fix(llama-cpp): terminate tensor/kv override vectors after passthrough

The tensor_buft_overrides padding and the kv/draft override terminators
ran before the generic option passthrough, so a passthrough flag
(--cpu-moe, --override-tensor, --override-kv, ...) appended a real entry
after the null sentinel - tripping the model loader's
back().pattern == nullptr assertion (crash) or being silently dropped.
Move all three termination/padding blocks to the end of params_parse,
after both the named-option loop and common_params_parse have pushed
their real entries. Also widen the exit()-flag skip list so --version,
--license, --list-devices and --cache-list cannot terminate the backend.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

.github/workflows/backend.yml

@@ -44,7 +44,7 @@ jobs:
       has-merges-singlearch: ${{ steps.set-matrix.outputs['has-merges-singlearch'] }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
 
       - name: Setup Bun
         uses: oven-sh/setup-bun@v2

.github/workflows/backend_build.yml

@@ -101,7 +101,7 @@ jobs:
     steps:
 
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
 

.github/workflows/backend_build_darwin.yml

@@ -57,7 +57,7 @@ jobs:
       HOMEBREW_NO_ANALYTICS: '1'
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
 

.github/workflows/backend_merge.yml

@@ -49,7 +49,7 @@ jobs:
       # Sparse checkout: the merge job needs `.github/scripts/` (for the
       # keepalive cleanup script) but none of the source tree.
       - name: Checkout (.github/scripts only)
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           sparse-checkout: |
             .github/scripts

.github/workflows/backend_pr.yml

@@ -23,7 +23,7 @@ jobs:
       has-merges-singlearch: ${{ steps.set-matrix.outputs['has-merges-singlearch'] }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
 
       - name: Setup Bun
         uses: oven-sh/setup-bun@v2

.github/workflows/base-images.yml

@@ -127,7 +127,7 @@ jobs:
             # the original l4t matrix entry which set skip-drivers: 'true'.
             skip-drivers: 'true'
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
         with:
           submodules: false
       - name: Free disk space

.github/workflows/build-test.yaml

@@ -11,7 +11,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           fetch-depth: 0
       - name: Set up Go
@@ -25,7 +25,7 @@ jobs:
     runs-on: macos-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           fetch-depth: 0
       - name: Set up Go
@@ -47,7 +47,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           fetch-depth: 0
       - name: Configure apt mirror on runner

.github/workflows/bump-inference-defaults.yml

@@ -14,7 +14,7 @@ jobs:
   bump:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
 
       - uses: actions/setup-go@v5
         with:

.github/workflows/bump_deps.yaml

@@ -92,7 +92,7 @@ jobs:
             file: "backend/go/vibevoice-cpp/Makefile"
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
       - name: Bump dependencies 🔧
         id: bump
         run: |
@@ -128,7 +128,7 @@ jobs:
     if: github.repository == 'mudler/LocalAI'
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
       - name: Bump vLLM cu130 wheel pin 🔧
         id: bump
         run: |

.github/workflows/bump_docs.yaml

@@ -13,7 +13,7 @@ jobs:
           - repository: "mudler/LocalAI"
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
       - name: Bump dependencies 🔧
         run: |
           bash .github/bump_docs.sh ${{ matrix.repository }}

.github/workflows/checksum_checker.yaml

@@ -8,7 +8,7 @@ jobs:
     if: github.repository == 'mudler/LocalAI'
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
       - name: Configure apt mirror on runner
         uses: ./.github/actions/configure-apt-mirror
       - name: Install dependencies

.github/workflows/deploy-explorer.yaml

@@ -16,7 +16,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - uses: actions/setup-go@v5

.github/workflows/gallery-agent.yaml

@@ -31,7 +31,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           token: ${{ secrets.GITHUB_TOKEN }}
 

.github/workflows/generate_intel_image.yaml

@@ -44,7 +44,7 @@ jobs:
         uses: docker/setup-buildx-action@master
 
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
 
       - name: Cache Intel images
         uses: docker/build-push-action@v7

.github/workflows/gh-pages.yml

@@ -28,7 +28,7 @@ jobs:
       HUGO_VERSION: "0.146.3"
     steps:
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           fetch-depth: 0  # needed for enableGitInfo
           submodules: true

.github/workflows/image_build.yml

@@ -80,7 +80,7 @@ jobs:
     steps:
 
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
 
       - name: Configure apt mirror on runner
         id: apt_mirror

.github/workflows/image_merge.yml

@@ -36,7 +36,7 @@ jobs:
       # Sparse checkout: needed for .github/scripts/ (the keepalive cleanup
       # script). Skips the rest of the source tree.
       - name: Checkout (.github/scripts only)
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           sparse-checkout: |
             .github/scripts

.github/workflows/lint.yml

@@ -20,7 +20,7 @@ jobs:
   golangci-lint:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
         with:
           # Full history so golangci-lint's new-from-merge-base can reach
           # origin/master and compute the diff against it.

.github/workflows/release.yaml

@@ -10,7 +10,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           fetch-depth: 0
       - name: Set up Go
@@ -28,7 +28,7 @@ jobs:
     runs-on: macos-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           fetch-depth: 0
       - name: Set up Go
@@ -46,7 +46,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           fetch-depth: 0
       - name: Configure apt mirror on runner

.github/workflows/secscan.yaml

@@ -14,7 +14,7 @@ jobs:
       GO111MODULE: on
     steps:
       - name: Checkout Source
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         if: ${{ github.actor != 'dependabot[bot]' }}
       - name: Run Gosec Security Scanner
         if: ${{ github.actor != 'dependabot[bot]' }}

.github/workflows/test-extra.yml

@@ -50,7 +50,7 @@ jobs:
       parakeet-cpp: ${{ steps.detect.outputs.parakeet-cpp }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
       - name: Setup Bun
         uses: oven-sh/setup-bun@v2
       - name: Install dependencies
@@ -67,7 +67,7 @@ jobs:
   #   runs-on: ubuntu-latest
   #   steps:
   #     - name: Clone
-  #       uses: actions/checkout@v6
+  #       uses: actions/checkout@v7
   #       with:
   #         submodules: true
   #     - name: Dependencies
@@ -90,7 +90,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -113,7 +113,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -137,7 +137,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -158,7 +158,7 @@ jobs:
   #  runs-on: ubuntu-latest
   #  steps:
   #    - name: Clone
-  #      uses: actions/checkout@v6
+  #      uses: actions/checkout@v7
   #      with:
   #        submodules: true
   #    - name: Dependencies
@@ -178,7 +178,7 @@ jobs:
   #   runs-on: ubuntu-latest
   #   steps:
   #     - name: Clone
-  #       uses: actions/checkout@v6
+  #       uses: actions/checkout@v7
   #       with:
   #         submodules: true
   #     - name: Dependencies
@@ -240,7 +240,7 @@ jobs:
   #           sudo rm -rf "$AGENT_TOOLSDIRECTORY" || true
   #           df -h
   #     - name: Clone
-  #       uses: actions/checkout@v6
+  #       uses: actions/checkout@v7
   #       with:
   #         submodules: true
   #     - name: Dependencies
@@ -265,7 +265,7 @@ jobs:
   #   runs-on: ubuntu-latest
   #   steps:
   #     - name: Clone
-  #       uses: actions/checkout@v6
+  #       uses: actions/checkout@v7
   #       with:
   #         submodules: true
   #     - name: Dependencies
@@ -288,7 +288,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -309,7 +309,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -330,7 +330,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -351,7 +351,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -373,7 +373,7 @@ jobs:
   #   timeout-minutes: 45
   #   steps:
   #     - name: Clone
-  #       uses: actions/checkout@v6
+  #       uses: actions/checkout@v7
   #       with:
   #         submodules: true
   #     - name: Dependencies
@@ -394,7 +394,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -415,7 +415,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -436,7 +436,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -462,7 +462,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -484,7 +484,7 @@ jobs:
     timeout-minutes: 30
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -513,7 +513,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -530,7 +530,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -552,7 +552,7 @@ jobs:
     timeout-minutes: 20
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -579,7 +579,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -604,7 +604,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -625,7 +625,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -645,7 +645,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -664,7 +664,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -681,7 +681,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -698,7 +698,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -741,7 +741,7 @@ jobs:
   #   timeout-minutes: 90
   #   steps:
   #     - name: Clone
-  #       uses: actions/checkout@v6
+  #       uses: actions/checkout@v7
   #       with:
   #         submodules: true
   #     - name: Dependencies
@@ -783,7 +783,7 @@ jobs:
   #   timeout-minutes: 90
   #   steps:
   #     - name: Clone
-  #       uses: actions/checkout@v6
+  #       uses: actions/checkout@v7
   #       with:
   #         submodules: true
   #     - name: Dependencies
@@ -808,7 +808,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -840,7 +840,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -876,7 +876,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -915,7 +915,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -952,7 +952,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -987,7 +987,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -1013,7 +1013,7 @@ jobs:
     timeout-minutes: 150
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -1042,7 +1042,7 @@ jobs:
     timeout-minutes: 60
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go
@@ -1058,7 +1058,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -1091,7 +1091,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -1114,7 +1114,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies
@@ -1140,7 +1140,7 @@ jobs:
     timeout-minutes: 90
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies

.github/workflows/test.yml

@@ -21,7 +21,7 @@ jobs:
         go-version: ['1.26.x']
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Free disk space
@@ -84,7 +84,7 @@ jobs:
         go-version: ['1.26.x']
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Setup Go ${{ matrix.go-version }}

.github/workflows/tests-aio.yml

@@ -62,7 +62,7 @@ jobs:
           sudo rm -rfv build || true
           df -h
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Dependencies

.github/workflows/tests-e2e.yml

@@ -21,7 +21,7 @@ jobs:
         go-version: ['1.25.x']
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Configure apt mirror on runner

.github/workflows/tests-pii-ner-e2e.yml

@@ -57,7 +57,7 @@ jobs:
         go-version: ['1.25.x']
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Free disk space

.github/workflows/tests-ui-e2e.yml

@@ -23,7 +23,7 @@ jobs:
         go-version: ['1.26.x']
     steps:
       - name: Clone
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
         with:
           submodules: true
       - name: Configure apt mirror on runner

.github/workflows/update_swagger.yaml

@@ -10,7 +10,7 @@ jobs:
       fail-fast: false
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
       - name: Configure apt mirror on runner
         uses: ./.github/actions/configure-apt-mirror
       - uses: actions/setup-go@v5

backend/cpp/ik-llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-IK_LLAMA_VERSION?=6c00e87ac84404af588ad2e65935bd6f079c696f
+IK_LLAMA_VERSION?=7ccf1d209588962b96eacca325b37e9b3e8faf5e
 LLAMA_REPO?=https://github.com/ikawrakow/ik_llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=7c082bc417bbe53210a83df4ba5b49e18ce6193c
+LLAMA_VERSION?=be4a6a63eb2b848e19c277bdcf2bd399e8af76d9
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/llama-cpp/grpc-server.cpp

@@ -37,6 +37,7 @@
 #include "backend.pb.h"
 #include "backend.grpc.pb.h"
 #include "common.h"
+#include "arg.h"
 #include "chat-auto-parser.h"
 #include <getopt.h>
 #include <grpcpp/ext/proto_server_reflection_plugin.h>
@@ -592,6 +593,10 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
     params.checkpoint_min_step = 256;
 #endif
 
+    // Raw upstream llama-server flags collected from any option entry that
+    // starts with '-'. Applied once after the loop via common_params_parse.
+    std::vector<std::string> extra_argv;
+
      // decode options. Options are in form optname:optvale, or if booleans only optname.
     for (int i = 0; i < request->options_size(); i++) {
         std::string opt = request->options(i);
@@ -1080,6 +1085,31 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
                 } catch (...) {}
             }
 
+        // --- main model MoE on CPU (upstream --cpu-moe / --n-cpu-moe) ---
+        } else if (!strcmp(optname, "cpu_moe")) {
+            // Bool-style flag: keep all MoE expert weights on CPU.
+            const bool enable = (optval == NULL) ||
+                optval_str == "true" || optval_str == "1" || optval_str == "yes" ||
+                optval_str == "on" || optval_str == "enabled";
+            if (enable) {
+                params.tensor_buft_overrides.push_back(llm_ffn_exps_cpu_override());
+            }
+        } else if (!strcmp(optname, "n_cpu_moe")) {
+            if (optval != NULL) {
+                try {
+                    int n = std::stoi(optval_str);
+                    if (n < 0) n = 0;
+                    // Keep override-name storage alive for the lifetime of the
+                    // params struct (mirrors upstream arg.cpp's function-local static).
+                    static std::list<std::string> buft_overrides_main;
+                    for (int i = 0; i < n; ++i) {
+                        buft_overrides_main.push_back(llm_ffn_exps_block_regex(i));
+                        params.tensor_buft_overrides.push_back(
+                            {buft_overrides_main.back().c_str(), ggml_backend_cpu_buffer_type()});
+                    }
+                } catch (...) {}
+            }
+
         // --- draft model tensor buffer overrides (upstream --spec-draft-override-tensor) ---
         } else if (!strcmp(optname, "draft_override_tensor") || !strcmp(optname, "spec_draft_override_tensor")) {
             // Format: <tensor regex>=<buffer type>,<tensor regex>=<buffer type>,...
@@ -1111,6 +1141,30 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
                 else { cur.push_back(c); }
             }
             if (!cur.empty()) flush(cur);
+
+        // --- generic passthrough: any entry starting with '-' is a raw
+        //     upstream llama-server flag, forwarded verbatim to the parser. ---
+        } else if (optname[0] == '-') {
+            std::string flag = optname;
+            // These flags make upstream's parser exit() (printing usage /
+            // completion), which would kill the backend process. Skip them.
+            if (flag == "-h" || flag == "--help" || flag == "--usage" ||
+                flag == "--version" || flag == "--license" ||
+                flag == "--list-devices" || flag == "-cl" ||
+                flag == "--cache-list" ||
+                flag.rfind("--completion", 0) == 0) {
+                fprintf(stderr,
+                    "[llama-cpp] ignoring passthrough flag that would exit: %s\n",
+                    flag.c_str());
+            } else {
+                extra_argv.push_back(flag);
+                // Preserve the whole value after the first ':' so embedded
+                // colons (e.g. host:port) survive strtok's truncation of optval.
+                auto colon = opt.find(':');
+                if (colon != std::string::npos) {
+                    extra_argv.push_back(opt.substr(colon + 1));
+                }
+            }
         }
     }
 
@@ -1146,27 +1200,6 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
         }
     }
 
-    if (!params.kv_overrides.empty()) {
-        params.kv_overrides.emplace_back();
-        params.kv_overrides.back().key[0] = 0;
-    }
-
-    // tensor_buft_overrides sentinel termination (mirrors upstream common/arg.cpp).
-    // Real entries are pushed during option parsing; here we pad/terminate so the
-    // model loader sees back().pattern == nullptr (GGML_ASSERT at common.cpp:1543)
-    // and so llama_params_fit has the placeholder slots it requires.
-    {
-        const size_t ntbo = llama_max_tensor_buft_overrides();
-        while (params.tensor_buft_overrides.size() < ntbo) {
-            params.tensor_buft_overrides.push_back({nullptr, nullptr});
-        }
-    }
-    // Terminate the draft tensor_buft_overrides list with a sentinel, mirroring
-    // the main-model handling above.
-    if (!params.speculative.draft.tensor_buft_overrides.empty()) {
-        params.speculative.draft.tensor_buft_overrides.push_back({nullptr, nullptr});
-    }
-
     // TODO: Add yarn
 
     if (!request->tensorsplit().empty()) {
@@ -1259,6 +1292,69 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
             params.sampling.grammar_triggers.push_back(std::move(trigger));
         }
     }
+
+    // Apply any raw upstream flags last so an explicit passthrough flag wins
+    // over the LocalAI-resolved field it maps to (e.g. --ctx-size beats
+    // context_size). This is the same parser llama-server itself uses.
+    if (!extra_argv.empty()) {
+        // common_params_parser_init resets a few fields for the SERVER example
+        // (n_parallel -> -1, use_color). Snapshot n_parallel so an unrelated
+        // passthrough flag can't silently clobber LocalAI's resolved value.
+        const int saved_n_parallel = params.n_parallel;
+
+        std::vector<char *> argv;
+        std::string prog = "llama-server";
+        argv.push_back(prog.data());
+        for (auto & a : extra_argv) {
+            argv.push_back(a.data());
+        }
+
+        // ctx_arg.params is a reference, so this overlays the given flags onto
+        // `params` in place. Returns false on a recoverable parse error (and
+        // self-restores params); may exit() on a hard error, exactly as
+        // passing the same bad flag to llama-server would.
+        if (!common_params_parse((int)argv.size(), argv.data(), params,
+                                 LLAMA_EXAMPLE_SERVER)) {
+            fprintf(stderr,
+                "[llama-cpp] failed to parse passthrough options; ignoring them\n");
+        }
+
+        // Restore n_parallel unless a passthrough flag explicitly set it
+        // (parser_init's reset sentinel for SERVER is -1).
+        if (params.n_parallel == -1) {
+            params.n_parallel = saved_n_parallel;
+        }
+    }
+
+    // Terminate/pad the override vectors only after BOTH the named-option loop
+    // and the generic passthrough (common_params_parse above) have pushed their
+    // real entries, so back() is the null sentinel the model loader asserts on.
+    // Running these before the passthrough let a passthrough flag (--cpu-moe,
+    // --override-tensor, --override-kv, ...) append a real entry after the
+    // sentinel: a GGML_ASSERT crash for tensor_buft_overrides, a silent drop for
+    // kv_overrides. Double-termination is harmless (the while is a no-op if the
+    // passthrough parse already padded; an extra trailing null is ignored).
+
+    if (!params.kv_overrides.empty()) {
+        params.kv_overrides.emplace_back();
+        params.kv_overrides.back().key[0] = 0;
+    }
+
+    // tensor_buft_overrides sentinel termination (mirrors upstream common/arg.cpp).
+    // Real entries are pushed during option parsing; here we pad/terminate so the
+    // model loader sees back().pattern == nullptr (GGML_ASSERT at common.cpp:1543)
+    // and so llama_params_fit has the placeholder slots it requires.
+    {
+        const size_t ntbo = llama_max_tensor_buft_overrides();
+        while (params.tensor_buft_overrides.size() < ntbo) {
+            params.tensor_buft_overrides.push_back({nullptr, nullptr});
+        }
+    }
+    // Terminate the draft tensor_buft_overrides list with a sentinel, mirroring
+    // the main-model handling above.
+    if (!params.speculative.draft.tensor_buft_overrides.empty()) {
+        params.speculative.draft.tensor_buft_overrides.push_back({nullptr, nullptr});
+    }
 }
 
 

backend/go/crispasr/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # CrispASR version (release tag)
 CRISPASR_REPO?=https://github.com/CrispStrobe/CrispASR
-CRISPASR_VERSION?=7a8cb80907341c0204bd0488c1244764f4163883
+CRISPASR_VERSION?=96b2a6ee31d30389fed8a7ef1a54239b75231ddc
 SO_TARGET?=libgocrispasr.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/go/parakeet-cpp/Makefile

@@ -1,6 +1,6 @@
 # parakeet-cpp backend Makefile.
 #
-# Upstream pin lives below as PARAKEET_VERSION?=db755a78d39f789bb7d4e3935158a9e8105dbe36
+# Upstream pin lives below as PARAKEET_VERSION?=89f5e2977b4d8bccd45e7bcc6f2ef7c4ed49e89a
 # (.github/bump_deps.sh) can find and update it - matches the
 # whisper.cpp / ds4 / vibevoice-cpp convention.
 #
@@ -15,7 +15,7 @@
 # That's what the L0 smoke test uses. The default target below does the
 # proper clone-at-pin + cmake build so CI doesn't need a side-checkout.
 
-PARAKEET_VERSION?=db755a78d39f789bb7d4e3935158a9e8105dbe36
+PARAKEET_VERSION?=89f5e2977b4d8bccd45e7bcc6f2ef7c4ed49e89a
 PARAKEET_REPO?=https://github.com/mudler/parakeet.cpp
 
 GOCMD?=go

backend/go/stablediffusion-ggml/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # stablediffusion.cpp (ggml)
 STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
-STABLEDIFFUSION_GGML_VERSION?=b12098f5d09fc83da36e65c784f7bdb16a5a5ebf
+STABLEDIFFUSION_GGML_VERSION?=f440ad9c29dd8bc34e5d1f4b863832b96d6ea05f
 
 CMAKE_ARGS+=-DGGML_MAX_NAME=128
 

backend/go/whisper/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggml-org/whisper.cpp
-WHISPER_CPP_VERSION?=5ed76e9a079962f1c85cfce44edd325c27ef1f97
+WHISPER_CPP_VERSION?=43d78af5be58f41d6ffbc227d608f104577741ea
 SO_TARGET?=libgowhisper.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

core/application/config_file_watcher.go

@@ -215,6 +215,7 @@ func readRuntimeSettingsJson(startupAppConfig config.ApplicationConfig) fileHand
 		envBackendGalleries := slices.Equal(appConfig.BackendGalleries, startupAppConfig.BackendGalleries)
 		envAutoloadGalleries := appConfig.AutoloadGalleries == startupAppConfig.AutoloadGalleries
 		envAutoloadBackendGalleries := appConfig.AutoloadBackendGalleries == startupAppConfig.AutoloadBackendGalleries
+		envPIIDefaultDetectors := slices.Equal(appConfig.PIIDefaultDetectors, startupAppConfig.PIIDefaultDetectors)
 		envAgentJobRetentionDays := appConfig.AgentJobRetentionDays == startupAppConfig.AgentJobRetentionDays
 		envForceEvictionWhenBusy := appConfig.ForceEvictionWhenBusy == startupAppConfig.ForceEvictionWhenBusy
 		envLRUEvictionMaxRetries := appConfig.LRUEvictionMaxRetries == startupAppConfig.LRUEvictionMaxRetries
@@ -335,6 +336,15 @@ func readRuntimeSettingsJson(startupAppConfig config.ApplicationConfig) fileHand
 			if settings.AutoloadBackendGalleries != nil && !envAutoloadBackendGalleries {
 				appConfig.AutoloadBackendGalleries = *settings.AutoloadBackendGalleries
 			}
+			if settings.PIIDefaultDetectors != nil && !envPIIDefaultDetectors {
+				// Request-side default redaction reads this live via
+				// ResolvePIIPolicy, so a file edit takes effect on the next chat
+				// request. The MITM listener resolves its per-host detector map
+				// once at start, so a raw file edit reaches cloud-proxy traffic
+				// only after a restart or a POST /api/settings (which rebuilds
+				// the listener) — the admin UI uses the latter.
+				appConfig.PIIDefaultDetectors = append([]string(nil), (*settings.PIIDefaultDetectors)...)
+			}
 			if settings.AutoUpgradeBackends != nil {
 				appConfig.AutoUpgradeBackends = *settings.AutoUpgradeBackends
 			}

core/application/runtime_settings_branding_test.go

@@ -109,6 +109,52 @@ var _ = Describe("loadRuntimeSettingsFromFile", func() {
 		})
 	})
 
+	// Instance-wide default PII detectors. The file is the only source (no
+	// env var), and the loader runs immediately before startMITMIfConfigured,
+	// so a regression here means the cloud-proxy MITM listener resolves an
+	// empty detector set at boot and forwards intercepted traffic unredacted —
+	// even though pii_default_detectors is on disk and the MITM model has PII
+	// enabled. It also breaks request-side default redaction the same way.
+	Describe("PII default detectors", func() {
+		It("loads pii_default_detectors from the file", func() {
+			cfg := &config.ApplicationConfig{DynamicConfigsDir: seedSettings(`{"pii_default_detectors": ["privacy-filter-nemotron", "secret-filter"]}`)}
+			loadRuntimeSettingsFromFile(cfg)
+			Expect(cfg.PIIDefaultDetectors).To(Equal([]string{"privacy-filter-nemotron", "secret-filter"}))
+		})
+
+		It("does not override an env/CLI-set value (LOCALAI_PII_DEFAULT_DETECTORS)", func() {
+			cfg := &config.ApplicationConfig{
+				DynamicConfigsDir:   seedSettings(`{"pii_default_detectors": ["from-file"]}`),
+				PIIDefaultDetectors: []string{"from-env"}, // simulate WithPIIDefaultDetectors(env)
+			}
+			loadRuntimeSettingsFromFile(cfg)
+			Expect(cfg.PIIDefaultDetectors).To(Equal([]string{"from-env"}), "env var must win over the persisted file value")
+		})
+	})
+
+	// The live file watcher applies pii_default_detectors on a runtime change
+	// the same way it handles galleries/threads/etc.: env-set values (current
+	// == startup snapshot) are left alone, otherwise the file value is applied
+	// to the live config so request-side default redaction picks it up without
+	// a restart.
+	Describe("file watcher: pii_default_detectors", func() {
+		It("applies a changed file value to the live config", func() {
+			startup := config.ApplicationConfig{} // no env baseline
+			live := &config.ApplicationConfig{PIIDefaultDetectors: []string{"old"}}
+			handler := readRuntimeSettingsJson(startup)
+			Expect(handler([]byte(`{"pii_default_detectors":["new-a","new-b"]}`), live)).To(Succeed())
+			Expect(live.PIIDefaultDetectors).To(Equal([]string{"new-a", "new-b"}))
+		})
+
+		It("leaves an env-controlled value untouched", func() {
+			startup := config.ApplicationConfig{PIIDefaultDetectors: []string{"from-env"}}
+			live := &config.ApplicationConfig{PIIDefaultDetectors: []string{"from-env"}}
+			handler := readRuntimeSettingsJson(startup)
+			Expect(handler([]byte(`{"pii_default_detectors":["from-file"]}`), live)).To(Succeed())
+			Expect(live.PIIDefaultDetectors).To(Equal([]string{"from-env"}), "env-controlled detectors must not be overwritten by the file")
+		})
+	})
+
 	// The Agent Pool block has a mix of zero and non-zero defaults
 	// (Enabled=true, EmbeddingModel="granite-...", MaxChunkingSize=400,
 	// VectorEngine="chromem", AgentHubURL="https://agenthub.localai.io").

core/application/startup.go

@@ -750,6 +750,20 @@ func loadRuntimeSettingsFromFile(options *config.ApplicationConfig) {
 		options.MITMListen = *settings.MITMListen
 	}
 
+	// Instance-wide default PII detectors. LOCALAI_PII_DEFAULT_DETECTORS (via
+	// WithPIIDefaultDetectors) wins when set; otherwise the file is the source
+	// — apply it only when the env/CLI left the value empty, mirroring the
+	// "env > file" precedence used for the other fields. This must land before
+	// startMITMIfConfigured (called right after this loader): the cloud-proxy
+	// listener resolves each intercept host's detectors once at start via
+	// ResolvePIIPolicy, and a MITM model that names no detectors of its own
+	// falls back to these defaults. Without it the listener (and request-side
+	// default redaction) starts with an empty detector set and forwards
+	// traffic unredacted even though pii_default_detectors is on disk.
+	if settings.PIIDefaultDetectors != nil && len(options.PIIDefaultDetectors) == 0 {
+		options.PIIDefaultDetectors = append([]string(nil), (*settings.PIIDefaultDetectors)...)
+	}
+
 	// Backend upgrade flags
 	if settings.AutoUpgradeBackends != nil {
 		if !options.AutoUpgradeBackends {

core/cli/run.go

@@ -181,6 +181,8 @@ type RunCMD struct {
 	// Cloud-proxy MITM listener (off by default).
 	MITMListen string `env:"LOCALAI_MITM_LISTEN" help:"Address (host:port) for the cloudproxy MITM listener. Empty = disabled. Clients set HTTPS_PROXY=http://<this>:<port>. Intercept hosts are declared per-model via the model YAML mitm.hosts: block; create one from the Add Model UI." group:"middleware"`
 	MITMCADir  string `env:"LOCALAI_MITM_CA_DIR" type:"path" help:"Directory holding the MITM proxy CA cert + key. Defaults to <data-path>/mitm-ca." group:"middleware"`
+
+	PIIDefaultDetectors []string `env:"LOCALAI_PII_DEFAULT_DETECTORS" help:"Instance-wide default PII/secret detector model names applied to any PII-enabled model (chiefly cloud-proxy / MITM models) that names no pii.detectors of its own. Comma-separated, e.g. privacy-filter-nemotron,secret-filter. Takes precedence over the value persisted via the Middleware UI." group:"middleware"`
 }
 
 func (r *RunCMD) Run(ctx *cliContext.Context) error {
@@ -243,6 +245,7 @@ func (r *RunCMD) Run(ctx *cliContext.Context) error {
 		config.WithAPIAddress(r.Address),
 		config.WithMITMListen(r.MITMListen),
 		config.WithMITMCADir(r.MITMCADir),
+		config.WithPIIDefaultDetectors(r.PIIDefaultDetectors),
 		config.WithAgentJobRetentionDays(r.AgentJobRetentionDays),
 		config.WithLlamaCPPTunnelCallback(func(tunnels []string) {
 			tunnelEnvVar := strings.Join(tunnels, ",")

core/config/application_config.go

@@ -712,6 +712,18 @@ func WithMITMCADir(dir string) AppOption {
 	}
 }
 
+// WithPIIDefaultDetectors sets the instance-wide default PII/secret detector
+// model names applied to any PII-enabled model (chiefly cloud-proxy / MITM
+// models) that names no pii.detectors of its own. CLI/env:
+// LOCALAI_PII_DEFAULT_DETECTORS. Empty leaves the value to
+// runtime_settings.json / the Middleware UI; a non-empty value takes
+// precedence over the file (env > file).
+func WithPIIDefaultDetectors(detectors []string) AppOption {
+	return func(o *ApplicationConfig) {
+		o.PIIDefaultDetectors = detectors
+	}
+}
+
 func WithDynamicConfigDir(dynamicConfigsDir string) AppOption {
 	return func(o *ApplicationConfig) {
 		o.DynamicConfigsDir = dynamicConfigsDir

core/config/meta/registry.go

@@ -537,6 +537,36 @@ func DefaultRegistry() map[string]FieldMetaOverride {
 			Component:   "number",
 			Order:       79,
 		},
+		"pipeline.compaction.enabled": {
+			Section:     "pipeline",
+			Label:       "Compaction Enabled",
+			Description: "Fold conversation items that age out of the live window (Max History Items) into a rolling summary instead of dropping them, so long realtime sessions stay cheap without losing earlier context. Off by default.",
+			Component:   "toggle",
+			Order:       80,
+		},
+		"pipeline.compaction.trigger_items": {
+			Section:     "pipeline",
+			Label:       "Compaction Trigger Items",
+			Description: "High-water mark: once the live conversation exceeds this many items, the overflow above Max History Items is summarized and evicted. Must be greater than Max History Items; defaults to twice it. The gap controls how often summarization runs.",
+			Component:   "number",
+			Order:       81,
+		},
+		"pipeline.compaction.summary_model": {
+			Section:     "pipeline",
+			Label:       "Compaction Summary Model",
+			Description: "Optional smaller/cheaper model used to produce the rolling summary. Empty reuses the pipeline's own LLM. On CPU, a tiny model here keeps compaction from competing with the conversation LLM.",
+			Component:   "input",
+			Advanced:    true,
+			Order:       82,
+		},
+		"pipeline.compaction.max_summary_tokens": {
+			Section:     "pipeline",
+			Label:       "Compaction Max Summary Tokens",
+			Description: "Advisory cap on the rolling summary length (fed to the summarizer prompt). Defaults to 512.",
+			Component:   "number",
+			Advanced:    true,
+			Order:       83,
+		},
 
 		// --- Functions ---
 		"function.grammar.parallel_calls": {

core/config/model_config.go

@@ -641,11 +641,32 @@ type Pipeline struct {
 	// context fills.
 	MaxHistoryItems *int `yaml:"max_history_items,omitempty" json:"max_history_items,omitempty"`
 
+	// Compaction folds conversation items that age out of the live window
+	// (max_history_items) into a rolling summary instead of dropping them, so
+	// long realtime sessions stay cheap without losing earlier context. Nil
+	// (block absent) means disabled, preserving existing behavior.
+	Compaction *PipelineCompaction `yaml:"compaction,omitempty" json:"compaction,omitempty"`
+
 	// VoiceRecognition gates the pipeline behind speaker verification. Nil
 	// (block absent) means no gate, preserving existing behavior.
 	VoiceRecognition *PipelineVoiceRecognition `yaml:"voice_recognition,omitempty" json:"voice_recognition,omitempty"`
 }
 
+// PipelineCompaction configures summarize-then-drop for a realtime pipeline.
+type PipelineCompaction struct {
+	// Enabled turns summarize-then-drop on. Default false.
+	Enabled bool `yaml:"enabled,omitempty" json:"enabled,omitempty"`
+	// TriggerItems is the high-water mark: once live items exceed it, overflow
+	// above max_history_items is summarized and evicted. Must exceed
+	// max_history_items; clamped up if not. Default: 2x max_history_items.
+	TriggerItems int `yaml:"trigger_items,omitempty" json:"trigger_items,omitempty"`
+	// SummaryModel optionally names a smaller/cheaper model for the summary
+	// call. Empty uses the pipeline's own LLM.
+	SummaryModel string `yaml:"summary_model,omitempty" json:"summary_model,omitempty"`
+	// MaxSummaryTokens advises the summary length (fed to the prompt). Default 512.
+	MaxSummaryTokens int `yaml:"max_summary_tokens,omitempty" json:"max_summary_tokens,omitempty"`
+}
+
 // ApplyReasoningEffort resolves the effective reasoning effort — a per-request
 // value (requestEffort) overrides the config's own ReasoningEffort default —
 // stores it on the config so gRPCPredictOpts forwards it to the backend as the

core/config/runtime_settings_persist.go

@@ -5,6 +5,7 @@ import (
 	"errors"
 	"os"
 	"path/filepath"
+	"reflect"
 )
 
 // runtimeSettingsFile is the on-disk filename inside DynamicConfigsDir.
@@ -33,6 +34,35 @@ func (o *ApplicationConfig) ReadPersistedSettings() (RuntimeSettings, error) {
 	return settings, nil
 }
 
+// MergeNonNil overlays every set (non-nil) field of overlay onto the
+// receiver, leaving the receiver's value untouched wherever overlay left a
+// field unset. Every RuntimeSettings field is a pointer precisely so "set"
+// can be told apart from "absent" (see the type doc), which makes this a
+// faithful partial update: a caller that submits only the field it owns
+// changes exactly that field and never clobbers unrelated settings.
+//
+// This is the read-modify-write contract the persistence helpers exist for.
+// UpdateSettingsEndpoint reads the on-disk settings, merges the request body
+// on top, and writes the result — so a focused admin page that POSTs only its
+// own field (the Middleware page sends only mitm_listen; the detector table
+// only pii_default_detectors) no longer nulls every other setting.
+//
+// Reflection keeps the merge total over the struct: a field added to
+// RuntimeSettings later is merged automatically, so the persistence path can
+// never silently drop a new setting the way a hand-maintained field list
+// would. Non-pointer fields (none today) are skipped — they cannot express
+// "absent", so the receiver wins.
+func (s *RuntimeSettings) MergeNonNil(overlay RuntimeSettings) {
+	dst := reflect.ValueOf(s).Elem()
+	src := reflect.ValueOf(overlay)
+	for i := 0; i < src.NumField(); i++ {
+		f := src.Field(i)
+		if f.Kind() == reflect.Pointer && !f.IsNil() {
+			dst.Field(i).Set(f)
+		}
+	}
+}
+
 // WritePersistedSettings serialises the given RuntimeSettings to
 // runtime_settings.json with restricted permissions (it may carry API
 // keys and P2P tokens).

core/config/runtime_settings_persist_test.go

@@ -12,6 +12,7 @@ import (
 )
 
 func strPtr(s string) *string { return &s }
+func boolPtr(b bool) *bool     { return &b }
 
 var _ = Describe("RuntimeSettings persistence helpers", func() {
 	var (
@@ -51,6 +52,47 @@ var _ = Describe("RuntimeSettings persistence helpers", func() {
 		})
 	})
 
+	// MergeNonNil is the partial-update primitive UpdateSettingsEndpoint
+	// relies on: a focused admin page POSTs only the field it owns, and the
+	// handler reads the on-disk settings and overlays the request on top.
+	// Without it, the body would be written verbatim and every field the
+	// caller omitted would be nulled (the reported regression: changing
+	// mitm_listen wiped the galleries, api keys, watchdog config, etc.).
+	Describe("MergeNonNil partial update", func() {
+		It("overlays set fields and preserves unset ones", func() {
+			base := config.RuntimeSettings{
+				MITMListen:          strPtr(":9000"),
+				Galleries:           &[]config.Gallery{{Name: "g1", URL: "http://example/g1"}},
+				WatchdogIdleEnabled: boolPtr(true),
+				ApiKeys:             &[]string{"persisted-key"},
+				PIIDefaultDetectors: &[]string{"det-a"},
+			}
+
+			// Simulate the Middleware proxy tab: only mitm_listen is sent.
+			overlay := config.RuntimeSettings{MITMListen: strPtr(":8443")}
+			base.MergeNonNil(overlay)
+
+			Expect(base.MITMListen).ToNot(BeNil())
+			Expect(*base.MITMListen).To(Equal(":8443"), "set field should be overlaid")
+			// Everything the overlay left unset must survive untouched.
+			Expect(base.Galleries).ToNot(BeNil(), "galleries were clobbered")
+			Expect(*base.Galleries).To(HaveLen(1))
+			Expect(base.WatchdogIdleEnabled).ToNot(BeNil())
+			Expect(*base.WatchdogIdleEnabled).To(BeTrue())
+			Expect(base.ApiKeys).ToNot(BeNil(), "api_keys were clobbered")
+			Expect(*base.ApiKeys).To(Equal([]string{"persisted-key"}))
+			Expect(base.PIIDefaultDetectors).ToNot(BeNil(), "pii_default_detectors were clobbered")
+			Expect(*base.PIIDefaultDetectors).To(Equal([]string{"det-a"}))
+		})
+
+		It("lets an explicit empty slice clear a field", func() {
+			base := config.RuntimeSettings{PIIDefaultDetectors: &[]string{"det-a"}}
+			base.MergeNonNil(config.RuntimeSettings{PIIDefaultDetectors: &[]string{}})
+			Expect(base.PIIDefaultDetectors).ToNot(BeNil())
+			Expect(*base.PIIDefaultDetectors).To(BeEmpty(), "an explicit empty slice should clear, not preserve")
+		})
+	})
+
 	// MITM round trip pins the contract that loadRuntimeSettingsFromFile
 	// MITM listener address must survive a write/read round trip so the
 	// next process restart can bring the listener back up. (Intercept

core/http/endpoints/localai/agent_collections.go

@@ -70,7 +70,7 @@ func UploadToCollectionEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		file, err := c.FormFile("file")
 		if err != nil {
 			return c.JSON(http.StatusBadRequest, map[string]string{"error": "file required"})
@@ -116,7 +116,7 @@ func ListCollectionEntriesEndpoint(app *application.Application) echo.HandlerFun
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		entries, err := svc.ListCollectionEntriesForUser(userID, c.Param("name"))
+		entries, err := svc.ListCollectionEntriesForUser(userID, decodedParam(c, "name"))
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
@@ -139,7 +139,7 @@ func GetCollectionEntryContentEndpoint(app *application.Application) echo.Handle
 		if err != nil {
 			entry = entryParam
 		}
-		content, chunkCount, err := svc.GetCollectionEntryContentForUser(userID, c.Param("name"), entry)
+		content, chunkCount, err := svc.GetCollectionEntryContentForUser(userID, decodedParam(c, "name"), entry)
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
@@ -164,7 +164,7 @@ func SearchCollectionEndpoint(app *application.Application) echo.HandlerFunc {
 		if err := c.Bind(&payload); err != nil {
 			return c.JSON(http.StatusBadRequest, map[string]string{"error": err.Error()})
 		}
-		results, err := svc.SearchCollectionForUser(userID, c.Param("name"), payload.Query, payload.MaxResults)
+		results, err := svc.SearchCollectionForUser(userID, decodedParam(c, "name"), payload.Query, payload.MaxResults)
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
@@ -182,7 +182,7 @@ func ResetCollectionEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		if err := svc.ResetCollectionForUser(userID, c.Param("name")); err != nil {
+		if err := svc.ResetCollectionForUser(userID, decodedParam(c, "name")); err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
 			}
@@ -202,7 +202,7 @@ func DeleteCollectionEntryEndpoint(app *application.Application) echo.HandlerFun
 		if err := c.Bind(&payload); err != nil {
 			return c.JSON(http.StatusBadRequest, map[string]string{"error": err.Error()})
 		}
-		remaining, err := svc.DeleteCollectionEntryForUser(userID, c.Param("name"), payload.Entry)
+		remaining, err := svc.DeleteCollectionEntryForUser(userID, decodedParam(c, "name"), payload.Entry)
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
@@ -230,7 +230,7 @@ func AddCollectionSourceEndpoint(app *application.Application) echo.HandlerFunc
 		if payload.UpdateInterval < 1 {
 			payload.UpdateInterval = 60
 		}
-		if err := svc.AddCollectionSourceForUser(userID, c.Param("name"), payload.URL, payload.UpdateInterval); err != nil {
+		if err := svc.AddCollectionSourceForUser(userID, decodedParam(c, "name"), payload.URL, payload.UpdateInterval); err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
 			}
@@ -250,7 +250,7 @@ func RemoveCollectionSourceEndpoint(app *application.Application) echo.HandlerFu
 		if err := c.Bind(&payload); err != nil {
 			return c.JSON(http.StatusBadRequest, map[string]string{"error": err.Error()})
 		}
-		if err := svc.RemoveCollectionSourceForUser(userID, c.Param("name"), payload.URL); err != nil {
+		if err := svc.RemoveCollectionSourceForUser(userID, decodedParam(c, "name"), payload.URL); err != nil {
 			return c.JSON(http.StatusInternalServerError, map[string]string{"error": err.Error()})
 		}
 		return c.JSON(http.StatusOK, map[string]string{"status": "ok"})
@@ -267,7 +267,7 @@ func GetCollectionEntryRawFileEndpoint(app *application.Application) echo.Handle
 		if err != nil {
 			entry = entryParam
 		}
-		fpath, err := svc.GetCollectionEntryFilePathForUser(userID, c.Param("name"), entry)
+		fpath, err := svc.GetCollectionEntryFilePathForUser(userID, decodedParam(c, "name"), entry)
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
@@ -282,7 +282,7 @@ func ListCollectionSourcesEndpoint(app *application.Application) echo.HandlerFun
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		sources, err := svc.ListCollectionSourcesForUser(userID, c.Param("name"))
+		sources, err := svc.ListCollectionSourcesForUser(userID, decodedParam(c, "name"))
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})

core/http/endpoints/localai/agent_collections_param_test.go

@@ -0,0 +1,49 @@
+package localai
+
+import (
+	"net/http"
+	"net/http/httptest"
+
+	"github.com/labstack/echo/v4"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+// Regression for #10443: agent/collection names carry a "legacy-api-key:"
+// prefix, so the ':' is percent-encoded as %3A in the request path. Echo routes
+// such paths via URL.RawPath and stores the path-param value still escaped, so
+// handlers must URL-decode it before looking the collection up in the store -
+// otherwise the lookup sees "legacy-api-key%3ALiteraryResearch" and 404s.
+var _ = Describe("decodedParam", func() {
+	var e *echo.Echo
+
+	BeforeEach(func() {
+		e = echo.New()
+	})
+
+	// route runs a request through Echo's real router so the path param is
+	// populated exactly as it would be in production, then returns the decoded
+	// value the handler would observe.
+	route := func(rawPath string) string {
+		var got string
+		e.GET("/api/agents/collections/:name/upload", func(c echo.Context) error {
+			got = decodedParam(c, "name")
+			return c.NoContent(http.StatusOK)
+		})
+		req := httptest.NewRequest(http.MethodGet, rawPath, nil)
+		rec := httptest.NewRecorder()
+		e.ServeHTTP(rec, req)
+		Expect(rec.Code).To(Equal(http.StatusOK))
+		return got
+	}
+
+	It("decodes a percent-encoded colon in the collection name", func() {
+		got := route("/api/agents/collections/legacy-api-key%3ALiteraryResearch/upload")
+		Expect(got).To(Equal("legacy-api-key:LiteraryResearch"))
+	})
+
+	It("leaves an unencoded name untouched", func() {
+		got := route("/api/agents/collections/PlainCollection/upload")
+		Expect(got).To(Equal("PlainCollection"))
+	})
+})

core/http/endpoints/localai/agents.go

@@ -6,6 +6,7 @@ import (
 	"io"
 	"maps"
 	"net/http"
+	"net/url"
 	"os"
 	"path/filepath"
 	"slices"
@@ -33,6 +34,22 @@ func getUserID(c echo.Context) string {
 	return user.ID
 }
 
+// decodedParam returns the named path parameter, URL-decoding it.
+//
+// Echo routes a request via URL.RawPath whenever the path contains
+// percent-encoded characters (e.g. %3A for ':'), and in that case stores the
+// matched path-param value raw/escaped. Agent and collection names carry a
+// "legacy-api-key:" prefix, so the ':' arrives as %3A and the raw param no
+// longer matches the stored name. Callers must unescape before lookups.
+// Falls back to the raw value if it isn't valid percent-encoding.
+func decodedParam(c echo.Context, name string) string {
+	raw := c.Param(name)
+	if decoded, err := url.PathUnescape(raw); err == nil {
+		return decoded
+	}
+	return raw
+}
+
 // isAdminUser returns true if the authenticated user has admin role.
 func isAdminUser(c echo.Context) bool {
 	user := auth.GetUser(c)
@@ -127,7 +144,7 @@ func GetAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 
 		statuses := svc.ListAgentsForUser(userID)
 		active, exists := statuses[name]
@@ -142,7 +159,7 @@ func UpdateAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		var cfg state.AgentConfig
 		if err := c.Bind(&cfg); err != nil {
 			return c.JSON(http.StatusBadRequest, map[string]string{"error": err.Error()})
@@ -161,7 +178,7 @@ func DeleteAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		if err := svc.DeleteAgentForUser(userID, name); err != nil {
 			return c.JSON(http.StatusInternalServerError, map[string]string{"error": err.Error()})
 		}
@@ -173,7 +190,7 @@ func GetAgentConfigEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		cfg := svc.GetAgentConfigForUser(userID, name)
 		if cfg == nil {
 			return c.JSON(http.StatusNotFound, map[string]string{"error": "Agent not found"})
@@ -186,7 +203,7 @@ func PauseAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		if err := svc.PauseAgentForUser(userID, c.Param("name")); err != nil {
+		if err := svc.PauseAgentForUser(userID, decodedParam(c, "name")); err != nil {
 			return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
 		}
 		return c.JSON(http.StatusOK, map[string]string{"status": "ok"})
@@ -197,7 +214,7 @@ func ResumeAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		if err := svc.ResumeAgentForUser(userID, c.Param("name")); err != nil {
+		if err := svc.ResumeAgentForUser(userID, decodedParam(c, "name")); err != nil {
 			return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
 		}
 		return c.JSON(http.StatusOK, map[string]string{"status": "ok"})
@@ -208,7 +225,7 @@ func GetAgentStatusEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 
 		history := svc.GetAgentStatusForUser(userID, name)
 		if history == nil {
@@ -241,7 +258,7 @@ func GetAgentObservablesEndpoint(app *application.Application) echo.HandlerFunc
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 
 		history, err := svc.GetAgentObservablesForUser(userID, name)
 		if err != nil {
@@ -261,7 +278,7 @@ func ClearAgentObservablesEndpoint(app *application.Application) echo.HandlerFun
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		if err := svc.ClearAgentObservablesForUser(userID, name); err != nil {
 			return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
 		}
@@ -273,7 +290,7 @@ func ChatWithAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		var payload struct {
 			Message string `json:"message"`
 		}
@@ -302,7 +319,7 @@ func AgentSSEEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 
 		// Try local SSE manager first
 		manager := svc.GetSSEManagerForUser(userID, name)
@@ -334,7 +351,7 @@ func ExportAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		data, err := svc.ExportAgentForUser(userID, name)
 		if err != nil {
 			return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})

core/http/endpoints/localai/settings.go

@@ -4,8 +4,6 @@ import (
 	"encoding/json"
 	"io"
 	"net/http"
-	"os"
-	"path/filepath"
 	"time"
 
 	"github.com/labstack/echo/v4"
@@ -110,6 +108,18 @@ func UpdateSettingsEndpoint(app *application.Application) echo.HandlerFunc {
 			})
 		}
 
+		// Read whatever is already persisted: it is both the source of truth
+		// for branding asset filenames (below) and the base we merge this
+		// request onto before writing. A read failure must not let a Save
+		// silently discard the existing settings — surface it instead.
+		persisted, err := appConfig.ReadPersistedSettings()
+		if err != nil {
+			return c.JSON(http.StatusInternalServerError, schema.SettingsResponse{
+				Success: false,
+				Error:   "Failed to read existing settings: " + err.Error(),
+			})
+		}
+
 		// Branding asset filenames are owned exclusively by
 		// /api/branding/asset/{kind} (upload/delete). The Settings page also
 		// round-trips them via GET /api/settings, but its local state is stale
@@ -118,11 +128,9 @@ func UpdateSettingsEndpoint(app *application.Application) echo.HandlerFunc {
 		// at page open. Replace whatever the body sent for these three fields
 		// with the values currently on disk so /api/settings can never
 		// regress them.
-		if existing, err := appConfig.ReadPersistedSettings(); err == nil {
-			settings.LogoFile = existing.LogoFile
-			settings.LogoHorizontalFile = existing.LogoHorizontalFile
-			settings.FaviconFile = existing.FaviconFile
-		}
+		settings.LogoFile = persisted.LogoFile
+		settings.LogoHorizontalFile = persisted.LogoHorizontalFile
+		settings.FaviconFile = persisted.FaviconFile
 
 		// The UI reads ApiKeys from GET /api/settings, which already returns the
 		// merged env+runtime list. When the user clicks Save, the same merged
@@ -145,16 +153,17 @@ func UpdateSettingsEndpoint(app *application.Application) echo.HandlerFunc {
 			settings.ApiKeys = &runtimeOnly
 		}
 
-		settingsFile := filepath.Join(appConfig.DynamicConfigsDir, "runtime_settings.json")
-		settingsJSON, err := json.MarshalIndent(settings, "", "  ")
-		if err != nil {
-			return c.JSON(http.StatusInternalServerError, schema.SettingsResponse{
-				Success: false,
-				Error:   "Failed to marshal settings: " + err.Error(),
-			})
-		}
-
-		if err := os.WriteFile(settingsFile, settingsJSON, 0600); err != nil {
+		// Persist as a partial update: overlay only the fields this request set
+		// onto the settings already on disk. Focused admin pages POST just the
+		// keys they own (the Middleware proxy tab sends only mitm_listen; the
+		// detector table only pii_default_detectors), so writing the request
+		// body verbatim would null every unrelated setting (the no-omitempty
+		// api_keys / pii_default_detectors fields even round-trip as JSON
+		// null). The full Settings page still round-trips every field, so its
+		// Save is unchanged.
+		toPersist := persisted
+		toPersist.MergeNonNil(settings)
+		if err := appConfig.WritePersistedSettings(toPersist); err != nil {
 			return c.JSON(http.StatusInternalServerError, schema.SettingsResponse{
 				Success: false,
 				Error:   "Failed to write settings file: " + err.Error(),
@@ -262,7 +271,14 @@ func UpdateSettingsEndpoint(app *application.Application) echo.HandlerFunc {
 			}
 		}
 
-		if settings.MITMListen != nil {
+		// Rebuild the MITM listener when its address OR the instance-wide
+		// default detectors change. The per-host detector map is resolved once
+		// at listener start (startMITMLocked → ResolvePIIPolicy), so a
+		// default-detector change is otherwise invisible to cloud-proxy traffic
+		// until the next restart — an admin toggling a default detector would
+		// see no redaction. RestartMITM is a no-op when the listener is
+		// disabled (empty address).
+		if settings.MITMListen != nil || settings.PIIDefaultDetectors != nil {
 			if err := app.RestartMITM(); err != nil {
 				xlog.Error("Failed to restart MITM proxy", "error", err)
 				return c.JSON(http.StatusInternalServerError, schema.SettingsResponse{

core/http/endpoints/localai/settings_test.go

@@ -52,6 +52,10 @@ var _ = Describe("Settings endpoints", func() {
 		// Settings are persisted here; set after construction since there's no
 		// dedicated AppOption for it.
 		app.ApplicationConfig().DynamicConfigsDir = tmp
+		// Contain the MITM CA inside tmp too. The partial-save spec flips
+		// mitm_listen, which starts the listener and writes a CA; without this
+		// it defaults to ./mitm-ca and litters the package source tree.
+		app.ApplicationConfig().MITMCADir = filepath.Join(tmp, "mitm-ca")
 
 		e = echo.New()
 		e.GET("/api/settings", GetSettingsEndpoint(app))
@@ -109,6 +113,57 @@ var _ = Describe("Settings endpoints", func() {
 		Expect(err).ToNot(HaveOccurred())
 	})
 
+	// Regression: a focused admin page (the Middleware proxy tab) POSTs only
+	// the one field it owns — mitm_listen. The old handler wrote the request
+	// body verbatim, so every other persisted setting was dropped (and
+	// api_keys / pii_default_detectors, which lack omitempty, were written as
+	// null). A partial POST must now merge onto what is already on disk.
+	It("preserves unrelated persisted settings when a partial POST sets only mitm_listen", func() {
+		// First save establishes a fuller settings file (as the full Settings
+		// page would): galleries, an API key, and the MITM listener. The
+		// listener restart binds a real socket, so use 127.0.0.1:0 for an
+		// ephemeral free port rather than a fixed one that may be in use.
+		rec := post(`{"mitm_listen":"127.0.0.1:0","galleries":[{"name":"g1","url":"http://example/g1"}],"api_keys":["k1"],"pii_default_detectors":["det-a"]}`)
+		Expect(rec.Code).To(Equal(http.StatusOK), rec.Body.String())
+
+		// The Middleware proxy tab then changes only the listen address — the
+		// exact partial body that nulled everything else before the fix.
+		rec = post(`{"mitm_listen":"127.0.0.1:0"}`)
+		Expect(rec.Code).To(Equal(http.StatusOK), rec.Body.String())
+
+		raw, err := os.ReadFile(filepath.Join(tmp, "runtime_settings.json"))
+		Expect(err).ToNot(HaveOccurred())
+		var ondisk config.RuntimeSettings
+		Expect(json.Unmarshal(raw, &ondisk)).To(Succeed())
+
+		Expect(ondisk.MITMListen).ToNot(BeNil())
+		Expect(*ondisk.MITMListen).To(Equal("127.0.0.1:0"), "the changed field should be saved")
+		Expect(ondisk.Galleries).ToNot(BeNil(), "galleries were clobbered by the partial save")
+		Expect(*ondisk.Galleries).To(HaveLen(1))
+		Expect(ondisk.ApiKeys).ToNot(BeNil(), "api_keys were nulled by the partial save")
+		Expect(*ondisk.ApiKeys).To(Equal([]string{"k1"}))
+		Expect(ondisk.PIIDefaultDetectors).ToNot(BeNil(), "pii_default_detectors were nulled by the partial save")
+		Expect(*ondisk.PIIDefaultDetectors).To(Equal([]string{"det-a"}))
+	})
+
+	// The MITM listener resolves its per-host PII detectors once at start
+	// (startMITMLocked → ResolvePIIPolicy), and the handler used to restart it
+	// only when mitm_listen changed. So an admin toggling a default detector
+	// (the Middleware detector table POSTs only pii_default_detectors) left
+	// cloud-proxy traffic unredacted until the next reboot. A
+	// pii_default_detectors change must now rebuild the listener.
+	It("rebuilds the MITM listener when only pii_default_detectors changes", func() {
+		rec := post(`{"mitm_listen":"127.0.0.1:0"}`)
+		Expect(rec.Code).To(Equal(http.StatusOK), rec.Body.String())
+		srv1 := app.MITMServer()
+		Expect(srv1).ToNot(BeNil(), "listener should be running after mitm_listen is set")
+
+		rec = post(`{"pii_default_detectors":["det-a"]}`)
+		Expect(rec.Code).To(Equal(http.StatusOK), rec.Body.String())
+		Expect(app.MITMServer()).ToNot(BeIdenticalTo(srv1),
+			"a default-detector change must restart the listener so it picks up the new detectors")
+	})
+
 	// Residual #9125: enabling the watchdog from a cold (off) state via the
 	// React master toggle must start the live watchdog immediately, without a
 	// restart. The toggle posts watchdog_idle_enabled/busy_enabled=true while

core/http/endpoints/openai/realtime.go

@@ -12,6 +12,7 @@ import (
 	"os"
 	"strconv"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"net/http"
@@ -134,6 +135,18 @@ type Session struct {
 	// pairs are kept together so we never feed an orphaned tool result.
 	MaxHistoryItems int
 
+	// Compaction settings resolved from pipeline.compaction (see resolveCompaction).
+	CompactionEnabled bool
+	CompactionTrigger int
+	SummaryModel      string
+	MaxSummaryTokens  int
+
+	// summarizerFactory lazily builds the model used for compaction summaries
+	// when summary_model is configured; nil means reuse the pipeline LLM.
+	summarizerFactory func() (Model, error)
+	summarizerOnce    sync.Once
+	summarizerCached  Model
+
 	// AssistantExecutor is non-nil when the session opted into the in-process
 	// LocalAI Assistant tool surface. Tool calls whose name matches this
 	// executor's catalog are run inproc and their output is fed back to the
@@ -241,6 +254,12 @@ type Conversation struct {
 	ID    string
 	Items []*types.MessageItemUnion
 	Lock  sync.Mutex
+	// Memory is the rolling summary of items already evicted by compaction. It
+	// is kept out of Items (so trimRealtimeItems never drops it) and rendered
+	// as a system message right after the session instructions.
+	Memory string
+	// compacting ensures at most one background compaction runs per conversation.
+	compacting atomic.Bool
 }
 
 func (c *Conversation) ToServer() types.Conversation {
@@ -540,13 +559,12 @@ func runRealtimeSession(application *application.Application, t Transport, model
 		SoundDetectionWindowMs:  cfg.Pipeline.SoundDetectionWindowMs,
 		SoundDetectionHopMs:     cfg.Pipeline.SoundDetectionHopMs,
 	}
+	session.CompactionEnabled, session.CompactionTrigger, session.MaxSummaryTokens, session.SummaryModel = resolveCompaction(cfg, session.MaxHistoryItems)
 
 	// Create a default conversation
 	conversationID := generateConversationID()
 	conversation := &Conversation{
-		ID: conversationID,
-		// TODO: We need to truncate the conversation items when a new item is added and we have run out of space. There are multiple places where items
-		//       can be added so we could use a datastructure here that enforces truncation upon addition
+		ID:    conversationID,
 		Items: []*types.MessageItemUnion{},
 	}
 	session.Conversations[conversationID] = conversation
@@ -577,6 +595,18 @@ func runRealtimeSession(application *application.Application, t Transport, model
 	}
 	session.ModelInterface = m
 
+	if session.SummaryModel != "" {
+		summaryModelName := session.SummaryModel
+		sid := sessionID
+		session.summarizerFactory = func() (Model, error) {
+			summaryCfg, lerr := application.ModelConfigLoader().LoadModelConfigFileByNameDefaultOptions(summaryModelName, application.ApplicationConfig())
+			if lerr != nil {
+				return nil, fmt.Errorf("load summary model config %q: %w", summaryModelName, lerr)
+			}
+			return newModel(&summaryCfg.Pipeline, application.ModelConfigLoader(), application.ModelLoader(), application.ApplicationConfig(), evaluator, buildRealtimeRoutingContext(application, sid))
+		}
+	}
+
 	if cfg.Pipeline.VoiceGateEnabled() {
 		gate, gerr := newVoiceGate(
 			*cfg.Pipeline.VoiceRecognition,
@@ -807,6 +837,15 @@ func runRealtimeSession(application *application.Application, t Transport, model
 				commitUtterance(respCtx, allAudio, session, conversation, t)
 			}()
 
+		case types.InputAudioBufferClearEvent:
+			xlog.Debug("recv", "message", string(msg))
+			// Discard a partially-captured utterance so the client can restart
+			// input cleanly without the stale buffer leaking into the next commit.
+			clearInputAudio(session)
+			sendEvent(t, types.InputAudioBufferClearedEvent{
+				ServerEventBase: types.ServerEventBase{EventID: e.EventID},
+			})
+
 		case types.ConversationItemCreateEvent:
 			xlog.Debug("recv", "message", string(msg))
 			// Add the item to the conversation
@@ -841,7 +880,39 @@ func runRealtimeSession(application *application.Application, t Transport, model
 			})
 
 		case types.ConversationItemDeleteEvent:
-			sendError(t, "not_implemented", "Deleting items not implemented", "", "event_TODO")
+			xlog.Debug("recv", "message", string(msg))
+			if e.ItemID == "" {
+				sendError(t, "invalid_item_id", "Need item_id, but none specified", "", "event_TODO")
+				continue
+			}
+			conversation.Lock.Lock()
+			updated, ok := deleteItem(conversation.Items, e.ItemID)
+			conversation.Items = updated
+			conversation.Lock.Unlock()
+			if !ok {
+				sendError(t, "invalid_item_id", "Item to delete not found", "", "event_TODO")
+				continue
+			}
+			sendEvent(t, types.ConversationItemDeletedEvent{
+				ServerEventBase: types.ServerEventBase{EventID: e.EventID},
+				ItemID:          e.ItemID,
+			})
+
+		case types.ConversationItemTruncateEvent:
+			xlog.Debug("recv", "message", string(msg))
+			conversation.Lock.Lock()
+			ok := truncateAssistantText(conversation.Items, e.ItemID, e.ContentIndex)
+			conversation.Lock.Unlock()
+			if !ok {
+				sendError(t, "invalid_item_id", "Item to truncate not found", "", "event_TODO")
+				continue
+			}
+			sendEvent(t, types.ConversationItemTruncatedEvent{
+				ServerEventBase: types.ServerEventBase{EventID: e.EventID},
+				ItemID:          e.ItemID,
+				ContentIndex:    e.ContentIndex,
+				AudioEndMs:      e.AudioEndMs,
+			})
 
 		case types.ConversationItemRetrieveEvent:
 			xlog.Debug("recv", "message", string(msg))
@@ -854,21 +925,7 @@ func runRealtimeSession(application *application.Application, t Transport, model
 			conversation.Lock.Lock()
 			var retrievedItem types.MessageItemUnion
 			for _, item := range conversation.Items {
-				// We need to check ID in the union
-				var id string
-				if item.System != nil {
-					id = item.System.ID
-				} else if item.User != nil {
-					id = item.User.ID
-				} else if item.Assistant != nil {
-					id = item.Assistant.ID
-				} else if item.FunctionCall != nil {
-					id = item.FunctionCall.ID
-				} else if item.FunctionCallOutput != nil {
-					id = item.FunctionCallOutput.ID
-				}
-
-				if id == e.ItemID {
+				if itemID(item) == e.ItemID {
 					retrievedItem = *item
 					break
 				}
@@ -1666,6 +1723,9 @@ const maxAssistantToolTurns = 10
 
 func triggerResponse(ctx context.Context, session *Session, conv *Conversation, t Transport, overrides *types.ResponseCreateParams) {
 	triggerResponseAtTurn(ctx, session, conv, t, overrides, 0)
+	// Fold aged-out turns into the rolling memory off the critical path; the
+	// next turn reaps the smaller buffer.
+	session.maybeCompact(conv)
 }
 
 func triggerResponseAtTurn(ctx context.Context, session *Session, conv *Conversation, t Transport, overrides *types.ResponseCreateParams, toolTurn int) {
@@ -1721,6 +1781,7 @@ func triggerResponseAtTurn(ctx context.Context, session *Session, conv *Conversa
 	var lastUserSpeaker *types.Speaker
 	personalize := session.voiceGate != nil && session.voiceGate.cfg.PersonalizeEnabled()
 	conv.Lock.Lock()
+	conversationHistory = withMemory(conversationHistory, conv.Memory)
 	items := trimRealtimeItems(conv.Items, session.MaxHistoryItems)
 	for _, item := range items {
 		if item.User != nil {

core/http/endpoints/openai/realtime_compaction.go

@@ -0,0 +1,326 @@
+package openai
+
+import (
+	"context"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/http/endpoints/openai/types"
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/pkg/reasoning"
+	"github.com/mudler/xlog"
+)
+
+const (
+	defaultMaxSummaryTokens = 512
+	memoryPrefix            = "Summary of earlier conversation:\n"
+	// compactionTimeout bounds the summarizer call so a stuck model can't pin the
+	// compacting flag (and thus block all further compaction) forever.
+	compactionTimeout = 60 * time.Second
+)
+
+// withMemory inserts the rolling summary as a system message after the existing
+// (instructions) history. No-op when memory is empty.
+func withMemory(history schema.Messages, memory string) schema.Messages {
+	if memory == "" {
+		return history
+	}
+	content := memoryPrefix + memory
+	return append(history, schema.Message{
+		Role:          string(types.MessageRoleSystem),
+		StringContent: content,
+		Content:       content,
+	})
+}
+
+// renderItemsTranscript renders conversation items as a plain "role: text"
+// transcript for summarization. Non-text items (bare tool calls) are labelled
+// so the summarizer keeps track of actions taken.
+func renderItemsTranscript(items []*types.MessageItemUnion) string {
+	var b strings.Builder
+	for _, item := range items {
+		switch {
+		case item.User != nil:
+			b.WriteString("user: ")
+			for _, c := range item.User.Content {
+				if c.Text != "" {
+					b.WriteString(c.Text)
+				}
+				if c.Transcript != "" {
+					b.WriteString(c.Transcript)
+				}
+			}
+			b.WriteString("\n")
+		case item.Assistant != nil:
+			b.WriteString("assistant: ")
+			// Realtime assistant *audio* turns store the spoken words in
+			// .Transcript (not .Text), so emit both or spoken turns are dropped.
+			for _, c := range item.Assistant.Content {
+				if c.Text != "" {
+					b.WriteString(c.Text)
+				}
+				if c.Transcript != "" {
+					b.WriteString(c.Transcript)
+				}
+			}
+			b.WriteString("\n")
+		case item.FunctionCall != nil:
+			b.WriteString(fmt.Sprintf("assistant called tool %s(%s)\n", item.FunctionCall.Name, item.FunctionCall.Arguments))
+		case item.FunctionCallOutput != nil:
+			b.WriteString(fmt.Sprintf("tool result: %s\n", item.FunctionCallOutput.Output))
+		}
+	}
+	return strings.TrimSpace(b.String())
+}
+
+// buildSummaryMessages builds the chat messages for the summarizer LLM: a system
+// instruction plus prior memory and the new transcript to fold in. maxTokens is
+// advisory (fed to the prompt; not hard-enforced in v1).
+func buildSummaryMessages(priorMemory, transcript string, maxTokens int) schema.Messages {
+	system := fmt.Sprintf("You maintain a running memory of a live voice conversation. "+
+		"Merge the prior memory with the new exchanges into an updated memory. "+
+		"Keep names, decisions, facts, preferences, and open threads. Be concise "+
+		"(under ~%d tokens). Output only the updated memory, with no reasoning or tags.", maxTokens)
+	var user strings.Builder
+	if priorMemory != "" {
+		user.WriteString("Prior memory:\n")
+		user.WriteString(priorMemory)
+		user.WriteString("\n\n")
+	}
+	user.WriteString("New exchanges to fold in:\n")
+	user.WriteString(transcript)
+	return schema.Messages{
+		{Role: string(types.MessageRoleSystem), StringContent: system, Content: system},
+		{Role: string(types.MessageRoleUser), StringContent: user.String(), Content: user.String()},
+	}
+}
+
+// clearInputAudio resets the session's pending input audio buffer (the raw
+// PCM and any buffered Opus frames). Used by the input_audio_buffer.clear
+// realtime event so a client can discard a partially-captured utterance.
+func clearInputAudio(s *Session) {
+	s.AudioBufferLock.Lock()
+	s.InputAudioBuffer = nil
+	s.AudioBufferLock.Unlock()
+	s.OpusFramesLock.Lock()
+	s.OpusFrames = nil
+	s.OpusFramesLock.Unlock()
+}
+
+// itemID extracts the id from any MessageItemUnion variant ("" if none).
+func itemID(item *types.MessageItemUnion) string {
+	switch {
+	case item == nil:
+		return ""
+	case item.System != nil:
+		return item.System.ID
+	case item.User != nil:
+		return item.User.ID
+	case item.Assistant != nil:
+		return item.Assistant.ID
+	case item.FunctionCall != nil:
+		return item.FunctionCall.ID
+	case item.FunctionCallOutput != nil:
+		return item.FunctionCallOutput.ID
+	default:
+		return ""
+	}
+}
+
+// deleteItem removes the item with id from items, returning the new slice and
+// whether it was found.
+func deleteItem(items []*types.MessageItemUnion, id string) ([]*types.MessageItemUnion, bool) {
+	for i, item := range items {
+		if itemID(item) == id {
+			return append(items[:i:i], items[i+1:]...), true
+		}
+	}
+	return items, false
+}
+
+// truncateAssistantText clears the text of the assistant item's content part at
+// contentIndex. Minimal truncate: used to discard an interrupted/barge-in
+// response tail. Both .Text and .Transcript are cleared because realtime audio
+// turns store the spoken words in .Transcript (clearing only .Text would no-op).
+func truncateAssistantText(items []*types.MessageItemUnion, id string, contentIndex int) bool {
+	for _, item := range items {
+		if itemID(item) != id || item.Assistant == nil {
+			continue
+		}
+		if contentIndex >= 0 && contentIndex < len(item.Assistant.Content) {
+			item.Assistant.Content[contentIndex].Text = ""
+			item.Assistant.Content[contentIndex].Transcript = ""
+		}
+		return true
+	}
+	return false
+}
+
+// compactionCut returns the index splitting items into overflow (items[:cut],
+// to be summarized+evicted) and the kept live tail (items[cut:]), keeping the
+// last `keep` items. It mirrors trimRealtimeItems' pair-safety: the cut is
+// pulled left so a function_call and its function_call_output are never split
+// across the boundary (the whole pair lands in the kept tail). Returns 0 when
+// there is nothing to cut.
+func compactionCut(items []*types.MessageItemUnion, keep int) int {
+	// keep <= 0 means no live-window cap (the "unlimited history" sentinel, as
+	// in trimRealtimeItems): there is nothing to evict, so cut nothing. This
+	// also avoids indexing items[len(items)] in the pair-safety loop below.
+	if keep <= 0 {
+		return 0
+	}
+	cut := len(items) - keep
+	if cut <= 0 {
+		return 0
+	}
+	for cut > 0 && items[cut] != nil && items[cut].FunctionCallOutput != nil {
+		cut--
+	}
+	return cut
+}
+
+// resolveCompaction reads the pipeline.compaction block, applying defaults and
+// the trigger>max_history invariant. maxHistory is the already-resolved live
+// window size. Returns enabled=false (and zero values) when compaction is off.
+func resolveCompaction(cfg *config.ModelConfig, maxHistory int) (enabled bool, trigger, maxSummaryTokens int, summaryModel string) {
+	if cfg == nil || cfg.Pipeline.Compaction == nil || !cfg.Pipeline.Compaction.Enabled {
+		return false, 0, 0, ""
+	}
+	c := cfg.Pipeline.Compaction
+	trigger = c.TriggerItems
+	if trigger <= 0 {
+		trigger = maxHistory * 2
+	}
+	if trigger <= maxHistory {
+		trigger = maxHistory + 1
+	}
+	maxSummaryTokens = c.MaxSummaryTokens
+	if maxSummaryTokens <= 0 {
+		maxSummaryTokens = defaultMaxSummaryTokens
+	}
+	return true, trigger, maxSummaryTokens, c.SummaryModel
+}
+
+// prefixMatches reports whether items begins with the same ids, in order, as
+// snapshot — i.e. the overflow we summarized is still at the head (no concurrent
+// client delete reshuffled it).
+func prefixMatches(items, snapshot []*types.MessageItemUnion) bool {
+	if len(items) < len(snapshot) {
+		return false
+	}
+	for i := range snapshot {
+		if itemID(items[i]) != itemID(snapshot[i]) {
+			return false
+		}
+	}
+	return true
+}
+
+// compact folds overflow items into conv.Memory and evicts them. It never holds
+// conv.Lock across the summarizer call: snapshot under lock, summarize unlocked,
+// commit under lock (re-validating the head is unchanged). On any error it
+// leaves the conversation untouched — items are never dropped without a summary.
+func (s *Session) compact(conv *Conversation, model Model) {
+	if model == nil {
+		return
+	}
+	// Snapshot.
+	conv.Lock.Lock()
+	if len(conv.Items) <= s.CompactionTrigger {
+		conv.Lock.Unlock()
+		return
+	}
+	cut := compactionCut(conv.Items, s.MaxHistoryItems)
+	if cut <= 0 {
+		conv.Lock.Unlock()
+		return
+	}
+	overflow := append([]*types.MessageItemUnion(nil), conv.Items[:cut]...)
+	prior := conv.Memory
+	conv.Lock.Unlock()
+
+	// Summarize (unlocked).
+	msgs := buildSummaryMessages(prior, renderItemsTranscript(overflow), s.MaxSummaryTokens)
+	ctx, cancel := context.WithTimeout(context.Background(), compactionTimeout)
+	defer cancel()
+	predFunc, err := model.Predict(ctx, msgs, nil, nil, nil, nil, nil, nil, nil, nil, nil)
+	if err != nil {
+		xlog.Warn("realtime compaction: summarizer predict failed", "error", err)
+		return
+	}
+	pred, err := predFunc()
+	if err != nil {
+		xlog.Warn("realtime compaction: summarizer inference failed", "error", err)
+		return
+	}
+	// Strip any leaked reasoning/thinking spans using the same extractor the
+	// rest of the realtime path uses, rather than a bespoke regex.
+	rcfg := reasoning.Config{}
+	if mc := model.PredictConfig(); mc != nil {
+		rcfg = spokenReasoningConfig(mc.ReasoningConfig)
+	}
+	_, summary := reasoning.ExtractReasoningComplete(pred.Response, "", rcfg)
+	summary = strings.TrimSpace(summary)
+	if summary == "" {
+		xlog.Warn("realtime compaction: empty summary, skipping eviction")
+		return
+	}
+
+	// Commit.
+	conv.Lock.Lock()
+	defer conv.Lock.Unlock()
+	if !prefixMatches(conv.Items, overflow) {
+		xlog.Debug("realtime compaction: head changed during summary, skipping")
+		return
+	}
+	conv.Memory = summary
+	conv.Items = conv.Items[len(overflow):]
+	xlog.Debug("realtime compaction: evicted items into memory", "evicted", len(overflow), "remaining", len(conv.Items))
+}
+
+// summarizerModel resolves the model used to produce compaction summaries.
+// Without a configured summary_model (or factory) it reuses the pipeline LLM.
+func (s *Session) summarizerModel() Model {
+	if s.SummaryModel == "" || s.summarizerFactory == nil {
+		return s.ModelInterface
+	}
+	s.summarizerOnce.Do(func() {
+		m, err := s.summarizerFactory()
+		if err != nil {
+			xlog.Warn("realtime compaction: summary_model load failed, falling back to pipeline LLM", "model", s.SummaryModel, "error", err)
+			m = s.ModelInterface
+		}
+		s.summarizerCached = m
+	})
+	return s.summarizerCached
+}
+
+// maybeCompact schedules a background compaction when the live buffer has grown
+// past the trigger and none is already running. Returns immediately.
+func (s *Session) maybeCompact(conv *Conversation) {
+	if !s.CompactionEnabled {
+		return
+	}
+	conv.Lock.Lock()
+	over := len(conv.Items) > s.CompactionTrigger
+	conv.Lock.Unlock()
+	if !over {
+		return
+	}
+	if !conv.compacting.CompareAndSwap(false, true) {
+		return
+	}
+	go func() {
+		defer conv.compacting.Store(false)
+		// Resolve (and, for a configured summary_model, lazily load) the
+		// summarizer only when a compaction actually runs, off the response
+		// path — so the model load never blocks a user turn.
+		model := s.summarizerModel()
+		if model == nil {
+			return
+		}
+		s.compact(conv, model)
+	}()
+}

core/http/endpoints/openai/realtime_compaction_test.go

@@ -0,0 +1,308 @@
+package openai
+
+import (
+	"errors"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/backend"
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/http/endpoints/openai/types"
+	"github.com/mudler/LocalAI/core/schema"
+)
+
+var _ = Describe("resolveCompaction", func() {
+	It("disables when the block is absent", func() {
+		enabled, _, _, _ := resolveCompaction(&config.ModelConfig{}, 6)
+		Expect(enabled).To(BeFalse())
+	})
+
+	It("defaults trigger to 2x max history and tokens to 512", func() {
+		cfg := &config.ModelConfig{Pipeline: config.Pipeline{Compaction: &config.PipelineCompaction{Enabled: true}}}
+		enabled, trigger, maxTok, _ := resolveCompaction(cfg, 6)
+		Expect(enabled).To(BeTrue())
+		Expect(trigger).To(Equal(12))
+		Expect(maxTok).To(Equal(512))
+	})
+
+	It("clamps trigger to max history + 1 when misconfigured", func() {
+		cfg := &config.ModelConfig{Pipeline: config.Pipeline{Compaction: &config.PipelineCompaction{Enabled: true, TriggerItems: 4}}}
+		_, trigger, _, _ := resolveCompaction(cfg, 6)
+		Expect(trigger).To(Equal(7))
+	})
+
+	It("honors explicit values", func() {
+		cfg := &config.ModelConfig{Pipeline: config.Pipeline{Compaction: &config.PipelineCompaction{
+			Enabled: true, TriggerItems: 20, MaxSummaryTokens: 256, SummaryModel: "tiny"}}}
+		enabled, trigger, maxTok, model := resolveCompaction(cfg, 6)
+		Expect(enabled).To(BeTrue())
+		Expect(trigger).To(Equal(20))
+		Expect(maxTok).To(Equal(256))
+		Expect(model).To(Equal("tiny"))
+	})
+})
+
+var _ = Describe("deleteItem", func() {
+	mk := func(ids ...string) []*types.MessageItemUnion {
+		out := make([]*types.MessageItemUnion, len(ids))
+		for i, id := range ids {
+			out[i] = &types.MessageItemUnion{User: &types.MessageItemUser{ID: id}}
+		}
+		return out
+	}
+
+	It("removes the item with the given id", func() {
+		items, ok := deleteItem(mk("a", "b", "c"), "b")
+		Expect(ok).To(BeTrue())
+		Expect(len(items)).To(Equal(2))
+		Expect(itemID(items[0])).To(Equal("a"))
+		Expect(itemID(items[1])).To(Equal("c"))
+	})
+
+	It("reports not found for an unknown id", func() {
+		_, ok := deleteItem(mk("a"), "zzz")
+		Expect(ok).To(BeFalse())
+	})
+})
+
+var _ = Describe("clearInputAudio", func() {
+	It("resets the pending PCM and buffered Opus frames", func() {
+		s := &Session{InputAudioBuffer: []byte{1, 2, 3}, OpusFrames: [][]byte{{9}}}
+		clearInputAudio(s)
+		Expect(s.InputAudioBuffer).To(BeNil())
+		Expect(s.OpusFrames).To(BeNil())
+	})
+})
+
+var _ = Describe("truncateAssistantText", func() {
+	It("clears the text of the assistant content part at the index", func() {
+		items := []*types.MessageItemUnion{{Assistant: &types.MessageItemAssistant{
+			ID:      "a1",
+			Content: []types.MessageContentOutput{{Type: types.MessageContentTypeText, Text: "hello world"}},
+		}}}
+		ok := truncateAssistantText(items, "a1", 0)
+		Expect(ok).To(BeTrue())
+		Expect(items[0].Assistant.Content[0].Text).To(Equal(""))
+	})
+
+	// Realtime assistant *audio* turns store the spoken words in .Transcript, not
+	// .Text, so a barge-in truncate must clear .Transcript too or it would no-op.
+	It("clears the transcript of an assistant audio content part", func() {
+		items := []*types.MessageItemUnion{{Assistant: &types.MessageItemAssistant{
+			ID:      "a1",
+			Content: []types.MessageContentOutput{{Type: types.MessageContentTypeAudio, Transcript: "hello world"}},
+		}}}
+		ok := truncateAssistantText(items, "a1", 0)
+		Expect(ok).To(BeTrue())
+		Expect(items[0].Assistant.Content[0].Transcript).To(Equal(""))
+	})
+
+	It("returns false for an unknown id", func() {
+		Expect(truncateAssistantText(nil, "nope", 0)).To(BeFalse())
+	})
+})
+
+var _ = Describe("compactionCut", func() {
+	user := func(id string) *types.MessageItemUnion {
+		return &types.MessageItemUnion{User: &types.MessageItemUser{ID: id}}
+	}
+	call := func(id string) *types.MessageItemUnion {
+		return &types.MessageItemUnion{FunctionCall: &types.MessageItemFunctionCall{ID: id}}
+	}
+	out := func(id string) *types.MessageItemUnion {
+		return &types.MessageItemUnion{FunctionCallOutput: &types.MessageItemFunctionCallOutput{ID: id}}
+	}
+
+	It("cuts exactly len-keep when no pairs straddle the boundary", func() {
+		items := []*types.MessageItemUnion{user("1"), user("2"), user("3"), user("4")}
+		Expect(compactionCut(items, 2)).To(Equal(2))
+	})
+
+	It("returns 0 when nothing to cut", func() {
+		Expect(compactionCut([]*types.MessageItemUnion{user("1")}, 2)).To(Equal(0))
+	})
+
+	It("returns 0 (cuts nothing) when keep is 0 — the unlimited-window sentinel", func() {
+		items := []*types.MessageItemUnion{user("1"), user("2"), user("3")}
+		Expect(compactionCut(items, 0)).To(Equal(0))
+	})
+
+	It("moves the boundary so a call/output pair is not split", func() {
+		// keep=2 -> naive cut=2, but items[2] is the output of items[1]'s call;
+		// pull the cut right so the whole pair stays in the kept tail.
+		items := []*types.MessageItemUnion{user("1"), call("c"), out("c"), user("4")}
+		Expect(compactionCut(items, 2)).To(Equal(1))
+	})
+})
+
+var _ = Describe("withMemory", func() {
+	It("inserts a memory system message when memory is non-empty", func() {
+		base := schema.Messages{{Role: "system", StringContent: "instructions"}}
+		out := withMemory(base, "user is Bob; wants pizza")
+		Expect(len(out)).To(Equal(2))
+		Expect(out[1].Role).To(Equal("system"))
+		Expect(out[1].StringContent).To(ContainSubstring("user is Bob"))
+		Expect(out[1].StringContent).To(ContainSubstring("Summary of earlier conversation"))
+	})
+
+	It("is a no-op when memory is empty", func() {
+		base := schema.Messages{{Role: "system", StringContent: "instructions"}}
+		Expect(withMemory(base, "")).To(HaveLen(1))
+	})
+})
+
+var _ = Describe("renderItemsTranscript", func() {
+	It("renders user and assistant text turns", func() {
+		items := []*types.MessageItemUnion{
+			{User: &types.MessageItemUser{Content: []types.MessageContentInput{{Type: types.MessageContentTypeInputText, Text: "hi"}}}},
+			{Assistant: &types.MessageItemAssistant{Content: []types.MessageContentOutput{{Type: types.MessageContentTypeText, Text: "hello"}}}},
+		}
+		out := renderItemsTranscript(items)
+		Expect(out).To(ContainSubstring("user: hi"))
+		Expect(out).To(ContainSubstring("assistant: hello"))
+	})
+
+	// Realtime assistant *audio* turns store the spoken words in .Transcript, not
+	// .Text, so the transcript builder must emit .Transcript too or spoken turns
+	// would be dropped from the summary.
+	It("renders an assistant audio turn from its transcript", func() {
+		items := []*types.MessageItemUnion{
+			{Assistant: &types.MessageItemAssistant{Content: []types.MessageContentOutput{{Type: types.MessageContentTypeAudio, Transcript: "spoken words"}}}},
+		}
+		Expect(renderItemsTranscript(items)).To(ContainSubstring("assistant: spoken words"))
+	})
+})
+
+var _ = Describe("buildSummaryMessages", func() {
+	It("includes prior memory and the new transcript", func() {
+		msgs := buildSummaryMessages("prior facts", "user: hi", 512)
+		Expect(len(msgs)).To(Equal(2))
+		Expect(msgs[0].Role).To(Equal("system"))
+		Expect(msgs[1].StringContent).To(ContainSubstring("prior facts"))
+		Expect(msgs[1].StringContent).To(ContainSubstring("user: hi"))
+	})
+})
+
+var _ = Describe("compact", func() {
+	user := func(id, text string) *types.MessageItemUnion {
+		return &types.MessageItemUnion{User: &types.MessageItemUser{ID: id,
+			Content: []types.MessageContentInput{{Type: types.MessageContentTypeInputText, Text: text}}}}
+	}
+
+	It("summarizes overflow into Memory and evicts it, keeping the live tail", func() {
+		conv := &Conversation{Items: []*types.MessageItemUnion{
+			user("1", "a"), user("2", "b"), user("3", "c"), user("4", "d"),
+			user("5", "e"), user("6", "f"), user("7", "g"), user("8", "h"),
+		}}
+		s := &Session{CompactionEnabled: true, CompactionTrigger: 7, MaxHistoryItems: 4, MaxSummaryTokens: 512}
+		m := &fakeModel{predictResp: backend.LLMResponse{Response: "ROLLED UP"}}
+
+		s.compact(conv, m)
+
+		Expect(conv.Memory).To(Equal("ROLLED UP"))
+		Expect(len(conv.Items)).To(Equal(4))
+		Expect(itemID(conv.Items[0])).To(Equal("5"))
+		// The summarizer saw the evicted turns.
+		Expect(m.lastMessages[1].StringContent).To(ContainSubstring("a"))
+	})
+
+	It("leaves Items and Memory untouched when the summarizer errors", func() {
+		items := []*types.MessageItemUnion{user("1", "a"), user("2", "b"), user("3", "c")}
+		conv := &Conversation{Items: items}
+		s := &Session{CompactionEnabled: true, CompactionTrigger: 2, MaxHistoryItems: 1, MaxSummaryTokens: 512}
+		m := &fakeModel{predictErr: errors.New("boom")}
+
+		s.compact(conv, m)
+
+		Expect(conv.Memory).To(Equal(""))
+		Expect(len(conv.Items)).To(Equal(3))
+	})
+
+	It("strips leaked reasoning tags from the summary via the shared extractor", func() {
+		conv := &Conversation{Items: []*types.MessageItemUnion{
+			user("1", "a"), user("2", "b"), user("3", "c"), user("4", "d"),
+			user("5", "e"), user("6", "f"), user("7", "g"), user("8", "h"),
+		}}
+		s := &Session{CompactionEnabled: true, CompactionTrigger: 7, MaxHistoryItems: 4, MaxSummaryTokens: 512}
+		m := &fakeModel{predictResp: backend.LLMResponse{Response: "<think>planning the summary</think>CLEAN SUMMARY"}}
+
+		s.compact(conv, m)
+
+		Expect(conv.Memory).To(Equal("CLEAN SUMMARY"))
+		Expect(conv.Memory).ToNot(ContainSubstring("planning"))
+	})
+
+	It("does nothing when items are at or below the trigger", func() {
+		conv := &Conversation{Items: []*types.MessageItemUnion{user("1", "a")}}
+		s := &Session{CompactionEnabled: true, CompactionTrigger: 7, MaxHistoryItems: 4}
+		s.compact(conv, &fakeModel{predictResp: backend.LLMResponse{Response: "x"}})
+		Expect(conv.Memory).To(Equal(""))
+		Expect(len(conv.Items)).To(Equal(1))
+	})
+})
+
+var _ = Describe("prefixMatches", func() {
+	user := func(id string) *types.MessageItemUnion {
+		return &types.MessageItemUnion{User: &types.MessageItemUser{ID: id}}
+	}
+
+	It("matches when items begins with the snapshot ids in order", func() {
+		items := []*types.MessageItemUnion{user("1"), user("2"), user("3")}
+		snap := []*types.MessageItemUnion{user("1"), user("2")}
+		Expect(prefixMatches(items, snap)).To(BeTrue())
+	})
+
+	It("matches an empty snapshot", func() {
+		Expect(prefixMatches([]*types.MessageItemUnion{user("1")}, nil)).To(BeTrue())
+	})
+
+	It("fails when items is shorter than the snapshot (a concurrent delete shrank the head)", func() {
+		items := []*types.MessageItemUnion{user("1")}
+		snap := []*types.MessageItemUnion{user("1"), user("2")}
+		Expect(prefixMatches(items, snap)).To(BeFalse())
+	})
+
+	It("fails when the head ids differ (a concurrent delete reordered the head)", func() {
+		items := []*types.MessageItemUnion{user("2"), user("3")}
+		snap := []*types.MessageItemUnion{user("1"), user("2")}
+		Expect(prefixMatches(items, snap)).To(BeFalse())
+	})
+})
+
+var _ = Describe("summarizerModel", func() {
+	It("returns the pipeline model when no summary_model is set", func() {
+		m := &fakeModel{}
+		s := &Session{ModelInterface: m}
+		Expect(s.summarizerModel()).To(Equal(m))
+	})
+
+	It("uses the factory (once) when summary_model is set", func() {
+		pipeline := &fakeModel{}
+		small := &fakeModel{}
+		calls := 0
+		s := &Session{ModelInterface: pipeline, SummaryModel: "tiny",
+			summarizerFactory: func() (Model, error) { calls++; return small, nil }}
+		Expect(s.summarizerModel()).To(Equal(small))
+		Expect(s.summarizerModel()).To(Equal(small))
+		Expect(calls).To(Equal(1))
+	})
+
+	It("falls back to the pipeline model when the factory errors", func() {
+		pipeline := &fakeModel{}
+		s := &Session{ModelInterface: pipeline, SummaryModel: "tiny",
+			summarizerFactory: func() (Model, error) { return nil, errors.New("nope") }}
+		Expect(s.summarizerModel()).To(Equal(pipeline))
+	})
+})
+
+var _ = Describe("itemID", func() {
+	It("returns the id for each variant and empty for nil", func() {
+		Expect(itemID(nil)).To(Equal(""))
+		Expect(itemID(&types.MessageItemUnion{User: &types.MessageItemUser{ID: "u1"}})).To(Equal("u1"))
+		Expect(itemID(&types.MessageItemUnion{Assistant: &types.MessageItemAssistant{ID: "a1"}})).To(Equal("a1"))
+		Expect(itemID(&types.MessageItemUnion{System: &types.MessageItemSystem{ID: "s1"}})).To(Equal("s1"))
+		Expect(itemID(&types.MessageItemUnion{FunctionCall: &types.MessageItemFunctionCall{ID: "f1"}})).To(Equal("f1"))
+		Expect(itemID(&types.MessageItemUnion{FunctionCallOutput: &types.MessageItemFunctionCallOutput{ID: "o1"}})).To(Equal("o1"))
+	})
+})

core/http/react-ui/public/locales/id/admin.json

@@ -45,7 +45,7 @@
   },
   "scheduling": {
     "title": "Penjadwalan",
-    "subtitle": "Aturan penempatan model dan replika di seluruh klaster"
+    "subtitle": "Aturan penempatan model dan replika di seluruh kluster"
   },
   "p2p": {
     "title": "Komputasi AI Terdistribusi",
@@ -86,4 +86,4 @@
     "title": "Penjelajah",
     "subtitle": "Jelajahi file dan konfigurasi"
   }
-}
+}

core/http/react-ui/public/locales/id/chat.json

@@ -72,7 +72,7 @@
   "actions": {
     "copy": "Salin",
     "regenerate": "Hasilkan ulang",
-    "jumpToLatest": "Jump to latest"
+    "jumpToLatest": "Lompat ke terbaru"
   },
   "streaming": {
     "transferring": "Mentransfer model...",
@@ -115,4 +115,4 @@
     "clearAll": "Hapus semua",
     "deleteAllTitle": "Hapus semua percakapan"
   }
-}
+}

core/http/react-ui/public/locales/id/common.json

@@ -1,8 +1,8 @@
 {
   "unsaved": {
-    "title": "Discard unsaved changes?",
-    "message": "You have unsaved changes that will be lost if you leave this page.",
-    "leave": "Leave"
+    "title": "Buang perubahan yang belum disimpan?",
+    "message": "Anda memiliki perubahan yang belum disimpan. Perubahan tersebut akan hilang jika Anda meninggalkan halaman ini.",
+    "leave": "Tinggalkan Halaman"
   },
   "actions": {
     "save": "Simpan",

core/http/react-ui/public/locales/id/home.json

@@ -7,15 +7,15 @@
   "resourceGpu": "GPU",
   "resourceRam": "RAM",
   "greeting": {
-    "morning": "Good morning",
-    "afternoon": "Good afternoon",
-    "evening": "Good evening",
-    "night": "Working late"
+    "morning": "Selamat pagi",
+    "afternoon": "Selamat siang",
+    "evening": "Selamat malam",
+    "night": "Selamat lembur"
   },
   "statusLine": {
-    "modelsLoaded_one": "{{count}} model loaded",
-    "modelsLoaded_other": "{{count}} models loaded",
-    "noModelsLoaded": "No models loaded",
+    "modelsLoaded_one": "{{count}} model dimuat",
+    "modelsLoaded_other": "{{count}} model dimuat",
+    "noModelsLoaded": "Tidak ada model yang dimuat",
     "nodes_one": "{{count}} node",
     "nodes_other": "{{count}} nodes"
   },
@@ -79,14 +79,14 @@
   },
   "connect": {
     "title": "Satu endpoint, semua API",
-    "subtitle": "LocalAI menyediakan API miliknya sendiri yang lengkap — pembuatan gambar & video, depth, deteksi objek, reranking, audio, pengenalan wajah & suara, serta suara realtime melalui WebRTC dan WebSocket. Di atas itu, lapisan kompatibilitas drop-in membuat aplikasi apa pun yang dibuat untuk OpenAI, Anthropic, Ollama, atau OpenAI Responses bekerja tanpa perubahan.",
+    "subtitle": "LocalAI menyediakan API miliknya sendiri yang lengkap — pembuatan gambar & video, depth, deteksi objek, reranking, audio, pengenalan wajah & suara, serta suara realtime melalui WebRTC dan WebSocket. Selain itu, lapisan kompatibilitas drop-in membuat aplikasi apa pun yang dibuat untuk OpenAI, Anthropic, Ollama, atau OpenAI Responses bekerja tanpa perubahan.",
     "nativeTitle": "API native",
     "compatTitle": "Kompatibilitas drop-in",
     "apiReference": "Referensi API lengkap",
     "copy": "Salin",
     "copied": "Disalin",
-    "browse": "Browse the API",
-    "hide": "Hide endpoints",
-    "dismiss": "Dismiss"
+    "browse": "Jelajahi API",
+    "hide": "Sembunyikan endpoint",
+    "dismiss": "Abaikan"
   }
 }

core/http/react-ui/public/locales/id/media.json

@@ -5,7 +5,7 @@
       "video": "Video",
       "tts": "TTS",
       "sound": "Suara",
-      "transform": "Transform"
+      "transform": "Transformasi"
     }
   },
   "image": {
@@ -30,7 +30,7 @@
       "refImagesAdded_other": "{{count}} gambar ditambahkan"
     },
     "actions": {
-      "view": "View",
+      "view": "Lihat",
       "generate": "Hasilkan",
       "generating": "Menghasilkan..."
     },
@@ -153,4 +153,4 @@
     "clearConfirm": "Hapus",
     "cleared": "Riwayat dihapus"
   }
-}
+}

core/http/react-ui/public/locales/id/nav.json

@@ -19,11 +19,11 @@
     "operate": "Operasikan"
   },
   "operate": {
-    "inference": "Inference",
-    "cluster": "Cluster",
-    "observability": "Observability",
-    "access": "Access",
-    "system": "System"
+    "inference": "Inferensi",
+    "cluster": "Kluster",
+    "observability": "Observabilitas",
+    "access": "Akses",
+    "system": "Sistem"
   },
   "items": {
     "home": "Beranda",
@@ -64,7 +64,7 @@
     "copyright": "© 2023-{{year}} {{author}}"
   },
   "console": {
-    "automation": "Otomasi",
+    "automation": "Automasi",
     "training": "Pelatihan"
   }
 }

core/services/distributed/gallery.go

@@ -79,21 +79,29 @@ func (s *GalleryStore) Create(op *GalleryOperationRecord) error {
 	}).Create(op).Error
 }
 
-// UpdateProgress updates progress for an operation.
-func (s *GalleryStore) UpdateProgress(id string, progress float64, message, downloadedSize string) error {
+// UpdateProgress updates progress for an operation. The cancellable flag is
+// persisted on every tick so a replica that restarts mid-install rehydrates the
+// op as still cancellable — otherwise the column keeps its Create-time zero
+// value (false), the UI hides the cancel button, and the orphaned op can only
+// be dismissed by waiting for the 30-minute stale reaper.
+func (s *GalleryStore) UpdateProgress(id string, progress float64, message, downloadedSize string, cancellable bool) error {
 	return s.db.Model(&GalleryOperationRecord{}).Where("id = ?", id).Updates(map[string]any{
 		"progress":             progress,
 		"message":              message,
 		"downloaded_file_size": downloadedSize,
+		"cancellable":          cancellable,
 		"updated_at":           time.Now(),
 	}).Error
 }
 
-// UpdateStatus updates the status of an operation.
+// UpdateStatus updates the status of an operation. A terminal status is never
+// cancellable, so the flag is cleared here to keep the persisted row consistent
+// with what the UI should offer.
 func (s *GalleryStore) UpdateStatus(id, status, errMsg string) error {
 	updates := map[string]any{
-		"status":     status,
-		"updated_at": time.Now(),
+		"status":      status,
+		"cancellable": false,
+		"updated_at":  time.Now(),
 	}
 	if errMsg != "" {
 		updates["error"] = errMsg

core/services/galleryop/cancellable_persist_test.go

@@ -0,0 +1,56 @@
+package galleryop_test
+
+import (
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/services/distributed"
+	"github.com/mudler/LocalAI/core/services/galleryop"
+	"github.com/mudler/LocalAI/core/services/testutil"
+)
+
+// Reproduces "an in-flight install can't be cancelled after a restart". The
+// live install path marks OpStatus.Cancellable=true on every progress tick, but
+// UpdateStatus persisted progress/status to the gallery store WITHOUT the
+// cancellable flag, and Create defaulted it to false. So after a replica
+// restart Hydrate rebuilt the op with Cancellable=false, /api/operations
+// reported cancellable:false, and the UI hid the cancel button — the orphaned
+// op lingered until the 30-minute stale reaper expired it. The cancellable
+// state must be persisted so a rehydrated in-flight op stays cancellable.
+var _ = Describe("GalleryService cancellable persistence across restart", func() {
+	It("rehydrates an in-flight op as still cancellable", func() {
+		db := testutil.SetupTestDB()
+		store, err := distributed.NewGalleryStore(db)
+		Expect(err).ToNot(HaveOccurred())
+
+		svc := galleryop.NewGalleryService(&config.ApplicationConfig{}, nil)
+		svc.SetGalleryStore(store)
+
+		// Seed the in-flight op row as the worker goroutine does on admission.
+		Expect(store.Create(&distributed.GalleryOperationRecord{
+			ID:                 "op-inflight",
+			GalleryElementName: "llama-cpp-development",
+			OpType:             "backend_install",
+			Status:             "pending",
+		})).To(Succeed())
+
+		// Simulate a progress tick: the live path always marks installs
+		// cancellable while they are downloading/processing.
+		svc.UpdateStatus("op-inflight", &galleryop.OpStatus{
+			Message:     "downloading",
+			Progress:    25,
+			Cancellable: true,
+		})
+
+		// A fresh replica boots and hydrates from the store.
+		fresh := galleryop.NewGalleryService(&config.ApplicationConfig{}, nil)
+		fresh.SetGalleryStore(store)
+		Expect(fresh.Hydrate()).To(Succeed())
+
+		st := fresh.GetStatus("op-inflight")
+		Expect(st).ToNot(BeNil(), "the in-flight op must hydrate after a restart")
+		Expect(st.Cancellable).To(BeTrue(),
+			"a still-active install must rehydrate as cancellable so the admin can dismiss it")
+	})
+})

core/services/galleryop/service.go

@@ -167,7 +167,7 @@ func (g *GalleryService) UpdateStatus(s string, op *OpStatus) {
 				xlog.Warn("Failed to persist gallery operation status", "op_id", s, "error", err)
 			}
 		} else {
-			if err := store.UpdateProgress(s, op.Progress, op.Message, op.DownloadedFileSize); err != nil {
+			if err := store.UpdateProgress(s, op.Progress, op.Message, op.DownloadedFileSize, op.Cancellable); err != nil {
 				xlog.Warn("Failed to persist gallery operation progress", "op_id", s, "error", err)
 			}
 		}
@@ -467,6 +467,7 @@ func (g *GalleryService) Start(c context.Context, cl *config.ModelConfigLoader,
 						GalleryElementName: op.GalleryElementName,
 						OpType:             "backend_install",
 						Status:             "pending",
+						Cancellable:        true,
 					})
 				}
 				err := g.backendHandler(&op, systemState)
@@ -499,6 +500,8 @@ func (g *GalleryService) Start(c context.Context, cl *config.ModelConfigLoader,
 						GalleryElementName: op.GalleryElementName,
 						OpType:             opType,
 						Status:             "pending",
+						// A delete is not cancellable; an install is.
+						Cancellable: !op.Delete,
 					})
 				}
 				err := g.modelHandler(&op, cl, systemState)

core/services/nodes/inflight.go

@@ -19,25 +19,40 @@ import (
 // Per-replica: a single tracker instance is bound to (nodeID, modelName, replicaIndex).
 // The router constructs one tracker per Route() result, so each in-flight tick lands
 // on the correct row even when multiple replicas of the same model live on the same node.
+//
+// Embedding only grpc.ControlBackend (not the whole grpc.Backend) is what makes
+// the in-flight accounting safe by construction: the control-plane methods pass
+// through untracked, while every grpc.InferenceBackend method must be declared
+// explicitly below to satisfy grpc.Backend. Adding an inference method to the
+// interface therefore breaks this file's build (see the var assertion below)
+// until it is wrapped with track() - so a new inference path can't be added
+// without an in-flight accounting decision.
 type InFlightTrackingClient struct {
-	grpc.Backend // embed for passthrough of untracked methods
-	registry     InFlightTracker
-	nodeID       string
-	modelName    string
-	replicaIndex int
+	grpc.ControlBackend                       // passthrough for control-plane / streaming-constructor methods
+	inner               grpc.InferenceBackend // tracked inference methods delegate here
+	registry            InFlightTracker
+	nodeID              string
+	modelName           string
+	replicaIndex        int
 
 	firstOnce       sync.Once // guards onFirstComplete
 	onFirstComplete func()    // called once after the first tracked inference call completes
 }
 
+// Compile-time contract: *InFlightTrackingClient must implement the FULL backend
+// surface. Because it embeds only ControlBackend, this fails to compile if any
+// InferenceBackend method is left unwrapped.
+var _ grpc.Backend = (*InFlightTrackingClient)(nil)
+
 // NewInFlightTrackingClient wraps a gRPC backend client with in-flight tracking.
 func NewInFlightTrackingClient(inner grpc.Backend, registry InFlightTracker, nodeID, modelName string, replicaIndex int) *InFlightTrackingClient {
 	return &InFlightTrackingClient{
-		Backend:      inner,
-		registry:     registry,
-		nodeID:       nodeID,
-		modelName:    modelName,
-		replicaIndex: replicaIndex,
+		ControlBackend: inner,
+		inner:          inner,
+		registry:       registry,
+		nodeID:         nodeID,
+		modelName:      modelName,
+		replicaIndex:   replicaIndex,
 	}
 }
 
@@ -91,154 +106,162 @@ func (c *InFlightTrackingClient) reconcile(err error) error {
 
 func (c *InFlightTrackingClient) Predict(ctx context.Context, in *pb.PredictOptions, opts ...ggrpc.CallOption) (*pb.Reply, error) {
 	defer c.track(ctx)()
-	reply, err := c.Backend.Predict(ctx, in, opts...)
+	reply, err := c.inner.Predict(ctx, in, opts...)
 	return reply, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) PredictStream(ctx context.Context, in *pb.PredictOptions, f func(reply *pb.Reply), opts ...ggrpc.CallOption) error {
 	defer c.track(ctx)()
-	return c.reconcile(c.Backend.PredictStream(ctx, in, f, opts...))
+	return c.reconcile(c.inner.PredictStream(ctx, in, f, opts...))
 }
 
 func (c *InFlightTrackingClient) Embeddings(ctx context.Context, in *pb.PredictOptions, opts ...ggrpc.CallOption) (*pb.EmbeddingResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Embeddings(ctx, in, opts...)
+	res, err := c.inner.Embeddings(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) GenerateImage(ctx context.Context, in *pb.GenerateImageRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.GenerateImage(ctx, in, opts...)
+	res, err := c.inner.GenerateImage(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) GenerateVideo(ctx context.Context, in *pb.GenerateVideoRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.GenerateVideo(ctx, in, opts...)
+	res, err := c.inner.GenerateVideo(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) TTS(ctx context.Context, in *pb.TTSRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.TTS(ctx, in, opts...)
+	res, err := c.inner.TTS(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) TTSStream(ctx context.Context, in *pb.TTSRequest, f func(reply *pb.Reply), opts ...ggrpc.CallOption) error {
 	defer c.track(ctx)()
-	return c.reconcile(c.Backend.TTSStream(ctx, in, f, opts...))
+	return c.reconcile(c.inner.TTSStream(ctx, in, f, opts...))
 }
 
 func (c *InFlightTrackingClient) SoundGeneration(ctx context.Context, in *pb.SoundGenerationRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.SoundGeneration(ctx, in, opts...)
+	res, err := c.inner.SoundGeneration(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) AudioTranscription(ctx context.Context, in *pb.TranscriptRequest, opts ...ggrpc.CallOption) (*pb.TranscriptResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.AudioTranscription(ctx, in, opts...)
+	res, err := c.inner.AudioTranscription(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) AudioTranscriptionStream(ctx context.Context, in *pb.TranscriptRequest, f func(chunk *pb.TranscriptStreamResponse), opts ...ggrpc.CallOption) error {
 	defer c.track(ctx)()
-	return c.reconcile(c.Backend.AudioTranscriptionStream(ctx, in, f, opts...))
+	return c.reconcile(c.inner.AudioTranscriptionStream(ctx, in, f, opts...))
 }
 
 func (c *InFlightTrackingClient) Detect(ctx context.Context, in *pb.DetectOptions, opts ...ggrpc.CallOption) (*pb.DetectResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Detect(ctx, in, opts...)
+	res, err := c.inner.Detect(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) Depth(ctx context.Context, in *pb.DepthRequest, opts ...ggrpc.CallOption) (*pb.DepthResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Depth(ctx, in, opts...)
+	res, err := c.inner.Depth(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) Rerank(ctx context.Context, in *pb.RerankRequest, opts ...ggrpc.CallOption) (*pb.RerankResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Rerank(ctx, in, opts...)
+	res, err := c.inner.Rerank(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) VAD(ctx context.Context, in *pb.VADRequest, opts ...ggrpc.CallOption) (*pb.VADResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.VAD(ctx, in, opts...)
+	res, err := c.inner.VAD(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) Diarize(ctx context.Context, in *pb.DiarizeRequest, opts ...ggrpc.CallOption) (*pb.DiarizeResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Diarize(ctx, in, opts...)
+	res, err := c.inner.Diarize(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) FaceVerify(ctx context.Context, in *pb.FaceVerifyRequest, opts ...ggrpc.CallOption) (*pb.FaceVerifyResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.FaceVerify(ctx, in, opts...)
+	res, err := c.inner.FaceVerify(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) FaceAnalyze(ctx context.Context, in *pb.FaceAnalyzeRequest, opts ...ggrpc.CallOption) (*pb.FaceAnalyzeResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.FaceAnalyze(ctx, in, opts...)
+	res, err := c.inner.FaceAnalyze(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) VoiceVerify(ctx context.Context, in *pb.VoiceVerifyRequest, opts ...ggrpc.CallOption) (*pb.VoiceVerifyResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.VoiceVerify(ctx, in, opts...)
+	res, err := c.inner.VoiceVerify(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) VoiceAnalyze(ctx context.Context, in *pb.VoiceAnalyzeRequest, opts ...ggrpc.CallOption) (*pb.VoiceAnalyzeResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.VoiceAnalyze(ctx, in, opts...)
+	res, err := c.inner.VoiceAnalyze(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) VoiceEmbed(ctx context.Context, in *pb.VoiceEmbedRequest, opts ...ggrpc.CallOption) (*pb.VoiceEmbedResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.VoiceEmbed(ctx, in, opts...)
+	res, err := c.inner.VoiceEmbed(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) TokenClassify(ctx context.Context, in *pb.TokenClassifyRequest, opts ...ggrpc.CallOption) (*pb.TokenClassifyResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.TokenClassify(ctx, in, opts...)
+	res, err := c.inner.TokenClassify(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) Score(ctx context.Context, in *pb.ScoreRequest, opts ...ggrpc.CallOption) (*pb.ScoreResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Score(ctx, in, opts...)
+	res, err := c.inner.Score(ctx, in, opts...)
+	return res, c.reconcile(err)
+}
+
+func (c *InFlightTrackingClient) SoundDetection(ctx context.Context, in *pb.SoundDetectionRequest, opts ...ggrpc.CallOption) (*pb.SoundDetectionResponse, error) {
+	defer c.track(ctx)()
+	res, err := c.inner.SoundDetection(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) AudioEncode(ctx context.Context, in *pb.AudioEncodeRequest, opts ...ggrpc.CallOption) (*pb.AudioEncodeResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.AudioEncode(ctx, in, opts...)
+	res, err := c.inner.AudioEncode(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) AudioDecode(ctx context.Context, in *pb.AudioDecodeRequest, opts ...ggrpc.CallOption) (*pb.AudioDecodeResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.AudioDecode(ctx, in, opts...)
+	res, err := c.inner.AudioDecode(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) AudioTransform(ctx context.Context, in *pb.AudioTransformRequest, opts ...ggrpc.CallOption) (*pb.AudioTransformResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.AudioTransform(ctx, in, opts...)
+	res, err := c.inner.AudioTransform(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
-// AudioTransformStream, AudioToAudioStream and Forward are deliberately left as
-// embedded passthrough: they return a stream client and the inference spans the
-// stream's lifetime, not the constructor call. Wrapping the constructor with
-// track() would increment and immediately decrement (and fire onFirstComplete)
-// before any audio flows. Tracking those correctly needs the done() func tied to
-// stream close, which the current Backend interface doesn't surface here.
+// AudioTransformStream, AudioToAudioStream and Forward live in grpc.ControlBackend
+// and are passed through via the embedded field, NOT tracked: they return a stream
+// client and the inference spans the stream's lifetime, not the constructor call.
+// Wrapping the constructor with track() would increment and immediately decrement
+// (and fire onFirstComplete) before any audio flows. Tracking those correctly needs
+// the done() func tied to stream close, which the Backend interface doesn't surface
+// here. If they ever need tracking, move them to grpc.InferenceBackend - the build
+// will then force an explicit wrapper here.

core/services/nodes/inflight_test.go

@@ -408,6 +408,13 @@ var _ = Describe("InFlightTrackingClient", func() {
 				return err
 			})
 		})
+
+		It("SoundDetection", func() {
+			assertTracked(func() error {
+				_, err := client.SoundDetection(context.Background(), &pb.SoundDetectionRequest{})
+				return err
+			})
+		})
 	})
 
 	Describe("stale model reload (self-heal)", func() {

docs/content/advanced/model-configuration.md

@@ -494,6 +494,39 @@ These llama.cpp options are passed through the `options:` array.
 | `direct_io` / `use_direct_io` | bool | `false` | Open the model with `O_DIRECT` (faster cold loads on NVMe; ignored if not supported). |
 | `verbosity` | int | `3` | llama.cpp internal log verbosity threshold. Higher = more verbose. |
 | `override_tensor` / `tensor_buft_overrides` | string | "" | Per-tensor buffer-type overrides for the main model. Format: `<tensor regex>=<buffer type>,<tensor regex>=<buffer type>,...`. Mirrors the existing `draft_override_tensor` syntax for the draft model. |
+| `cpu_moe` | bool | false | Keep all MoE expert weights of the main model on CPU (upstream `--cpu-moe`). Frees VRAM on large MoE models (DeepSeek, Qwen3 `*-A3B`). |
+| `n_cpu_moe` | int | 0 | Keep MoE expert weights of the first N main-model layers on CPU (upstream `--n-cpu-moe`). |
+
+#### Generic option passthrough
+
+Any `options:` entry whose name starts with `-` is forwarded **verbatim** to
+upstream llama.cpp's own `llama-server` argument parser. This means any flag the
+bundled llama.cpp supports works without LocalAI needing a dedicated option,
+even ones added after your LocalAI version was built. See the upstream
+[server flags reference](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md).
+
+Format mirrors the rest of the array - `--flag` for a boolean, or `--flag:value`
+for a flag that takes a value. Everything after the first `:` is the value, so
+embedded colons (e.g. `host:port`) are preserved:
+
+```yaml
+options:
+  - "--cpu-moe"                 # boolean flag
+  - "--n-cpu-moe:4"             # flag with a value
+  - "--override-tensor:exps=CPU"
+```
+
+Notes:
+
+- **Precedence:** passthrough flags are applied last, so an explicit flag
+  overrides the LocalAI option it maps to (e.g. `--ctx-size:8192` overrides
+  `context_size`).
+- **Power-user territory:** an invalid flag or value is rejected by the upstream
+  parser exactly as it would be by `llama-server`, which can fail model loading.
+  Prefer the named options above when one exists.
+- Flags that would terminate the process (such as `--help`, `--usage`,
+  `--version`, `--license`, `--list-devices`, `--cache-list`, and
+  `--completion*`) are ignored.
 
 ### Prompt Caching
 

docs/content/features/middleware.md

@@ -185,6 +185,13 @@ It is persisted through `POST /api/settings` and read live, so a change takes
 effect on the next request without a restart. A default that names a model no
 longer loaded still appears (marked *not loaded*) so it can be toggled off.
 
+The default set can also be supplied out-of-band with the
+`LOCALAI_PII_DEFAULT_DETECTORS` environment variable (comma-separated model
+names, e.g. `privacy-filter-nemotron,secret-filter`). When set it takes
+precedence over the value persisted via the UI (env > file), which is the
+right behaviour for immutable container deployments that pin filtering policy
+at boot rather than via the admin UI.
+
 This is what makes `cloud-proxy` / MITM redaction work out of the box: those
 backends default to PII-enabled but ship no detector list, so without a
 default detector the filter runs with nothing to scan. Set one here and

docs/content/features/openai-realtime.md

@@ -68,6 +68,33 @@ pipeline:
 
 This is applied only to the realtime session's copy of the LLM config, so it does not affect other users of the same model. Leave it unset to use the LLM model config's own reasoning settings.
 
+### Conversation compaction (long sessions on CPU)
+
+By default a realtime session feeds only the last `max_history_items` turns to the LLM; older turns are dropped and forgotten. On CPU, long calls also grow expensive as the prompt fills with verbatim history. Enable `compaction` to instead fold older turns into a rolling summary, so long calls stay cheap without losing earlier context.
+
+Compaction works with two numbers:
+
+- **`max_history_items`** is the *live window* — the recent turns kept verbatim in the prompt.
+- **`compaction.trigger_items`** is the *high-water mark* — let the buffer grow to here, then summarize the overflow (everything above `max_history_items`) into a rolling memory and evict it. It must be greater than `max_history_items`; if it is not, it is clamped up.
+
+The gap between the two controls how often summarization runs: a summary call fires roughly every `(trigger_items - max_history_items)` turns (here, about every 6 turns).
+
+```yaml
+pipeline:
+  max_history_items: 6        # live window — recent turns kept verbatim
+  compaction:
+    enabled: true
+    trigger_items: 12         # summarize overflow back down to max_history_items
+    summary_model: ""         # optional: a small model for the summary (CPU); default = pipeline LLM
+    max_summary_tokens: 512
+```
+
+{{% notice tip %}}
+On CPU, set `summary_model` to a small, fast model so compaction never competes with the conversation LLM for compute. Left empty, the pipeline's own LLM produces the summary.
+{{% /notice %}}
+
+Clients can also manage history directly via the now-supported `conversation.item.delete`, `conversation.item.truncate`, and `input_audio_buffer.clear` realtime events.
+
 ## Transports
 
 The Realtime API supports two transports: **WebSocket** and **WebRTC**.

gallery/index.yaml

@@ -1,4 +1,225 @@
 ---
+- name: "lfm2.5-1.2b-instruct"
+  url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
+  urls:
+    - https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct-GGUF
+  description: |
+    Try LFM • Docs • LEAP • Discord
+
+    # LFM2.5-1.2B-Instruct
+
+    LFM2.5 is a new family of hybrid models designed for **on-device deployment**. It builds on the LFM2 architecture with extended pre-training and reinforcement learning.
+
+      - **Best-in-class performance**: A 1.2B model rivaling much larger models, bringing high-quality AI to your pocket.
+      - **Fast edge inference**: 239 tok/s decode on AMD CPU, 82 tok/s on mobile NPU. Runs under 1GB of memory with day-one support for llama.cpp, MLX, and vLLM.
+      - **Scaled training**: Extended pre-training from 10T to 28T tokens and large-scale multi-stage reinforcement learning.
+
+    Find more information about LFM2.5 in our blog post.
+
+    ## 🗒️ Model Details
+
+    LFM2.5-1.2B-Instruct is a general-purpose text-only model with the following features:
+
+    ...
+  license: "other"
+  tags:
+    - llm
+    - gguf
+  icon: https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/dxnYF2fuLpulismtFSGFi.png
+  overrides:
+    backend: llama-cpp
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+    options:
+      - use_jinja:true
+    parameters:
+      min_p: 0.15
+      model: llama-cpp/models/LFM2.5-1.2B-Instruct-GGUF/LFM2.5-1.2B-Instruct-Q4_K_M.gguf
+      repeat_penalty: 1.05
+      temperature: 0.1
+      top_k: 50
+      top_p: 0.1
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/LFM2.5-1.2B-Instruct-GGUF/LFM2.5-1.2B-Instruct-Q4_K_M.gguf
+      sha256: b1b3de114215d9507409a662a501a631095a479a419584e8a2ded6304b19b4f5
+      uri: https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct-GGUF/resolve/main/LFM2.5-1.2B-Instruct-Q4_K_M.gguf
+- name: "qwopus3.6-27b-coder-compat-mtp"
+  url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
+  urls:
+    - https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF
+  description: "\U0001FA90 Qwopus-3.6-27B-Coder\nCoder SFT Release\n\nAgentic Coding & Tool-Use Reasoning Model Fine-Tuned on Qwopus3.6-27B-v2\n\n\U0001F9EC Trace Inversion & Negentropy\n\U0001F9E0 27B Dense Model\n⚡ Agentic Coding\n\U0001F6E0️ Tool Calling & Agent\n\U0001F3C6 SWE-bench Verified: 67.0% (off-thinking)\n\n\U0001F4A1 What is Qwopus-3.6-27B-Coder?\n\U0001FA90 Qwopus-3.6-27B-Coder is a reasoning-enhanced agentic coding model built on top of Qwopus3.6-27B-v2. It inherits the powerful reasoning foundation of the v2 base — which achieved 87.43% MMLU-Pro and 75.25% SWE-bench Verified — and further specializes it for agentic code generation, structured tool calling, debugging, and instruction-following in developer workflows. The model is designed to excel at repository-level coding tasks, multi-turn tool orchestration, and complex logical reasoning under realistic agent environments.\n\n\U0001F9E9 Agentic Coding\nOptimized for repository-level coding, debugging, patch generation, and structured multi-step development workflows.\n\n\U0001F6E0️ Tool Calling\nLearns from real agent trajectories with tool definitions, tool calls, and environment feedback for robust multi-turn execution.\n\n...\n"
+  license: "apache-2.0"
+  tags:
+    - llm
+    - gguf
+    - vision
+    - multimodal
+    - reasoning
+  icon: https://cdn-uploads.huggingface.co/production/uploads/66309bd090589b7c65950665/sGQKmrMc6L6guMoaB5_Y2.png
+  overrides:
+    backend: llama-cpp
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+    mmproj: llama-cpp/mmproj/Qwopus3.6-27B-Coder-Compat-MTP-GGUF/mmproj-F32.gguf
+    options:
+      - use_jinja:true
+      - spec_type:draft-mtp
+      - spec_n_max:6
+      - spec_p_min:0.75
+    parameters:
+      model: llama-cpp/models/Qwopus3.6-27B-Coder-Compat-MTP-GGUF/Qwopus3.6-27B-Coder-Compat-MTP-Q4_K_M.gguf
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/Qwopus3.6-27B-Coder-Compat-MTP-GGUF/Qwopus3.6-27B-Coder-Compat-MTP-Q4_K_M.gguf
+      sha256: f893632170124da60e159b7bcc9d91e1cda3014b2c6b8ad9c6cde38a1fcd2f6f
+      uri: https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF/resolve/main/Qwopus3.6-27B-Coder-Compat-MTP-Q4_K_M.gguf
+    - filename: llama-cpp/mmproj/Qwopus3.6-27B-Coder-Compat-MTP-GGUF/mmproj-F32.gguf
+      sha256: 32f7ea0600c07272547da401d460f8abbd980f3a57b69d6df87be0e2505e0b9c
+      uri: https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF/resolve/main/mmproj-F32.gguf
+- name: "kimi-k2.7-code"
+  url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
+  urls:
+    - https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF
+  description: |
+    ## 1. Model Introduction
+
+    Kimi K2.7 Code is a coding-focused agentic model built upon Kimi K2.6. With substantial improvements on real-world long-horizon coding tasks, it strengthens end-to-end task completion across complex software engineering workflows while improving token efficiency, reducing thinking-token usage by approximately 30% compared with Kimi K2.6.
+
+    ## 2. Model Summary
+
+    ## 3. Evaluation Results
+
+    Benchmark
+    Kimi K2.6
+    Kimi K2.7 Code
+    GPT-5.5
+    Claude Opus 4.8
+
+    Coding
+
+    Kimi Code Bench v2
+    50.9
+    62.0
+    69.0
+    67.4
+
+    Program Bench
+    48.3
+    53.6
+    69.1
+    63.8
+
+    MLS Bench Lite
+    26.7
+    35.1
+    35.5
+    42.8
+
+    Agentic
+
+    Kimi Claw 24/7 Bench
+    42.9
+    46.9
+    52.8
+    50.4
+
+    MCP Atlas
+    69.4
+    76.0
+    79.4
+    81.3
+
+    MCP Mark Verified
+    72.8
+    81.1
+    92.9
+    76.4
+
+    Footnotes
+
+    ...
+  license: "other"
+  tags:
+    - llm
+    - gguf
+  icon: https://huggingface.co/moonshotai/Kimi-K2.7-Code/resolve/main/figures/kimi-logo.png
+  overrides:
+    backend: llama-cpp
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+    mmproj: llama-cpp/mmproj/Kimi-K2.7-Code-GGUF/mmproj-F32.gguf
+    options:
+      - use_jinja:true
+    parameters:
+      min_p: 0.01
+      model: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00001-of-00014.gguf
+      repeat_penalty: 1
+      temperature: 0.6
+      top_k: -1
+      top_p: 0.95
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00001-of-00014.gguf
+      sha256: 65f0aca336f876902323a90e2aff32cac76d071b2cdd818c6a8d78be8fc2c680
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00001-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00002-of-00014.gguf
+      sha256: 40f4416c130827a11502778891f4ef95b2144db90f51d63aa3548d0952a39683
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00002-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00003-of-00014.gguf
+      sha256: ba2ba0b5168784ace7c752ecadfc3631279b2bb023824cb0fe9e2dab3dd28f22
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00003-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00004-of-00014.gguf
+      sha256: 10298a6c98b13ef49be286fefbea8663e16473fb69bbeabe153bc80c60ae116e
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00004-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00005-of-00014.gguf
+      sha256: 8e9e4c8e35d34fc4fef6bfb65a715ad7defbd196970d833c1df6924d701c88b3
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00005-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00006-of-00014.gguf
+      sha256: ccff6e7f299742f82cf6f51a871e3eb3167511efaee967477cc8387f54d16442
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00006-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00007-of-00014.gguf
+      sha256: 1a3b639633a2d22f71156a9f643ded2329cdd969cc21177b644b5741bac1af8e
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00007-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00008-of-00014.gguf
+      sha256: bde28f682a1eab973538b2102007d952f37a13c1f7d55e2ed99177445ddc4282
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00008-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00009-of-00014.gguf
+      sha256: b6a23a95b61e100f7593fa75e2363966323fa767b7e4fdf45d963b59e8fdc69f
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00009-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00010-of-00014.gguf
+      sha256: fb10231c2e6d76921d40f22690f4aa08a8090c708edeaf7e581abafc24d3b25c
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00010-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00011-of-00014.gguf
+      sha256: d2290be7ed1a22ac1f9f8a4813389689e075ce2ab8abc3aaaa1157a3cb1462d8
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00011-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00012-of-00014.gguf
+      sha256: ce0d028314aa3fc783082dbca097e1055d69686a17ab8306574e2949568f26a5
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00012-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00013-of-00014.gguf
+      sha256: 217864ce63a1d130ab39dcb0996b6097e1aa78eb896e38efaefdbbac3a00b7ec
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00013-of-00014.gguf
+    - filename: llama-cpp/models/Kimi-K2.7-Code-GGUF/Kimi-K2.7-Code-UD-Q8_K_XL-00014-of-00014.gguf
+      sha256: eb7582ad7066c5eaa01bde95acb00b4ad9cd7b07cd50a6cf5c9ee427258bc9dd
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/UD-Q8_K_XL/Kimi-K2.7-Code-UD-Q8_K_XL-00014-of-00014.gguf
+    - filename: llama-cpp/mmproj/Kimi-K2.7-Code-GGUF/mmproj-F32.gguf
+      sha256: b2cc50c8c13fe70fc4968a83332f31e9007ea09ebb9ae91d46a4e4cd2a3053cd
+      uri: https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/resolve/main/mmproj-F32.gguf
 - name: "qwythos-9b-claude-mythos-5-1m"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
@@ -49,33 +270,7 @@
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
     - https://huggingface.co/unsloth/GLM-5.2-GGUF
-  description: |
-    # GLM-5.2
-
-    👋 Join our WeChat or Discord community.
-
-    📖 Check out the GLM-5.2 blog and GLM-5 Technical report.
-
-    📍 Use GLM-5.2 API services on Z.ai API Platform.
-
-    🔜 Try GLM-5.2 here.
-
-    [Paper]
-    [GitHub]
-
-    ## Introduction
-
-    We're introducing GLM-5.2, our latest flagship model for long-horizon tasks. It marks a substantial leap in long-horizon task capability over its predecessor GLM-5.1 and, for the first time, delivers that capability on a **solid 1M-token context**. GLM-5.2's new capabilities include:
-      - **Solid 1M Context:** A solid 1M-token context that stably sustains long-horizon work
-      - **Advanced Coding with Flexible Effort**: Stronger coding capabilities with multiple thinking effort levels to balance performance and latency
-      - **Improved Architecture**: We propose IndexShare, which reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length. We also improve GLM-5.2’s MTP layer for speculative decoding, increasing the acceptance length by up to 20%
-      - **Pure Open**: An MIT open-source license — no regional limits, technical access without borders
-
-    ## Benchmark
-
-    ## Serve GLM-5.2 Locally
-
-    ...
+  description: "# GLM-5.2\n\n\U0001F44B Join our WeChat or Discord community.\n\n\U0001F4D6 Check out the GLM-5.2 blog and GLM-5 Technical report.\n\n\U0001F4CD Use GLM-5.2 API services on Z.ai API Platform.\n\n\U0001F51C Try GLM-5.2 here.\n\n[Paper]\n[GitHub]\n\n## Introduction\n\nWe're introducing GLM-5.2, our latest flagship model for long-horizon tasks. It marks a substantial leap in long-horizon task capability over its predecessor GLM-5.1 and, for the first time, delivers that capability on a **solid 1M-token context**. GLM-5.2's new capabilities include:\n  - **Solid 1M Context:** A solid 1M-token context that stably sustains long-horizon work\n  - **Advanced Coding with Flexible Effort**: Stronger coding capabilities with multiple thinking effort levels to balance performance and latency\n  - **Improved Architecture**: We propose IndexShare, which reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length. We also improve GLM-5.2’s MTP layer for speculative decoding, increasing the acceptance length by up to 20%\n  - **Pure Open**: An MIT open-source license — no regional limits, technical access without borders\n\n## Benchmark\n\n## Serve GLM-5.2 Locally\n\n...\n"
   license: "mit"
   tags:
     - llm
@@ -198,26 +393,7 @@
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
     - https://huggingface.co/michaelw9999/Qwopus3.6-27B-v2-MTP-NVFP4-GGUF
-  description: |
-    🪐 Qwopus3.6-27B-v2-MTP
-    MTP Release
-
-    Multi-Token Prediction reasoning model fine-tuned from Qwen3.6-27B
-
-    🧬 Trace Inversion & Negentropy
-    🧠 27B Parameters
-    ⚡ Speculative Decoding
-    🛠️ Coding / DevOps / Math
-
-    💡 What is Qwopus3.6-27B-v2-MTP?
-    🪐 Qwopus3.6-27B-v2-MTP is a speed-oriented reasoning release built on top of Qwen3.6-27B. It keeps the Qwopus line's focus on reconstructed reasoning traces, coding discipline, DevOps procedures, and mathematical derivations, while adding Multi-Token Prediction for faster generation. The goal is simple: preserve the depth and structure of a 27B reasoning model while making real interactive use noticeably faster.
-
-    ⚡ MTP DecodingAuxiliary future-token prediction improves throughput on long reasoning, code, math, and strict-format prompts.
-    🧩 Structured ReasoningInherits the Qwopus training recipe built around reconstructed step-by-step reasoning trajectories.
-    🧪 GB10 TestedValidated on a 30-question local benchmark across Logic, Coding, DevOps, Math, and Edge tasks.
-    🚀 Practical SpeedDesigned for workflows where strong answers matter, but waiting several extra minutes per task does not.
-
-    ...
+  description: "\U0001FA90 Qwopus3.6-27B-v2-MTP\nMTP Release\n\nMulti-Token Prediction reasoning model fine-tuned from Qwen3.6-27B\n\n\U0001F9EC Trace Inversion & Negentropy\n\U0001F9E0 27B Parameters\n⚡ Speculative Decoding\n\U0001F6E0️ Coding / DevOps / Math\n\n\U0001F4A1 What is Qwopus3.6-27B-v2-MTP?\n\U0001FA90 Qwopus3.6-27B-v2-MTP is a speed-oriented reasoning release built on top of Qwen3.6-27B. It keeps the Qwopus line's focus on reconstructed reasoning traces, coding discipline, DevOps procedures, and mathematical derivations, while adding Multi-Token Prediction for faster generation. The goal is simple: preserve the depth and structure of a 27B reasoning model while making real interactive use noticeably faster.\n\n⚡ MTP DecodingAuxiliary future-token prediction improves throughput on long reasoning, code, math, and strict-format prompts.\n\U0001F9E9 Structured ReasoningInherits the Qwopus training recipe built around reconstructed step-by-step reasoning trajectories.\n\U0001F9EA GB10 TestedValidated on a 30-question local benchmark across Logic, Coding, DevOps, Math, and Edge tasks.\n\U0001F680 Practical SpeedDesigned for workflows where strong answers matter, but waiting several extra minutes per task does not.\n\n...\n"
   tags:
     - llm
     - gguf
@@ -243,28 +419,7 @@
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
     - https://huggingface.co/michaelw9999/Qwopus3.6-27B-Coder-MTP-NVFP4-GGUF
-  description: |
-    🪐 Qwopus-3.6-27B-Coder
-    Coder SFT Release
-
-    Agentic Coding & Tool-Use Reasoning Model Fine-Tuned on Qwopus3.6-27B-v2
-
-    🧬 Trace Inversion & Negentropy
-    🧠 27B Dense Model
-    ⚡ Agentic Coding
-    🛠️ Tool Calling & Agent
-    🏆 SWE-bench Verified: 67.0% (off-thinking)
-
-    💡 What is Qwopus-3.6-27B-Coder?
-    🪐 Qwopus-3.6-27B-Coder is a reasoning-enhanced agentic coding model built on top of Qwopus3.6-27B-v2. It inherits the powerful reasoning foundation of the v2 base — which achieved 87.43% MMLU-Pro (300ex) and 75.25% SWE-bench Verified — and further specializes it for agentic code generation, structured tool calling, debugging, and instruction-following in developer workflows. The model is designed to excel at repository-level coding tasks, multi-turn tool orchestration, and complex logical reasoning under realistic agent environments.
-
-    🧩 Agentic Coding
-    Optimized for repository-level coding, debugging, patch generation, and structured multi-step development workflows.
-
-    🛠️ Tool Calling
-    Learns from real agent trajectories with tool definitions, tool calls, and environment feedback for robust multi-turn execution.
-
-    ...
+  description: "\U0001FA90 Qwopus-3.6-27B-Coder\nCoder SFT Release\n\nAgentic Coding & Tool-Use Reasoning Model Fine-Tuned on Qwopus3.6-27B-v2\n\n\U0001F9EC Trace Inversion & Negentropy\n\U0001F9E0 27B Dense Model\n⚡ Agentic Coding\n\U0001F6E0️ Tool Calling & Agent\n\U0001F3C6 SWE-bench Verified: 67.0% (off-thinking)\n\n\U0001F4A1 What is Qwopus-3.6-27B-Coder?\n\U0001FA90 Qwopus-3.6-27B-Coder is a reasoning-enhanced agentic coding model built on top of Qwopus3.6-27B-v2. It inherits the powerful reasoning foundation of the v2 base — which achieved 87.43% MMLU-Pro (300ex) and 75.25% SWE-bench Verified — and further specializes it for agentic code generation, structured tool calling, debugging, and instruction-following in developer workflows. The model is designed to excel at repository-level coding tasks, multi-turn tool orchestration, and complex logical reasoning under realistic agent environments.\n\n\U0001F9E9 Agentic Coding\nOptimized for repository-level coding, debugging, patch generation, and structured multi-step development workflows.\n\n\U0001F6E0️ Tool Calling\nLearns from real agent trajectories with tool definitions, tool calls, and environment feedback for robust multi-turn execution.\n\n...\n"
   tags:
     - llm
     - gguf
@@ -1484,8 +1639,8 @@
       use_tokenizer_template: true
   files:
     - filename: llama-cpp/models/Qwopus3.6-27B-v2-MTP-GGUF/Qwopus3.6-27B-v2-MTP-Q4_K_M.gguf
-      sha256: 818d68223be4d8518dac0b3b5604dde633cbbcbae1f491d842a3e26711c6606d
       uri: https://huggingface.co/Jackrong/Qwopus3.6-27B-v2-MTP-GGUF/resolve/main/Qwopus3.6-27B-v2-MTP-Q4_K_M.gguf
+      sha256: 31cf5fc2406a0c7aaebcc26d440bf0df94e215d0589d5205bf319649c052b50a
 - name: "qwen3.6-40b-claude-4.6-opus-deckard-heretic-uncensored-thinking-neo-code-di-imatrix-max"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:

pkg/grpc/backend.go

@@ -41,11 +41,34 @@ func buildClient(address string, parallel bool, wd WatchDog, enableWatchDog bool
 	}
 }
 
+// Backend is the full client surface of a model backend. It is deliberately
+// composed of two sub-interfaces so that wrappers can get a COMPILE-TIME
+// guarantee about which methods they must account for:
+//
+//   - InferenceBackend - methods that each perform one discrete inference call
+//     (the call begins on entry and ends on return). A wrapper that does
+//     per-call accounting - e.g. the distributed router's in-flight tracker,
+//     core/services/nodes.InFlightTrackingClient - embeds only ControlBackend
+//     and implements every InferenceBackend method explicitly. Adding a method
+//     to InferenceBackend therefore breaks that wrapper's build until it is
+//     implemented: inference can't be added without an accounting decision.
+//   - ControlBackend - everything that is NOT a discrete inference call:
+//     lifecycle/control-plane operations and the streaming constructors whose
+//     work spans the returned stream rather than the constructor call. These
+//     are safe to pass through untracked.
+//
+// Keep the two sets disjoint; every backend method belongs to exactly one.
 type Backend interface {
-	IsBusy() bool
-	HealthCheck(ctx context.Context) (bool, error)
+	InferenceBackend
+	ControlBackend
+}
+
+// InferenceBackend is the subset of Backend whose methods each map to a single
+// inference call. Wrappers that account for in-flight work must implement these
+// explicitly (see Backend). Do NOT add methods that return a stream client or
+// that are control-plane only - those belong in ControlBackend.
+type InferenceBackend interface {
 	Embeddings(ctx context.Context, in *pb.PredictOptions, opts ...grpc.CallOption) (*pb.EmbeddingResult, error)
-	LoadModel(ctx context.Context, in *pb.ModelOptions, opts ...grpc.CallOption) (*pb.Result, error)
 	PredictStream(ctx context.Context, in *pb.PredictOptions, f func(reply *pb.Reply), opts ...grpc.CallOption) error
 	Predict(ctx context.Context, in *pb.PredictOptions, opts ...grpc.CallOption) (*pb.Reply, error)
 	GenerateImage(ctx context.Context, in *pb.GenerateImageRequest, opts ...grpc.CallOption) (*pb.Result, error)
@@ -53,6 +76,8 @@ type Backend interface {
 	TTS(ctx context.Context, in *pb.TTSRequest, opts ...grpc.CallOption) (*pb.Result, error)
 	TTSStream(ctx context.Context, in *pb.TTSRequest, f func(reply *pb.Reply), opts ...grpc.CallOption) error
 	SoundGeneration(ctx context.Context, in *pb.SoundGenerationRequest, opts ...grpc.CallOption) (*pb.Result, error)
+	AudioTranscription(ctx context.Context, in *pb.TranscriptRequest, opts ...grpc.CallOption) (*pb.TranscriptResult, error)
+	AudioTranscriptionStream(ctx context.Context, in *pb.TranscriptRequest, f func(chunk *pb.TranscriptStreamResponse), opts ...grpc.CallOption) error
 	Detect(ctx context.Context, in *pb.DetectOptions, opts ...grpc.CallOption) (*pb.DetectResponse, error)
 	Depth(ctx context.Context, in *pb.DepthRequest, opts ...grpc.CallOption) (*pb.DepthResponse, error)
 	FaceVerify(ctx context.Context, in *pb.FaceVerifyRequest, opts ...grpc.CallOption) (*pb.FaceVerifyResponse, error)
@@ -60,8 +85,25 @@ type Backend interface {
 	VoiceVerify(ctx context.Context, in *pb.VoiceVerifyRequest, opts ...grpc.CallOption) (*pb.VoiceVerifyResponse, error)
 	VoiceAnalyze(ctx context.Context, in *pb.VoiceAnalyzeRequest, opts ...grpc.CallOption) (*pb.VoiceAnalyzeResponse, error)
 	VoiceEmbed(ctx context.Context, in *pb.VoiceEmbedRequest, opts ...grpc.CallOption) (*pb.VoiceEmbedResponse, error)
-	AudioTranscription(ctx context.Context, in *pb.TranscriptRequest, opts ...grpc.CallOption) (*pb.TranscriptResult, error)
-	AudioTranscriptionStream(ctx context.Context, in *pb.TranscriptRequest, f func(chunk *pb.TranscriptStreamResponse), opts ...grpc.CallOption) error
+	Rerank(ctx context.Context, in *pb.RerankRequest, opts ...grpc.CallOption) (*pb.RerankResult, error)
+	TokenClassify(ctx context.Context, in *pb.TokenClassifyRequest, opts ...grpc.CallOption) (*pb.TokenClassifyResponse, error)
+	Score(ctx context.Context, in *pb.ScoreRequest, opts ...grpc.CallOption) (*pb.ScoreResponse, error)
+	VAD(ctx context.Context, in *pb.VADRequest, opts ...grpc.CallOption) (*pb.VADResponse, error)
+	Diarize(ctx context.Context, in *pb.DiarizeRequest, opts ...grpc.CallOption) (*pb.DiarizeResponse, error)
+	SoundDetection(ctx context.Context, in *pb.SoundDetectionRequest, opts ...grpc.CallOption) (*pb.SoundDetectionResponse, error)
+	AudioEncode(ctx context.Context, in *pb.AudioEncodeRequest, opts ...grpc.CallOption) (*pb.AudioEncodeResult, error)
+	AudioDecode(ctx context.Context, in *pb.AudioDecodeRequest, opts ...grpc.CallOption) (*pb.AudioDecodeResult, error)
+	AudioTransform(ctx context.Context, in *pb.AudioTransformRequest, opts ...grpc.CallOption) (*pb.AudioTransformResult, error)
+}
+
+// ControlBackend is the subset of Backend that is NOT per-call inference:
+// lifecycle/control-plane operations and the streaming constructors whose work
+// spans the returned stream rather than the constructor call. In-flight-tracking
+// wrappers embed this directly and pass it through untracked (see Backend).
+type ControlBackend interface {
+	IsBusy() bool
+	HealthCheck(ctx context.Context) (bool, error)
+	LoadModel(ctx context.Context, in *pb.ModelOptions, opts ...grpc.CallOption) (*pb.Result, error)
 	TokenizeString(ctx context.Context, in *pb.PredictOptions, opts ...grpc.CallOption) (*pb.TokenizationResponse, error)
 	Status(ctx context.Context) (*pb.StatusResponse, error)
 
@@ -70,24 +112,11 @@ type Backend interface {
 	StoresGet(ctx context.Context, in *pb.StoresGetOptions, opts ...grpc.CallOption) (*pb.StoresGetResult, error)
 	StoresFind(ctx context.Context, in *pb.StoresFindOptions, opts ...grpc.CallOption) (*pb.StoresFindResult, error)
 
-	Rerank(ctx context.Context, in *pb.RerankRequest, opts ...grpc.CallOption) (*pb.RerankResult, error)
-
-	TokenClassify(ctx context.Context, in *pb.TokenClassifyRequest, opts ...grpc.CallOption) (*pb.TokenClassifyResponse, error)
-
-	Score(ctx context.Context, in *pb.ScoreRequest, opts ...grpc.CallOption) (*pb.ScoreResponse, error)
-
 	GetTokenMetrics(ctx context.Context, in *pb.MetricsRequest, opts ...grpc.CallOption) (*pb.MetricsResponse, error)
 
-	VAD(ctx context.Context, in *pb.VADRequest, opts ...grpc.CallOption) (*pb.VADResponse, error)
-
-	Diarize(ctx context.Context, in *pb.DiarizeRequest, opts ...grpc.CallOption) (*pb.DiarizeResponse, error)
-
-	SoundDetection(ctx context.Context, in *pb.SoundDetectionRequest, opts ...grpc.CallOption) (*pb.SoundDetectionResponse, error)
-
-	AudioEncode(ctx context.Context, in *pb.AudioEncodeRequest, opts ...grpc.CallOption) (*pb.AudioEncodeResult, error)
-	AudioDecode(ctx context.Context, in *pb.AudioDecodeRequest, opts ...grpc.CallOption) (*pb.AudioDecodeResult, error)
-
-	AudioTransform(ctx context.Context, in *pb.AudioTransformRequest, opts ...grpc.CallOption) (*pb.AudioTransformResult, error)
+	// Streaming constructors: these return a stream client immediately; the
+	// actual inference spans the stream's lifetime, not this call, so they are
+	// NOT tracked as a single in-flight unit.
 	AudioTransformStream(ctx context.Context, opts ...grpc.CallOption) (AudioTransformStreamClient, error)
 	AudioToAudioStream(ctx context.Context, opts ...grpc.CallOption) (AudioToAudioStreamClient, error)
 

tests/e2e/distributed/gallery_distributed_test.go

@@ -53,12 +53,13 @@ var _ = Describe("Gallery Distributed", Label("Distributed"), func() {
 			Expect(retrieved.Status).To(Equal("downloading"))
 			Expect(retrieved.FrontendID).To(Equal("f1"))
 
-			// Update progress
-			Expect(galleryStore.UpdateProgress(op.ID, 0.75, "75% complete", "6GB")).To(Succeed())
+			// Update progress (cancellable: a downloading install can be cancelled)
+			Expect(galleryStore.UpdateProgress(op.ID, 0.75, "75% complete", "6GB", true)).To(Succeed())
 
 			updated, _ := galleryStore.Get(op.ID)
 			Expect(updated.Progress).To(BeNumerically("~", 0.75, 0.01))
 			Expect(updated.Message).To(Equal("75% complete"))
+			Expect(updated.Cancellable).To(BeTrue())
 
 			// Complete
 			Expect(galleryStore.UpdateStatus(op.ID, "completed", "")).To(Succeed())

tests/e2e/distributed/phase4_test.go

@@ -104,11 +104,12 @@ var _ = Describe("Phase 4: MCP, Skills, Gallery, Fine-Tuning", Label("Distribute
 			}
 			stores.Gallery.Create(op)
 
-			Expect(stores.Gallery.UpdateProgress(op.ID, 0.5, "50% complete", "2GB")).To(Succeed())
+			Expect(stores.Gallery.UpdateProgress(op.ID, 0.5, "50% complete", "2GB", true)).To(Succeed())
 
 			updated, _ := stores.Gallery.Get(op.ID)
 			Expect(updated.Progress).To(BeNumerically("~", 0.5, 0.01))
 			Expect(updated.Message).To(Equal("50% complete"))
+			Expect(updated.Cancellable).To(BeTrue())
 		})
 
 		It("should deduplicate concurrent downloads", func() {