server: remove the verbose_prompt parameter (#21059 )

* server: respect the verbose_prompt parameter * Revert "server: respect the verbose_prompt parameter" This reverts commit 8ed885cf37. * Remove --verbose-prompt parameter from llama-server * Using set_examples instead of set_excludes
mtmd: add more sanity checks (#21047 )
2026-04-30 16:47:31 +03:00 · 2026-03-27 13:36:13 +02:00 · 2026-03-27 11:00:52 +01:00 · 2026-03-27 10:07:11 +01:00 · 2026-03-27 10:59:35 +02:00 · 2026-03-27 09:25:58 +01:00
33 changed files with 1351 additions and 90 deletions
--- a/.devops/cann.Dockerfile
+++ b/.devops/cann.Dockerfile
@@ -4,7 +4,7 @@

 # Define the CANN base image for easier version updates later
 ARG CHIP_TYPE=910b
-ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.3.rc2-${CHIP_TYPE}-openeuler24.03-py3.11
+ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.5.0-${CHIP_TYPE}-openeuler24.03-py3.11

 # ==============================================================================
 # BUILD STAGE
--- a/.devops/llama-cli-cann.Dockerfile
+++ b/.devops/llama-cli-cann.Dockerfile
@@ -1,4 +1,4 @@
-ARG ASCEND_VERSION=8.1.RC1.alpha001-910b-openeuler22.03-py3.10
+ARG ASCEND_VERSION=8.5.0-910b-openeuler22.03-py3.10

 FROM ascendai/cann:$ASCEND_VERSION AS build

--- a/.github/workflows/build-cann.yml
+++ b/.github/workflows/build-cann.yml
@@ -63,7 +63,7 @@ jobs:
      - name: Set container image
        id: cann-image
        run: |
-          image="ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc2-910b-openeuler24.03-py3.11' || '8.3.rc2-310p-openeuler24.03-py3.11' }}"
+          image="ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.5.0-910b-openeuler24.03-py3.11' || '8.5.0-310p-openeuler24.03-py3.11' }}"
          echo "image=${image}" >> "${GITHUB_OUTPUT}"

      - name: Pull container image
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -907,7 +907,7 @@ jobs:
      - name: Set container image
        id: cann-image
        run: |
-          image="ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc2-910b-openeuler24.03-py3.11' || '8.3.rc2-310p-openeuler24.03-py3.11' }}"
+          image="ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.5.0-910b-openeuler24.03-py3.11' || '8.5.0-310p-openeuler24.03-py3.11' }}"
          echo "image=${image}" >> "${GITHUB_OUTPUT}"

      - name: Pull container image
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -1079,7 +1079,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params) {
            params.verbose_prompt = true;
        }
-    ));
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL}));
    add_opt(common_arg(
        {"--display-prompt"},
        {"--no-display-prompt"},
@@ -2843,6 +2843,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.webui_mcp_proxy = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_MCP_PROXY"));
+    add_opt(common_arg(
+        {"--tools"}, "TOOL1,TOOL2,...",
+        "experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)\n"
+        "specify \"all\" to enable all tools\n"
+        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, apply_diff",
+        [](common_params & params, const std::string & value) {
+            params.server_tools = parse_csv_row(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TOOLS"));
    add_opt(common_arg(
        {"--webui"},
        {"--no-webui"},
--- a/common/common.h
+++ b/common/common.h
@@ -613,6 +613,9 @@ struct common_params {
    bool endpoint_props   = false; // only control POST requests, not GET
    bool endpoint_metrics = false;

+    // enable built-in tools
+    std::vector<std::string> server_tools;
+
    // router server configs
    std::string models_dir    = ""; // directory containing models for the router server
    std::string models_preset = ""; // directory containing model presets for the router server
--- a/docs/backend/CANN.md
+++ b/docs/backend/CANN.md
@@ -42,12 +42,22 @@ The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the abi

 ### Ascend NPU

-**Verified devices**
+You can retrieve your Ascend device IDs using the following command:

-| Ascend NPU                    | Status  |
-|:-----------------------------:|:-------:|
-| Atlas 300T A2                 | Support |
-| Atlas 300I Duo                | Support |
+```sh
+lspci -n | grep -Eo '19e5:d[0-9a-f]{3}' | cut -d: -f2
+```
+
+**Devices**
+
+| Device Id | Product Series | Product Models | Chip Model | Verified Status |
+|:---------:|----------------|----------------|:----------:|:---------------:|
+|    d803   | Atlas A3 Train |                |    910C    |                 |
+|    d803   | Atlas A3 Infer |                |    910C    |                 |
+|    d802   | Atlas A2 Train |                |    910B    |                 |
+|    d802   | Atlas A2 Infer | Atlas 300I A2  |    910B    |     Support     |
+|    d801   | Atlas Train    |                |     910    |                 |
+|    d500   | Atlas Infer    | Atlas 300I Duo |    310P    |     Support     |

 *Notes:*

@@ -57,6 +67,9 @@ The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the abi

 ## Model Supports

+<details>
+<summary>Text-only</summary>
+
 | Model Name                  | FP16  | Q4_0 | Q8_0 |
 |:----------------------------|:-----:|:----:|:----:|
 | Llama-2                     |   √   |   √  |   √  |
@@ -118,8 +131,11 @@ The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the abi
 | Trillion-7B-preview         |   √   |   √  |   √  |
 | Ling models                 |   √   |   √  |   √  |

+</details>
+
+<details>
+<summary>Multimodal</summary>

-**Multimodal**
 | Model Name                  | FP16  | Q4_0 | Q8_0 |
 |:----------------------------|:-----:|:----:|:----:|
 | LLaVA 1.5 models, LLaVA 1.6 models      |   x   |   x  |   x  |
@@ -134,15 +150,22 @@ The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the abi
 |  GLM-EDGE                   |   √   |   √  |   √  |
 |  Qwen2-VL                   |   √   |   √  |   √  |

+</details>
+


 ## DataType Supports

-| DataType               | Status  |
-|:----------------------:|:-------:|
-| FP16                   | Support |
-| Q8_0                   | Support |
-| Q4_0                   | Support |
+| DataType               | 910B    | 310P    |
+|:----------------------:|:-------:|:-------:|
+| FP16                   | Support | Support |
+| Q8_0                   | Support | Partial |
+| Q4_0                   | Support | Partial |
+| BF16                   | Support |         |
+
+> **310P note**
+> - `Q8_0`: data transform / buffer path is implemented, and `GET_ROWS` is supported, but quantized `MUL_MAT` / `MUL_MAT_ID` are not supported.
+> - `Q4_0`: data transform / buffer path is implemented, but quantized `MUL_MAT` / `MUL_MAT_ID` are not supported.

 ## Docker

@@ -160,7 +183,20 @@ npu-smi info

 # Select the cards that you want to use, make sure these cards are not used by someone.
 # Following using cards of device0.
-docker run --name llamacpp --device /dev/davinci0  --device /dev/davinci_manager --device /dev/devmm_svm --device /dev/hisi_hdc -v /usr/local/dcmi:/usr/local/dcmi -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info -v /PATH_TO_YOUR_MODELS/:/app/models -it llama-cpp-cann -m /app/models/MODEL_PATH -ngl 32 -p "Building a website can be done in 10 simple steps:"
+docker run --name llamacpp \
+  --device /dev/davinci0 \
+  --device /dev/davinci_manager \
+  --device /dev/devmm_svm \
+  --device /dev/hisi_hdc \
+  -v /usr/local/dcmi:/usr/local/dcmi \
+  -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
+  -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
+  -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
+  -v /PATH_TO_YOUR_MODELS/:/app/models \
+  -it llama-cpp-cann \
+  -m /app/models/MODEL_PATH \
+  -ngl 32 \
+  -p "Building a website can be done in 10 simple steps:"
 ```

 *Notes:*
@@ -171,69 +207,57 @@ docker run --name llamacpp --device /dev/davinci0  --device /dev/davinci_manager

 ### I. Setup Environment

-1. **Install Ascend Driver and firmware**
+1. **Configure Ascend user and group**

    ```sh
-    # create driver running user.
-    sudo groupadd -g HwHiAiUser
+    sudo groupadd HwHiAiUser
    sudo useradd -g HwHiAiUser -d /home/HwHiAiUser -m HwHiAiUser -s /bin/bash
    sudo usermod -aG HwHiAiUser $USER
-
-    # download driver from https://www.hiascend.com/hardware/firmware-drivers/community according to your system
-    # and install driver.
-    sudo sh Ascend-hdk-910b-npu-driver_x.x.x_linux-{arch}.run --full --install-for-all
    ```

-    Once installed, run `npu-smi info` to check whether driver is installed successfully.
+2. **Install dependencies**
+
+    **Ubuntu/Debian:**
    ```sh
-    +-------------------------------------------------------------------------------------------+
-    | npu-smi 24.1.rc2               Version: 24.1.rc2                                          |
-    +----------------------+---------------+----------------------------------------------------+
-    | NPU   Name           | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|
-    | Chip                 | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
-    +======================+===============+====================================================+
-    | 2     xxx            | OK            | 64.4        51                15   / 15            |
-    | 0                    | 0000:01:00.0  | 0           1873 / 15077      0    / 32768         |
-    +======================+===============+====================================================+
-    | 5     xxx            | OK            | 64.0        52                15   / 15            |
-    | 0                    | 0000:81:00.0  | 0           1874 / 15077      0    / 32768         |
-    +======================+===============+====================================================+
-    | No running processes found in NPU 2                                                       |
-    +======================+===============+====================================================+
-    | No running processes found in NPU 5                                                       |
-    +======================+===============+====================================================+
+    sudo apt-get update
+    sudo apt-get install -y gcc python3 python3-pip linux-headers-$(uname -r)
    ```

-2. **Install Ascend Firmware**
+    **RHEL/CentOS:**
    ```sh
-    # download driver from https://www.hiascend.com/hardware/firmware-drivers/community according to your system
-    # and install driver.
-    sudo sh Ascend-hdk-910b-npu-firmware_x.x.x.x.X.run --full
+    sudo yum makecache
+    sudo yum install -y gcc python3 python3-pip kernel-headers-$(uname -r) kernel-devel-$(uname -r)
    ```
-    If the following message appears, firmware is installed successfully.
+
+3. **Install CANN (driver + toolkit)**
+
+    > The `Ascend-cann` package includes both the driver and toolkit.
+    > `$ARCH` can be `x86_64` or `aarch64`, `$CHIP` can be `910b` or `310p`.
+
    ```sh
-    Firmware package installed successfully!
+    wget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/CANN/CANN%208.5.T63/Ascend-cann_8.5.0_linux-$ARCH.run
+    sudo bash ./Ascend-cann_8.5.0_linux-$ARCH.run --install
+
+    wget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/CANN/CANN%208.5.T63/Ascend-cann-$CHIP-ops_8.5.0_linux-$ARCH.run
+    sudo bash ./Ascend-cann-$CHIP-ops_8.5.0_linux-$ARCH.run --install
    ```

+4. **Verify installation**

-3. **Install CANN toolkit and kernels**
-
-    CANN toolkit and kernels can be obtained from the official [CANN Toolkit](https://www.hiascend.com/zh/developer/download/community/result?module=cann) page.
-
-    Please download the corresponding version that satified your system. The minimum version required is 8.0.RC2.alpha002 and here is the install command.
    ```sh
-    pip3 install attrs numpy decorator sympy cffi pyyaml pathlib2 psutil protobuf scipy requests absl-py wheel typing_extensions
-    sh Ascend-cann-toolkit_8.0.RC2.alpha002_linux-aarch64.run --install
-    sh Ascend-cann-kernels-910b_8.0.RC2.alpha002_linux.run --install
+    npu-smi info
    ```

-    Set Ascend Variables:
+    If device information is displayed correctly, the driver is functioning properly.
+
    ```sh
-    echo "source ~/Ascend/ascend-toolkit/set_env.sh" >> ~/.bashrc
-    source ~/.bashrc
+    # Set environment variables (adjust path if needed)
+    source /usr/local/Ascend/cann/set_env.sh
+
+    python3 -c "import acl; print(acl.get_soc_name())"
    ```

-Upon a successful installation, CANN is enabled for the available ascend devices.
+    If the command outputs the chip model, the installation was successful.

 ### II. Build llama.cpp

--- a/ggml/src/ggml-metal/ggml-metal-device.m
+++ b/ggml/src/ggml-metal/ggml-metal-device.m
@@ -690,7 +690,7 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
                    "    auto tB = B.slice((int)tgid.x, 0); \n"
                    " \n"
                    "    matmul2d< \n"
-                    "        matmul2d_descriptor(8, 8, dynamic_extent), \n"
+                    "        matmul2d_descriptor(16, 16, dynamic_extent), \n"
                    "        execution_simdgroups<4>> mm; \n"
                    " \n"
                    "    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>(); \n"
@@ -740,7 +740,7 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
                    "    auto tB = B.slice((int)tgid.x, 0); \n"
                    " \n"
                    "    matmul2d< \n"
-                    "        matmul2d_descriptor(8, 8, dynamic_extent), \n"
+                    "        matmul2d_descriptor(16, 16, dynamic_extent), \n"
                    "        execution_simdgroups<4>> mm; \n"
                    " \n"
                    "    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>(); \n"
--- a/ggml/src/ggml-rpc/ggml-rpc.cpp
+++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
@@ -589,8 +589,10 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
        ggml_backend_buffer_t buffer = tensor->buffer;
        ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
        result.buffer = ctx != nullptr ? ctx->remote_ptr : 0;
+        result.data = reinterpret_cast<uint64_t>(tensor->data);
    } else {
        result.buffer = 0;
+        result.data   = 0;
    }
    for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
        result.ne[i] = tensor->ne[i];
@@ -606,7 +608,6 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
    }
    result.view_src = reinterpret_cast<uint64_t>(tensor->view_src);
    result.view_offs = tensor->view_offs;
-    result.data = reinterpret_cast<uint64_t>(tensor->data);

    // Avoid sending uninitialized data over the wire
    memset(result.name, 0, sizeof(result.name));
@@ -1443,9 +1444,11 @@ ggml_tensor * rpc_server::create_node(uint64_t id,
    const rpc_tensor * tensor = it_ptr->second;

    struct ggml_tensor * result = deserialize_tensor(ctx, tensor);
-    if (result == nullptr || result->buffer == nullptr) {
-        GGML_LOG_ERROR("[%s] invalid tensor: null %s (id=%" PRIu64 ")\n",
-                       __func__, result == nullptr ? "tensor" : "buffer", id);
+    if (result == nullptr) {
+        return nullptr;
+    }
+    if (result->buffer == nullptr && result->data != nullptr) {
+        GGML_LOG_ERROR("[%s] invalid data ptr", __func__);
        return nullptr;
    }
    tensor_map[id] = result;
--- a/tools/completion/completion.cpp
+++ b/tools/completion/completion.cpp
@@ -146,13 +146,19 @@ int main(int argc, char ** argv) {

    ctx   = llama_init->context();
    model = llama_init->model();
-    smpl  = llama_init->sampler(0);

    if (ctx == NULL) {
        LOG_ERR("%s: error: unable to create context\n", __func__);
        return 1;
    }

+    if (model == NULL) {
+        LOG_ERR("%s: error: unable to load model\n", __func__);
+        return 1;
+    }
+
+    smpl = llama_init->sampler(0);
+
    llama_memory_t mem = llama_get_memory(ctx);
    const llama_vocab * vocab = llama_model_get_vocab(model);

@@ -695,7 +701,7 @@ int main(int argc, char ** argv) {
                if (!common_prompt_batch_decode(ctx, embd, n_past, params.n_batch, path_session, save_now)) {
                    return 1;
                }
-                session_tokens.insert(session_tokens.end(), embd.begin(), embd.begin());
+                session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
                n_session_consumed = session_tokens.size();
                session_do_save = false;

--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -1377,6 +1377,16 @@ struct clip_model_loader {

            // sanity check
            {
+                if (hparams.image_size < 0) {
+                    // note: some models having hparams.image_size == 0, which means the image size is dynamic
+                    throw std::runtime_error(string_format("%s: image_size (%d) cannot be negative\n", __func__, hparams.image_size));
+                }
+                if (hparams.patch_size <= 0) {
+                    throw std::runtime_error(string_format("%s: patch_size (%d) must be greater than 0\n", __func__, hparams.patch_size));
+                }
+                if (hparams.n_embd <= 0) {
+                    throw std::runtime_error(string_format("%s: n_embd (%d) must be greater than 0\n", __func__, hparams.n_embd));
+                }
                if (hparams.image_max_pixels < hparams.image_min_pixels) {
                    throw std::runtime_error(string_format("%s: image_max_pixels (%d) is less than image_min_pixels (%d)\n", __func__, hparams.image_max_pixels, hparams.image_min_pixels));
                }
--- a/tools/mtmd/mtmd-audio.cpp
+++ b/tools/mtmd/mtmd-audio.cpp
@@ -13,23 +13,20 @@

 constexpr bool DEBUG = false;

-void mtmd_audio_cache::fill_sin_cos_table(int n) {
+void mtmd_audio_cache::fill_sin_cos_table(uint32_t n) {
    sin_vals.resize(n);
    cos_vals.resize(n);
-    for (int i = 0; i < n; i++) {
+    for (uint32_t i = 0; i < n; i++) {
        double theta = (2 * M_PI * i) / n;
        sin_vals[i]  = sinf(theta);
        cos_vals[i]  = cosf(theta);
    }
 }

-void mtmd_audio_cache::fill_hann_window(int length, bool periodic) {
+void mtmd_audio_cache::fill_hann_window(uint32_t length, bool periodic) {
    hann_window.resize(length);
-    int offset = -1;
-    if (periodic) {
-        offset = 0;
-    }
-    for (int i = 0; i < length; i++) {
+    int offset = periodic ? 0 : -1;
+    for (uint32_t i = 0; i < length; i++) {
        hann_window[i] = 0.5 * (1.0 - cosf((2.0 * M_PI * i) / (length + offset)));
    }
 }
@@ -165,6 +162,7 @@ static void dft_impl(const mtmd_audio_cache & cache, const float * in, int N, fl
 //              false = input is complex-valued (interleaved real/imag, stride 2)
 template <bool Inverse, bool RealInput>
 static void fft_impl(const mtmd_audio_cache & cache, float * in, int N, float * out) {
+    GGML_ASSERT(N > 0);
    const int n_sin_cos_vals = cache.sin_vals.size();

    if (N == 1) {
@@ -407,6 +405,8 @@ static bool log_mel_spectrogram(
    }


+    GGML_ASSERT(params.n_fft_bins > 0);
+    GGML_ASSERT(params.hop_length > 0);
    out.n_mel = params.n_mel;
    out.n_len = (n_samples - frame_size) / frame_step + 1;
    // TODO: handle these checks better
@@ -438,6 +438,7 @@ static bool log_mel_spectrogram(

    const int effective_n_len = n_samples_in / frame_step;
    if (params.norm_per_feature) {
+        GGML_ASSERT(effective_n_len > 1);
        for (int i = 0; i < out.n_mel; i++) {
            double mean = 0;
            for (int j = 0; j < effective_n_len; ++j) {
@@ -639,6 +640,7 @@ mtmd_audio_streaming_istft::mtmd_audio_streaming_istft(int n_fft, int hop_length
    padding_to_remove((n_fft - hop_length) / 2),
    ifft_in(n_fft * 2 * 4, 0.0f),  // extra space for recursive IFFT
    ifft_out(n_fft * 2 * 4, 0.0f) {
+    GGML_ASSERT(n_fft > 0 && hop_length > 0 && hop_length <= n_fft);
    cache.fill_sin_cos_table(n_fft);
    cache.fill_hann_window(n_fft, true);
 }
--- a/tools/mtmd/mtmd-audio.h
+++ b/tools/mtmd/mtmd-audio.h
@@ -33,9 +33,9 @@ struct mtmd_audio_cache {

    mtmd_audio_mel_filters filters;

-    void fill_sin_cos_table(int n);
+    void fill_sin_cos_table(uint32_t n);

-    void fill_hann_window(int length, bool periodic);
+    void fill_hann_window(uint32_t length, bool periodic);

    // Build mel filterbank matrix [n_mel × n_fft_bins] at runtime.
    // n_fft_bins must be (N_fft / 2 + 1). Example: if N_fft=512 -> n_fft_bins=257.
--- a/tools/mtmd/mtmd-helper.cpp
+++ b/tools/mtmd/mtmd-helper.cpp
@@ -127,6 +127,7 @@ struct decode_embd_batch {
    std::vector<int8_t>         logits;
    llama_batch batch;
    decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
+        GGML_ASSERT(n_tokens > 0 && n_pos_per_embd > 0 && n_mmproj_embd > 0);
        pos     .resize(n_tokens * n_pos_per_embd);
        n_seq_id.resize(n_tokens);
        seq_ids .resize(n_tokens + 1);
@@ -157,6 +158,7 @@ struct decode_embd_batch {
    // M-RoPE for image
    void set_position_mrope_2d(llama_pos pos_0, int nx, int ny, llama_seq_id seq_id) {
        GGML_ASSERT(n_pos_per_embd == 4);
+        GGML_ASSERT(nx > 0 && ny > 0 && nx * ny == batch.n_tokens);
        seq_id_0[0] = seq_id;
        for (int y = 0; y < ny; y++) {
            for (int x = 0; x < nx; x++) {
@@ -192,6 +194,7 @@ struct decode_embd_batch {
    }

    llama_batch get_view(int offset, int n_tokens) {
+        GGML_ASSERT(offset >= 0 && n_tokens > 0 && offset + n_tokens <= batch.n_tokens);
        llama_pos * pos_ptr;
        pos_view.clear();
        pos_view.reserve(n_tokens * n_pos_per_embd);
@@ -235,6 +238,7 @@ int32_t mtmd_helper_decode_image_chunk(
        llama_seq_id seq_id,
        int32_t n_batch,
        llama_pos * new_n_past) {
+    GGML_ASSERT(n_batch > 0);
    auto chunk_type = mtmd_input_chunk_get_type(chunk);
    const char * name = chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE ? "image" : "audio";
    if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
@@ -312,6 +316,7 @@ int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
        int32_t n_batch,
        bool logits_last,
        llama_pos * new_n_past) {
+    GGML_ASSERT(n_batch > 0);
    int32_t ret;
    llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
    auto chunk_type = mtmd_input_chunk_get_type(chunk);
@@ -508,6 +513,11 @@ mtmd_bitmap * mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char *
    fseek(f, 0, SEEK_END);
    long file_size = ftell(f);
    fseek(f, 0, SEEK_SET);
+    if (file_size < 0) {
+        LOG_ERR("Failed to get file size of %s\n", fname);
+        fclose(f);
+        return nullptr;
+    }
    buf.resize(file_size);

    size_t n_read = fread(buf.data(), 1, file_size, f);
--- a/tools/mtmd/mtmd-image.cpp
+++ b/tools/mtmd/mtmd-image.cpp
@@ -99,6 +99,8 @@ struct img_tool {
    }

    static void crop(const clip_image_u8 & image, clip_image_u8 & dst, int x, int y, int w, int h) {
+        GGML_ASSERT(x >= 0 && y >= 0 && w > 0 && h > 0);
+        GGML_ASSERT(x + w <= image.nx && y + h <= image.ny);
        dst.nx = w;
        dst.ny = h;
        dst.buf.resize(3 * w * h);
@@ -196,6 +198,7 @@ struct img_tool {
 private:
    // Bilinear resize function
    static void resize_bilinear(const clip_image_u8 & src, clip_image_u8 & dst, int target_width, int target_height) {
+        GGML_ASSERT(src.nx >= 2 && src.ny >= 2);
        dst.nx = target_width;
        dst.ny = target_height;
        dst.buf.resize(3 * target_width * target_height);
@@ -207,8 +210,8 @@ private:
            for (int x = 0; x < target_width; x++) {
                float px = x_ratio * x;
                float py = y_ratio * y;
-                int x_floor = static_cast<int>(px);
-                int y_floor = static_cast<int>(py);
+                int x_floor = std::min(static_cast<int>(px), src.nx - 2);
+                int y_floor = std::min(static_cast<int>(py), src.ny - 2);
                float x_lerp = px - x_floor;
                float y_lerp = py - y_floor;

@@ -347,6 +350,7 @@ private:
        // Returns: kernel size (ksize) - number of input pixels that contribute to each output pixel
        auto precompute_weights = [&](int inSize, int outSize,
                                     std::vector<int> & bounds, std::vector<int32_t> & weights) -> int {
+            GGML_ASSERT(inSize > 0 && outSize > 0);
            double support, scale, filterscale;
            double center, ww, ss;
            int xx, x, ksize, xmin, xmax, xcnt;
--- a/tools/mtmd/mtmd.cpp
+++ b/tools/mtmd/mtmd.cpp
@@ -641,6 +641,11 @@ struct mtmd_tokenizer {
                add_text(ctx->img_beg, true); // add image begin token
            }

+            // sanity check
+            GGML_ASSERT(bitmap->nx > 0 && bitmap->ny > 0);
+            GGML_ASSERT(bitmap->data.size() == (size_t)bitmap->nx * bitmap->ny * 3);
+            GGML_ASSERT(ctx->image_preproc != nullptr);
+
            // convert mtmd_bitmap to clip_image_u8
            clip_image_u8_ptr img_u8(clip_image_u8_init());
            img_u8->nx = bitmap->nx;
@@ -649,7 +654,6 @@ struct mtmd_tokenizer {
            std::memcpy(img_u8->buf.data(), bitmap->data.data(), img_u8->nx * img_u8->ny * 3);

            // preprocess image
-            GGML_ASSERT(ctx->image_preproc != nullptr);
            clip_image_f32_batch batch_f32;
            bool ok = ctx->image_preproc->preprocess(*img_u8, batch_f32);
            if (!ok) {
@@ -773,6 +777,11 @@ struct mtmd_tokenizer {
                add_text(ctx->aud_beg, true); // add audio begin token
            }

+            // sanity check
+            GGML_ASSERT(ctx->audio_preproc != nullptr);
+            GGML_ASSERT(bitmap->data.size() > sizeof(float));
+            GGML_ASSERT(bitmap->data.size() % sizeof(float) == 0);
+
            // preprocess audio
            std::vector<mtmd_audio_mel> mel_spec_chunks;
            const float * samples = (const float *)bitmap->data.data();
--- a/tools/server/CMakeLists.txt
+++ b/tools/server/CMakeLists.txt
@@ -13,6 +13,8 @@ add_library(${TARGET} STATIC
    server-common.h
    server-context.cpp
    server-context.h
+    server-tools.cpp
+    server-tools.h
 )

 if (BUILD_SHARED_LIBS)
--- a/tools/server/README-dev.md
+++ b/tools/server/README-dev.md
@@ -125,6 +125,61 @@ The framework automatically starts a `llama-server` instance, sends requests, an

 For detailed instructions, see the [test documentation](./tests/README.md).

+### API for tools
+
+This endpoint is intended to be used internally by the Web UI and subject to change or to be removed in the future.
+
+**GET /tools**
+
+Get a list of tools, each tool has these fields:
+- `tool` (string): the ID name of the tool, to be used in POST call. Example: `read_file`
+- `display_name` (string): the name to be displayed on UI. Example: `Read file`
+- `type` (string): always be `"builtin"` for now
+- `permissions` (object): a mapping string --> boolean that indicates the permission required by this tool. This is useful for the UI to ask the user before calling the tool. For now, the only permission supported is `"write"`
+- `definition` (object): the OAI-compat definition of this tool
+
+**POST /tools**
+
+Invoke a tool call, request body is a JSON object with:
+- `tool` (string): the name of the tool
+- `params` (object): a mapping from argument name (string) to argument value
+
+Returns JSON object. There are two response formats:
+
+Format 1: Plain text. The text will be placed into a field called `plain_text_response`, example:
+
+```json
+{
+    "plain_text_response": "this is a text response"
+}
+```
+
+The client should extract this value and place it inside message content (note: content is no longer a JSON), example
+
+```json
+{
+    "role": "tool",
+    "content": "this is a text response"
+}
+```
+
+Format 2: Normal JSON response, example:
+
+```json
+{
+    "error": "cannot open this file"
+}
+```
+
+That requires `JSON.stringify` when formatted to message content:
+
+```json
+{
+    "role": "tool",
+    "content": "{\"error\":\"cannot open this file\"}"
+}
+```
+
 ### Notable Related PRs

 - Initial server implementation: https://github.com/ggml-org/llama.cpp/pull/1443
--- a/tools/server/README.md
+++ b/tools/server/README.md
@@ -36,7 +36,6 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `--license` | show source code license and dependencies |
 | `-cl, --cache-list` | show list of models in cache |
 | `--completion-bash` | print source-able bash completion script for llama.cpp |
-| `--verbose-prompt` | print a verbose prompt before generation (default: false) |
 | `-t, --threads N` | number of CPU threads to use during generation (default: -1)<br/>(env: LLAMA_ARG_THREADS) |
 | `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) |
 | `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") |
@@ -194,6 +193,7 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `--webui-config JSON` | JSON that provides default WebUI settings (overrides WebUI defaults)<br/>(env: LLAMA_ARG_WEBUI_CONFIG) |
 | `--webui-config-file PATH` | JSON file that provides default WebUI settings (overrides WebUI defaults)<br/>(env: LLAMA_ARG_WEBUI_CONFIG_FILE) |
 | `--webui-mcp-proxy, --no-webui-mcp-proxy` | experimental: whether to enable MCP CORS proxy - do not enable in untrusted environments (default: disabled)<br/>(env: LLAMA_ARG_WEBUI_MCP_PROXY) |
+| `--tools TOOL1,TOOL2,...` | experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)<br/>specify "all" to enable all tools<br/>available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, apply_diff<br/>(env: LLAMA_ARG_TOOLS) |
 | `--webui, --no-webui` | whether to enable the Web UI (default: enabled)<br/>(env: LLAMA_ARG_WEBUI) |
 | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
 | `--rerank, --reranking` | enable reranking endpoint on server (default: disabled)<br/>(env: LLAMA_ARG_RERANKING) |
@@ -293,6 +293,12 @@ It is currently available in the following endpoints:

 For more details, please refer to [multimodal documentation](../../docs/multimodal.md)

+### Built-in tools support
+
+The server includes a set of built-in tools that enable the LLM to access the local file system directly from the Web UI.
+
+To use this feature, start the server with `--tools all`. You can also enable only specific tools by passing a comma-separated list: `--tools name1,name2,...`. Run `--help` for the full list of available tool names.
+
 ## Build

 `llama-server` is built alongside everything else from the root of the project
@@ -1438,6 +1444,14 @@ curl http://localhost:8080/v1/messages/count_tokens \
 {"input_tokens": 10}
 ```

+## Server built-in tools
+
+The server exposes a REST API under `/tools` that allows the Web UI to call built-in tools. This endpoint is intended to be used internally by the Web UI and subject to change or to be removed in the future.
+
+**Please do NOT use this endpoint in a downstream application**
+
+For further documentation about this endpoint, please refer to [server internal documentation](./README-dev.md)
+
 ## Using multiple models

 `llama-server` can be launched in a **router mode** that exposes an API for dynamically loading and unloading models. The main process (the "router") automatically forwards each request to the appropriate model instance.
--- a/tools/server/public/index.html.gz
+++ b/tools/server/public/index.html.gz
--- a/tools/server/server-tools.cpp
+++ b/tools/server/server-tools.cpp
@@ -0,0 +1,800 @@
+#include "server-tools.h"
+
+#include <sheredom/subprocess.h>
+
+#include <filesystem>
+#include <fstream>
+#include <regex>
+#include <thread>
+#include <chrono>
+#include <atomic>
+#include <cstring>
+#include <climits>
+
+namespace fs = std::filesystem;
+
+//
+// internal helpers
+//
+
+static std::vector<char *> to_cstr_vec(const std::vector<std::string> & v) {
+    std::vector<char *> r;
+    r.reserve(v.size() + 1);
+    for (const auto & s : v) {
+        r.push_back(const_cast<char *>(s.c_str()));
+    }
+    r.push_back(nullptr);
+    return r;
+}
+
+struct run_proc_result {
+    std::string output;
+    int  exit_code = -1;
+    bool timed_out = false;
+};
+
+static run_proc_result run_process(
+        const std::vector<std::string> & args,
+        size_t max_output,
+        int timeout_secs) {
+    run_proc_result res;
+
+    subprocess_s proc;
+    auto argv = to_cstr_vec(args);
+
+    int options = subprocess_option_no_window
+                | subprocess_option_combined_stdout_stderr
+                | subprocess_option_inherit_environment
+                | subprocess_option_search_user_path;
+
+    if (subprocess_create(argv.data(), options, &proc) != 0) {
+        res.output = "failed to spawn process";
+        return res;
+    }
+
+    std::atomic<bool> done{false};
+    std::atomic<bool> timed_out{false};
+
+    std::thread timeout_thread([&]() {
+        auto deadline = std::chrono::steady_clock::now() + std::chrono::seconds(timeout_secs);
+        while (!done.load()) {
+            if (std::chrono::steady_clock::now() >= deadline) {
+                timed_out.store(true);
+                subprocess_terminate(&proc);
+                return;
+            }
+            std::this_thread::sleep_for(std::chrono::milliseconds(100));
+        }
+    });
+
+    FILE * f = subprocess_stdout(&proc);
+    std::string output;
+    bool truncated = false;
+    if (f) {
+        char buf[4096];
+        while (fgets(buf, sizeof(buf), f) != nullptr) {
+            if (!truncated) {
+                size_t len = strlen(buf);
+                if (output.size() + len <= max_output) {
+                    output.append(buf, len);
+                } else {
+                    output.append(buf, max_output - output.size());
+                    truncated = true;
+                }
+            }
+        }
+    }
+
+    done.store(true);
+    if (timeout_thread.joinable()) {
+        timeout_thread.join();
+    }
+
+    subprocess_join(&proc, &res.exit_code);
+    subprocess_destroy(&proc);
+
+    res.output    = output;
+    res.timed_out = timed_out.load();
+    if (truncated) {
+        res.output += "\n[output truncated]";
+    }
+    return res;
+}
+
+// simple glob: * matches non-/ chars, ** matches anything including /
+static bool glob_match(const char * pattern, const char * str) {
+    if (*pattern == '\0') {
+        return *str == '\0';
+    }
+    if (pattern[0] == '*' && pattern[1] == '*') {
+        const char * p = pattern + 2;
+        if (*p == '/') p++;
+        if (glob_match(p, str)) return true;
+        if (*str != '\0') return glob_match(pattern, str + 1);
+        return false;
+    }
+    if (*pattern == '*') {
+        const char * p = pattern + 1;
+        for (; *str != '\0' && *str != '/'; str++) {
+            if (glob_match(p, str)) return true;
+        }
+        return glob_match(p, str);
+    }
+    if (*pattern == '?' && *str != '\0' && *str != '/') {
+        return glob_match(pattern + 1, str + 1);
+    }
+    if (*pattern == *str) {
+        return glob_match(pattern + 1, str + 1);
+    }
+    return false;
+}
+
+static bool glob_match(const std::string & pattern, const std::string & str) {
+    return glob_match(pattern.c_str(), str.c_str());
+}
+
+json server_tool::to_json() {
+    return {
+        {"display_name", display_name},
+        {"tool", name},
+        {"type", "builtin"},
+        {"permissions", json{
+            {"write", permission_write}
+        }},
+        {"definition", get_definition()},
+    };
+}
+
+//
+// read_file: read a file with optional line range and line-number prefix
+//
+
+static constexpr size_t SERVER_TOOL_READ_FILE_MAX_SIZE = 16 * 1024; // 16 KB
+
+struct server_tool_read_file : server_tool {
+    server_tool_read_file() {
+        name = "read_file";
+        display_name = "Read file";
+        permission_write = false;
+    }
+
+    json get_definition() override {
+        return {
+            {"type", "function"},
+            {"function", {
+                {"name", name},
+                {"description", "Read the contents of a file. Optionally specify a 1-based line range. "
+                                "If append_loc is true, each line is prefixed with its line number (e.g. \"1\u2192 ...\")."},
+                {"parameters", {
+                    {"type", "object"},
+                    {"properties", {
+                        {"path",       {{"type", "string"},  {"description", "Path to the file"}}},
+                        {"start_line", {{"type", "integer"}, {"description", "First line to read, 1-based (default: 1)"}}},
+                        {"end_line",   {{"type", "integer"}, {"description", "Last line to read, 1-based inclusive (default: end of file)"}}},
+                        {"append_loc", {{"type", "boolean"}, {"description", "Prefix each line with its line number"}}},
+                    }},
+                    {"required", json::array({"path"})},
+                }},
+            }},
+        };
+    }
+
+    json invoke(json params) override {
+        std::string path  = params.at("path").get<std::string>();
+        int  start_line   = json_value(params, "start_line", 1);
+        int  end_line     = json_value(params, "end_line",  -1); // -1 = no limit
+        bool append_loc   = json_value(params, "append_loc", false);
+
+        std::error_code ec;
+        uintmax_t file_size = fs::file_size(path, ec);
+        if (ec) {
+            return {{"error", "cannot stat file: " + ec.message()}};
+        }
+        if (file_size > SERVER_TOOL_READ_FILE_MAX_SIZE && end_line == -1) {
+            return {{"error", string_format(
+                "file too large (%zu bytes, max %zu). Use start_line/end_line to read a portion.",
+                (size_t)file_size, SERVER_TOOL_READ_FILE_MAX_SIZE)}};
+        }
+
+        std::ifstream f(path);
+        if (!f) {
+            return {{"error", "failed to open file: " + path}};
+        }
+
+        std::string result;
+        std::string line;
+        int lineno = 0;
+
+        while (std::getline(f, line)) {
+            lineno++;
+            if (lineno < start_line) continue;
+            if (end_line != -1 && lineno > end_line) break;
+
+            std::string out_line;
+            if (append_loc) {
+                out_line = std::to_string(lineno) + "\u2192 " + line + "\n";
+            } else {
+                out_line = line + "\n";
+            }
+
+            if (result.size() + out_line.size() > SERVER_TOOL_READ_FILE_MAX_SIZE) {
+                result += "[output truncated]";
+                break;
+            }
+            result += out_line;
+        }
+
+        return {{"plain_text_response", result}};
+    }
+};
+
+//
+// file_glob_search: find files matching a glob pattern under a base directory
+//
+
+static constexpr size_t SERVER_TOOL_FILE_SEARCH_MAX_RESULTS = 100;
+
+struct server_tool_file_glob_search : server_tool {
+    server_tool_file_glob_search() {
+        name = "file_glob_search";
+        display_name = "File search";
+        permission_write = false;
+    }
+
+    json get_definition() override {
+        return {
+            {"type", "function"},
+            {"function", {
+                {"name", name},
+                {"description", "Recursively search for files matching a glob pattern under a directory."},
+                {"parameters", {
+                    {"type", "object"},
+                    {"properties", {
+                        {"path",    {{"type", "string"}, {"description", "Base directory to search in"}}},
+                        {"include", {{"type", "string"}, {"description", "Glob pattern for files to include (e.g. \"**/*.cpp\"). Default: **"}}},
+                        {"exclude", {{"type", "string"}, {"description", "Glob pattern for files to exclude"}}},
+                    }},
+                    {"required", json::array({"path"})},
+                }},
+            }},
+        };
+    }
+
+    json invoke(json params) override {
+        std::string base    = params.at("path").get<std::string>();
+        std::string include = json_value(params, "include", std::string("**"));
+        std::string exclude = json_value(params, "exclude", std::string(""));
+
+        std::ostringstream output_text;
+        size_t count = 0;
+
+        std::error_code ec;
+        for (const auto & entry : fs::recursive_directory_iterator(base,
+                fs::directory_options::skip_permission_denied, ec)) {
+            if (!entry.is_regular_file()) continue;
+
+            std::string rel = fs::relative(entry.path(), base, ec).string();
+            if (ec) continue;
+            std::replace(rel.begin(), rel.end(), '\\', '/');
+
+            if (!glob_match(include, rel)) continue;
+            if (!exclude.empty() && glob_match(exclude, rel)) continue;
+
+            output_text << entry.path().string() << "\n";
+            if (++count >= SERVER_TOOL_FILE_SEARCH_MAX_RESULTS) {
+                break;
+            }
+        }
+
+        output_text << "\n---\nTotal matches: " << count << "\n";
+
+        return {{"plain_text_response", output_text.str()}};
+    }
+};
+
+//
+// grep_search: search for a regex pattern in files
+//
+
+static constexpr size_t SERVER_TOOL_GREP_SEARCH_MAX_RESULTS = 100;
+
+struct server_tool_grep_search : server_tool {
+    server_tool_grep_search() {
+        name = "grep_search";
+        display_name = "Grep search";
+        permission_write = false;
+    }
+
+    json get_definition() override {
+        return {
+            {"type", "function"},
+            {"function", {
+                {"name", name},
+                {"description", "Search for a regex pattern in files under a path. Returns matching lines."},
+                {"parameters", {
+                    {"type", "object"},
+                    {"properties", {
+                        {"path",                {{"type", "string"},  {"description", "File or directory to search in"}}},
+                        {"pattern",             {{"type", "string"},  {"description", "Regular expression pattern to search for"}}},
+                        {"include",             {{"type", "string"},  {"description", "Glob pattern to filter files (default: **)"}}},
+                        {"exclude",             {{"type", "string"},  {"description", "Glob pattern to exclude files"}}},
+                        {"return_line_numbers", {{"type", "boolean"}, {"description", "If true, include line numbers in results"}}},
+                    }},
+                    {"required", json::array({"path", "pattern"})},
+                }},
+            }},
+        };
+    }
+
+    json invoke(json params) override {
+        std::string path    = params.at("path").get<std::string>();
+        std::string pat_str = params.at("pattern").get<std::string>();
+        std::string include = json_value(params, "include", std::string("**"));
+        std::string exclude = json_value(params, "exclude", std::string(""));
+        bool show_lineno    = json_value(params, "return_line_numbers", false);
+
+        std::regex pattern;
+        try {
+            pattern = std::regex(pat_str);
+        } catch (const std::regex_error & e) {
+            return {{"error", std::string("invalid regex: ") + e.what()}};
+        }
+
+        std::ostringstream output_text;
+        size_t total = 0;
+
+        auto search_file = [&](const fs::path & fpath) {
+            std::ifstream f(fpath);
+            if (!f) return;
+            std::string line;
+            int lineno = 0;
+            while (std::getline(f, line) && total < SERVER_TOOL_GREP_SEARCH_MAX_RESULTS) {
+                lineno++;
+                if (std::regex_search(line, pattern)) {
+                    output_text << fpath.string() << ":";
+                    if (show_lineno) {
+                        output_text << lineno << ":";
+                    }
+                    output_text << line << "\n";
+                    total++;
+                }
+            }
+        };
+
+        std::error_code ec;
+        if (fs::is_regular_file(path, ec)) {
+            search_file(path);
+        } else if (fs::is_directory(path, ec)) {
+            for (const auto & entry : fs::recursive_directory_iterator(path,
+                    fs::directory_options::skip_permission_denied, ec)) {
+                if (!entry.is_regular_file()) continue;
+                if (total >= SERVER_TOOL_GREP_SEARCH_MAX_RESULTS) break;
+
+                std::string rel = fs::relative(entry.path(), path, ec).string();
+                if (ec) continue;
+                std::replace(rel.begin(), rel.end(), '\\', '/');
+
+                if (!glob_match(include, rel)) continue;
+                if (!exclude.empty() && glob_match(exclude, rel)) continue;
+
+                search_file(entry.path());
+            }
+        } else {
+            return {{"error", "path does not exist: " + path}};
+        }
+
+        output_text << "\n\n---\nTotal matches: " << total << "\n";
+
+        return {{"plain_text_response", output_text.str()}};
+    }
+};
+
+//
+// exec_shell_command: run an arbitrary shell command
+//
+
+static constexpr size_t SERVER_TOOL_EXEC_SHELL_COMMAND_MAX_OUTPUT_SIZE = 16 * 1024; // 16 KB
+static constexpr int    SERVER_TOOL_EXEC_SHELL_COMMAND_MAX_TIMEOUT     = 60;        // seconds
+
+struct server_tool_exec_shell_command : server_tool {
+    server_tool_exec_shell_command() {
+        name = "exec_shell_command";
+        display_name = "Execute shell command";
+        permission_write = true;
+    }
+
+    json get_definition() override {
+        return {
+            {"type", "function"},
+            {"function", {
+                {"name", name},
+                {"description", "Execute a shell command and return its output (stdout and stderr combined)."},
+                {"parameters", {
+                    {"type", "object"},
+                    {"properties", {
+                        {"command",         {{"type", "string"},  {"description", "Shell command to execute"}}},
+                        {"timeout",         {{"type", "integer"}, {"description", string_format("Timeout in seconds (default 10, max %d)", SERVER_TOOL_EXEC_SHELL_COMMAND_MAX_TIMEOUT)}}},
+                        {"max_output_size", {{"type", "integer"}, {"description", string_format("Maximum output size in bytes (default %zu)", SERVER_TOOL_EXEC_SHELL_COMMAND_MAX_OUTPUT_SIZE)}}},
+                    }},
+                    {"required", json::array({"command"})},
+                }},
+            }},
+        };
+    }
+
+    json invoke(json params) override {
+        std::string command   = params.at("command").get<std::string>();
+        int    timeout        = json_value(params, "timeout",         10);
+        size_t max_output     = (size_t) json_value(params, "max_output_size", (int) SERVER_TOOL_EXEC_SHELL_COMMAND_MAX_OUTPUT_SIZE);
+
+        timeout    = std::min(timeout,    SERVER_TOOL_EXEC_SHELL_COMMAND_MAX_TIMEOUT);
+        max_output = std::min(max_output, SERVER_TOOL_EXEC_SHELL_COMMAND_MAX_OUTPUT_SIZE);
+
+#ifdef _WIN32
+        std::vector<std::string> args = {"cmd", "/c", command};
+#else
+        std::vector<std::string> args = {"sh", "-c", command};
+#endif
+
+        auto res = run_process(args, max_output, timeout);
+
+        std::string text_output = res.output;
+        text_output += string_format("\n[exit code: %d]", res.exit_code);
+        if (res.timed_out) {
+            text_output += " [exit due to timed out]";
+        }
+
+        return {{"plain_text_response", text_output}};
+    }
+};
+
+//
+// write_file: create or overwrite a file
+//
+
+struct server_tool_write_file : server_tool {
+    server_tool_write_file() {
+        name = "write_file";
+        display_name = "Write file";
+        permission_write = true;
+    }
+
+    json get_definition() override {
+        return {
+            {"type", "function"},
+            {"function", {
+                {"name", name},
+                {"description", "Write content to a file, creating it (including parent directories) if it does not exist. May use with edit_file for more complex edits."},
+                {"parameters", {
+                    {"type", "object"},
+                    {"properties", {
+                        {"path",    {{"type", "string"}, {"description", "Path of the file to write"}}},
+                        {"content", {{"type", "string"}, {"description", "Content to write"}}},
+                    }},
+                    {"required", json::array({"path", "content"})},
+                }},
+            }},
+        };
+    }
+
+    json invoke(json params) override {
+        std::string path    = params.at("path").get<std::string>();
+        std::string content = params.at("content").get<std::string>();
+
+        std::error_code ec;
+        fs::path fpath(path);
+        if (fpath.has_parent_path()) {
+            fs::create_directories(fpath.parent_path(), ec);
+            if (ec) {
+                return {{"error", "failed to create directories: " + ec.message()}};
+            }
+        }
+
+        std::ofstream f(path, std::ios::binary);
+        if (!f) {
+            return {{"error", "failed to open file for writing: " + path}};
+        }
+        f << content;
+        if (!f) {
+            return {{"error", "failed to write file: " + path}};
+        }
+
+        return {{"result", "file written successfully"}, {"path", path}, {"bytes", content.size()}};
+    }
+};
+
+//
+// edit_file: edit file content via line-based changes
+//
+
+struct server_tool_edit_file : server_tool {
+    server_tool_edit_file() {
+        name = "edit_file";
+        display_name = "Edit file";
+        permission_write = true;
+    }
+
+    json get_definition() override {
+        return {
+            {"type", "function"},
+            {"function", {
+                {"name", name},
+                {"description",
+                    "Edit a file by applying a list of line-based changes. "
+                    "Each change targets a 1-based inclusive line range and has a mode: "
+                    "\"replace\" (replace lines with content), "
+                    "\"delete\" (remove lines, content must be empty string), "
+                    "\"append\" (insert content after line_end). "
+                    "Set line_start to -1 to target the end of file (line_end is ignored in that case). "
+                    "Changes must not overlap. They are applied in reverse line order automatically."},
+                {"parameters", {
+                    {"type", "object"},
+                    {"properties", {
+                        {"path",    {{"type", "string"}, {"description", "Path to the file to edit"}}},
+                        {"changes", {
+                            {"type", "array"},
+                            {"description", "List of changes to apply"},
+                            {"items", {
+                                {"type", "object"},
+                                {"properties", {
+                                    {"mode",       {{"type", "string"},  {"description", "\"replace\", \"delete\", or \"append\""}}},
+                                    {"line_start", {{"type", "integer"}, {"description", "First line of the range (1-based); use -1 for end of file"}}},
+                                    {"line_end",   {{"type", "integer"}, {"description", "Last line of the range (1-based, inclusive); ignored when line_start is -1"}}},
+                                    {"content",    {{"type", "string"},  {"description", "Content to insert; must be empty string for delete mode"}}},
+                                }},
+                                {"required", json::array({"mode", "line_start", "line_end", "content"})},
+                            }},
+                        }},
+                    }},
+                    {"required", json::array({"path", "changes"})},
+                }},
+            }},
+        };
+    }
+
+    json invoke(json params) override {
+        std::string path = params.at("path").get<std::string>();
+        const json & changes = params.at("changes");
+
+        if (!changes.is_array()) {
+            return {{"error", "\"changes\" must be an array"}};
+        }
+
+        // read file into lines
+        std::ifstream fin(path);
+        if (!fin) {
+            return {{"error", "failed to open file: " + path}};
+        }
+        std::vector<std::string> lines;
+        {
+            std::string line;
+            while (std::getline(fin, line)) {
+                lines.push_back(line);
+            }
+        }
+        fin.close();
+
+        // validate and collect changes, then sort descending by line_start
+        struct change_entry {
+            std::string mode;
+            int line_start; // 1-based
+            int line_end;   // 1-based inclusive
+            std::string content;
+        };
+        std::vector<change_entry> entries;
+        entries.reserve(changes.size());
+
+        for (const auto & ch : changes) {
+            change_entry e;
+            e.mode       = ch.at("mode").get<std::string>();
+            e.line_start = ch.at("line_start").get<int>();
+            e.line_end   = ch.at("line_end").get<int>();
+            e.content    = ch.at("content").get<std::string>();
+
+            if (e.mode != "replace" && e.mode != "delete" && e.mode != "append") {
+                return {{"error", "invalid mode \"" + e.mode + "\"; must be replace, delete, or append"}};
+            }
+            if (e.mode == "delete" && !e.content.empty()) {
+                return {{"error", "content must be empty string for delete mode"}};
+            }
+            int n = (int) lines.size();
+            if (e.line_start == -1) {
+                // -1 means end of file; line_end is ignored — normalize to point past last line
+                e.line_start = n + 1;
+                e.line_end   = n + 1;
+            } else {
+                if (e.line_start < 1 || e.line_end < e.line_start) {
+                    return {{"error", string_format("invalid line range [%d, %d]", e.line_start, e.line_end)}};
+                }
+                if (e.line_end > n) {
+                    return {{"error", string_format("line_end %d exceeds file length %d", e.line_end, n)}};
+                }
+            }
+            entries.push_back(std::move(e));
+        }
+
+        // sort descending so earlier-indexed changes don't shift later ones
+        std::sort(entries.begin(), entries.end(), [](const change_entry & a, const change_entry & b) {
+            return a.line_start > b.line_start;
+        });
+
+        // apply changes (0-based indices internally)
+        for (const auto & e : entries) {
+            int idx_start = e.line_start - 1; // 0-based
+            int idx_end   = e.line_end   - 1; // 0-based inclusive
+
+            // split content into lines (preserve trailing newline awareness)
+            std::vector<std::string> new_lines;
+            if (!e.content.empty()) {
+                std::istringstream ss(e.content);
+                std::string ln;
+                while (std::getline(ss, ln)) {
+                    new_lines.push_back(ln);
+                }
+                // if content ends with \n, getline consumed it — no extra empty line needed
+                // if content does NOT end with \n, last line is still captured correctly
+            }
+
+            if (e.mode == "replace") {
+                // erase [idx_start, idx_end] and insert new_lines
+                lines.erase(lines.begin() + idx_start, lines.begin() + idx_end + 1);
+                lines.insert(lines.begin() + idx_start, new_lines.begin(), new_lines.end());
+            } else if (e.mode == "delete") {
+                lines.erase(lines.begin() + idx_start, lines.begin() + idx_end + 1);
+            } else { // append
+                // idx_end + 1 may equal lines.size() when line_start == -1 (end of file)
+                lines.insert(lines.begin() + idx_end + 1, new_lines.begin(), new_lines.end());
+            }
+        }
+
+        // write file back
+        std::ofstream fout(path, std::ios::binary);
+        if (!fout) {
+            return {{"error", "failed to open file for writing: " + path}};
+        }
+        for (size_t i = 0; i < lines.size(); i++) {
+            fout << lines[i];
+            if (i + 1 < lines.size()) {
+                fout << "\n";
+            }
+        }
+        if (!lines.empty()) {
+            fout << "\n";
+        }
+        if (!fout) {
+            return {{"error", "failed to write file: " + path}};
+        }
+
+        return {{"result", "file edited successfully"}, {"path", path}, {"lines", (int) lines.size()}};
+    }
+};
+
+//
+// apply_diff: apply a unified diff via git apply
+//
+
+struct server_tool_apply_diff : server_tool {
+    server_tool_apply_diff() {
+        name = "apply_diff";
+        display_name = "Apply diff";
+        permission_write = true;
+    }
+
+    json get_definition() override {
+        return {
+            {"type", "function"},
+            {"function", {
+                {"name", name},
+                {"description", "Apply a unified diff to edit one or more files using git apply. Use this instead of edit_file when the changes are complex."},
+                {"parameters", {
+                    {"type", "object"},
+                    {"properties", {
+                        {"diff", {{"type", "string"}, {"description", "Unified diff content in git diff format"}}},
+                    }},
+                    {"required", json::array({"diff"})},
+                }},
+            }},
+        };
+    }
+
+    json invoke(json params) override {
+        std::string diff = params.at("diff").get<std::string>();
+
+        // write diff to a temporary file
+        static std::atomic<int> counter{0};
+        std::string tmp_path = (fs::temp_directory_path() /
+            ("llama_patch_" + std::to_string(++counter) + ".patch")).string();
+
+        {
+            std::ofstream f(tmp_path, std::ios::binary);
+            if (!f) {
+                return {{"error", "failed to create temp patch file"}};
+            }
+            f << diff;
+        }
+
+        auto res = run_process({"git", "apply", tmp_path}, 4096, 10);
+
+        std::error_code ec;
+        fs::remove(tmp_path, ec);
+
+        if (res.exit_code != 0) {
+            return {{"error", "git apply failed (exit " + std::to_string(res.exit_code) + "): " + res.output}};
+        }
+        return {{"result", "patch applied successfully"}};
+    }
+};
+
+//
+// public API
+//
+
+static std::vector<std::unique_ptr<server_tool>> build_tools() {
+    std::vector<std::unique_ptr<server_tool>> tools;
+    tools.push_back(std::make_unique<server_tool_read_file>());
+    tools.push_back(std::make_unique<server_tool_file_glob_search>());
+    tools.push_back(std::make_unique<server_tool_grep_search>());
+    tools.push_back(std::make_unique<server_tool_exec_shell_command>());
+    tools.push_back(std::make_unique<server_tool_write_file>());
+    tools.push_back(std::make_unique<server_tool_edit_file>());
+    tools.push_back(std::make_unique<server_tool_apply_diff>());
+    return tools;
+}
+
+void server_tools::setup(const std::vector<std::string> & enabled_tools) {
+    if (!enabled_tools.empty()) {
+        std::unordered_set<std::string> enabled_set(enabled_tools.begin(), enabled_tools.end());
+        auto all_tools = build_tools();
+
+        tools.clear();
+        for (auto & t : all_tools) {
+            if (enabled_set.count(t->name) > 0 || enabled_set.count("all") > 0) {
+                tools.push_back(std::move(t));
+            }
+        }
+    }
+
+    handle_get = [this](const server_http_req &) -> server_http_res_ptr {
+        auto res = std::make_unique<server_http_res>();
+        try {
+            json result = json::array();
+            for (const auto & t : tools) {
+                result.push_back(t->to_json());
+            }
+            res->data = safe_json_to_str(result);
+        } catch (const std::exception & e) {
+            SRV_ERR("got exception: %s\n", e.what());
+            res->status = 500;
+            res->data   = safe_json_to_str(format_error_response(e.what(), ERROR_TYPE_SERVER));
+        }
+        return res;
+    };
+
+    handle_post = [this](const server_http_req & req) -> server_http_res_ptr {
+        auto res = std::make_unique<server_http_res>();
+        try {
+            json body = json::parse(req.body);
+            std::string tool_name = body.at("tool").get<std::string>();
+            json params = body.value("params", json::object());
+            json result = invoke(tool_name, params);
+            res->data   = safe_json_to_str(result);
+        } catch (const json::exception & e) {
+            res->status = 400;
+            res->data   = safe_json_to_str(format_error_response(e.what(), ERROR_TYPE_INVALID_REQUEST));
+        } catch (const std::exception & e) {
+            SRV_ERR("got exception: %s\n", e.what());
+            res->status = 500;
+            res->data   = safe_json_to_str(format_error_response(e.what(), ERROR_TYPE_SERVER));
+        }
+        return res;
+    };
+}
+
+json server_tools::invoke(const std::string & name, const json & params) {
+    for (auto & t : tools) {
+        if (t->name == name) {
+            return t->invoke(params);
+        }
+    }
+    return {{"error", "unknown tool: " + name}};
+}
--- a/tools/server/server-tools.h
+++ b/tools/server/server-tools.h
@@ -0,0 +1,26 @@
+#pragma once
+
+#include "server-common.h"
+#include "server-http.h"
+
+struct server_tool {
+    std::string name;
+    std::string display_name;
+    bool permission_write = false;
+
+    virtual ~server_tool() = default;
+    virtual json get_definition() = 0;
+    virtual json invoke(json params) = 0;
+
+    json to_json();
+};
+
+struct server_tools {
+    std::vector<std::unique_ptr<server_tool>> tools;
+
+    void setup(const std::vector<std::string> & enabled_tools);
+    json invoke(const std::string & name, const json & params);
+
+    server_http_context::handler_t handle_get;
+    server_http_context::handler_t handle_post;
+};
--- a/tools/server/server.cpp
+++ b/tools/server/server.cpp
@@ -2,6 +2,7 @@
 #include "server-http.h"
 #include "server-models.h"
 #include "server-cors-proxy.h"
+#include "server-tools.h"

 #include "arg.h"
 #include "common.h"
@@ -124,6 +125,7 @@ int main(int argc, char ** argv) {

    // register API routes
    server_routes routes(params, ctx_server);
+    server_tools tools;

    bool is_router_server = params.model.path.empty();
    std::optional<server_models_routes> models_routes{};
@@ -211,6 +213,16 @@ int main(int argc, char ** argv) {
        ctx_http.get ("/cors-proxy",      ex_wrapper(proxy_handler_get));
        ctx_http.post("/cors-proxy",      ex_wrapper(proxy_handler_post));
    }
+    // EXPERIMENTAL built-in tools
+    if (!params.server_tools.empty()) {
+        tools.setup(params.server_tools);
+        SRV_WRN("%s", "-----------------\n");
+        SRV_WRN("%s", "Built-in tools are enabled, do not expose server to untrusted environments\n");
+        SRV_WRN("%s", "This feature is EXPERIMENTAL and may be changed in the future\n");
+        SRV_WRN("%s", "-----------------\n");
+        ctx_http.get ("/tools",           ex_wrapper(tools.handle_get));
+        ctx_http.post("/tools",           ex_wrapper(tools.handle_post));
+    }

    //
    // Start the server
--- a/tools/server/webui/src/lib/components/app/chat/ChatSettings/ChatSettings.svelte
+++ b/tools/server/webui/src/lib/components/app/chat/ChatSettings/ChatSettings.svelte
@@ -296,6 +296,11 @@
 					label: 'Disable reasoning content parsing',
 					type: SettingsFieldType.CHECKBOX
 				},
+				{
+					key: SETTINGS_KEYS.EXCLUDE_REASONING_FROM_CONTEXT,
+					label: 'Exclude reasoning from context',
+					type: SettingsFieldType.CHECKBOX
+				},
 				{
 					key: SETTINGS_KEYS.SHOW_RAW_OUTPUT_SWITCH,
 					label: 'Enable raw output toggle',
--- a/tools/server/webui/src/lib/constants/agentic.ts
+++ b/tools/server/webui/src/lib/constants/agentic.ts
@@ -50,6 +50,8 @@ export const AGENTIC_REGEX = {
 	PARTIAL_MARKER: /<<<[A-Za-z_]*$/,
 	// Matches reasoning content blocks (including tags)
 	REASONING_BLOCK: /<<<reasoning_content_start>>>[\s\S]*?<<<reasoning_content_end>>>/g,
+	// Captures the reasoning text between start/end tags
+	REASONING_EXTRACT: /<<<reasoning_content_start>>>([\s\S]*?)<<<reasoning_content_end>>>/,
 	// Matches an opening reasoning tag and any remaining content (unterminated)
 	REASONING_OPEN: /<<<reasoning_content_start>>>[\s\S]*$/,
 	// Matches a complete agentic tool call display block (start to end marker)
--- a/tools/server/webui/src/lib/constants/settings-config.ts
+++ b/tools/server/webui/src/lib/constants/settings-config.ts
@@ -10,6 +10,7 @@ export const SETTING_CONFIG_DEFAULT: Record<string, string | number | boolean |
 	theme: ColorMode.SYSTEM,
 	showThoughtInProgress: false,
 	disableReasoningParsing: false,
+	excludeReasoningFromContext: false,
 	showRawOutputSwitch: false,
 	keepStatsVisible: false,
 	showMessageStats: true,
@@ -106,6 +107,8 @@ export const SETTING_CONFIG_INFO: Record<string, string> = {
 	showThoughtInProgress: 'Expand thought process by default when generating messages.',
 	disableReasoningParsing:
 		'Send reasoning_format=none to prevent server-side extraction of reasoning tokens into separate field',
+	excludeReasoningFromContext:
+		'Strip reasoning content from previous messages before sending to the model. When unchecked, reasoning is sent back via the reasoning_content field so the model can see its own chain-of-thought across turns.',
 	showRawOutputSwitch:
 		'Show toggle button to display messages as plain text instead of Markdown-formatted content',
 	keepStatsVisible: 'Keep processing statistics visible after generation finishes.',
--- a/tools/server/webui/src/lib/constants/settings-keys.ts
+++ b/tools/server/webui/src/lib/constants/settings-keys.ts
@@ -54,6 +54,7 @@ export const SETTINGS_KEYS = {
 	SHOW_TOOL_CALL_IN_PROGRESS: 'showToolCallInProgress',
 	// Developer
 	DISABLE_REASONING_PARSING: 'disableReasoningParsing',
+	EXCLUDE_REASONING_FROM_CONTEXT: 'excludeReasoningFromContext',
 	SHOW_RAW_OUTPUT_SWITCH: 'showRawOutputSwitch',
 	CUSTOM: 'custom'
 } as const;
--- a/tools/server/webui/src/lib/services/chat.service.ts
+++ b/tools/server/webui/src/lib/services/chat.service.ts
@@ -57,6 +57,46 @@ export class ChatService {
 	 *
 	 */

+	/**
+	 * Extracts reasoning text from content that contains internal reasoning tags.
+	 * Returns the concatenated reasoning content or undefined if none found.
+	 */
+	private static extractReasoningFromContent(
+		content: ApiChatMessageData['content'] | null | undefined
+	): string | undefined {
+		if (!content) return undefined;
+
+		const extractFromString = (text: string): string => {
+			const parts: string[] = [];
+			// Use a fresh regex instance to avoid shared lastIndex state
+			const re = new RegExp(AGENTIC_REGEX.REASONING_EXTRACT.source);
+			let match = re.exec(text);
+			while (match) {
+				parts.push(match[1]);
+				// advance past the matched portion and retry
+				text = text.slice(match.index + match[0].length);
+				match = re.exec(text);
+			}
+			return parts.join('');
+		};
+
+		if (typeof content === 'string') {
+			const result = extractFromString(content);
+			return result || undefined;
+		}
+
+		if (!Array.isArray(content)) return undefined;
+
+		const parts: string[] = [];
+		for (const part of content) {
+			if (part.type === ContentPartType.TEXT && part.text) {
+				const result = extractFromString(part.text);
+				if (result) parts.push(result);
+			}
+		}
+		return parts.length > 0 ? parts.join('') : undefined;
+	}
+
 	/**
 	 * Sends a chat completion request to the llama.cpp server.
 	 * Supports both streaming and non-streaming responses with comprehensive parameter configuration.
@@ -111,7 +151,8 @@ export class ChatService {
 			custom,
 			timings_per_token,
 			// Config options
-			disableReasoningParsing
+			disableReasoningParsing,
+			excludeReasoningFromContext
 		} = options;

 		const normalizedMessages: ApiChatMessageData[] = messages
@@ -159,14 +200,24 @@ export class ChatService {
 		}

 		const requestBody: ApiChatCompletionRequest = {
-			messages: normalizedMessages.map((msg: ApiChatMessageData) => ({
-				role: msg.role,
-				// Strip reasoning tags/content from the prompt to avoid polluting KV cache.
-				// TODO: investigate backend expectations for reasoning tags and add a toggle if needed.
-				content: ChatService.stripReasoningContent(msg.content),
-				tool_calls: msg.tool_calls,
-				tool_call_id: msg.tool_call_id
-			})),
+			messages: normalizedMessages.map((msg: ApiChatMessageData) => {
+				// Always strip internal reasoning/agentic tags from content
+				const cleanedContent = ChatService.stripReasoningContent(msg.content);
+				const mapped: ApiChatCompletionRequest['messages'][0] = {
+					role: msg.role,
+					content: cleanedContent,
+					tool_calls: msg.tool_calls,
+					tool_call_id: msg.tool_call_id
+				};
+				// When preserving reasoning, extract it from raw content and send as separate field
+				if (!excludeReasoningFromContext) {
+					const reasoning = ChatService.extractReasoningFromContent(msg.content);
+					if (reasoning) {
+						mapped.reasoning_content = reasoning;
+					}
+				}
+				return mapped;
+			}),
 			stream,
 			return_progress: stream ? true : undefined,
 			tools: tools && tools.length > 0 ? tools : undefined
--- a/tools/server/webui/src/lib/services/parameter-sync.service.ts
+++ b/tools/server/webui/src/lib/services/parameter-sync.service.ts
@@ -227,6 +227,12 @@ export const SYNCABLE_PARAMETERS: SyncableParameter[] = [
 		serverKey: 'alwaysShowAgenticTurns',
 		type: SyncableParameterType.BOOLEAN,
 		canSync: true
+	},
+	{
+		key: 'excludeReasoningFromContext',
+		serverKey: 'excludeReasoningFromContext',
+		type: SyncableParameterType.BOOLEAN,
+		canSync: true
 	}
 ];

--- a/tools/server/webui/src/lib/stores/chat.svelte.ts
+++ b/tools/server/webui/src/lib/stores/chat.svelte.ts
@@ -1479,6 +1479,8 @@ class ChatStore {

 		if (currentConfig.disableReasoningParsing) apiOptions.disableReasoningParsing = true;

+		if (currentConfig.excludeReasoningFromContext) apiOptions.excludeReasoningFromContext = true;
+
 		if (hasValue(currentConfig.temperature))
 			apiOptions.temperature = Number(currentConfig.temperature);

--- a/tools/server/webui/src/lib/types/api.d.ts
+++ b/tools/server/webui/src/lib/types/api.d.ts
@@ -45,6 +45,7 @@ export interface ApiErrorResponse {
 export interface ApiChatMessageData {
 	role: ChatRole;
 	content: string | ApiChatMessageContentPart[];
+	reasoning_content?: string;
 	tool_calls?: ApiChatCompletionToolCall[];
 	tool_call_id?: string;
 	timestamp?: number;
@@ -201,6 +202,9 @@ export interface ApiChatCompletionRequest {
 	messages: Array<{
 		role: ChatRole;
 		content: string | ApiChatMessageContentPart[];
+		reasoning_content?: string;
+		tool_calls?: ApiChatCompletionToolCall[];
+		tool_call_id?: string;
 	}>;
 	stream?: boolean;
 	model?: string;
--- a/tools/server/webui/src/lib/types/settings.d.ts
+++ b/tools/server/webui/src/lib/types/settings.d.ts
@@ -24,6 +24,8 @@ export interface SettingsChatServiceOptions {
 	systemMessage?: string;
 	// Disable reasoning parsing (use 'none' instead of 'auto')
 	disableReasoningParsing?: boolean;
+	// Strip reasoning content from context before sending
+	excludeReasoningFromContext?: boolean;
 	tools?: OpenAIToolDefinition[];
 	// Generation parameters
 	temperature?: number;
--- a/tools/server/webui/tests/unit/reasoning-context.test.ts
+++ b/tools/server/webui/tests/unit/reasoning-context.test.ts
@@ -0,0 +1,196 @@
+import { describe, it, expect } from 'vitest';
+import { AGENTIC_REGEX, REASONING_TAGS } from '$lib/constants/agentic';
+import { ContentPartType } from '$lib/enums';
+
+// Replicate ChatService.extractReasoningFromContent (private static)
+function extractReasoningFromContent(
+	content: string | Array<{ type: string; text?: string }> | null | undefined
+): string | undefined {
+	if (!content) return undefined;
+
+	const extractFromString = (text: string): string => {
+		const parts: string[] = [];
+		const re = new RegExp(AGENTIC_REGEX.REASONING_EXTRACT.source);
+		let match = re.exec(text);
+		while (match) {
+			parts.push(match[1]);
+			text = text.slice(match.index + match[0].length);
+			match = re.exec(text);
+		}
+		return parts.join('');
+	};
+
+	if (typeof content === 'string') {
+		const result = extractFromString(content);
+		return result || undefined;
+	}
+
+	if (!Array.isArray(content)) return undefined;
+
+	const parts: string[] = [];
+	for (const part of content) {
+		if (part.type === ContentPartType.TEXT && part.text) {
+			const result = extractFromString(part.text);
+			if (result) parts.push(result);
+		}
+	}
+	return parts.length > 0 ? parts.join('') : undefined;
+}
+
+// Replicate ChatService.stripReasoningContent (private static)
+function stripReasoningContent(
+	content: string | Array<{ type: string; text?: string }> | null | undefined
+): typeof content {
+	if (!content) return content;
+
+	if (typeof content === 'string') {
+		return content
+			.replace(AGENTIC_REGEX.REASONING_BLOCK, '')
+			.replace(AGENTIC_REGEX.REASONING_OPEN, '')
+			.replace(AGENTIC_REGEX.AGENTIC_TOOL_CALL_BLOCK, '')
+			.replace(AGENTIC_REGEX.AGENTIC_TOOL_CALL_OPEN, '');
+	}
+
+	if (!Array.isArray(content)) return content;
+
+	return content.map((part) => {
+		if (part.type !== ContentPartType.TEXT || !part.text) return part;
+		return {
+			...part,
+			text: part.text
+				.replace(AGENTIC_REGEX.REASONING_BLOCK, '')
+				.replace(AGENTIC_REGEX.REASONING_OPEN, '')
+				.replace(AGENTIC_REGEX.AGENTIC_TOOL_CALL_BLOCK, '')
+				.replace(AGENTIC_REGEX.AGENTIC_TOOL_CALL_OPEN, '')
+		};
+	});
+}
+
+// Simulate the message mapping logic from ChatService.sendMessage
+function buildApiMessage(
+	content: string,
+	excludeReasoningFromContext: boolean
+): { role: string; content: string; reasoning_content?: string } {
+	const cleaned = stripReasoningContent(content) as string;
+	const mapped: { role: string; content: string; reasoning_content?: string } = {
+		role: 'assistant',
+		content: cleaned
+	};
+	if (!excludeReasoningFromContext) {
+		const reasoning = extractReasoningFromContent(content);
+		if (reasoning) {
+			mapped.reasoning_content = reasoning;
+		}
+	}
+	return mapped;
+}
+
+// Helper: wrap reasoning the same way the chat store does during streaming
+function wrapReasoning(reasoning: string, content: string): string {
+	return `${REASONING_TAGS.START}${reasoning}${REASONING_TAGS.END}${content}`;
+}
+
+describe('reasoning content extraction', () => {
+	it('extracts reasoning from tagged string content', () => {
+		const input = wrapReasoning('step 1, step 2', 'The answer is 42.');
+		const result = extractReasoningFromContent(input);
+		expect(result).toBe('step 1, step 2');
+	});
+
+	it('returns undefined when no reasoning tags present', () => {
+		expect(extractReasoningFromContent('Just a normal response.')).toBeUndefined();
+	});
+
+	it('returns undefined for null/empty input', () => {
+		expect(extractReasoningFromContent(null)).toBeUndefined();
+		expect(extractReasoningFromContent(undefined)).toBeUndefined();
+		expect(extractReasoningFromContent('')).toBeUndefined();
+	});
+
+	it('extracts reasoning from content part arrays', () => {
+		const input = [
+			{
+				type: ContentPartType.TEXT,
+				text: wrapReasoning('thinking hard', 'result')
+			}
+		];
+		expect(extractReasoningFromContent(input)).toBe('thinking hard');
+	});
+
+	it('handles multiple reasoning blocks', () => {
+		const input =
+			REASONING_TAGS.START +
+			'block1' +
+			REASONING_TAGS.END +
+			'middle' +
+			REASONING_TAGS.START +
+			'block2' +
+			REASONING_TAGS.END +
+			'end';
+		expect(extractReasoningFromContent(input)).toBe('block1block2');
+	});
+
+	it('ignores non-text content parts', () => {
+		const input = [{ type: 'image_url', text: wrapReasoning('hidden', 'img') }];
+		expect(extractReasoningFromContent(input)).toBeUndefined();
+	});
+});
+
+describe('strip reasoning content', () => {
+	it('removes reasoning tags from string content', () => {
+		const input = wrapReasoning('internal thoughts', 'visible answer');
+		expect(stripReasoningContent(input)).toBe('visible answer');
+	});
+
+	it('removes reasoning from content part arrays', () => {
+		const input = [
+			{
+				type: ContentPartType.TEXT,
+				text: wrapReasoning('thoughts', 'answer')
+			}
+		];
+		const result = stripReasoningContent(input) as Array<{ type: string; text?: string }>;
+		expect(result[0].text).toBe('answer');
+	});
+});
+
+describe('API message building with reasoning preservation', () => {
+	const storedContent = wrapReasoning('Let me think: 2+2=4, basic arithmetic.', 'The answer is 4.');
+
+	it('preserves reasoning_content when excludeReasoningFromContext is false', () => {
+		const msg = buildApiMessage(storedContent, false);
+		expect(msg.content).toBe('The answer is 4.');
+		expect(msg.reasoning_content).toBe('Let me think: 2+2=4, basic arithmetic.');
+		// no internal tags leak into either field
+		expect(msg.content).not.toContain('<<<');
+		expect(msg.reasoning_content).not.toContain('<<<');
+	});
+
+	it('strips reasoning_content when excludeReasoningFromContext is true', () => {
+		const msg = buildApiMessage(storedContent, true);
+		expect(msg.content).toBe('The answer is 4.');
+		expect(msg.reasoning_content).toBeUndefined();
+	});
+
+	it('handles content with no reasoning in both modes', () => {
+		const plain = 'No reasoning here.';
+		const msgPreserve = buildApiMessage(plain, false);
+		const msgExclude = buildApiMessage(plain, true);
+		expect(msgPreserve.content).toBe(plain);
+		expect(msgPreserve.reasoning_content).toBeUndefined();
+		expect(msgExclude.content).toBe(plain);
+		expect(msgExclude.reasoning_content).toBeUndefined();
+	});
+
+	it('cleans agentic tool call blocks from content even when preserving reasoning', () => {
+		const input =
+			wrapReasoning('plan', 'text') +
+			'\n\n<<<AGENTIC_TOOL_CALL_START>>>\n' +
+			'<<<TOOL_NAME:bash>>>\n' +
+			'<<<TOOL_ARGS_START>>>\n{}\n<<<TOOL_ARGS_END>>>\nout\n' +
+			'<<<AGENTIC_TOOL_CALL_END>>>\n';
+		const msg = buildApiMessage(input, false);
+		expect(msg.content).not.toContain('<<<');
+		expect(msg.reasoning_content).toBe('plan');
+	});
+});
Author	SHA1	Message	Date
AN Long	48cda24c11	server: remove the verbose_prompt parameter (#21059 ) * server: respect the verbose_prompt parameter * Revert "server: respect the verbose_prompt parameter" This reverts commit `8ed885cf37`. * Remove --verbose-prompt parameter from llama-server * Using set_examples instead of set_excludes	2026-03-27 13:36:13 +02:00
Xuan-Son Nguyen	871f1a2d2f	mtmd: add more sanity checks (#21047 )	2026-03-27 11:00:52 +01:00
Xuan-Son Nguyen	20197b6fe3	server: add built-in tools backend support (#20898 ) * wip: server_tools * refactor * displayName -> display_name * snake_case everywhere * rm redundant field * change arg to --tools all * add readme mention * llama-gen-docs	2026-03-27 10:07:11 +01:00
Radoslav Gerganov	ba38f3becc	rpc : proper handling of data pointers to CPU buffers (#21030 ) The compute graph may contain tensors pointing to CPU buffers. In these cases the buffer address is serialized as 0 and sent over the wire. However, the data pointer is serialized as-is and this prevents proper validation on the server side. This patches fixes this by serializing the data pointer as 0 for non-RPC buffers and doing proper validation on the server side. closes: #21006	2026-03-27 10:59:35 +02:00
mtmcp	37f230dd7c	completion : session_tokens insert range in completion tool (no-op → correct) (#20917 ) The embd.begin(), embd.begin() range is empty and inserts nothing, so session_tokens never gets updated after decoding. Should be embd.begin(), embd.end(). Introduced in commit `2b6dfe8`.	2026-03-27 09:25:58 +01:00
mtmcp	a308e584ca	completion : Fix segfault on model load failure (#21049 )	2026-03-27 10:01:13 +02:00
Pascal	d0fa2c9fbb	Send reasoning content back to the model across turns via the reasoning_content API field (#21036 ) * webui: send reasoning_content back to model in context Preserve assistant reasoning across turns by extracting it from internal tags and sending it as a separate reasoning_content field in the API payload. The server and Jinja templates handle native formatting (e.g. <think> tags for Qwen, GLM, DeepSeek...). Adds "Exclude reasoning from context" toggle in Settings > Developer (off by default, so reasoning is preserved). Includes unit tests. * webui: add syncable parameter for excludeReasoningFromContext * chore: update webui build output	2026-03-27 08:17:35 +01:00
ren	9bcb4eff4d	metal : Fix dimension constraint violation in matmul2d descriptor (#21048 ) Updates Metal tensor API test probe to fix the dimension constraint violation in the matmul2d descriptor (at least one value must be a multiple of 16).	2026-03-27 09:05:21 +02:00
KokerZhou	6861f6509a	CANN: update docker images to 8.5.0 and improve CANN.md (#20801 ) * cann: update docker images to 8.5.0 - bump CANN base image from 8.3.rc2 to 8.5.0 - bump ASCEND_VERSION from 8.1.RC1.alpha001 to 8.5.0 Move to newer stable releases. * cann: update CANN.md * Update CANN.md to include BF16 support Added BF16 support information to the CANN documentation and corrected formatting for the installation instructions. * Fix formatting issues in CANN.md Fix 234: Trailing whitespace	2026-03-27 08:53:00 +08:00