HarmonyOS AI Coding: Dual-Engine Tooling, Three-Loop Harness, and Four-Layer Verification

The article outlines Huawei's engineering response to the unique challenges of AI-assisted coding for HarmonyOS: ArkTS has limited public training corpus compared to Java/Kotlin or Swift/Obj-C, and Huawei's diverse device form factors (straight, dual-fold, tri-fold) multiply adaptation complexity. Traditional IDEs are built for human GUI interaction, not AI agent workflows.

Dual-Engine Layout: DevEco Code + DevEco CLI + 格物市场

Huawei splits the tooling into two layers plus an open marketplace:

DevEco Code — a one-stop HarmonyOS AI IDE covering project creation, requirements design, code generation, functional verification, integration testing, and ops. It ships with Zhipu models and supports custom/third-party model integration.

DevEco CLI — an atomic command-line suite (create project, syntax check, compile/build, get device/emulator, run, knowledge retrieval, Skills management) that turns GUI interactions into programmable primitives for AI agents. A single devecocli init bootstraps the environment and installs Skills.

格物市场 — an open platform aggregating official, community, and partner Skills, MCP tools, and knowledge bases.

The decoupling lets DevEco Code serve developers who want zero-config, while DevEco CLI serves teams with existing AI pipelines needing HarmonyOS atomic capabilities.

DevEco Code: Three-Loop Harness & Three Development Modes

Three-Loop Quality Assurance (Harness)

DevEco Code's Harness runs four progressive verification loops:

Lightweight syntax check — LSP-based dependency analysis, call-chain analysis, reference lookup to help the agent understand project structure quickly.

Compile diagnosis — invokes atomic CLI compile tools; errors feed back to the agent for repair.

UI intent verification — a multimodal agent launches the emulator, simulates human interactions (tap, swipe, long-press, input), captures screenshots, and produces a visual diff report. Failures auto-feedback to the agent for repair.

Auto-fix agent — covers syntax, build errors, runtime crashes, UI display and interaction issues. For runtime crashes, a dedicated Skill collects logs and analyzes; built on 100+ crash localization cases, fix success rate exceeds 80%, achieving minute-level crash resolution.

Benchmarks across GLM5.1, DeepSeek-V4-Pro, and MiniMax-M2.7 show DevEco Code consistently outperforms OpenCode on HarmonyOS task completion rates when Harness is enabled.

Three Development Modes

Build mode — maximum autonomy; agent writes and iterates code with minimal human interaction. Suits clear, fast-iteration tasks.

Plan + Build mode — two-phase: Plan Agent scans the codebase, analyzes dependencies, emits a structured task plan for user confirmation; Build Agent then executes each task with auto build-verify-fix loops.

Goal mode (Spec-Driven Development) — developer states natural-language requirements; agent interacts briefly to confirm acceptance criteria, then runs the full autonomous loop: code gen → syntax check → auto-fix → build → push to emulator → functional self-verification. Developer only does final sign-off.

DevEco CLI: Atomic Tools, 20M-Token Knowledge Base, 70+ Skills

Knowledge Base

DevEco CLI bundles a 20-million-character official knowledge base (best practices, guides, API refs, FAQs, version notes, change previews) with four engineering advantages: local response (no network), progressive loading (avoids context overflow), dedicated vocabulary (precise HarmonyOS term mapping), and structured API output (agent-friendly parsing). This directly compensates for ArkTS's scarce public corpus.

70+ Skills & Multi-Device Adaptation

Skills cover ArkTS syntax, ArkUI, multi-device adaptation, engineering tools, auto bug-fix, log analysis, testing, and app quality. The multi-device Skill hmos-multidevice-scenario-entry demonstrates concrete impact: without it, the agent generates single-column layouts for dual-fold and dual-column for tri-fold; with the Skill, dual-fold correctly yields two columns and tri-fold three columns.

Four-Level Skill Routing (L0–L3)

As Skill count grows, model selection degrades. Huawei designed a hierarchical router:

L0 (Syntax/Framework) — routes to ArkTS syntax rules, preventing TS-style code.

L1 (Component/System API) — routes to specific HarmonyOS API docs for correct system calls.

L2 (Engineering Context) — routes to project-specific build conventions, common error fix paths.

L3 (Business Rules) — routes to domain-level constraints and standards.

Principle: "core first, then scenario" — solve language and API universals before contextual and business specifics, avoiding context overload.

HarmonyOS Verify: Four-Layer Verification Stack

Complementing Skill routing ("can write"), Verify ensures "writes correctly":

ETS syntax check + compile/build repair — static analysis and compilation baseline.

UI automation + multimodal image verification — agent executes simulated interactions per spec, screenshots compared by multimodal model against expectations.

Diagnostic tooling — Hidumper dumps UI element trees (coordinates, attributes); log collectors stream compile/runtime logs for root-cause analysis.

One-to-many adaptation verification — auto-runs on straight, dual-fold, tri-fold, tablet form factors to validate multi-device consistency.

Layers form a progressive chain: static → compile → runtime UI → multi-device, covering code-correct → functional-correct → cross-device-consistent.

Conclusion

The framework is in place: DevEco Code's Harness guarantees quality, DevEco CLI's atomic tools and knowledge base enable end-to-end automation, HarmonyOS Verify's four layers close the reliability loop, and Native Flow (mentioned briefly) handles multi-demand isolation and checkpoint recovery. These are already validated in partner scenarios. Next-phase ecosystem priorities: HarmonyOS-specific high-quality code corpus, global understanding of ultra-large complex projects, and goal anchoring/segmented checkpoints for long-horizon tasks. HarmonyOS's latecomer status becomes a structural advantage — tooling designed from zero for AI workflows, not retrofitted onto human-centric chains.