Changelog

All notable changes to the Adsum IoT Coder extension will be documented in this file.

[0.1.0] - 2026-05-26

The first release built around the skill-first architecture: domain expertise lives in versioned Markdown modules (workflows, actions, rules, board specs) that are loaded into the system prompt on demand based on what the agent detects in your workspace — not baked into a fixed prompt. Same model, smaller context, fewer wrong turns. Backed by an open hardware-in-the-loop benchmark.

Architecture — iot-knowledge/ skill library

• New iot-knowledge/ tree. Every piece of nRF/Zephyr expertise the agent uses is now a Markdown file under iot-knowledge/, organized as rules/, platforms/nrf/{rules,boards,sdks,workflows,actions}/. Author once, ship as part of the extension, regenerate snapshots when changed. Adding a new SoC, protocol, or debug procedure is a Markdown PR — not a code change.
• Workflows are entry points, Actions are subroutines. Three primary workflows ship: log-analyzer.md, log-generator.md, debug-loop.md. Four atomic actions: build.md, flash.md, capture-logs.md, analyze-logs.md. The agent is hard-prohibited from loading an Action as the first read of a task — it must enter through a Workflow, which then pulls Actions via MANDATORY SKILL LOAD directives. This is enforced by platforms/nrf/rules/skill-loading.md.
• Operation-Gating Principle. Before performing any complex hardware operation (build / flash / capture / analyze / inject logs) the agent must read_file the matching Workflow first — regardless of how it arrived at that decision. Closes the failure mode where smaller models would skip the workflow load and improvise from pre-training.
• Dynamic context loading. iot_context.ts reads prj.conf, scans for build_info.yml in every build directory, and only injects the modules that match: BLE protocol guide when CONFIG_BT=y, the matching board file (nRF52832/52840/5340) for each detected build target, NCS SDK reference always. The system prompt fits what your project actually is.
• Scope Gate + Context Budget Protection. AGENT.md refuses non-NCS work in-extension (politely redirects), and rules/core.md Rule 9 forbids speculative reads outside the workspace and outside the NCS file pattern. No more agents wandering into ~/Documents to “understand the workspace”.
• Device-Identity rule loaded by default. platforms/nrf/rules/device-identity.md was listed mandatory in PLATFORM.md but never actually loaded in 0.0.x; now it is. Multi-device captures use generic device1/device2 labels until role is confirmed by config or logs — no more guessing Central vs Peripheral from board type.
• platforms/nrf/PLATFORM.md rewritten as a proper directory index with full triggerNordicAction documentation, board-target → board-file map, and the skill library index.
• Board Target Resolution Protocol in actions/build.md: decision matrix that cross-references build_info.yml history with live nrfutil device device-info --serial-number output instead of guessing.

Agent reliability — works across “any model”

• Tool-call parsing hardened against four real-world failure modes: DeepSeek DSML token stream (<｜｜DSML｜｜...>), Markdown code fences wrapping tool XML, literal template mimicry from small models (emitting <tool_name>read_file</tool_name> instead of substituting), and mid-task model switches that cross the native-tools ↔ XML-tools boundary. New normalize-assistant-message.ts stream-safe rewriter, +30 unit tests.
• Model-family detection broadened: DeepSeek V4-class included in native-tools allowlist; Anthropic -latest aliases recognized as Claude 4+; new getToolCallReliabilityTier() so the UI can warn once when a low-tier model fails to produce tool calls three turns in a row.
• Windows process-cleanup fix. The taskkill /F /IM JLink.exe & taskkill /F /IM nrfutil.exe form silently failed under PowerShell (& is reserved). Now emitted as cmd /c "..." from three sources — the action manuals and the tool’s TECHNICAL_REFERENCE — so it works in PowerShell, bash, and cmd.exe identically.
• Cross-platform J-Link binary resolution (RTT Plan B). connectRTTPlanB previously hardcoded /bin/bash and JLinkExe. New jlinkResolver.ts walks deterministic SEGGER install paths, versioned JLink_V* dirs, and PATH — fails loudly with a useful error instead of silently launching a broken terminal. 33 unit tests cover Win32 / macOS Apple Silicon / Linux.
• Read-file error guidance. Missing logs/<transport>/*.log files now return a structured “run capture first” hint instead of raw ENOENT, and the analyze workflow is required to list_files on the logs directory before reading.

Hardware-in-the-loop benchmark — IoT-FirmwareDebugBench v0.1

• Published the first open benchmark for AI agents on embedded IoT firmware. Six BLE-focused tasks across three difficulty levels, run on real nRF52840 DK and nRF52832 DK boards with NCS v3.2.1, same model on both sides (Claude Haiku 4.5). Full report under docs/benchmarks/v0.1-report.md with regenerable SVG figures and a stdlib-only generate_figures.py script.
• Headline result: Adsum IoT Coder resolved 5/6 tasks vs the general-agent baseline’s 3/6, at 3.8× lower token cost. Architecture is the only variable.

UI / UX

• Welcome & home screen redesigned with the new Adsum brand: coral logo, two prominent entry-point cards (“Debug Live Device Logs” promoted to first with brighter border), recent task cards on the home screen, brand orange (#d76947) on the focus-chain progress bar with green retained for success states.
• Claude Code-style ThinkingBlock replaces ThinkingRow: italic “Thought for Xs” label, animated dots while streaming, grid-rows collapse animation, auto-collapse when the response begins, expandable on demand. +21 unit tests covering label states, duration formatting, and expand/collapse.
• Codicons + code font now load reliably in the installed extension. Vite was emitting absolute @font-face URLs that 404’d silently under the webview’s vscode-resource:// scheme; fixed by setting base: "./" and dropping a redundant <link> to node_modules/@vscode/codicons/.
• Mode icons rendered as base64 PNGs (CSP-safe), with a theme-aware CSS filter (white in dark mode, black in light mode) via a new useVSCodeTheme hook that no longer misidentifies high-contrast light as dark.

Telemetry — own project, real opt-out

• Own PostHog project, separate from upstream Cline. ADSUM_TELEMETRY_SERVICE_API_KEY injected at build time; esbuild.mjs auto-loads .env so F5 and npm run package both work.
• Opt-out setting (adsum-iot-coder.telemetry.enabled) AND-combined with VS Code’s global telemetry.telemetryLevel — either being off stops all collection.
• Fork-attribution fields on every event (extension_name, extension_publisher, is_fork, upstream, arch) so upstream-vs-fork usage is unambiguous in analytics.
• New extension_installed (once) and extension_activated (every session) lifecycle events for clean DAU/install tracking.
• User-facing privacy document: TELEMETRY.md.

Rebrand

• Display name and UI now Adsum IoT Coder – for nRF; activity-bar icon, sidebar logo, welcome screen, social cards, hero GIF all updated. Marketplace identity (AdsumNetwork.nrf-ai-debugger) preserved to keep installs intact.
• All command, view, and config keys migrated to the adsum-iot-coder.* namespace.
• README rewritten around the AI-agent positioning, the skill-first architecture, and the v0.1 benchmark; supported SoCs broadened from “nRF52840” to nRF52 / nRF53 / nRF54.

Other fixes

• macOS symlink resolution in iot-knowledge auto-approval allowlist.
• Onboarding reset button restored in About section; terminal icon fixed.
• Chat freeze on certain large reasoning streams resolved.
• DeepSeek reasoning content now passes through the OpenAI-compatible provider correctly.
• build-proto.mjs no longer breaks on macOS paths containing spaces.

[0.0.4] - 2026-03-23

Changed

• Major Rebrand: Extension renamed from “nRF AI Debugger” to Adsum IoT Coder – for nRF.
• Repository Move: All internal links and configuration updated to point to the new repository at https://github.com/adsumnetworks/SoC-AI-Debugger.

Added

• PostHog Analytics: Integrated PostHog to track anonymous usage data and Nordic toolchain errors, helping us identify missing dependencies or environmental issues automatically.
• Compliance: Added official trademark disclaimer for nRF and Nordic Semiconductor compliance.

Fixed

• Log Analyzer Reliability: Significant improvements to cross-platform UART and RTT log capture stability.
• Terminal Routing: Fixed a bug where named terminals (nRF Connect) were incorrectly routed to hidden cmd.exe processes in background execution mode.

[0.0.2] - 2026-03-02

Fixed

• Terminal Warning Suppression: Removed the annoying “Shell Integration Unavailable” warning for nRF Connect terminals.
• Background Execution: Fixed a critical bug where named terminals (e.g., nRF Connect) were routed to hidden cmd.exe processes instead of the proper PowerShell terminal when the terminal execution mode was set to “Background Exec”. This ensures nrfutil and west commands work reliably.
• Terminal Timeout: Increased the shell integration timeout to ensure slower PCs (e.g., Windows 10) have enough time to initialize the nRF Connect SDK environment before executing commands.

[0.0.1] - Initial Release

Added

• Initial release of Adsum IoT Coder!
• Seamless integration with the nRF Connect SDK terminal in VS Code.
• AI-powered assistant for Zephyr-based projects capable of automatically analyzing UAR/RTT logs, executing Nordic toolchain commands, and debugging code.