@dvai-bridge/capacitor-llama 4.0.0

package/DVAICapacitorLlama.podspec

@@ -0,0 +1,21 @@
+require 'json'
+package = JSON.parse(File.read(File.join(__dir__, 'package.json')))
+
+Pod::Spec.new do |s|
+  s.name             = 'DVAICapacitorLlama'
+  s.version          = package['version']
+  s.summary          = package['description']
+  s.license          = 'Custom (See LICENSE)'
+  s.homepage         = package['repository']['url']
+  s.author           = package['author']
+  s.source           = { :git => package['repository']['url'], :tag => s.version.to_s }
+  s.source_files     = [
+    'ios/Sources/**/*.{swift,h,m,mm}',
+    '../dvai-bridge-ios-llama-core/ios/Sources/**/*.{swift,h,m,mm}',
+  ]
+  s.public_header_files = '../dvai-bridge-ios-llama-core/ios/Sources/DVAILlamaCoreObjC/include/*.h'
+  s.ios.deployment_target = '14.0'
+  s.swift_version    = '5.9'
+  s.dependency 'Capacitor'
+  s.dependency 'Telegraph', '~> 0.40'
+end

package/LICENSE

@@ -0,0 +1,51 @@
+# Deep Voice Ai Limited - Software License Agreement
+
+**Version 1.0.0**
+
+This License Agreement governs the use of the DVAI-Bridge software (the "Software"). By downloading, installing, or using the Software, you agree to be bound by the terms of this License.
+
+---
+
+## 1. LICENSE GRANTS
+
+### 1.1 Development and Personal Use (Free Tier)
+Deep Voice Ai Limited ("Licensor") grants you a non-exclusive, non-transferable, royalty-free license to use the Software solely for:
+- Internal development and testing purposes.
+- Non-commercial personal projects.
+- Academic and non-profit research.
+
+### 1.2 Commercial Use (Paid Tier)
+Any use of the Software for **Commercial Purposes** requires a separate, paid Commercial License from Licensor. "Commercial Purposes" include:
+- Use in production environments.
+- Integration into revenue-generating products or services.
+- Distribution to third-party customers for a fee.
+- Use by an entity with more than $100,000 USD in annual revenue.
+
+To obtain a Commercial License, contact `info@deepvoiceai.co` or visit `https://deepvoiceai.co/licensing`.
+
+---
+
+## 2. RESTRICTIONS
+Except as expressly permitted, you may not:
+- Sublicense, rent, lease, or resell the Software without express permission.
+- Remove any proprietary notices or branding from the Software.
+- Use the Software for any illegal or malicious purposes.
+
+---
+
+## 3. INTELLECTUAL PROPERTY
+The Software is owned by **Deep Voice Ai Limited** and is protected by copyright and intellectual property laws. This agreement does not transfer ownership of the Software.
+
+---
+
+## 4. NO WARRANTY
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED. IN NO EVENT SHALL THE LICENSOR BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE.
+
+---
+
+## 5. GOVERNING LAW
+This License shall be governed by and construed in accordance with the laws of the jurisdiction where Deep Voice Ai Limited is registered.
+
+---
+
+© 2026 Deep Voice Ai Limited. All rights reserved.

package/README.md

@@ -0,0 +1,199 @@
+![DVAI-Bridge](/assets/banner.png)
+
+# DVAI-Bridge
+
+<!-- [![Smoke — real models](https://github.com/Westenets/dvai-bridge/actions/workflows/smoke-real-models.yml/badge.svg?branch=main)](https://github.com/Westenets/dvai-bridge/actions/workflows/smoke-real-models.yml) -->
+
+[![License](https://img.shields.io/badge/License-Commercial-blue.svg)](LICENSE) ![Node.js](https://img.shields.io/badge/Node.js-22+-green?logo=node.js) ![TypeScript](https://img.shields.io/badge/TypeScript-5.6+-blue?logo=typescript) ![Swift](https://img.shields.io/badge/Swift-5.9+-F05138?logo=swift) ![Kotlin](https://img.shields.io/badge/Kotlin-2.0+-7F52FF?logo=kotlin) ![Flutter](https://img.shields.io/badge/Flutter-3.39+-02569B?logo=flutter) ![.NET](https://img.shields.io/badge/.NET-10.0_LTS-512BD4?logo=dotnet)
+
+> **The local OpenAI server you embed inside your app.**
+> One library. One HTTP wire. Every platform. Zero install for your users.
+
+**Docs:** [dvai-bridge.deepvoiceai.co](https://dvai-bridge.deepvoiceai.co)
+
+```ts
+import { DVAI } from "@dvai-bridge/core";
+import OpenAI from "openai";
+
+const dvai = new DVAI({ backend: "transformers" });
+await dvai.initialize();
+
+const openai = new OpenAI({ baseURL: dvai.baseUrl, apiKey: "ignored" });
+await openai.chat.completions.create({
+  model: dvai.transformersModelId,
+  messages: [{ role: "user", content: "Hello!" }],
+});
+```
+
+That's it. A real OpenAI-compatible server is now running inside your app's
+own process. Point any OpenAI client — LangChain, the OpenAI SDK, the Vercel
+AI SDK, anything — at `dvai.baseUrl` and your agent code keeps working.
+
+Built by **[Deep Voice AI](https://deepvoiceai.co)**.
+
+---
+
+## Why it exists
+
+Local AI works beautifully on a laptop with **Ollama + LangChain**. Then you
+try to ship the app and your users don't have Ollama. Mobile can't run it.
+Corporate IT won't add another daemon. So you reinvent the same plumbing —
+spawn an inference engine, bind a port, translate to OpenAI HTTP, handle
+CORS, manage lifecycle, wrap the accelerator of the day per platform — and
+do it all over again for every target OS.
+
+DVAI-Bridge is that plumbing, packaged as a library, for every client
+platform.
+
+---
+
+## What you get
+
+- **One OpenAI HTTP surface.** Bound on `127.0.0.1` (or `0.0.0.0` for
+  device-to-device). Streaming, embeddings, models, recovery — all built in.
+- **Six SDKs.** `@dvai-bridge/core` + `react` + `vanilla` + `capacitor`,
+  `DVAIBridge` (Swift / iOS), `co.deepvoiceai:dvai-bridge` (Kotlin / Android),
+  `@dvai-bridge/react-native`, `dvai_bridge` (Flutter), `co.deepvoiceai.dvai-bridge` (.NET).
+- **Nine backends.** WebLLM, Transformers.js, llama.cpp, Apple Foundation
+  Models, MLX, CoreML / ANE, MediaPipe LLM, LiteRT, ONNX Runtime GenAI —
+  selected per-platform, invisible to your agent code.
+- **Native acceleration** wherever it runs: WebGPU in browsers, CUDA / Metal
+  / Vulkan / DirectML on desktop, ANE / Metal / MLX on iOS, NNAPI / QNN
+  Hexagon / GPU delegate on Android.
+- **Multimodal.** Text, image, audio, video — declarative loader for
+  cutting-edge models (Gemma 4, LLaVA, Idefics) without waiting for library
+  updates.
+- **Distributed inference (v3.0+).** Phone too slow? Offload to your laptop
+  on the same Wi-Fi via mDNS pairing — same OpenAI wire, transparent to
+  your code. Internet path via a self-hostable rendezvous server.
+- **DVAI Hub (v3.1+).** A first-party desktop utility that turns any device
+  into a strong-peer for the rest of your fleet. Brand-neutral install via
+  Homebrew / winget / GitHub Releases, OR fork it for your own branded
+  companion. Routes through Ollama / LM Studio / vLLM / llama-server /
+  llamafile if you've already got those running.
+- **Zero user install.** It's a library, not a daemon. `npm install`,
+  `cocoapods`, gradle — your CI already has the muscle for it.
+
+---
+
+## Supported platforms
+
+| Stack | Package | Backends |
+| --- | --- | --- |
+| Browser (React, Vue, Svelte, vanilla JS) | `@dvai-bridge/core` + `react` / `vanilla` | WebLLM (WebGPU), Transformers.js (WebGPU / WASM SIMD) |
+| Node / Bun / Electron | `@dvai-bridge/core` | Transformers.js, native llama.cpp |
+| Capacitor hybrid mobile | `@dvai-bridge/capacitor` + backend slice | Native llama.cpp (Metal iOS, Vulkan / CPU Android) |
+| iOS native (Swift) | `DVAIBridge` (SPM / CocoaPods) | llama.cpp (Metal), CoreML / ANE, Apple Foundation Models, MLX |
+| Android native (Kotlin / Java) | `co.deepvoiceai:dvai-bridge` (AAR) | llama.cpp, MediaPipe LLM, LiteRT, NNAPI / QNN |
+| React Native (≥0.77, TurboModule) | `@dvai-bridge/react-native` | All iOS + Android backends (delegates) |
+| Flutter (≥3.39) | `dvai_bridge` (pub.dev) | All iOS + Android backends (Pigeon channels) |
+| .NET 10 LTS (MAUI / Avalonia / WinUI / Catalyst / desktop) | `co.deepvoiceai.dvai-bridge*` (NuGet) | iOS / Android delegate to native; desktop = llama.cpp + ONNX Runtime GenAI + ML.NET |
+
+Full quickstart per platform: [dvai-bridge.deepvoiceai.co/guide/getting-started](https://dvai-bridge.deepvoiceai.co/guide/getting-started)
+
+---
+
+## Examples
+
+```ts
+// React
+import { DVAIProvider, useDVAI } from "@dvai-bridge/react";
+<DVAIProvider config={{ backend: "transformers" }}>
+  <Chat />
+</DVAIProvider>;
+function Chat() {
+  const { isReady, baseUrl } = useDVAI();
+  return isReady ? <div>Local AI live at {baseUrl}</div> : <Loading />;
+}
+```
+
+```swift
+// iOS
+let server = try await DVAIBridge.shared.start()
+// server.baseUrl = "http://127.0.0.1:38883/v1"
+```
+
+```kotlin
+// Android
+val server = DVAIBridge.start(context)
+// server.baseUrl = "http://127.0.0.1:38883/v1"
+```
+
+```dart
+// Flutter
+final state = await DVAIBridge.instance.start(
+  backend: BackendKind.auto,
+  modelPath: '/path/to/model.gguf',
+);
+// state.baseUrl = "http://127.0.0.1:38883/v1"
+```
+
+```csharp
+// .NET
+var server = await DVAIBridge.Shared.StartAsync(new StartOptions {
+    Backend = BackendKind.Auto,
+    ModelPath = "/path/to/model.gguf",
+});
+// server.BaseUrl = "http://127.0.0.1:38883/v1"
+```
+
+Multimodal, streaming, embeddings, distributed offload, the Hub —
+everything's at the [docs site](https://dvai-bridge.deepvoiceai.co).
+
+---
+
+## What's new in v3.1
+
+- **DVAI Hub** — Tauri desktop utility that's the strong-peer side of v3
+  distributed inference. `brew install deepvoiceai/dvai-hub/dvai-hub` (or
+  `winget install DeepVoiceAI.DVAIHub`) → mobile apps on the same Wi-Fi
+  pair with it and offload heavy inference. [Guide →](https://dvai-bridge.deepvoiceai.co/guide/dvai-hub)
+- **External-engine bridge.** Hub surfaces Ollama / LM Studio / vLLM /
+  llama-server / llamafile as additional backend pools so paired apps
+  serve from whatever's already cached. Opt-in per engine.
+- **Strict substitution policy.** Models with mismatched family / version /
+  size / type are refused by default; quant-only mismatches gated behind a
+  per-pairing `preferBetterQuant` flag. No silent mis-routing.
+- **HMAC-signed identity** on `/v1/chat/completions`. Per-app audit logs
+  surface who served what, with structured `(appId, peerDeviceId,
+  engine, requestedModel, servedModel, outcome)` rows.
+- **Library finalization.** `httpBindHost` (LAN bind), `chatCompletionInterceptor`
+  (extension point), HMAC primitives re-exported, `/v1/dvai/*` routes
+  actually dispatched, TransformersBackend Node-mode device fix.
+  [Migration v3.0 → v3.1 →](https://dvai-bridge.deepvoiceai.co/migration/v3.0-to-v3.1)
+
+---
+
+## Robustness
+
+Streaming-correct (SSE passthrough + blank-chunk detection), generation
+timeout, automatic engine-state recovery on fatal errors, port fallback,
+worker offloading, Private Network Access ready, CORS configured. The
+boring substrate so your agent code never has to think about it.
+
+---
+
+## Licensing
+
+Dual: **free for development & personal use** on `localhost` (verified at
+runtime). **Commercial use** requires a license key — `info@deepvoiceai.co`.
+
+---
+
+## Contributing
+
+PRs welcome.
+
+```bash
+pnpm install
+pnpm build
+bash scripts/build-all.sh   # full matrix (auto-skips per-host)
+```
+
+[`CONTRIBUTING.md`](./CONTRIBUTING.md) for the PR flow. Per-platform
+contributor docs (iOS / Android / RN / Flutter / .NET) under
+[`docs/development/`](./docs/development/).
+
+---
+
+© Deep Voice AI Limited. All rights reserved.

package/android/build.gradle

@@ -0,0 +1,91 @@
+buildscript {
+    ext {
+        kotlinVersion = '2.3.21'
+    }
+    repositories {
+        google()
+        mavenCentral()
+    }
+    dependencies {
+        classpath 'com.android.tools.build:gradle:9.2.0'
+        classpath "org.jetbrains.kotlin:kotlin-gradle-plugin:$kotlinVersion"
+    }
+}
+
+ext {
+    junitVersion = '4.13.2'
+    androidxAppCompatVersion = '1.7.1'
+    capacitorVersion = '8.3.1'
+    coroutinesVersion = '1.10.2'
+    // Pinned within Ktor 2.x — Ktor 3 has breaking API changes around server
+    // configuration that would require non-trivial refactors. Stay on the
+    // latest 2.x patch (2.3.13) until a deliberate Ktor 3 migration.
+    ktorVersion = '2.3.13'
+}
+
+allprojects {
+    repositories {
+        google()
+        mavenCentral()
+        // Phase 3D: android-llama-core's transitive dep on shared-core resolves
+        // here in dev (after `bash scripts/android-publish-local.sh`).
+        // Production consumers add the GitHub Packages Maven repo per the
+        // consumer-integration docs.
+        mavenLocal()
+    }
+}
+
+// AGP 9+ ships built-in Kotlin support; the standalone 'kotlin-android'
+// plugin is no longer needed (and is rejected by the build).
+// See https://kotl.in/gradle/agp-built-in-kotlin
+apply plugin: 'com.android.library'
+
+android {
+    namespace 'co.deepvoiceai.bridge.llama'
+    compileSdk 36
+
+    defaultConfig {
+        minSdk 24
+        targetSdk 36
+        testInstrumentationRunner "androidx.test.runner.AndroidJUnitRunner"
+        // NDK / externalNativeBuild live in dvai-bridge-android-llama-core; not here.
+    }
+
+    compileOptions {
+        sourceCompatibility JavaVersion.VERSION_17
+        targetCompatibility JavaVersion.VERSION_17
+    }
+    testOptions {
+        unitTests {
+            includeAndroidResources = true
+            returnDefaultValues = true
+        }
+    }
+}
+
+// AGP 9 / Kotlin 2.x: the legacy android.kotlinOptions block is gone. The
+// replacement is the standalone `kotlin { compilerOptions { ... } }` extension
+// (provided by AGP's built-in Kotlin integration).
+kotlin {
+    compilerOptions {
+        jvmTarget = org.jetbrains.kotlin.gradle.dsl.JvmTarget.JVM_17
+    }
+}
+
+dependencies {
+    // The core re-exports its Kotlin API to wrapper consumers via `api`.
+    api project(':dvai-bridge-android-llama-core')
+
+    implementation "com.capacitorjs:core:$capacitorVersion"
+    implementation "androidx.appcompat:appcompat:$androidxAppCompatVersion"
+    implementation "org.jetbrains.kotlinx:kotlinx-coroutines-core:$coroutinesVersion"
+    implementation "org.jetbrains.kotlinx:kotlinx-coroutines-android:$coroutinesVersion"
+
+    testImplementation "junit:junit:$junitVersion"
+    testImplementation "org.jetbrains.kotlinx:kotlinx-coroutines-test:$coroutinesVersion"
+    testImplementation 'org.robolectric:robolectric:4.13'
+
+    androidTestImplementation 'androidx.test:runner:1.6.2'
+    androidTestImplementation 'androidx.test:core:1.6.1'
+    androidTestImplementation 'androidx.test.ext:junit:1.2.1'
+}

package/android/gradle.properties

@@ -0,0 +1,4 @@
+android.useAndroidX=true
+kotlin.code.style=official
+android.nonTransitiveRClass=true
+org.gradle.jvmargs=-Xmx4096m

package/android/settings.gradle

@@ -0,0 +1,4 @@
+include ':dvai-bridge-android-llama-core'
+project(':dvai-bridge-android-llama-core').projectDir = file('../../dvai-bridge-android-llama-core/android')
+
+rootProject.name = 'dvai-bridge-capacitor-llama'

package/android/src/androidTest/java/co/deepvoiceai/bridge/llama/RealModelSmokeTest.kt

@@ -0,0 +1,238 @@
+package co.deepvoiceai.bridge.llama
+
+import androidx.test.ext.junit.runners.AndroidJUnit4
+import androidx.test.platform.app.InstrumentationRegistry
+import co.deepvoiceai.bridge.llama.core.LlamaCppBridge
+import co.deepvoiceai.bridge.llama.core.ModelDownloader
+import co.deepvoiceai.bridge.llama.core.MTMD_MEDIA_MARKER
+import org.junit.After
+import org.junit.Assert.assertFalse
+import org.junit.Assert.assertNotNull
+import org.junit.Assert.assertTrue
+import org.junit.Assume.assumeTrue
+import org.junit.Test
+import org.junit.runner.RunWith
+import java.io.File
+
+/**
+ * End-to-end smoke test against a small public GGUF model. Verifies
+ * mechanics (download → load → respond → free) only, not output quality.
+ *
+ * Reads `smoke_model_url` / `smoke_model_sha256` from the instrumentation
+ * arguments — the workflow forwards them via:
+ *
+ *     ./gradlew connectedAndroidTest \
+ *       -Pandroid.testInstrumentationRunnerArguments.smoke_model_url=$URL \
+ *       -Pandroid.testInstrumentationRunnerArguments.smoke_model_sha256=$SHA
+ *
+ * When either is missing the test is skipped via `Assume.assumeTrue`,
+ * so it stays safe to run locally without those args.
+ */
+@RunWith(AndroidJUnit4::class)
+class RealModelSmokeTest {
+    private val ctx = InstrumentationRegistry.getInstrumentation().targetContext
+    private val args = InstrumentationRegistry.getArguments()
+    private var bridge: LlamaCppBridge? = null
+    private var tempDir: File? = null
+
+    @After
+    fun tearDown() {
+        bridge?.unload()
+        bridge = null
+        tempDir?.deleteRecursively()
+        tempDir = null
+    }
+
+    /**
+     * Vision smoke: download model + mmproj, load both, run a multimodal
+     * completion against the tiny test image asset. Skips cleanly if any of
+     * smoke_vision_model_url / smoke_vision_model_sha256 /
+     * smoke_vision_mmproj_url / smoke_vision_mmproj_sha256 are missing.
+     */
+    @Test
+    fun smokeVisionEndToEnd() {
+        val modelUrl = args.getString("smoke_vision_model_url")
+        val modelSha = args.getString("smoke_vision_model_sha256")
+        val mmprojUrl = args.getString("smoke_vision_mmproj_url")
+        val mmprojSha = args.getString("smoke_vision_mmproj_sha256")
+        assumeTrue(
+            "smoke_vision_* not all provided as instrumentation args; skipping",
+            !modelUrl.isNullOrEmpty() && !modelSha.isNullOrEmpty() &&
+                !mmprojUrl.isNullOrEmpty() && !mmprojSha.isNullOrEmpty()
+        )
+
+        val cacheRoot = File(ctx.cacheDir, "dvai-vision-${System.nanoTime()}")
+        cacheRoot.mkdirs()
+        tempDir = cacheRoot
+
+        val downloader = ModelDownloader(ctx, cacheDirOverride = cacheRoot)
+        val (modelPath, _) = downloader.downloadModel(
+            url = modelUrl!!,
+            expectedSha256 = modelSha!!.lowercase(),
+            destFilename = "smoke-vision-model.gguf",
+            headers = emptyMap(),
+            onProgress = { _, _ -> },
+        )
+        val (mmprojPath, _) = downloader.downloadModel(
+            url = mmprojUrl!!,
+            expectedSha256 = mmprojSha!!.lowercase(),
+            destFilename = "smoke-vision-mmproj.gguf",
+            headers = emptyMap(),
+            onProgress = { _, _ -> },
+        )
+
+        val bridge = LlamaCppBridge()
+        this.bridge = bridge
+        val loaded = bridge.loadModel(
+            path = modelPath,
+            mmprojPath = null,
+            gpuLayers = 99,
+            contextSize = 4096,
+            threads = 4,
+            embeddingMode = false,
+        )
+        assertTrue("model load should succeed", loaded)
+        val mmOk = bridge.loadMmproj(mmprojPath)
+        assertTrue("mmproj load should succeed", mmOk)
+        assertTrue("bridge should report mmproj loaded", bridge.isMmprojLoaded())
+
+        // Read the tiny PNG from assets (1x1 transparent pixel).
+        val imageBytes = ctx.assets.open("images/tiny-test.png").use { it.readBytes() }
+
+        val messages = listOf(mapOf("role" to "user", "content" to "Describe this image: $MTMD_MEDIA_MARKER"))
+        val chatPrompt = bridge.applyChatTemplate(
+            templateOverride = null,
+            messages = messages,
+            addAssistant = true,
+        )
+        assertNotNull("chat template should render", chatPrompt)
+
+        val completion = bridge.completeMultimodalPrompt(
+            prompt = chatPrompt!!,
+            media = listOf(imageBytes),
+            maxTokens = 32,
+            temperature = 0.0f,
+            topP = 1.0f,
+        )
+        assertNotNull("vision completion should not be null", completion)
+        assertFalse("vision completion should not be empty", completion!!.isEmpty())
+    }
+
+    /**
+     * Audio smoke: same as vision, but with the WAV fixture. Skipped if the
+     * loaded mmproj has no audio encoder.
+     */
+    @Test
+    fun smokeAudioEndToEnd() {
+        val modelUrl = args.getString("smoke_vision_model_url")
+        val modelSha = args.getString("smoke_vision_model_sha256")
+        val mmprojUrl = args.getString("smoke_vision_mmproj_url")
+        val mmprojSha = args.getString("smoke_vision_mmproj_sha256")
+        assumeTrue(
+            "smoke_vision_* not all provided as instrumentation args; skipping",
+            !modelUrl.isNullOrEmpty() && !modelSha.isNullOrEmpty() &&
+                !mmprojUrl.isNullOrEmpty() && !mmprojSha.isNullOrEmpty()
+        )
+
+        val cacheRoot = File(ctx.cacheDir, "dvai-audio-${System.nanoTime()}")
+        cacheRoot.mkdirs()
+        tempDir = cacheRoot
+
+        val downloader = ModelDownloader(ctx, cacheDirOverride = cacheRoot)
+        val (modelPath, _) = downloader.downloadModel(
+            url = modelUrl!!,
+            expectedSha256 = modelSha!!.lowercase(),
+            destFilename = "smoke-audio-model.gguf",
+            headers = emptyMap(),
+            onProgress = { _, _ -> },
+        )
+        val (mmprojPath, _) = downloader.downloadModel(
+            url = mmprojUrl!!,
+            expectedSha256 = mmprojSha!!.lowercase(),
+            destFilename = "smoke-audio-mmproj.gguf",
+            headers = emptyMap(),
+            onProgress = { _, _ -> },
+        )
+
+        val bridge = LlamaCppBridge()
+        this.bridge = bridge
+        bridge.loadModel(
+            path = modelPath,
+            mmprojPath = null,
+            gpuLayers = 99,
+            contextSize = 4096,
+            threads = 4,
+            embeddingMode = false,
+        )
+        bridge.loadMmproj(mmprojPath)
+        assumeTrue("Loaded mmproj reports no audio encoder; skipping audio smoke", bridge.hasAudioEncoder())
+
+        // mtmd accepts wav/mp3/flac for audio. Use the WAV fixture from assets.
+        val audioBytes = ctx.assets.open("audio/wav-1s-16khz-mono.wav").use { it.readBytes() }
+
+        val messages = listOf(mapOf("role" to "user", "content" to "Transcribe this: $MTMD_MEDIA_MARKER"))
+        val chatPrompt = bridge.applyChatTemplate(
+            templateOverride = null,
+            messages = messages,
+            addAssistant = true,
+        )!!
+
+        val completion = bridge.completeMultimodalPrompt(
+            prompt = chatPrompt,
+            media = listOf(audioBytes),
+            maxTokens = 32,
+            temperature = 0.0f,
+            topP = 1.0f,
+        )
+        assertNotNull("audio completion should not be null", completion)
+        assertFalse("audio completion should not be empty", completion!!.isEmpty())
+    }
+
+    @Test
+    fun smokeRealModelEndToEnd() {
+        val url = args.getString("smoke_model_url")
+        val sha = args.getString("smoke_model_sha256")
+        assumeTrue(
+            "smoke_model_url/smoke_model_sha256 not provided as instrumentation args; skipping",
+            !url.isNullOrEmpty() && !sha.isNullOrEmpty()
+        )
+
+        val cacheRoot = File(ctx.cacheDir, "dvai-smoke-${System.nanoTime()}")
+        cacheRoot.mkdirs()
+        tempDir = cacheRoot
+
+        val downloader = ModelDownloader(ctx, cacheDirOverride = cacheRoot)
+        val (path, cached) = downloader.downloadModel(
+            url = url!!,
+            expectedSha256 = sha!!.lowercase(),
+            destFilename = "smoke-model.gguf",
+            headers = emptyMap(),
+            onProgress = { _, _ -> /* no-op for smoke */ },
+        )
+        assertFalse("first download into a fresh temp dir should not be cached", cached)
+        assertTrue("downloaded file should exist at $path", File(path).exists())
+
+        val bridge = LlamaCppBridge()
+        this.bridge = bridge
+        val loaded = bridge.loadModel(
+            path = path,
+            mmprojPath = null,
+            gpuLayers = 99,
+            contextSize = 2048,
+            threads = 4,
+            embeddingMode = false,
+        )
+        assertTrue("model load should succeed", loaded)
+        assertTrue("bridge should report loaded", bridge.isLoaded())
+
+        val completion = bridge.completePrompt(
+            prompt = "<|begin_of_text|>What is 2+2?",
+            maxTokens = 32,
+            temperature = 0.0f,
+            topP = 1.0f,
+        )
+        // Don't assert specific content — that's quality testing, not smoke.
+        assertNotNull("completion should not be null", completion)
+        assertFalse("completion should not be empty", completion!!.isEmpty())
+    }
+}

package/android/src/main/AndroidManifest.xml

@@ -0,0 +1,7 @@
+<?xml version="1.0" encoding="utf-8"?>
+<manifest xmlns:android="http://schemas.android.com/apk/res/android"
+    xmlns:tools="http://schemas.android.com/tools">
+    <application
+        android:networkSecurityConfig="@xml/dvai_network_security_config"
+        tools:replace="android:networkSecurityConfig" />
+</manifest>