物联网AI编码

Requirements

Getting Started

Open the VS Code Extensions panel and search for Adsum IoT Coder, then click Install. Or install from the VS Code Marketplace directly.

See CHANGELOG.md for release notes.

Configure an AI provider, and open your NCS project. The agent starts with two entry-point workflows:

<p><img src="assets/docs/home.png" width="100%" alt="Adsum IoT Coder Home" /></p>

Analyze nRF Device Logs — captures live RTT/UART logs from connected boards, runs code-aware analysis, produces structured reports. Auto-detects boards via J-Link, supports multi-device simultaneous capture, correlates output with your source code and configuration.

Generate Logging Code — reads your NCS project, understands the BLE stack, and injects LOG_* macros following Zephyr best practices. The agent that writes the log statements knows the context when it later parses them.

From analysis results, the agent can enter a Debug Loop — iterative Build → Flash → Capture → Analyze → Fix cycle — continuing until the bug is resolved or you stop it.

Modules loaded on demand

This release inverts that. Domain knowledge and tool-use skills are structured as a framework of discrete, composable modules — each scoped to a specific chip family, protocol stack, or debug capability. At session start, the agent assesses what the project is and what the task requires, then fetches the relevant modules on demand.

User task ──► Agent assesses project ──► Loads scoped modules ──► Executes
              (chip family,                (only what's needed     (Build → Flash →
               protocol stack,              from iot-knowledge/)    Capture → Analyze
               debug category)                                      → Fix, looped)

The module tree on disk:

iot-knowledge/
├── rules/                        # Platform-agnostic agent constraints
│   ├── core.md                   # Universal embedded development rules
│   ├── tool-routing.md           # When to use nRF terminal vs standard shell
│   └── device-identity.md        # Never guess device roles from board type
├── platforms/nrf/                # Adsum IoT Coder – for nRF (shipping)
│   ├── PLATFORM.md               # Master index — what to load and when
│   ├── boards/                   # Per-SoC: nRF52840, nRF52832, nRF5340
│   ├── sdks/ncs/                 # NCS project structure, Kconfig, BLE stack
│   │   ├── protocols/BLE.md      # BLE-specific modules
│   │   └── SDK.md                # NCS-specific modules
│   ├── workflows/                # Entry-point sequences (start here)
│   │   ├── log-analyzer.md       # Capture → Analyze → Report
│   │   ├── log-generator.md      # Instrument firmware with LOG_* macros
│   │   └── debug-loop.md         # Build → Flash → Capture → Analyze → Fix
│   └── actions/                  # Subroutines (loaded by workflows only)
│       ├── capture-logs.md
│       ├── analyze-logs.md
│       ├── build.md
│       └── flash.md
└── platforms/esp/                # Adsum IoT Coder – for ESP (roadmap)

Analyzing a UART log drop loads log-analyzer.md + capture-logs.md + sdks/ncs/SDK.md. Debugging a failed BLE connection on a two-board setup also pulls in BLE.md, device-identity.md, and the relevant board file — and nothing else. The model gets exactly what the task requires, no more.

The bigger payoff isn't just avoiding context overflow — it's context quality. Even when a full static bundle would technically fit, loading only the relevant modules keeps domain knowledge in the model's effective working set rather than letting it get buried under unrelated material as the session grows. This is the "lost in the middle" failure mode the benchmark caught Claude Code hitting on L1-T2 — same model, full 200k window, lost the original symptom by debug cycle four.

Troubleshooting

Shell integration warning on first run — restart VS Code and open a new terminal session.

Linux notifications — if ENOENT errors appear when tasks complete: sudo apt install libnotify-bin

J-Link not detected / board not auto-detected — confirm the SEGGER J-Link drivers are installed and the board enumerates in nrfjprog --ids. Re-plug the board and reload the VS Code window.

Flash command fails — make sure no other tool (nRF Connect for Desktop, OpenOCD) holds the J-Link interface. Only one process can flash at a time.

AI provider authentication errors — verify your API key in the extension settings and that the endpoint URL matches your provider (e.g. https://openrouter.ai/api/v1 for OpenRouter).

Model refuses tool calls / returns plain text — the configured model must support native tool-calling. Models without function-calling support cannot drive hardware workflows. See Tested Models.

Still stuck? Open a Discussion — we read every one.

---