create-litert-lm-android-demo-app

name	create-litert-lm-android-demo-app
description	Guides the agent to orchestrate and implement a standalone LiteRT-LM Android demo app with backend selection and multi-modality support using Bazel 9. Use when asked to create a demo app for LiteRT-LM on Android. Don't use for general Android app development without LiteRT-LM, or for building the core library alone without an app.

Create a LiteRT-LM Android Demo App

This skill ensures that the generated LiteRT-LM Android demo app captures all requirements on the first pass, strictly follows platform constraints, and implements the user interface, state machine, and feature specifications.

Core Constraints & Technical Rules

• Monitoring Subagent: You MUST maintain a dedicated, active monitoring subagent. This subagent is responsible for monitoring task.md to detect any loss of context or deviation from plan, and for ensuring strict adherence to all listed compliance items. Do not create duplicate subagents during context refreshes.
• Airtight Compliance & Gating: To successfully declare a task complete, you MUST satisfy both of these strict verification gates:
  • No Heuristic Pass: You are STRICTLY FORBIDDEN from using recollection or memory to pass items. Every Pass item MUST list a direct markdown link to the live workspace file with exact starting and ending line numbers in the Evidence column, or print the exact terminal log block. Blank or generic claims will be flagged as an immediate compliance failure.
  • No Pending/Fail Items: You MUST NOT notify of task completion if any checklist item remains in Pending or Fail status, or if any detailed checkbox [ ] is left unchecked. Every single row and checkbox MUST be verified and marked Pass or checked [x] with active evidence.
• No Custom Planning: You are STRICTLY FORBIDDEN from proposing, drafting, or generating any custom implementation plans or alternative planning files. The verbatim execution steps copied directly into task.md during Step 1 serve as your SOLE, absolute plan. You MUST follow the tasks inside task.md sequentially step-by-step and track progress exclusively by checking off its checkboxes.
• Just-In-Time (JIT) Reference Reading: To prevent context pollution and early phase-leakage, you MUST NOT read reference guides (like references/*.md) ahead of time. Load and study each guide sheet strictly just-in-time as required by the active execution step. When loading or copying these files, resolve their path dynamically relative to where this SKILL.md is stored to prevent directory drift errors.
• Strict Sequential Execution: You MUST follow the execution steps in the exact numerical order listed in SKILL.md or task.md. You are NOT allowed to skip steps, jump ahead, or reorder operations (such as pulling dependencies early) based on active developer preference, unless you first ask for and receive explicit approval from the user.
• Stick to the Chosen Build Path: Once the build path (Source vs. Maven) is decided, stick strictly to it. Do NOT swap paths mid-task even upon compilation failures.
• WORKSPACE Modification Ban: Modifying WORKSPACE is banned. Compile exclusively via Bzlmod using MODULE.bazel for all extensions.
• No Mock Inference: Ensure the app starts in a clean, uninitialized state. Do not use hardcoded paths for the model picker. Real inference is required.
• Verify Before Declaring Failure: Always inspect active stdout/stderr logs before declaring a background command failed; ensure it isn't simply processing slowly.
• Strict Asynchronous Verification: For all background/async operations (cloning, Bazel builds), do NOT assume task completion based on partial or progress logs. You MUST actively check command status and verify successful completion before beginning any dependent steps.
• Headless Environment: Always assume there is no device or emulator available for execution. Inspect the implementation statically to ensure compliance requirements are met, and mark runtime tests in the compliance reports as "Pass (Static/Build Only)".
• Tool Path & Reference Persistence: You MUST resolve and persistently record the absolute paths of all local/external tools (Android SDK/NDK, bazelisk) and all relative skill references/*.md files under an Environment, Tools & Reference Paths section in the top of task.md during Step 1 (before your execution Cwd changes). For all subsequent steps, you MUST use these recorded absolute paths for all view, read, and copy operations.

---

Context Scenarios

Scenario: Source Build

• Prerequisites: You have a local clone of the LiteRT-LM repository, or intend to build the project from source.
• Path Context: The source code is available. All paths are relative to the cloned repository root. You will build dependencies from source using Bazel.
• Dependency Action: You MUST read references/dependency_source_build.md for instructions on how to obtain the LiteRT-LM shared library and Kotlin interface.

Scenario: Maven Integration

• Prerequisites: You are creating a standalone Android application without the LiteRT-LM source code.
• Path Context: Only the requirement and skill markdowns are available. You will resolve the LiteRT-LM dependencies via Maven.
• Dependency Action: You MUST read references/dependency_maven_integration.md for instructions on how to obtain the LiteRT-LM shared library and Kotlin interface.

---

Execution Steps

1. Create task.md Checklist: Create a task.md file to track progress and constraints.

   • 1.1 Copy Execution Steps: You MUST copy ALL execution steps, including all numbered sub-steps and Context Refresh instructions, verbatim into task.md as a checklist. Do not summarize or omit any steps.
   • 1.2 Copy Constraints & Rules: You MUST copy the complete Core Constraints & Technical Rules section from SKILL.md verbatim into the top of task.md to ensure they remain persistently active constraints in file system memory.
   • 1.3 Context Refresh Safeguard: You MUST copy instructions into the top of task.md directing you to perform the following actions before beginning any next step (x.) or sub-step (x.y), or if the active history is compacted:
     • 1.3.1 Reload and read the ENTIRE SKILL.md file from line 1 to the end.
     • 1.3.2 Fully reload and read any referenced markdown documents associated with that step (e.g., references/ui_layout_and_state.md before initiating 3.2).
     • 1.3.3 Check if a monitoring subagent is active for this task, and if not, instantiate one as required by the Monitoring Subagent Constraint.
   • 1.4 Instantiate Monitoring Subagent: Instantiate and launch the dedicated monitoring subagent immediately as required by the Monitoring Subagent Constraint.
   • 1.5 Record Absolute Reference Paths: You MUST immediately resolve the absolute path onto each relative references/*.md file listed in this skill from your starting trigger directory, and record them under an Environment, Tools & Reference Paths section in the top of task.md.

2. Gather Inputs & Initial Environment Setup:

   • 2.1 Project Root: Ask the user for the root path to store the sample code and all dependencies.
   • 2.2 Target Scenario: Ask the user to choose their target scenario (from the Context Scenarios above) to determine the build path, but only record the choice for use in step Obtain LiteRT-LM Shared Library and Kotlin Interface.