = '\bstd::is_volatile\b', include = 'type_traits' },
+    { pattern = '\bstd::is_trivial\b', include = 'type_traits' },
+    { pattern = '\bstd::is_trivially_copyable\b', include = 'type_traits' },
+    { pattern = '\bstd::is_standard_layout\b', include = 'type_traits' },
+    { pattern = '\bstd::is_pod\b', include = 'type_traits' },
+    { pattern = '\bstd::is_literal_type\b', include = 'type_traits' },
+    { pattern = '\bstd::is_empty\b', include = 'type_traits' },
+    { pattern = '\bstd::is_polymorphic\b', include = 'type_traits' },
+    { pattern = '\bstd::is_abstract\b', include = 'type_traits' },
+    { pattern = '\bstd::is_signed\b', include = 'type_traits' },
+    { pattern = '\bstd::is_unsigned\b', include = 'type_traits' },
+    { pattern = '\bstd::is_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_trivially_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_nothrow_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_default_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_trivially_default_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_nothrow_default_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_copy_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_trivially_copy_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_nothrow_copy_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_move_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_trivially_move_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_nothrow_move_constructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_assignable\b', include = 'type_traits' },
+    { pattern = '\bstd::is_trivially_assignable\b', include = 'type_traits' },
+    { pattern = '\bstd::is_nothrow_assignable\b', include = 'type_traits' },
+    { pattern = '\bstd::is_copy_assignable\b', include = 'type_traits' },
+    { pattern = '\bstd::is_trivially_copy_assignable\b', include = 'type_traits' },
+    { pattern = '\bstd::is_nothrow_copy_assignable\b', include = 'type_traits' },
+    { pattern = '\bstd::is_move_assignable\b', include = 'type_traits' },
+    { pattern = '\bstd::is_trivially_move_assignable\b', include = 'type_traits' },
+    { pattern = '\bstd::is_nothrow_move_assignable\b', include = 'type_traits' },
+    { pattern = '\bstd::is_destructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_trivially_destructible\b', include = 'type_traits' },
+    { pattern = '\bstd::is_same\b', include = 'type_traits' },
+    { pattern = '\bstd::is_base_of\b', include = 'type_traits' },
+    { pattern = '\bstd::is_convertible\b', include = 'type_traits' },
+    { pattern = '\bstd::has_virtual_destructor\b', include = 'type_traits' },
+    { pattern = '\bstd::alignment_of\b', include = 'type_traits' },
+    { pattern = '\bstd::aligned_storage\b', include = 'type_traits' },
+    { pattern = '\bstd::aligned_union\b', include = 'type_traits' },
+    { pattern = '\bstd::rank\b', include = 'type_traits' },
+    { pattern = '\bstd::extent\b', include = 'type_traits' },
+    { pattern = '\bstd::add_(cv|const|pointer|volatile|lvalue_reference|rvalue_reference)\b', include = 'type_traits' },
+    { pattern = '\bstd::make_signed\b', include = 'type_traits' },
+    { pattern = '\bstd::make_unsigned\b', include = 'type_traits' },
+    { pattern = '\bstd::remove_(cv|const|volatile|reference|pointer|extent|all_extents)\b', include = 'type_traits' },
+    { pattern = '\bstd::decay\b', include = 'type_traits' },
+    { pattern = '\bstd::enable_if\b', include = 'type_traits' },
+    { pattern = '\bstd::conditional\b', include = 'type_traits' },
+    { pattern = '\bstd::common_type\b', include = 'type_traits' },
+    { pattern = '\bstd::underlying_type\b', include = 'type_traits' },
+    { pattern = '\bstd::result_of\b', include = 'type_traits' },
+    { pattern = '\bstd::is_invocable([a-z_]*)\b', include = 'type_traits' },
+    { pattern = '\btypeid\b', include = 'typeinfo' },
+    { pattern = '\bstd::type_info\b', include = 'typeinfo' },
+    { pattern = '\bstd::(mem((set)|(cpy)|(move)))\b', include = 'cstring' },
+    { pattern = '\bstd::swap\b', include = 'utility' },
+    { pattern = '\bstd::move\b', include = 'utility' },
+    { pattern = '\bstd::forward\b', include = 'utility' },
+    { pattern = '\bstd::declval\b', include = 'utility' },
+    { pattern = '\bstd::pair\b', include = 'utility' },
+    { pattern = '\bstd::make_pair\b', include = 'utility' },
+    { pattern = '\bstd::in_place\b', include = 'utility' },
+    { pattern = '\bstd::exception_ptr\b', include = 'exception' },
+    { pattern = '\bstd::terminate\b', include = 'exception' },
+    { pattern = '\bstd::(s|g)et_terminate\b', include = 'exception' },
+    { pattern = '\bstd::swap_ranges\b', include = 'algorithm' },
+    { pattern = '\bstd::iter_swap\b', include = 'algorithm' },
+    { pattern = '\bstd::atomic\b', include = 'atomic' },
+    { pattern = '\bstd::(memory_order_[a-z_]+)\b', include = 'atomic' },
+    { pattern = '\bstd::(minstd_rand(0)?)\b', include = 'random' },
+    { pattern = '\bstd::(mt19937(_64)?)\b', include = 'random' },
+    { pattern = '\bstd::random_device\b', include = 'random' },
+    { pattern = '\bstd::uniform_(int|uniform)_distribution\b', include = 'random' },
+    { pattern = '\bstd::apply\b', include = 'tuple' },
+    { pattern = '\bstd::(forward_as_|make_|)tuple\b', include = 'tuple' },
+    { pattern = '\bstd::ignore\b', include = 'tuple' },
+    { pattern = '\bstd::tie\b', include = 'tuple' },
+    { pattern = '\bstd::tuple_(cat|element|size)\b', include = 'tuple' },
+    { pattern = '\bstd::(make_)optional\b', include = 'optional' },
+    { pattern = '\bstd::nullopt\b', include = 'optional' },
+    { pattern = '\bstd::any(|_cast)\b', include = 'any' },
+    { pattern = '\bstd::error_(code|condition|category)\b', include = 'system_error' },
+    { pattern = '\bstd::system_error\b', include = 'system_error' },
+    { pattern = '\bstd::chrono\b', include = 'chrono' },
+]
+
+[copyright]
+enable = false
+
+[license]
+enable = false
+
+[spdx]
+enable = false
+
+[pragma_once]
+enable = false
+
+[deprecated_includes]
+enable = false
+
+[deprecated_names]
+enable = false
+
+[disallowed_macros]
+enable = false
diff --git a/include/dlaf/auxiliary/norm/mc.h b/include/dlaf/auxiliary/norm/mc.h
index cf887f0d2c..b4c60286ef 100644
--- a/include/dlaf/auxiliary/norm/mc.h
+++ b/include/dlaf/auxiliary/norm/mc.h
@@ -9,6 +9,9 @@
 //
 #pragma once
 
+#include <utility>
+#include <vector>
+
 #include <dlaf/auxiliary/norm/api.h>
 #include <dlaf/common/range2d.h>
 #include <dlaf/common/vector.h>
diff --git a/include/dlaf/blas/tile.h b/include/dlaf/blas/tile.h
index c9098f788e..d534f93d5d 100644
--- a/include/dlaf/blas/tile.h
+++ b/include/dlaf/blas/tile.h
@@ -13,6 +13,7 @@
 /// Provides `Tile` wrappers for BLAS operations.
 
 #include <cstddef>
+#include <utility>
 
 #include <blas.hh>
 
diff --git a/include/dlaf/common/assert.h b/include/dlaf/common/assert.h
index a07e011010..5244f73303 100644
--- a/include/dlaf/common/assert.h
+++ b/include/dlaf/common/assert.h
@@ -14,6 +14,8 @@
 #include <iostream>
 #include <memory>
 #include <sstream>
+#include <string>
+#include <utility>
 
 #include <dlaf/common/source_location.h>
 
diff --git a/include/dlaf/common/callable_object.h b/include/dlaf/common/callable_object.h
index b167e8a63e..7b253d34ae 100644
--- a/include/dlaf/common/callable_object.h
+++ b/include/dlaf/common/callable_object.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <utility>
+
 /// @file
 
 /// Given a function name @fname generates a constexpr object with name fname_o
@@ -23,6 +25,7 @@
 /// The function name is wrapped in parentheses to disable ADL. It is assumed
 /// that all required overloads are found relative to the namespace where the
 /// macro is used.
+
 #define DLAF_MAKE_CALLABLE_OBJECT(fname)                                                    \
   constexpr struct fname##_t {                                                              \
     template <typename... Ts>                                                               \
diff --git a/include/dlaf/common/data_descriptor.h b/include/dlaf/common/data_descriptor.h
index 87cfddc7cb..951a9c3d11 100644
--- a/include/dlaf/common/data_descriptor.h
+++ b/include/dlaf/common/data_descriptor.h
@@ -13,6 +13,7 @@
 #pragma once
 
 #include <algorithm>
+#include <cstddef>
 #include <memory>
 #include <type_traits>
 #include <utility>
diff --git a/include/dlaf/common/index2d.h b/include/dlaf/common/index2d.h
index 59bf55efd2..42677666ba 100644
--- a/include/dlaf/common/index2d.h
+++ b/include/dlaf/common/index2d.h
@@ -17,6 +17,7 @@
 #include <ostream>
 #include <string>
 #include <type_traits>
+#include <utility>
 
 #include <dlaf/common/assert.h>
 #include <dlaf/types.h>
diff --git a/include/dlaf/common/pipeline.h b/include/dlaf/common/pipeline.h
index b325982462..57a99dbcb5 100644
--- a/include/dlaf/common/pipeline.h
+++ b/include/dlaf/common/pipeline.h
@@ -12,6 +12,7 @@
 
 /// @file
 
+#include <optional>
 #include <utility>
 
 #include <pika/async_rw_mutex.hpp>
diff --git a/include/dlaf/common/round_robin.h b/include/dlaf/common/round_robin.h
index d02924bf1e..4a430036a8 100644
--- a/include/dlaf/common/round_robin.h
+++ b/include/dlaf/common/round_robin.h
@@ -11,6 +11,7 @@
 
 /// @file
 
+#include <cstddef>
 #include <vector>
 
 #include <dlaf/common/assert.h>
diff --git a/include/dlaf/communication/broadcast_panel.h b/include/dlaf/communication/broadcast_panel.h
index ce465c49e2..e83e35a98d 100644
--- a/include/dlaf/communication/broadcast_panel.h
+++ b/include/dlaf/communication/broadcast_panel.h
@@ -13,6 +13,7 @@
 /// @file
 
 #include <type_traits>
+#include <utility>
 
 #include <pika/execution.hpp>
 
diff --git a/include/dlaf/communication/communicator_grid.h b/include/dlaf/communication/communicator_grid.h
index 6f6f033cb1..4099fb5945 100644
--- a/include/dlaf/communication/communicator_grid.h
+++ b/include/dlaf/communication/communicator_grid.h
@@ -13,6 +13,7 @@
 #pragma once
 
 #include <array>
+#include <cstddef>
 
 #include <dlaf/common/index2d.h>
 #include <dlaf/communication/communicator.h>
diff --git a/include/dlaf/communication/communicator_pipeline.h b/include/dlaf/communication/communicator_pipeline.h
index 11e527797c..80d8c66de4 100644
--- a/include/dlaf/communication/communicator_pipeline.h
+++ b/include/dlaf/communication/communicator_pipeline.h
@@ -12,6 +12,8 @@
 
 #pragma once
 
+#include <utility>
+
 #include <dlaf/common/index2d.h>
 #include <dlaf/common/pipeline.h>
 #include <dlaf/common/round_robin.h>
diff --git a/include/dlaf/communication/kernels/internal/all_reduce.h b/include/dlaf/communication/kernels/internal/all_reduce.h
index 78d91fed2d..6ba6be7693 100644
--- a/include/dlaf/communication/kernels/internal/all_reduce.h
+++ b/include/dlaf/communication/kernels/internal/all_reduce.h
@@ -11,6 +11,7 @@
 #pragma once
 
 #include <complex>
+#include <functional>
 #include <utility>
 
 #include <mpi.h>
diff --git a/include/dlaf/communication/kernels/internal/broadcast.h b/include/dlaf/communication/kernels/internal/broadcast.h
index bc615d549f..aab56c910a 100644
--- a/include/dlaf/communication/kernels/internal/broadcast.h
+++ b/include/dlaf/communication/kernels/internal/broadcast.h
@@ -11,6 +11,7 @@
 #pragma once
 
 #include <complex>
+#include <functional>
 #include <utility>
 
 #include <mpi.h>
diff --git a/include/dlaf/communication/kernels/internal/p2p.h b/include/dlaf/communication/kernels/internal/p2p.h
index fb4d2bdae4..d2672b0d0d 100644
--- a/include/dlaf/communication/kernels/internal/p2p.h
+++ b/include/dlaf/communication/kernels/internal/p2p.h
@@ -13,6 +13,7 @@
 /// @file
 
 #include <complex>
+#include <functional>
 #include <utility>
 
 #include <mpi.h>
diff --git a/include/dlaf/communication/kernels/internal/reduce.h b/include/dlaf/communication/kernels/internal/reduce.h
index 76bd154560..bd8fa8d1f7 100644
--- a/include/dlaf/communication/kernels/internal/reduce.h
+++ b/include/dlaf/communication/kernels/internal/reduce.h
@@ -11,6 +11,7 @@
 #pragma once
 
 #include <complex>
+#include <functional>
 #include <utility>
 
 #include <mpi.h>
diff --git a/include/dlaf/communication/sync/basic.h b/include/dlaf/communication/sync/basic.h
index 4cf7e1b547..3ae5722502 100644
--- a/include/dlaf/communication/sync/basic.h
+++ b/include/dlaf/communication/sync/basic.h
@@ -12,6 +12,8 @@
 
 /// @file
 
+#include <utility>
+
 #include <dlaf/common/data.h>
 #include <dlaf/communication/communicator.h>
 #include <dlaf/communication/error.h>
diff --git a/include/dlaf/communication/sync/broadcast.h b/include/dlaf/communication/sync/broadcast.h
index dbcb3b4268..eb352d75e7 100644
--- a/include/dlaf/communication/sync/broadcast.h
+++ b/include/dlaf/communication/sync/broadcast.h
@@ -12,6 +12,8 @@
 
 /// @file
 
+#include <utility>
+
 #include <dlaf/common/assert.h>
 #include <dlaf/common/callable_object.h>
 #include <dlaf/common/data.h>
diff --git a/include/dlaf/eigensolver/band_to_tridiag/mc.h b/include/dlaf/eigensolver/band_to_tridiag/mc.h
index 86146d88d7..ca680884ac 100644
--- a/include/dlaf/eigensolver/band_to_tridiag/mc.h
+++ b/include/dlaf/eigensolver/band_to_tridiag/mc.h
@@ -11,7 +11,11 @@
 #pragma once
 
 #include <atomic>
+#include <cstddef>
+#include <memory>
 #include <sstream>
+#include <utility>
+#include <vector>
 
 #include <pika/execution.hpp>
 #include <pika/semaphore.hpp>
diff --git a/include/dlaf/eigensolver/bt_band_to_tridiag/impl.h b/include/dlaf/eigensolver/bt_band_to_tridiag/impl.h
index 864253c4b4..946668b8b2 100644
--- a/include/dlaf/eigensolver/bt_band_to_tridiag/impl.h
+++ b/include/dlaf/eigensolver/bt_band_to_tridiag/impl.h
@@ -10,8 +10,11 @@
 
 #pragma once
 
+#include <cstddef>
 #include <tuple>
 #include <type_traits>
+#include <utility>
+#include <vector>
 
 #include <pika/execution.hpp>
 #include <pika/thread.hpp>
diff --git a/include/dlaf/eigensolver/bt_reduction_to_band/impl.h b/include/dlaf/eigensolver/bt_reduction_to_band/impl.h
index 4bc42cb69a..d268ab7158 100644
--- a/include/dlaf/eigensolver/bt_reduction_to_band/impl.h
+++ b/include/dlaf/eigensolver/bt_reduction_to_band/impl.h
@@ -9,6 +9,9 @@
 //
 #pragma once
 
+#include <cstddef>
+#include <utility>
+
 #include <pika/execution.hpp>
 #include <pika/thread.hpp>
 
diff --git a/include/dlaf/eigensolver/eigensolver.h b/include/dlaf/eigensolver/eigensolver.h
index bddf4ddc6e..1167f86193 100644
--- a/include/dlaf/eigensolver/eigensolver.h
+++ b/include/dlaf/eigensolver/eigensolver.h
@@ -11,6 +11,8 @@
 
 /// @file
 
+#include <utility>
+
 #include <blas.hh>
 
 #include <dlaf/communication/communicator_grid.h>
diff --git a/include/dlaf/eigensolver/gen_eigensolver.h b/include/dlaf/eigensolver/gen_eigensolver.h
index 41842bc2ec..495018f7c3 100644
--- a/include/dlaf/eigensolver/gen_eigensolver.h
+++ b/include/dlaf/eigensolver/gen_eigensolver.h
@@ -11,6 +11,8 @@
 
 /// @file
 
+#include <utility>
+
 #include <blas.hh>
 
 #include <dlaf/eigensolver/gen_eigensolver/api.h>
diff --git a/include/dlaf/eigensolver/gen_to_std/impl.h b/include/dlaf/eigensolver/gen_to_std/impl.h
index df04ce281f..83a5e8ee0d 100644
--- a/include/dlaf/eigensolver/gen_to_std/impl.h
+++ b/include/dlaf/eigensolver/gen_to_std/impl.h
@@ -7,8 +7,12 @@
 // Please, refer to the LICENSE file in the root directory.
 // SPDX-License-Identifier: BSD-3-Clause
 //
+
 #pragma once
 
+#include <cstddef>
+#include <utility>
+
 #include <pika/execution.hpp>
 
 #include <dlaf/blas/tile.h>
diff --git a/include/dlaf/eigensolver/internal/get_red2band_panel_nworkers.h b/include/dlaf/eigensolver/internal/get_red2band_panel_nworkers.h
index 1bf112b10f..f269232d18 100644
--- a/include/dlaf/eigensolver/internal/get_red2band_panel_nworkers.h
+++ b/include/dlaf/eigensolver/internal/get_red2band_panel_nworkers.h
@@ -10,6 +10,7 @@
 #pragma once
 
 #include <algorithm>
+#include <cstddef>
 #include <cstdint>
 
 #include <pika/runtime.hpp>
diff --git a/include/dlaf/eigensolver/internal/get_tridiag_rank1_nworkers.h b/include/dlaf/eigensolver/internal/get_tridiag_rank1_nworkers.h
index a3ab3d7569..8ab4269bbe 100644
--- a/include/dlaf/eigensolver/internal/get_tridiag_rank1_nworkers.h
+++ b/include/dlaf/eigensolver/internal/get_tridiag_rank1_nworkers.h
@@ -10,6 +10,7 @@
 #pragma once
 
 #include <algorithm>
+#include <cstddef>
 #include <cstdint>
 
 #include <pika/runtime.hpp>
diff --git a/include/dlaf/eigensolver/reduction_to_band/impl.h b/include/dlaf/eigensolver/reduction_to_band/impl.h
index 148078d422..2a7882fd82 100644
--- a/include/dlaf/eigensolver/reduction_to_band/impl.h
+++ b/include/dlaf/eigensolver/reduction_to_band/impl.h
@@ -9,10 +9,12 @@
 //
 #pragma once
 
+#include <array>
 #include <atomic>
 #include <cmath>
 #include <cstddef>
 #include <sstream>
+#include <utility>
 #include <vector>
 
 #include <pika/barrier.hpp>
diff --git a/include/dlaf/eigensolver/tridiag_solver/impl.h b/include/dlaf/eigensolver/tridiag_solver/impl.h
index 0e1f7c4c94..47ed683421 100644
--- a/include/dlaf/eigensolver/tridiag_solver/impl.h
+++ b/include/dlaf/eigensolver/tridiag_solver/impl.h
@@ -12,6 +12,9 @@
 #include <algorithm>
 #include <atomic>
 #include <sstream>
+#include <tuple>
+#include <utility>
+#include <vector>
 
 #include <pika/execution.hpp>
 #include <pika/thread.hpp>
diff --git a/include/dlaf/eigensolver/tridiag_solver/index_manipulation.h b/include/dlaf/eigensolver/tridiag_solver/index_manipulation.h
index 7e6dc5912b..9c9e848865 100644
--- a/include/dlaf/eigensolver/tridiag_solver/index_manipulation.h
+++ b/include/dlaf/eigensolver/tridiag_solver/index_manipulation.h
@@ -10,6 +10,7 @@
 #pragma once
 
 #include <algorithm>
+#include <utility>
 
 #include <pika/execution.hpp>
 
diff --git a/include/dlaf/eigensolver/tridiag_solver/kernels_async.h b/include/dlaf/eigensolver/tridiag_solver/kernels_async.h
index 72bb3207eb..b6bbd2cf21 100644
--- a/include/dlaf/eigensolver/tridiag_solver/kernels_async.h
+++ b/include/dlaf/eigensolver/tridiag_solver/kernels_async.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <utility>
+
 #include <dlaf/eigensolver/tridiag_solver/kernels.h>
 #include <dlaf/lapack/tile.h>
 #include <dlaf/matrix/tile.h>
diff --git a/include/dlaf/eigensolver/tridiag_solver/merge.h b/include/dlaf/eigensolver/tridiag_solver/merge.h
index da75305c48..b6387a76d2 100644
--- a/include/dlaf/eigensolver/tridiag_solver/merge.h
+++ b/include/dlaf/eigensolver/tridiag_solver/merge.h
@@ -11,9 +11,12 @@
 
 #include <algorithm>
 #include <array>
+#include <cstddef>
 #include <functional>
 #include <numeric>
 #include <tuple>
+#include <utility>
+#include <vector>
 
 #include <pika/barrier.hpp>
 #include <pika/execution.hpp>
diff --git a/include/dlaf/eigensolver/tridiag_solver/rot.h b/include/dlaf/eigensolver/tridiag_solver/rot.h
index e2b8eb1b56..18adca31cb 100644
--- a/include/dlaf/eigensolver/tridiag_solver/rot.h
+++ b/include/dlaf/eigensolver/tridiag_solver/rot.h
@@ -13,6 +13,10 @@
 #include <whip.hpp>
 #endif
 
+#include <cstddef>
+#include <utility>
+#include <vector>
+
 #include <dlaf/common/assert.h>
 #include <dlaf/common/index2d.h>
 #include <dlaf/common/range2d.h>
diff --git a/include/dlaf/eigensolver/tridiag_solver/tile_collector.h b/include/dlaf/eigensolver/tridiag_solver/tile_collector.h
index 50a6cbc220..4cf19eb098 100644
--- a/include/dlaf/eigensolver/tridiag_solver/tile_collector.h
+++ b/include/dlaf/eigensolver/tridiag_solver/tile_collector.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+#include <utility>
+
 #include <dlaf/common/range2d.h>
 #include <dlaf/types.h>
 #include <dlaf/util_matrix.h>
diff --git a/include/dlaf/factorization/cholesky/impl.h b/include/dlaf/factorization/cholesky/impl.h
index 0834682d4e..4eb9280b5a 100644
--- a/include/dlaf/factorization/cholesky/impl.h
+++ b/include/dlaf/factorization/cholesky/impl.h
@@ -7,8 +7,12 @@
 // Please, refer to the LICENSE file in the root directory.
 // SPDX-License-Identifier: BSD-3-Clause
 //
+
 #pragma once
 
+#include <cstddef>
+#include <utility>
+
 #include <pika/execution.hpp>
 
 #include <dlaf/blas/tile.h>
diff --git a/include/dlaf/factorization/qr.h b/include/dlaf/factorization/qr.h
index ff0dae09e2..7337412e0e 100644
--- a/include/dlaf/factorization/qr.h
+++ b/include/dlaf/factorization/qr.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+#include <utility>
+
 /// @file
 
 #include <dlaf/communication/communicator_pipeline.h>
diff --git a/include/dlaf/factorization/qr/t_factor_impl.h b/include/dlaf/factorization/qr/t_factor_impl.h
index 4523ee7a28..87ed14e845 100644
--- a/include/dlaf/factorization/qr/t_factor_impl.h
+++ b/include/dlaf/factorization/qr/t_factor_impl.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <utility>
+
 #include <blas.hh>
 
 #include <pika/execution.hpp>
diff --git a/include/dlaf/init.h b/include/dlaf/init.h
index 7ca7a6605e..c04ee0d947 100644
--- a/include/dlaf/init.h
+++ b/include/dlaf/init.h
@@ -11,7 +11,9 @@
 
 /// @file
 
+#include <cstddef>
 #include <iostream>
+#include <string>
 
 #include <pika/program_options.hpp>
 #include <pika/runtime.hpp>
diff --git a/include/dlaf/lapack/tile.h b/include/dlaf/lapack/tile.h
index c90364c371..8f36976731 100644
--- a/include/dlaf/lapack/tile.h
+++ b/include/dlaf/lapack/tile.h
@@ -13,6 +13,7 @@
 /// @file
 
 #include <cstddef>
+#include <utility>
 
 #include <lapack.hh>
 // LAPACKPP includes complex.h which defines the macro I.
diff --git a/include/dlaf/matrix/copy_tile.h b/include/dlaf/matrix/copy_tile.h
index 0fb2645c23..64a887f5b9 100644
--- a/include/dlaf/matrix/copy_tile.h
+++ b/include/dlaf/matrix/copy_tile.h
@@ -12,7 +12,9 @@
 
 /// @file
 
+#include <cstddef>
 #include <type_traits>
+#include <utility>
 
 #if DLAF_WITH_GPU
 #include <whip.hpp>
diff --git a/include/dlaf/matrix/create_matrix.h b/include/dlaf/matrix/create_matrix.h
index 18c8d43fc8..4d071cfcd2 100644
--- a/include/dlaf/matrix/create_matrix.h
+++ b/include/dlaf/matrix/create_matrix.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+#include <utility>
+
 #include <dlaf/matrix/distribution.h>
 #include <dlaf/matrix/layout_info.h>
 #include <dlaf/matrix/matrix.h>
diff --git a/include/dlaf/matrix/hdf5.h b/include/dlaf/matrix/hdf5.h
index 30be778aec..65c6f6bcdd 100644
--- a/include/dlaf/matrix/hdf5.h
+++ b/include/dlaf/matrix/hdf5.h
@@ -13,10 +13,12 @@
 #ifdef DLAF_WITH_HDF5
 
 #include <complex>
+#include <cstddef>
 #include <cstdint>
 #include <string>
 #include <string_view>
 #include <typeinfo>
+#include <utility>
 
 #include <H5Cpp.h>
 #include <mpi.h>
diff --git a/include/dlaf/matrix/internal/tile_pipeline.h b/include/dlaf/matrix/internal/tile_pipeline.h
index ee27bfe6fd..b7e1db41c1 100644
--- a/include/dlaf/matrix/internal/tile_pipeline.h
+++ b/include/dlaf/matrix/internal/tile_pipeline.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <utility>
+
 #include <dlaf/common/assert.h>
 #include <dlaf/matrix/tile.h>
 
diff --git a/include/dlaf/matrix/matrix.h b/include/dlaf/matrix/matrix.h
index 58194daaff..61d884d960 100644
--- a/include/dlaf/matrix/matrix.h
+++ b/include/dlaf/matrix/matrix.h
@@ -12,7 +12,9 @@
 
 /// @file
 
+#include <cstddef>
 #include <exception>
+#include <utility>
 #include <vector>
 
 #include <pika/execution.hpp>
diff --git a/include/dlaf/matrix/matrix_base.h b/include/dlaf/matrix/matrix_base.h
index 7a1a225631..c2f3e42bbd 100644
--- a/include/dlaf/matrix/matrix_base.h
+++ b/include/dlaf/matrix/matrix_base.h
@@ -14,6 +14,7 @@
 
 #include <cstddef>
 #include <ostream>
+#include <utility>
 
 #include <dlaf/matrix/distribution.h>
 
diff --git a/include/dlaf/matrix/matrix_ref.h b/include/dlaf/matrix/matrix_ref.h
index 685acf9b81..b77275c271 100644
--- a/include/dlaf/matrix/matrix_ref.h
+++ b/include/dlaf/matrix/matrix_ref.h
@@ -12,6 +12,8 @@
 
 /// @file
 
+#include <utility>
+
 #include <dlaf/matrix/distribution.h>
 #include <dlaf/matrix/matrix.h>
 #include <dlaf/matrix/matrix_base.h>
diff --git a/include/dlaf/matrix/panel.h b/include/dlaf/matrix/panel.h
index 6db6608bf4..498ed60697 100644
--- a/include/dlaf/matrix/panel.h
+++ b/include/dlaf/matrix/panel.h
@@ -11,6 +11,9 @@
 
 /// @file
 
+#include <set>
+#include <utility>
+
 #include <pika/execution.hpp>
 
 #include <dlaf/common/index2d.h>
diff --git a/include/dlaf/matrix/print_csv.h b/include/dlaf/matrix/print_csv.h
index 284ed69d93..a5850c2463 100644
--- a/include/dlaf/matrix/print_csv.h
+++ b/include/dlaf/matrix/print_csv.h
@@ -7,10 +7,15 @@
 // Please, refer to the LICENSE file in the root directory.
 // SPDX-License-Identifier: BSD-3-Clause
 //
+
 #pragma once
 
 /// @file
 
+#include <iomanip>
+#include <iostream>
+#include <string>
+
 #include <dlaf/common/index2d.h>
 #include <dlaf/common/range2d.h>
 #include <dlaf/matrix/matrix.h>
diff --git a/include/dlaf/matrix/print_numpy.h b/include/dlaf/matrix/print_numpy.h
index fc2acb94ff..c7a3a98db1 100644
--- a/include/dlaf/matrix/print_numpy.h
+++ b/include/dlaf/matrix/print_numpy.h
@@ -12,6 +12,8 @@
 
 /// @file
 
+#include <iostream>
+#include <sstream>
 #include <string>
 
 #include <dlaf/common/index2d.h>
diff --git a/include/dlaf/matrix/tile.h b/include/dlaf/matrix/tile.h
index 589d4cbb50..95bba545d9 100644
--- a/include/dlaf/matrix/tile.h
+++ b/include/dlaf/matrix/tile.h
@@ -13,9 +13,12 @@
 /// @file
 
 #include <exception>
+#include <memory>
 #include <ostream>
 #include <tuple>
 #include <type_traits>
+#include <utility>
+#include <vector>
 
 #include <pika/async_rw_mutex.hpp>
 #include <pika/execution.hpp>
diff --git a/include/dlaf/memory/memory_chunk.h b/include/dlaf/memory/memory_chunk.h
index 6ff2e6f68f..c837d54f49 100644
--- a/include/dlaf/memory/memory_chunk.h
+++ b/include/dlaf/memory/memory_chunk.h
@@ -12,7 +12,10 @@
 
 /// @file
 
+#include <cstddef>
 #include <cstdlib>
+#include <exception>
+#include <iostream>
 #include <memory>
 
 #include <umpire/Allocator.hpp>
diff --git a/include/dlaf/multiplication/general/impl.h b/include/dlaf/multiplication/general/impl.h
index c27eae942c..8542fc7d03 100644
--- a/include/dlaf/multiplication/general/impl.h
+++ b/include/dlaf/multiplication/general/impl.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <cstddef>
+
 #include <dlaf/blas/tile.h>
 #include <dlaf/blas/tile_extensions.h>
 #include <dlaf/common/assert.h>
diff --git a/include/dlaf/multiplication/hermitian/impl.h b/include/dlaf/multiplication/hermitian/impl.h
index fb522b8e33..4f3b4ae8d8 100644
--- a/include/dlaf/multiplication/hermitian/impl.h
+++ b/include/dlaf/multiplication/hermitian/impl.h
@@ -7,8 +7,12 @@
 // Please, refer to the LICENSE file in the root directory.
 // SPDX-License-Identifier: BSD-3-Clause
 //
+
 #pragma once
 
+#include <cstddef>
+#include <utility>
+
 #include <pika/execution.hpp>
 #include <pika/thread.hpp>
 
diff --git a/include/dlaf/multiplication/triangular/impl.h b/include/dlaf/multiplication/triangular/impl.h
index 81629608ac..8132c377c3 100644
--- a/include/dlaf/multiplication/triangular/impl.h
+++ b/include/dlaf/multiplication/triangular/impl.h
@@ -7,8 +7,12 @@
 // Please, refer to the LICENSE file in the root directory.
 // SPDX-License-Identifier: BSD-3-Clause
 //
+
 #pragma once
 
+#include <cstddef>
+#include <utility>
+
 #include <pika/execution.hpp>
 #include <pika/thread.hpp>
 
diff --git a/include/dlaf/permutations/general/impl.h b/include/dlaf/permutations/general/impl.h
index 6876ed46ed..4991ccf90d 100644
--- a/include/dlaf/permutations/general/impl.h
+++ b/include/dlaf/permutations/general/impl.h
@@ -10,7 +10,11 @@
 
 #pragma once
 
+#include <cstddef>
 #include <numeric>
+#include <tuple>
+#include <utility>
+#include <vector>
 
 #include <mpi.h>
 
diff --git a/include/dlaf/sender/transform.h b/include/dlaf/sender/transform.h
index 202a526216..78b66eb001 100644
--- a/include/dlaf/sender/transform.h
+++ b/include/dlaf/sender/transform.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+#include <utility>
+
 #include <pika/execution.hpp>
 
 #include <dlaf/common/consume_rvalues.h>
diff --git a/include/dlaf/sender/transform_mpi.h b/include/dlaf/sender/transform_mpi.h
index 9f1f4b93fd..97bb6bb380 100644
--- a/include/dlaf/sender/transform_mpi.h
+++ b/include/dlaf/sender/transform_mpi.h
@@ -10,6 +10,7 @@
 #pragma once
 
 #include <type_traits>
+#include <utility>
 
 #include <pika/execution.hpp>
 
diff --git a/include/dlaf/sender/with_temporary_tile.h b/include/dlaf/sender/with_temporary_tile.h
index 3263206487..476b885de8 100644
--- a/include/dlaf/sender/with_temporary_tile.h
+++ b/include/dlaf/sender/with_temporary_tile.h
@@ -10,6 +10,9 @@
 
 #pragma once
 
+#include <functional>
+#include <utility>
+
 #include <pika/execution.hpp>
 
 #include <dlaf/common/unwrap.h>
diff --git a/include/dlaf/solver/triangular/impl.h b/include/dlaf/solver/triangular/impl.h
index 6353650385..f1fa273753 100644
--- a/include/dlaf/solver/triangular/impl.h
+++ b/include/dlaf/solver/triangular/impl.h
@@ -7,8 +7,12 @@
 // Please, refer to the LICENSE file in the root directory.
 // SPDX-License-Identifier: BSD-3-Clause
 //
+
 #pragma once
 
+#include <cstddef>
+#include <utility>
+
 #include <pika/execution.hpp>
 #include <pika/thread.hpp>
 
diff --git a/include/dlaf/tune.h b/include/dlaf/tune.h
index f5450d7cfb..17d418b351 100644
--- a/include/dlaf/tune.h
+++ b/include/dlaf/tune.h
@@ -11,6 +11,7 @@
 
 /// @file
 
+#include <cstddef>
 #include <cstdint>
 #include <iosfwd>
 
diff --git a/include/dlaf/types.h b/include/dlaf/types.h
index 3e8400120b..9d50a7c57f 100644
--- a/include/dlaf/types.h
+++ b/include/dlaf/types.h
@@ -16,6 +16,7 @@
 #include <cstddef>
 #include <limits>
 #include <ostream>
+#include <type_traits>
 
 #include <dlaf/common/assert.h>
 
diff --git a/include/dlaf/util_math.h b/include/dlaf/util_math.h
index 411089e908..d26e0ba527 100644
--- a/include/dlaf/util_math.h
+++ b/include/dlaf/util_math.h
@@ -13,6 +13,7 @@
 #include <cstddef>
 #include <functional>
 #include <type_traits>
+#include <utility>
 #include <vector>
 
 #include <dlaf/types.h>
diff --git a/include/dlaf/util_matrix.h b/include/dlaf/util_matrix.h
index 130bce86b9..e01285b900 100644
--- a/include/dlaf/util_matrix.h
+++ b/include/dlaf/util_matrix.h
@@ -10,9 +10,12 @@
 #pragma once
 
 #include <cmath>
+#include <cstddef>
 #include <exception>
+#include <optional>
 #include <random>
 #include <string>
+#include <utility>
 
 #ifndef M_PI
 constexpr double M_PI = 3.141592;
diff --git a/miniapp/include/dlaf/miniapp/kernel_runner.h b/miniapp/include/dlaf/miniapp/kernel_runner.h
index 99ae9ebf97..1223f344ea 100644
--- a/miniapp/include/dlaf/miniapp/kernel_runner.h
+++ b/miniapp/include/dlaf/miniapp/kernel_runner.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <utility>
+
 #ifdef DLAF_WITH_GPU
 #include <whip.hpp>
 #endif
diff --git a/miniapp/include/dlaf/miniapp/options.h b/miniapp/include/dlaf/miniapp/options.h
index 5de4e97213..935d2fbe03 100644
--- a/miniapp/include/dlaf/miniapp/options.h
+++ b/miniapp/include/dlaf/miniapp/options.h
@@ -12,10 +12,13 @@
 
 #include <algorithm>
 #include <cctype>
+#include <cstddef>
 #include <exception>
 #include <iostream>
+#include <sstream>
 #include <string>
 #include <type_traits>
+#include <vector>
 
 #include <blas.hh>
 
diff --git a/miniapp/include/dlaf/miniapp/work_tiles.h b/miniapp/include/dlaf/miniapp/work_tiles.h
index eb0ccc2f85..e04ff0b8d4 100644
--- a/miniapp/include/dlaf/miniapp/work_tiles.h
+++ b/miniapp/include/dlaf/miniapp/work_tiles.h
@@ -11,6 +11,7 @@
 #pragma once
 
 #include <limits>
+#include <utility>
 
 #include <dlaf/common/single_threaded_blas.h>
 #include <dlaf/common/vector.h>
diff --git a/miniapp/kernel/miniapp_laset.cpp b/miniapp/kernel/miniapp_laset.cpp
index e24d967ed8..e640d1e5e5 100644
--- a/miniapp/kernel/miniapp_laset.cpp
+++ b/miniapp/kernel/miniapp_laset.cpp
@@ -9,6 +9,9 @@
 //
 
 #include <future>
+#include <iostream>
+#include <string>
+#include <utility>
 
 #ifdef DLAF_WITH_GPU
 #include <whip.hpp>
diff --git a/miniapp/miniapp_band_to_tridiag.cpp b/miniapp/miniapp_band_to_tridiag.cpp
index f3d81c047a..bace372c48 100644
--- a/miniapp/miniapp_band_to_tridiag.cpp
+++ b/miniapp/miniapp_band_to_tridiag.cpp
@@ -11,6 +11,7 @@
 #include <cstdlib>
 #include <iostream>
 #include <limits>
+#include <string>
 
 #include <pika/init.hpp>
 #include <pika/program_options.hpp>
diff --git a/miniapp/miniapp_cholesky.cpp b/miniapp/miniapp_cholesky.cpp
index b082334c8b..947835f955 100644
--- a/miniapp/miniapp_cholesky.cpp
+++ b/miniapp/miniapp_cholesky.cpp
@@ -10,6 +10,8 @@
 
 #include <cstdlib>
 #include <iostream>
+#include <string>
+#include <utility>
 
 #include <blas/util.hh>
 #include <mpi.h>
diff --git a/miniapp/miniapp_communication.cpp b/miniapp/miniapp_communication.cpp
index 9e4582ffa6..dd147d6d26 100644
--- a/miniapp/miniapp_communication.cpp
+++ b/miniapp/miniapp_communication.cpp
@@ -10,6 +10,9 @@
 
 #include <cstdlib>
 #include <iostream>
+#include <sstream>
+#include <string>
+#include <utility>
 
 #include <blas/util.hh>
 #include <mpi.h>
diff --git a/miniapp/miniapp_gen_eigensolver.cpp b/miniapp/miniapp_gen_eigensolver.cpp
index 3c63dc766d..40659b7c49 100644
--- a/miniapp/miniapp_gen_eigensolver.cpp
+++ b/miniapp/miniapp_gen_eigensolver.cpp
@@ -11,6 +11,7 @@
 #include <cstdlib>
 #include <iostream>
 #include <limits>
+#include <string>
 
 #include <pika/init.hpp>
 #include <pika/program_options.hpp>
diff --git a/miniapp/miniapp_gen_to_std.cpp b/miniapp/miniapp_gen_to_std.cpp
index ae93c667b3..8fa3a8c70d 100644
--- a/miniapp/miniapp_gen_to_std.cpp
+++ b/miniapp/miniapp_gen_to_std.cpp
@@ -10,6 +10,7 @@
 
 #include <cstdlib>
 #include <iostream>
+#include <string>
 
 #include <blas/util.hh>
 #include <mpi.h>
diff --git a/miniapp/miniapp_reduction_to_band.cpp b/miniapp/miniapp_reduction_to_band.cpp
index 6856a73396..299af230b0 100644
--- a/miniapp/miniapp_reduction_to_band.cpp
+++ b/miniapp/miniapp_reduction_to_band.cpp
@@ -11,6 +11,7 @@
 #include <cstdlib>
 #include <iostream>
 #include <limits>
+#include <string>
 
 #include <pika/init.hpp>
 #include <pika/program_options.hpp>
diff --git a/miniapp/miniapp_triangular_multiplication.cpp b/miniapp/miniapp_triangular_multiplication.cpp
index 1341121d03..006dd65eb9 100644
--- a/miniapp/miniapp_triangular_multiplication.cpp
+++ b/miniapp/miniapp_triangular_multiplication.cpp
@@ -11,6 +11,7 @@
 #include <cstdlib>
 #include <iostream>
 #include <limits>
+#include <string>
 #include <type_traits>
 
 #include <blas/util.hh>
diff --git a/miniapp/miniapp_triangular_solver.cpp b/miniapp/miniapp_triangular_solver.cpp
index 8c0b35fa48..f2b98b8f4b 100644
--- a/miniapp/miniapp_triangular_solver.cpp
+++ b/miniapp/miniapp_triangular_solver.cpp
@@ -11,6 +11,7 @@
 #include <cstdlib>
 #include <iostream>
 #include <limits>
+#include <string>
 #include <type_traits>
 
 #include <blas/util.hh>
diff --git a/src/c_api/factorization/cholesky.h b/src/c_api/factorization/cholesky.h
index 71d57821ac..407f77bcb3 100644
--- a/src/c_api/factorization/cholesky.h
+++ b/src/c_api/factorization/cholesky.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <utility>
+
 #include <blas.hh>
 #include <blas/util.hh>
 #include <mpi.h>
diff --git a/src/c_api/grid.cpp b/src/c_api/grid.cpp
index e286e79127..3e041f4021 100644
--- a/src/c_api/grid.cpp
+++ b/src/c_api/grid.cpp
@@ -8,7 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <iostream>
 #include <limits>
+#include <unordered_map>
 
 #include <mpi.h>
 
diff --git a/src/c_api/utils.cpp b/src/c_api/utils.cpp
index 3aa256461b..cf8b5f2723 100644
--- a/src/c_api/utils.cpp
+++ b/src/c_api/utils.cpp
@@ -8,8 +8,12 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <exception>
 #include <iostream>
+#include <optional>
+#include <sstream>
 #include <stdexcept>
+#include <tuple>
 
 #include <dlaf/communication/communicator_grid.h>
 #include <dlaf_c/desc.h>
diff --git a/src/c_api/utils.h b/src/c_api/utils.h
index 9526b27d06..e2db52ecfb 100644
--- a/src/c_api/utils.h
+++ b/src/c_api/utils.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <tuple>
+
 #include <dlaf/communication/communicator_grid.h>
 #include <dlaf/matrix/distribution.h>
 #include <dlaf/matrix/layout_info.h>
diff --git a/src/communication/communicator_grid.cpp b/src/communication/communicator_grid.cpp
index 9cc4c02060..61a1222ded 100644
--- a/src/communication/communicator_grid.cpp
+++ b/src/communication/communicator_grid.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <cstddef>
+
 #include <dlaf/common/with_result_of.h>
 #include <dlaf/communication/communicator_grid.h>
 #include <dlaf/communication/index.h>
diff --git a/src/eigensolver/tridiag_solver/gpu/kernels.cu b/src/eigensolver/tridiag_solver/gpu/kernels.cu
index 6c93fa1baa..3ef6cf918d 100644
--- a/src/eigensolver/tridiag_solver/gpu/kernels.cu
+++ b/src/eigensolver/tridiag_solver/gpu/kernels.cu
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <type_traits>
+
 #include <whip.hpp>
 
 #include <dlaf/eigensolver/tridiag_solver/gpu/kernels.h>
diff --git a/src/init.cpp b/src/init.cpp
index 2eb38d7f38..c03b2b8b63 100644
--- a/src/init.cpp
+++ b/src/init.cpp
@@ -11,11 +11,13 @@
 #include <algorithm>
 #include <array>
 #include <cctype>
+#include <cstddef>
 #include <cstdlib>
 #include <exception>
 #include <iostream>
 #include <memory>
 #include <optional>
+#include <string>
 
 #include <pika/runtime.hpp>
 
diff --git a/src/matrix/distribution.cpp b/src/matrix/distribution.cpp
index 8203b4ac4c..b5bb9b2779 100644
--- a/src/matrix/distribution.cpp
+++ b/src/matrix/distribution.cpp
@@ -9,6 +9,7 @@
 //
 
 #include <cstdlib>
+#include <utility>
 
 #include <dlaf/matrix/distribution.h>
 
diff --git a/src/memory/memory_chunk.cpp b/src/memory/memory_chunk.cpp
index 3b20a58e64..a2b0c33885 100644
--- a/src/memory/memory_chunk.cpp
+++ b/src/memory/memory_chunk.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <cstddef>
+
 #include <umpire/ResourceManager.hpp>
 #include <umpire/strategy/QuickPool.hpp>
 #include <umpire/strategy/ThreadSafeAllocator.hpp>
diff --git a/src/permutations/general/perms.cu b/src/permutations/general/perms.cu
index 37d519e32e..c0d1aca3c1 100644
--- a/src/permutations/general/perms.cu
+++ b/src/permutations/general/perms.cu
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <vector>
+
 #include <whip.hpp>
 
 #include <dlaf/permutations/general/perms.h>
diff --git a/test/include/dlaf_c_test/c_api_helpers.h b/test/include/dlaf_c_test/c_api_helpers.h
index 9806276dcf..312bb18127 100644
--- a/test/include/dlaf_c_test/c_api_helpers.h
+++ b/test/include/dlaf_c_test/c_api_helpers.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <utility>
+
 #include <mpi.h>
 
 #include <pika/execution.hpp>
diff --git a/test/include/dlaf_test/blas/invoke.h b/test/include/dlaf_test/blas/invoke.h
index 0962f7b8c4..f9ec26bd81 100644
--- a/test/include/dlaf_test/blas/invoke.h
+++ b/test/include/dlaf_test/blas/invoke.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <utility>
+
 #ifdef DLAF_WITH_GPU
 #include <whip.hpp>
 #endif
diff --git a/test/include/dlaf_test/comm_grids/grids_6_ranks.h b/test/include/dlaf_test/comm_grids/grids_6_ranks.h
index babc194fd1..d2b22beb76 100644
--- a/test/include/dlaf_test/comm_grids/grids_6_ranks.h
+++ b/test/include/dlaf_test/comm_grids/grids_6_ranks.h
@@ -10,6 +10,9 @@
 
 #pragma once
 
+#include <cstddef>
+#include <vector>
+
 #include <dlaf/communication/communicator_grid.h>
 #include <dlaf/communication/error.h>
 
diff --git a/test/include/dlaf_test/lapack/invoke.h b/test/include/dlaf_test/lapack/invoke.h
index 211fd5bbf0..cea16ea4e0 100644
--- a/test/include/dlaf_test/lapack/invoke.h
+++ b/test/include/dlaf_test/lapack/invoke.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+#include <utility>
+
 #ifdef DLAF_WITH_GPU
 #include <whip.hpp>
 #endif
diff --git a/test/include/dlaf_test/matrix/matrix_local.h b/test/include/dlaf_test/matrix/matrix_local.h
index 8cd565ecf9..009f7a5c8a 100644
--- a/test/include/dlaf_test/matrix/matrix_local.h
+++ b/test/include/dlaf_test/matrix/matrix_local.h
@@ -10,7 +10,9 @@
 
 #pragma once
 
+#include <cstddef>
 #include <type_traits>
+#include <vector>
 
 #include <dlaf/common/range2d.h>
 #include <dlaf/matrix/layout_info.h>
diff --git a/test/include/dlaf_test/matrix/util_matrix.h b/test/include/dlaf_test/matrix/util_matrix.h
index 8c3a65011f..3af1fe0513 100644
--- a/test/include/dlaf_test/matrix/util_matrix.h
+++ b/test/include/dlaf_test/matrix/util_matrix.h
@@ -12,6 +12,7 @@
 
 /// @file
 
+#include <cstddef>
 #include <functional>
 #include <iomanip>
 #include <iostream>
diff --git a/test/include/dlaf_test/matrix/util_matrix_local.h b/test/include/dlaf_test/matrix/util_matrix_local.h
index aa9b37cb1b..53b3c58f34 100644
--- a/test/include/dlaf_test/matrix/util_matrix_local.h
+++ b/test/include/dlaf_test/matrix/util_matrix_local.h
@@ -12,8 +12,12 @@
 
 /// @file
 
+#include <cstddef>
 #include <functional>
+#include <iostream>
+#include <string>
 #include <type_traits>
+#include <utility>
 
 #include <dlaf/common/range2d.h>
 #include <dlaf/communication/communicator_grid.h>
diff --git a/test/include/dlaf_test/matrix/util_matrix_senders.h b/test/include/dlaf_test/matrix/util_matrix_senders.h
index 9e6aee6f8f..f4c9f23ec5 100644
--- a/test/include/dlaf_test/matrix/util_matrix_senders.h
+++ b/test/include/dlaf_test/matrix/util_matrix_senders.h
@@ -12,7 +12,9 @@
 
 /// @file
 
+#include <cstddef>
 #include <sstream>
+#include <utility>
 #include <vector>
 
 #include <pika/execution.hpp>
diff --git a/test/include/dlaf_test/matrix/util_tile.h b/test/include/dlaf_test/matrix/util_tile.h
index 39d4e7b805..02f22a6912 100644
--- a/test/include/dlaf_test/matrix/util_tile.h
+++ b/test/include/dlaf_test/matrix/util_tile.h
@@ -12,10 +12,13 @@
 
 /// @file
 
+#include <atomic>
 #include <functional>
 #include <iomanip>
 #include <iostream>
+#include <memory>
 #include <sstream>
+#include <utility>
 
 #include <dlaf/matrix/copy_tile.h>
 #include <dlaf/matrix/tile.h>
diff --git a/test/src/gtest_mpi_listener.cpp b/test/src/gtest_mpi_listener.cpp
index c97fe93432..4c8f78366b 100644
--- a/test/src/gtest_mpi_listener.cpp
+++ b/test/src/gtest_mpi_listener.cpp
@@ -9,7 +9,12 @@
 //
 
 #include <algorithm>
+#include <cstddef>
 #include <functional>
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
 
 #include "gtest_mpi_listener.h"
 
diff --git a/test/src/gtest_mpi_listener.h b/test/src/gtest_mpi_listener.h
index 53ca3b07f6..7916f2bc8d 100644
--- a/test/src/gtest_mpi_listener.h
+++ b/test/src/gtest_mpi_listener.h
@@ -11,6 +11,7 @@
 #pragma once
 
 #include <memory>
+#include <string>
 #include <vector>
 
 #include <mpi.h>
diff --git a/test/unit/auxiliary/mc/test_norm.cpp b/test/unit/auxiliary/mc/test_norm.cpp
index c23c02e388..e596ca4eed 100644
--- a/test/unit/auxiliary/mc/test_norm.cpp
+++ b/test/unit/auxiliary/mc/test_norm.cpp
@@ -9,6 +9,8 @@
 //
 
 #include <limits>
+#include <utility>
+#include <vector>
 
 #include <dlaf/auxiliary/norm.h>
 #include <dlaf/communication/communicator_grid.h>
diff --git a/test/unit/c_api/eigensolver/test_eigensolver_c_api.cpp b/test/unit/c_api/eigensolver/test_eigensolver_c_api.cpp
index 07644be243..429a322b22 100644
--- a/test/unit/c_api/eigensolver/test_eigensolver_c_api.cpp
+++ b/test/unit/c_api/eigensolver/test_eigensolver_c_api.cpp
@@ -10,6 +10,9 @@
 
 #include <complex>
 #include <tuple>
+#include <typeinfo>
+#include <utility>
+#include <vector>
 
 #include <blas/util.hh>
 
diff --git a/test/unit/c_api/eigensolver/test_gen_eigensolver_c_api.cpp b/test/unit/c_api/eigensolver/test_gen_eigensolver_c_api.cpp
index f6b936ba03..806678571f 100644
--- a/test/unit/c_api/eigensolver/test_gen_eigensolver_c_api.cpp
+++ b/test/unit/c_api/eigensolver/test_gen_eigensolver_c_api.cpp
@@ -10,6 +10,9 @@
 
 #include <functional>
 #include <tuple>
+#include <typeinfo>
+#include <utility>
+#include <vector>
 
 #include <pika/init.hpp>
 
diff --git a/test/unit/c_api/factorization/test_cholesky_c_api.cpp b/test/unit/c_api/factorization/test_cholesky_c_api.cpp
index 79286f5e17..8ff3f3be83 100644
--- a/test/unit/c_api/factorization/test_cholesky_c_api.cpp
+++ b/test/unit/c_api/factorization/test_cholesky_c_api.cpp
@@ -10,6 +10,9 @@
 
 #include <complex>
 #include <tuple>
+#include <typeinfo>
+#include <utility>
+#include <vector>
 
 #include <blas/util.hh>
 
diff --git a/test/unit/common/test_data_descriptor.cpp b/test/unit/common/test_data_descriptor.cpp
index 4f17b16879..c0418ec4c6 100644
--- a/test/unit/common/test_data_descriptor.cpp
+++ b/test/unit/common/test_data_descriptor.cpp
@@ -8,8 +8,11 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <cstddef>
+#include <iostream>
 #include <memory>
 #include <type_traits>
+#include <vector>
 
 #include <dlaf/common/data.h>
 #include <dlaf/common/data_descriptor.h>
diff --git a/test/unit/common/test_pipeline.cpp b/test/unit/common/test_pipeline.cpp
index 75f6b72015..daf69ee79e 100644
--- a/test/unit/common/test_pipeline.cpp
+++ b/test/unit/common/test_pipeline.cpp
@@ -13,6 +13,7 @@
 #include <cstddef>
 #include <mutex>
 #include <random>
+#include <utility>
 #include <vector>
 
 #include <pika/execution.hpp>
diff --git a/test/unit/common/test_size2d.cpp b/test/unit/common/test_size2d.cpp
index cab07eccc1..a630bcfef6 100644
--- a/test/unit/common/test_size2d.cpp
+++ b/test/unit/common/test_size2d.cpp
@@ -9,6 +9,7 @@
 //
 
 #include <array>
+#include <sstream>
 
 #include <dlaf/common/index2d.h>
 
diff --git a/test/unit/communication/test_all_reduce.cpp b/test/unit/communication/test_all_reduce.cpp
index 112302ddb1..cab786c2fc 100644
--- a/test/unit/communication/test_all_reduce.cpp
+++ b/test/unit/communication/test_all_reduce.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <utility>
+
 #include <dlaf/common/data_descriptor.h>
 #include <dlaf/communication/communicator_grid.h>
 #include <dlaf/communication/sync/all_reduce.h>
diff --git a/test/unit/communication/test_broadcast_panel.cpp b/test/unit/communication/test_broadcast_panel.cpp
index eb686dd464..bfcf26e2fe 100644
--- a/test/unit/communication/test_broadcast_panel.cpp
+++ b/test/unit/communication/test_broadcast_panel.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <vector>
+
 #include <dlaf/common/range2d.h>
 #include <dlaf/communication/broadcast_panel.h>
 #include <dlaf/communication/communicator.h>
diff --git a/test/unit/communication/test_collective_async.cpp b/test/unit/communication/test_collective_async.cpp
index 707b094694..7d0afcc634 100644
--- a/test/unit/communication/test_collective_async.cpp
+++ b/test/unit/communication/test_collective_async.cpp
@@ -8,6 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <string>
+#include <utility>
+
 #include <mpi.h>
 
 #include <dlaf/common/data.h>
diff --git a/test/unit/communication/test_comm_p2p.cpp b/test/unit/communication/test_comm_p2p.cpp
index 13c38fc744..f84643db33 100644
--- a/test/unit/communication/test_comm_p2p.cpp
+++ b/test/unit/communication/test_comm_p2p.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <utility>
+
 #include <dlaf/common/data.h>
 #include <dlaf/common/index2d.h>
 #include <dlaf/common/range2d.h>
diff --git a/test/unit/communication/test_comm_sender.cpp b/test/unit/communication/test_comm_sender.cpp
index c34fd41526..3083301660 100644
--- a/test/unit/communication/test_comm_sender.cpp
+++ b/test/unit/communication/test_comm_sender.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <cstddef>
+#include <utility>
 #include <vector>
 
 #include <mpi.h>
diff --git a/test/unit/communication/test_communicator_grid.cpp b/test/unit/communication/test_communicator_grid.cpp
index dd16c4b806..81585720b1 100644
--- a/test/unit/communication/test_communicator_grid.cpp
+++ b/test/unit/communication/test_communicator_grid.cpp
@@ -8,6 +8,11 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <array>
+#include <cstddef>
+#include <set>
+#include <vector>
+
 #include <mpi.h>
 
 #include <pika/execution.hpp>
diff --git a/test/unit/communication/test_message.cpp b/test/unit/communication/test_message.cpp
index 8c8e6f7b8f..f598b490d1 100644
--- a/test/unit/communication/test_message.cpp
+++ b/test/unit/communication/test_message.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <utility>
+
 #include <dlaf/common/data_descriptor.h>
 #include <dlaf/communication/message.h>
 #include <dlaf/types.h>
diff --git a/test/unit/communication/test_reduce.cpp b/test/unit/communication/test_reduce.cpp
index 92c9120e85..6e10b47c19 100644
--- a/test/unit/communication/test_reduce.cpp
+++ b/test/unit/communication/test_reduce.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <utility>
+
 #include <dlaf/common/data_descriptor.h>
 #include <dlaf/communication/communicator_grid.h>
 #include <dlaf/communication/sync/reduce.h>
diff --git a/test/unit/communication/test_transform_mpi.cpp b/test/unit/communication/test_transform_mpi.cpp
index 47d81c9c47..bf9f857e0d 100644
--- a/test/unit/communication/test_transform_mpi.cpp
+++ b/test/unit/communication/test_transform_mpi.cpp
@@ -11,6 +11,7 @@
 #include <atomic>
 #include <chrono>
 #include <thread>
+#include <utility>
 
 #include <dlaf/common/pipeline.h>
 #include <dlaf/communication/communicator.h>
diff --git a/test/unit/eigensolver/test_band_to_tridiag.cpp b/test/unit/eigensolver/test_band_to_tridiag.cpp
index f2caef3adf..108bd53789 100644
--- a/test/unit/eigensolver/test_band_to_tridiag.cpp
+++ b/test/unit/eigensolver/test_band_to_tridiag.cpp
@@ -12,6 +12,8 @@
 #include <functional>
 #include <sstream>
 #include <tuple>
+#include <utility>
+#include <vector>
 
 #include <dlaf/common/single_threaded_blas.h>
 #include <dlaf/eigensolver/band_to_tridiag.h>
diff --git a/test/unit/eigensolver/test_bt_band_to_tridiag.cpp b/test/unit/eigensolver/test_bt_band_to_tridiag.cpp
index 39432947ad..d947554cd0 100644
--- a/test/unit/eigensolver/test_bt_band_to_tridiag.cpp
+++ b/test/unit/eigensolver/test_bt_band_to_tridiag.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <vector>
+
 #include <pika/init.hpp>
 
 #include <dlaf/common/single_threaded_blas.h>
diff --git a/test/unit/eigensolver/test_bt_reduction_to_band.cpp b/test/unit/eigensolver/test_bt_reduction_to_band.cpp
index 8648908878..d85ca65357 100644
--- a/test/unit/eigensolver/test_bt_reduction_to_band.cpp
+++ b/test/unit/eigensolver/test_bt_reduction_to_band.cpp
@@ -12,6 +12,7 @@
 #include <optional>
 #include <sstream>
 #include <tuple>
+#include <utility>
 #include <vector>
 
 #include <pika/init.hpp>
diff --git a/test/unit/eigensolver/test_eigensolver.cpp b/test/unit/eigensolver/test_eigensolver.cpp
index cc9d7386b4..81fff7d42b 100644
--- a/test/unit/eigensolver/test_eigensolver.cpp
+++ b/test/unit/eigensolver/test_eigensolver.cpp
@@ -10,6 +10,8 @@
 
 #include <functional>
 #include <tuple>
+#include <utility>
+#include <vector>
 
 #include <dlaf/communication/communicator_grid.h>
 #include <dlaf/eigensolver/eigensolver.h>
diff --git a/test/unit/eigensolver/test_gen_eigensolver.cpp b/test/unit/eigensolver/test_gen_eigensolver.cpp
index 19fcbbb056..e7df53fcb0 100644
--- a/test/unit/eigensolver/test_gen_eigensolver.cpp
+++ b/test/unit/eigensolver/test_gen_eigensolver.cpp
@@ -10,6 +10,8 @@
 
 #include <functional>
 #include <tuple>
+#include <utility>
+#include <vector>
 
 #include <pika/init.hpp>
 
diff --git a/test/unit/eigensolver/test_gen_to_std.cpp b/test/unit/eigensolver/test_gen_to_std.cpp
index b17f662f24..becbc779c0 100644
--- a/test/unit/eigensolver/test_gen_to_std.cpp
+++ b/test/unit/eigensolver/test_gen_to_std.cpp
@@ -10,6 +10,8 @@
 
 #include <functional>
 #include <tuple>
+#include <utility>
+#include <vector>
 
 #include <pika/init.hpp>
 
diff --git a/test/unit/eigensolver/test_reduction_to_band.cpp b/test/unit/eigensolver/test_reduction_to_band.cpp
index 6700b793e2..c4bee9db09 100644
--- a/test/unit/eigensolver/test_reduction_to_band.cpp
+++ b/test/unit/eigensolver/test_reduction_to_band.cpp
@@ -9,6 +9,7 @@
 //
 
 #include <cmath>
+#include <vector>
 
 #include <pika/execution.hpp>
 #include <pika/init.hpp>
diff --git a/test/unit/eigensolver/test_tridiag_solver_distributed.cpp b/test/unit/eigensolver/test_tridiag_solver_distributed.cpp
index 8cd17aa51b..7871aa9938 100644
--- a/test/unit/eigensolver/test_tridiag_solver_distributed.cpp
+++ b/test/unit/eigensolver/test_tridiag_solver_distributed.cpp
@@ -8,6 +8,10 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <tuple>
+#include <utility>
+#include <vector>
+
 #include <pika/init.hpp>
 
 #include <dlaf/communication/communicator_pipeline.h>
diff --git a/test/unit/eigensolver/test_tridiag_solver_local.cpp b/test/unit/eigensolver/test_tridiag_solver_local.cpp
index b3ea23ac40..aca9ee77ac 100644
--- a/test/unit/eigensolver/test_tridiag_solver_local.cpp
+++ b/test/unit/eigensolver/test_tridiag_solver_local.cpp
@@ -8,6 +8,10 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <tuple>
+#include <utility>
+#include <vector>
+
 #include <dlaf/eigensolver/tridiag_solver.h>
 #include <dlaf/eigensolver/tridiag_solver/impl.h>
 #include <dlaf/matrix/matrix_mirror.h>
diff --git a/test/unit/eigensolver/test_tridiag_solver_merge.cpp b/test/unit/eigensolver/test_tridiag_solver_merge.cpp
index ccf01c349e..d4813d45f7 100644
--- a/test/unit/eigensolver/test_tridiag_solver_merge.cpp
+++ b/test/unit/eigensolver/test_tridiag_solver_merge.cpp
@@ -8,8 +8,11 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <array>
 #include <cstddef>
 #include <numeric>
+#include <utility>
+#include <vector>
 
 #include <dlaf/eigensolver/tridiag_solver/coltype.h>
 #include <dlaf/eigensolver/tridiag_solver/merge.h>
diff --git a/test/unit/eigensolver/test_tridiag_solver_rot.cpp b/test/unit/eigensolver/test_tridiag_solver_rot.cpp
index 1f9fd4f754..a1bb0d9375 100644
--- a/test/unit/eigensolver/test_tridiag_solver_rot.cpp
+++ b/test/unit/eigensolver/test_tridiag_solver_rot.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <vector>
+
 #include <dlaf/common/single_threaded_blas.h>
 #include <dlaf/communication/communicator.h>
 #include <dlaf/communication/communicator_grid.h>
diff --git a/test/unit/factorization/test_cholesky.cpp b/test/unit/factorization/test_cholesky.cpp
index 572f304e2a..301678e947 100644
--- a/test/unit/factorization/test_cholesky.cpp
+++ b/test/unit/factorization/test_cholesky.cpp
@@ -10,6 +10,8 @@
 
 #include <functional>
 #include <tuple>
+#include <utility>
+#include <vector>
 
 #include <pika/init.hpp>
 
diff --git a/test/unit/factorization/test_compute_t_factor.cpp b/test/unit/factorization/test_compute_t_factor.cpp
index 62a623dad8..7cafc2fedd 100644
--- a/test/unit/factorization/test_compute_t_factor.cpp
+++ b/test/unit/factorization/test_compute_t_factor.cpp
@@ -9,6 +9,8 @@
 //
 
 #include <tuple>
+#include <utility>
+#include <vector>
 
 #include <blas.hh>
 
diff --git a/test/unit/init/test_init.cpp b/test/unit/init/test_init.cpp
index c17b9c16de..491c5251da 100644
--- a/test/unit/init/test_init.cpp
+++ b/test/unit/init/test_init.cpp
@@ -8,7 +8,11 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <cstddef>
 #include <cstdlib>
+#include <memory>
+#include <string>
+#include <utility>
 
 #include <pika/init.hpp>
 
diff --git a/test/unit/lapack/gpu/test_lacpy.cpp b/test/unit/lapack/gpu/test_lacpy.cpp
index 16a73a5827..6f34a774a8 100644
--- a/test/unit/lapack/gpu/test_lacpy.cpp
+++ b/test/unit/lapack/gpu/test_lacpy.cpp
@@ -8,6 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <tuple>
+#include <vector>
+
 #include <whip.hpp>
 
 #include <dlaf/blas/enum_output.h>
diff --git a/test/unit/matrix/test_copy.cpp b/test/unit/matrix/test_copy.cpp
index 72d6236e18..df728df86a 100644
--- a/test/unit/matrix/test_copy.cpp
+++ b/test/unit/matrix/test_copy.cpp
@@ -8,6 +8,7 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <utility>
 #include <vector>
 
 #include <dlaf/matrix/copy.h>
diff --git a/test/unit/matrix/test_distribution.cpp b/test/unit/matrix/test_distribution.cpp
index 9f347bfa1e..8f7201ac07 100644
--- a/test/unit/matrix/test_distribution.cpp
+++ b/test/unit/matrix/test_distribution.cpp
@@ -10,6 +10,9 @@
 
 #include <array>
 #include <stdexcept>
+#include <tuple>
+#include <utility>
+#include <vector>
 
 #include <dlaf/matrix/distribution.h>
 
diff --git a/test/unit/matrix/test_distribution_extensions.cpp b/test/unit/matrix/test_distribution_extensions.cpp
index e201196563..3253f5b73c 100644
--- a/test/unit/matrix/test_distribution_extensions.cpp
+++ b/test/unit/matrix/test_distribution_extensions.cpp
@@ -10,6 +10,7 @@
 
 #include <array>
 #include <stdexcept>
+#include <vector>
 
 #include <dlaf/matrix/distribution.h>
 
diff --git a/test/unit/matrix/test_layout_info.cpp b/test/unit/matrix/test_layout_info.cpp
index e00836c342..935507853b 100644
--- a/test/unit/matrix/test_layout_info.cpp
+++ b/test/unit/matrix/test_layout_info.cpp
@@ -8,6 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <tuple>
+#include <vector>
+
 #include <dlaf/matrix/index.h>
 #include <dlaf/matrix/layout_info.h>
 #include <dlaf/util_math.h>
diff --git a/test/unit/matrix/test_matrix.cpp b/test/unit/matrix/test_matrix.cpp
index f2603fa936..ecaa53a8e5 100644
--- a/test/unit/matrix/test_matrix.cpp
+++ b/test/unit/matrix/test_matrix.cpp
@@ -10,6 +10,9 @@
 
 #include <atomic>
 #include <chrono>
+#include <sstream>
+#include <tuple>
+#include <utility>
 #include <vector>
 
 #include <pika/execution.hpp>
diff --git a/test/unit/matrix/test_matrix_hdf5.cpp b/test/unit/matrix/test_matrix_hdf5.cpp
index 511fac83e3..7b7f056a40 100644
--- a/test/unit/matrix/test_matrix_hdf5.cpp
+++ b/test/unit/matrix/test_matrix_hdf5.cpp
@@ -11,6 +11,8 @@
 #include <cstdlib>
 #include <filesystem>
 #include <string>
+#include <tuple>
+#include <utility>
 
 #include <dlaf/common/index2d.h>
 #include <dlaf/communication/communicator.h>
diff --git a/test/unit/matrix/test_matrix_local.cpp b/test/unit/matrix/test_matrix_local.cpp
index 32f6b6a1ec..29eef594b6 100644
--- a/test/unit/matrix/test_matrix_local.cpp
+++ b/test/unit/matrix/test_matrix_local.cpp
@@ -9,6 +9,8 @@
 //
 
 #include <sstream>
+#include <string>
+#include <utility>
 #include <vector>
 
 #include <dlaf/matrix/distribution.h>
diff --git a/test/unit/matrix/test_matrix_mirror.cpp b/test/unit/matrix/test_matrix_mirror.cpp
index b32ba83c6c..9b36620a14 100644
--- a/test/unit/matrix/test_matrix_mirror.cpp
+++ b/test/unit/matrix/test_matrix_mirror.cpp
@@ -8,6 +8,7 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <utility>
 #include <vector>
 
 #include <dlaf/communication/communicator_grid.h>
diff --git a/test/unit/matrix/test_matrix_output.cpp b/test/unit/matrix/test_matrix_output.cpp
index c37ec80992..34cf1466b0 100644
--- a/test/unit/matrix/test_matrix_output.cpp
+++ b/test/unit/matrix/test_matrix_output.cpp
@@ -9,7 +9,9 @@
 //
 
 #include <sstream>
+#include <string>
 #include <tuple>
+#include <utility>
 
 #include <dlaf/communication/communicator_grid.h>
 #include <dlaf/matrix/matrix.h>
diff --git a/test/unit/matrix/test_panel.cpp b/test/unit/matrix/test_panel.cpp
index 6057f8817c..8e3c4fa927 100644
--- a/test/unit/matrix/test_panel.cpp
+++ b/test/unit/matrix/test_panel.cpp
@@ -9,6 +9,7 @@
 //
 
 #include <functional>
+#include <string>
 #include <type_traits>
 #include <vector>
 
diff --git a/test/unit/matrix/test_retiled_matrix.cpp b/test/unit/matrix/test_retiled_matrix.cpp
index e9987974c3..bdddbeeeec 100644
--- a/test/unit/matrix/test_retiled_matrix.cpp
+++ b/test/unit/matrix/test_retiled_matrix.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <tuple>
+#include <utility>
 #include <vector>
 
 #include <dlaf/communication/communicator_grid.h>
diff --git a/test/unit/matrix/test_tile.cpp b/test/unit/matrix/test_tile.cpp
index 62f31179da..0fd80723ce 100644
--- a/test/unit/matrix/test_tile.cpp
+++ b/test/unit/matrix/test_tile.cpp
@@ -8,6 +8,12 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <cstddef>
+#include <string>
+#include <tuple>
+#include <utility>
+#include <vector>
+
 #include <pika/execution.hpp>
 
 #include <dlaf/matrix/index.h>
diff --git a/test/unit/matrix/test_util_distribution.cpp b/test/unit/matrix/test_util_distribution.cpp
index 52a3d71663..9ae71f3502 100644
--- a/test/unit/matrix/test_util_distribution.cpp
+++ b/test/unit/matrix/test_util_distribution.cpp
@@ -8,6 +8,8 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <vector>
+
 #include <dlaf/matrix/util_distribution.h>
 #include <dlaf/types.h>
 
diff --git a/test/unit/matrix/test_util_matrix.cpp b/test/unit/matrix/test_util_matrix.cpp
index e65a5209a1..84d7aa493f 100644
--- a/test/unit/matrix/test_util_matrix.cpp
+++ b/test/unit/matrix/test_util_matrix.cpp
@@ -8,6 +8,7 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <utility>
 #include <vector>
 
 #include <pika/execution.hpp>
diff --git a/test/unit/matrix/test_views.cpp b/test/unit/matrix/test_views.cpp
index 4116908c9e..c1fc5ceb45 100644
--- a/test/unit/matrix/test_views.cpp
+++ b/test/unit/matrix/test_views.cpp
@@ -8,7 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <list>
 #include <sstream>
+#include <vector>
 
 #include <dlaf/common/index2d.h>
 #include <dlaf/communication/communicator_grid.h>
diff --git a/test/unit/memory/test_memory_chunk.cpp b/test/unit/memory/test_memory_chunk.cpp
index aef85c73cc..f638eacd9b 100644
--- a/test/unit/memory/test_memory_chunk.cpp
+++ b/test/unit/memory/test_memory_chunk.cpp
@@ -8,6 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <memory>
+#include <utility>
+
 #include <dlaf/memory/memory_chunk.h>
 
 #include <gtest/gtest.h>
diff --git a/test/unit/memory/test_memory_view.cpp b/test/unit/memory/test_memory_view.cpp
index de0d8b5338..88ccdb7830 100644
--- a/test/unit/memory/test_memory_view.cpp
+++ b/test/unit/memory/test_memory_view.cpp
@@ -8,6 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <memory>
+#include <utility>
+
 #include <dlaf/memory/memory_view.h>
 
 #include <gtest/gtest.h>
diff --git a/test/unit/multiplication/test_multiplication_general.cpp b/test/unit/multiplication/test_multiplication_general.cpp
index 0a48f5d37f..644f67a7ec 100644
--- a/test/unit/multiplication/test_multiplication_general.cpp
+++ b/test/unit/multiplication/test_multiplication_general.cpp
@@ -8,6 +8,10 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <tuple>
+#include <utility>
+#include <vector>
+
 #include <pika/init.hpp>
 
 #include <dlaf/blas/enum_output.h>
diff --git a/test/unit/multiplication/test_multiplication_hermitian.cpp b/test/unit/multiplication/test_multiplication_hermitian.cpp
index 941dae2516..0f21eab622 100644
--- a/test/unit/multiplication/test_multiplication_hermitian.cpp
+++ b/test/unit/multiplication/test_multiplication_hermitian.cpp
@@ -10,6 +10,8 @@
 
 #include <functional>
 #include <tuple>
+#include <utility>
+#include <vector>
 
 #include <pika/init.hpp>
 
diff --git a/test/unit/multiplication/test_multiplication_triangular.cpp b/test/unit/multiplication/test_multiplication_triangular.cpp
index 84fe2c6330..8910987834 100644
--- a/test/unit/multiplication/test_multiplication_triangular.cpp
+++ b/test/unit/multiplication/test_multiplication_triangular.cpp
@@ -10,6 +10,8 @@
 
 #include <functional>
 #include <tuple>
+#include <utility>
+#include <vector>
 
 #include <pika/init.hpp>
 
diff --git a/test/unit/permutations/test_permutations_distributed.cpp b/test/unit/permutations/test_permutations_distributed.cpp
index a4b3f388ce..e0ff529a62 100644
--- a/test/unit/permutations/test_permutations_distributed.cpp
+++ b/test/unit/permutations/test_permutations_distributed.cpp
@@ -9,6 +9,10 @@
 //
 
 #include <algorithm>
+#include <cstddef>
+#include <tuple>
+#include <utility>
+#include <vector>
 
 #include <pika/init.hpp>
 
diff --git a/test/unit/permutations/test_permutations_local.cpp b/test/unit/permutations/test_permutations_local.cpp
index ffb79e4bac..9ebae3ae00 100644
--- a/test/unit/permutations/test_permutations_local.cpp
+++ b/test/unit/permutations/test_permutations_local.cpp
@@ -8,6 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <tuple>
+#include <vector>
+
 #include <dlaf/matrix/matrix_mirror.h>
 #include <dlaf/permutations/general.h>
 #include <dlaf/util_matrix.h>
diff --git a/test/unit/sender/test_with_temporary_tile.cpp b/test/unit/sender/test_with_temporary_tile.cpp
index 4ae6547da4..fa9f8f6350 100644
--- a/test/unit/sender/test_with_temporary_tile.cpp
+++ b/test/unit/sender/test_with_temporary_tile.cpp
@@ -8,6 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <functional>
+#include <utility>
+
 #include <dlaf/lapack/tile.h>
 #include <dlaf/matrix/matrix.h>
 #include <dlaf/sender/policy.h>
diff --git a/test/unit/solver/test_triangular.cpp b/test/unit/solver/test_triangular.cpp
index 128a774895..409e162403 100644
--- a/test/unit/solver/test_triangular.cpp
+++ b/test/unit/solver/test_triangular.cpp
@@ -10,6 +10,8 @@
 
 #include <functional>
 #include <tuple>
+#include <utility>
+#include <vector>
 
 #include <pika/init.hpp>
 
diff --git a/test/unit/test_blas_tile.cpp b/test/unit/test_blas_tile.cpp
index 65d8d6c072..8b43f8711e 100644
--- a/test/unit/test_blas_tile.cpp
+++ b/test/unit/test_blas_tile.cpp
@@ -8,6 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <tuple>
+#include <vector>
+
 #include <dlaf/blas/tile.h>
 
 #include "test_blas_tile/test_gemm.h"
diff --git a/test/unit/test_blas_tile_extensions.cpp b/test/unit/test_blas_tile_extensions.cpp
index 67fe11023d..faf29d6ab4 100644
--- a/test/unit/test_blas_tile_extensions.cpp
+++ b/test/unit/test_blas_tile_extensions.cpp
@@ -8,6 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 
+#include <tuple>
+#include <vector>
+
 #include <dlaf/blas/tile.h>
 
 #include "test_blas_tile/test_add.h"
diff --git a/test/unit/test_lapack_tile.cpp b/test/unit/test_lapack_tile.cpp
index cfd8e0a2e0..fb1ef2e4b7 100644
--- a/test/unit/test_lapack_tile.cpp
+++ b/test/unit/test_lapack_tile.cpp
@@ -8,6 +8,9 @@
 // SPDX-License-Identifier: BSD-3-Clause
 //
 #include <cmath>
+#include <tuple>
+#include <utility>
+#include <vector>
 
 #include <dlaf/lapack/tile.h>
 
diff --git a/test/unit/test_util_math.cpp b/test/unit/test_util_math.cpp
index 588be1369a..ab10a21557 100644
--- a/test/unit/test_util_math.cpp
+++ b/test/unit/test_util_math.cpp
@@ -10,6 +10,8 @@
 
 #include <cstddef>
 #include <limits>
+#include <type_traits>
+#include <vector>
 
 #include <dlaf/util_math.h>