Add built-in on-device vision, language, and inference

.github/workflows/ai-cn1lib-native-check.yml

@@ -1,23 +1,23 @@
-name: AI cn1lib iOS xcodebuild check
+name: Large AI cn1lib iOS xcodebuild check
 
-# Per cn1lib: synthesise a tiny Xcode project that links the bundled
-# `nativeios/*.m` against the real `GoogleMLKit/*` (or TFLite / whisper)
-# pod, run `pod install`, then `xcodebuild build`. Catches API drift,
-# missing imports, and missing pod hints early -- on the same macOS
-# runners we already use for build-ios.
+# The lightweight vision, language, and LiteRT bridges now live in core and
+# are exercised by the normal port/build tests. Keep this isolated native
+# check for the two deliberately large cn1libs.
 
 on:
   workflow_dispatch:
   pull_request:
     branches: [master]
     paths:
-      - 'maven/cn1-ai-**'
+      - 'maven/cn1-ai-whisper/**'
+      - 'maven/cn1-ai-stablediffusion/**'
       - 'scripts/gen-ai-cn1libs.py'
       - '.github/workflows/ai-cn1lib-native-check.yml'
   push:
     branches: [master]
     paths:
-      - 'maven/cn1-ai-**'
+      - 'maven/cn1-ai-whisper/**'
+      - 'maven/cn1-ai-stablediffusion/**'
       - 'scripts/gen-ai-cn1libs.py'
 
 jobs:
@@ -32,17 +32,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - { lib: cn1-ai-mlkit-text,         pod: 'GoogleMLKit/TextRecognition'    }
-          - { lib: cn1-ai-mlkit-barcode,      pod: 'GoogleMLKit/BarcodeScanning'    }
-          - { lib: cn1-ai-mlkit-face,         pod: 'GoogleMLKit/FaceDetection'      }
-          - { lib: cn1-ai-mlkit-labeling,     pod: 'GoogleMLKit/ImageLabeling'      }
-          - { lib: cn1-ai-mlkit-translate,    pod: 'GoogleMLKit/Translate'          }
-          - { lib: cn1-ai-mlkit-smartreply,   pod: 'GoogleMLKit/SmartReply'         }
-          - { lib: cn1-ai-mlkit-langid,       pod: 'GoogleMLKit/LanguageID'         }
-          - { lib: cn1-ai-mlkit-pose,         pod: 'GoogleMLKit/PoseDetection'      }
-          - { lib: cn1-ai-mlkit-segmentation, pod: 'GoogleMLKit/SegmentationSelfie' }
-          - { lib: cn1-ai-mlkit-docscan,      pod: ''                               }   # VisionKit/CoreImage only
-          - { lib: cn1-ai-tflite,             pod: 'TensorFlowLiteObjC'             }
           - { lib: cn1-ai-whisper,            pod: ''                               }   # links static libwhisper.a
           - { lib: cn1-ai-stablediffusion,    pod: ''                               }   # links Swift runner
     steps:

.github/workflows/pr.yml

@@ -184,17 +184,6 @@ jobs:
           git config --global --add safe.directory "$GITHUB_WORKSPACE"
 
           all_libs=(
-            cn1-ai-mlkit-text
-            cn1-ai-mlkit-barcode
-            cn1-ai-mlkit-face
-            cn1-ai-mlkit-labeling
-            cn1-ai-mlkit-translate
-            cn1-ai-mlkit-smartreply
-            cn1-ai-mlkit-langid
-            cn1-ai-mlkit-pose
-            cn1-ai-mlkit-segmentation
-            cn1-ai-mlkit-docscan
-            cn1-ai-tflite
             cn1-ai-whisper
             cn1-ai-stablediffusion
             cn1-admob
@@ -203,17 +192,6 @@ jobs:
             cn1-ads-mock
           )
           ai_libs=(
-            cn1-ai-mlkit-text
-            cn1-ai-mlkit-barcode
-            cn1-ai-mlkit-face
-            cn1-ai-mlkit-labeling
-            cn1-ai-mlkit-translate
-            cn1-ai-mlkit-smartreply
-            cn1-ai-mlkit-langid
-            cn1-ai-mlkit-pose
-            cn1-ai-mlkit-segmentation
-            cn1-ai-mlkit-docscan
-            cn1-ai-tflite
             cn1-ai-whisper
             cn1-ai-stablediffusion
           )

.github/workflows/scripts-ios.yml

@@ -240,6 +240,13 @@ jobs:
           CN1SS_FAIL_ON_MISMATCH: '1'
           CN1SS_ALLOWED_MISSING: '0'
           CN1SS_PORT_ID: ios-gl
+          # Google ML Kit excludes arm64 simulators, so this job uses ParparVM's
+          # x86 path. On a contended runner the allocation benchmark remained
+          # active after the generic 720s idle limit, with less than three
+          # minutes left under the 35-minute absolute cap. Keep strict missing-
+          # screenshot gating, but allow the active benchmark to complete.
+          CN1SS_SUITE_IDLE_TIMEOUT_SECONDS: '1200'
+          CN1SS_SUITE_TIMEOUT_SECONDS: '3000'
         run: |
           set -euo pipefail
           mkdir -p "${ARTIFACTS_DIR}"
@@ -251,9 +258,9 @@ jobs:
             "${{ steps.build-ios-app.outputs.workspace }}" \
             "" \
             "${{ steps.build-ios-app.outputs.scheme }}"
-        # The script enforces both a 35-minute absolute suite limit and a
-        # 12-minute no-progress limit. This outer cap also includes its
-        # native rebuild and cleanup, so it must leave those phases headroom.
+        # This job uses a 50-minute absolute suite limit and a 20-minute
+        # no-progress limit. The outer cap also includes its native rebuild
+        # and cleanup, so it leaves those phases headroom.
         timeout-minutes: 65
 
       - name: Upload iOS OpenGL port status
@@ -464,6 +471,14 @@ jobs:
           CN1SS_REPORT_TITLE: 'iOS Metal screenshot updates'
           CN1SS_SUCCESS_MESSAGE: '✅ Native iOS Metal screenshot tests passed.'
           CN1SS_COMMENT_LOG_PREFIX: '[run-ios-device-tests-metal]'
+          # The x86 simulator runs CommonWorkloadBenchmarkTest through
+          # ParparVM's non-NEON path. Its object-allocation block took 555s on
+          # an otherwise identical successful run, but exceeded the generic
+          # 720s idle gate on a contended runner while the hang sample showed
+          # active GC/allocation work. Allow that benchmark variance here;
+          # the suite's independent 35-minute absolute limit still catches a
+          # genuinely stuck app.
+          CN1SS_SUITE_IDLE_TIMEOUT_SECONDS: '1200'
           # Strict screenshot gating: fail on any pixel mismatch and on any
           # missing screenshot. The Metal backend must produce the full
           # baseline set; a test that hangs (e.g. the DialogTheme texture

CodenameOne/src/com/codename1/ai/ChatMessage.java

@@ -39,10 +39,19 @@ public final class ChatMessage {
     private final String name;
     private final String toolCallId;
 
+    /// Creates a message with no assistant tool calls or provider name.
+    /// @param role speaker role; required
+    /// @param parts ordered multimodal content; {@code null} becomes empty
     public ChatMessage(Role role, List<MessagePart> parts) {
         this(role, parts, null, null, null);
     }
 
+    /// Creates a fully specified normalized message.
+    /// @param role speaker role; required
+    /// @param parts ordered content parts; {@code null} becomes empty
+    /// @param toolCalls assistant tool calls; {@code null} becomes empty
+    /// @param name optional provider-visible participant name
+    /// @param toolCallId tool call answered by a {@link Role#TOOL} message
     public ChatMessage(Role role, List<MessagePart> parts, List<ToolCall> toolCalls,
                        String name, String toolCallId) {
         if (role == null) {
@@ -57,20 +66,29 @@ public ChatMessage(Role role, List<MessagePart> parts, List<ToolCall> toolCalls,
         this.toolCallId = toolCallId;
     }
 
+    /// @param text system instruction text
+    /// @return a single-part system message
     public static ChatMessage system(String text) {
         return single(Role.SYSTEM, new TextPart(text));
     }
 
+    /// @param text user input text
+    /// @return a single-part user message
     public static ChatMessage user(String text) {
         return single(Role.USER, new TextPart(text));
     }
 
+    /// @param text assistant output text
+    /// @return a single-part assistant message
     public static ChatMessage assistant(String text) {
         return single(Role.ASSISTANT, new TextPart(text));
     }
 
     /// Builds a USER message containing both a text and image part --
     /// the common multi-modal pattern.
+    /// @param text optional text accompanying the image
+    /// @param image image content sent to the model
+    /// @return multimodal user message
     public static ChatMessage userWithImage(String text, ImagePart image) {
         List<MessagePart> parts = new ArrayList<MessagePart>(2);
         if (text != null && text.length() > 0) {
@@ -81,6 +99,9 @@ public static ChatMessage userWithImage(String text, ImagePart image) {
     }
 
     /// Builds a TOOL message wrapping the result of a previous tool call.
+    /// @param toolCallId provider id of the answered tool call
+    /// @param resultJson application result encoded as JSON
+    /// @return tool-role result message
     public static ChatMessage toolResult(String toolCallId, String resultJson) {
         return new ChatMessage(Role.TOOL,
                 Arrays.<MessagePart>asList(new ToolResultPart(toolCallId, resultJson)),
@@ -93,29 +114,35 @@ private static ChatMessage single(Role r, MessagePart p) {
         return new ChatMessage(r, parts);
     }
 
+    /// @return speaker role
     public Role getRole() {
         return role;
     }
 
+    /// @return immutable content parts in message order
     public List<MessagePart> getParts() {
         return parts;
     }
 
+    /// @return immutable assistant tool calls
     public List<ToolCall> getToolCalls() {
         return toolCalls;
     }
 
+    /// @return optional participant name, or {@code null}
     public String getName() {
         return name;
     }
 
+    /// @return answered tool-call id for tool messages, or {@code null}
     public String getToolCallId() {
         return toolCallId;
     }
 
     /// Convenience: concatenates the text of every [TextPart]. Image
     /// and tool-result parts are skipped. Useful for `ChatView`
     /// rendering when you don't care about multi-modal content.
+    /// @return text parts joined with newlines
     public String getText() {
         StringBuilder sb = new StringBuilder();
         for (MessagePart p : parts) {