Skip to content

Releases

v0.5.2

Choose a tag to compare

@sssadsum sssadsum released this 19 Oct 10:57

[0.5.2] - 2026-10-18

Fixed

  • Windows: the Hardware page finds BLG20x and LBG20x gateways over Bluetooth again. The Python included for Windows lacked a module its Bluetooth part needs, so every scan failed.
  • The Hardware page's firmware download: with no folder open it asks where to keep the release (it looked like nothing happened), and a release downloaded into a folder you chose is listed and can be flashed (the page kept offering Download).

Changed

  • One package for every system. The first time you use a gateway, Adsum IoT Coder sets up the Python its gateway tool needs on your computer (once, about a minute), checks it before each use and repairs it by itself if it breaks. If it cannot, the Hardware page says why and offers Retry; Adsum: Repair Bluetooth support rebuilds it.

v0.5.1

Choose a tag to compare

@sssadsum sssadsum released this 13 Oct 12:47

[0.5.1] - 2026-10-11

Added

  • Gateways over Bluetooth: the Hardware page finds a Fanstel BLG20x or LBG20x in range, pairs with it, claims it and checks its firmware against the latest release.
  • Firmware updates over Bluetooth for both halves of a gateway, keeping its settings.
  • Commission a gateway from a session. Each setting is written to the unit and read back, and a site profile reuses your answers for the next unit.
  • Demo and licensed firmware download into your workspace with their licence, and flash from there.

Changed

  • The extension includes the Python its tools need on every platform.

v0.4.3

Choose a tag to compare

@sssadsum sssadsum released this 01 Oct 09:29

[0.4.3] - 2026-09-30

Added

  • LoRa and LoRaWAN knowledge, included with a free account: point-to-point LoRa and LoRaWAN on nRF and ESP32 boards
    with a Semtech radio. The modulation and its parameters, regions, duty cycles and link budgets, the Semtech SX126x,
    SX127x, LR2021 and SX130x parts, LoRaWAN internals (join, class A receive windows, frame counters, MAC commands,
    payload codecs, keys), and how to build a node on nRF Connect SDK or ESP-IDF. Verified on an nRF52840 with an
    SX1262. Knowledge contributed by Omar Morceli.
  • LoRa is detected from the project, as cellular and GNSS are: CONFIG_LORA in the configuration or a Semtech
    sx12xx devicetree node on nRF Connect SDK, a LoRa driver or RadioLib among the components or in the configuration
    on ESP-IDF. The LoRa knowledge then loads on its own.
  • Tool bits for LoRa: time on air and the duty-cycle budget of a packet, to check before a spreading factor is
    chosen; an SX126x status, IRQ or error word decoded into plain words, with what to try next; a capture turned into
    a report that says which devices were heard and how much of the band was listened to.

Changed

  • A free registered account now opens LoRa knowledge beside cellular, satellite and on-device inference. The gate
    panel and the unlocked card say so.
  • The account screen names the sets your account opens; a set this build has no word for is left off rather than
    shown as a raw id.

v0.4.2

Choose a tag to compare

@sssadsum sssadsum released this 17 Sep 12:38

[0.4.2] - 2026-09-17

Changed

  • Packaging cleanup: a smaller, tidier extension package.

[0.4.1] - 2026-09-17

Added

  • A guided first run for the Fanstel BLG20x gateway. It starts with the plan and the parts to have on the desk,
    then identifies the board from its probes, flashes the prebuilt image pair onto both halves, commissions the unit
    from a browser, shows the unit's own view, and brings the first reading to your own tenant, with an update and a
    command back as optional steps. Every step says where you are, what it ends with, and how to pick it up again.
    Needs a free registered account.
  • Advanced knowledge for building on the BLG20x (bearer switching, satellite payloads, updates between the two halves,
    commissioning internals and more), opened per account on request.
  • The Fanstel LEW840x cellular images: the demo pair (the card's bearer, in windows, and the ESP32 application with
    the cellular rung) comes with a free registered account; the production bearer, with no limit, is opened per account.
  • The LEW840x BLE scanner source and the ESP32 application source without the cellular rung are free, like their
    prebuilt images.
  • An account item in the panel header, next to history and settings: sign in when you are signed out; your account
    and sign out when you are signed in. Every window of the editor follows the same session.
  • OpenRouter: choose which sellers may serve a model, with price ceilings, from Settings.

Changed

  • Sign-in finishes in the window that started it, even when the editor hands the browser's link to another window.
    If it still cannot, the sign-in page shows the link once, and the sign-in window has a field to paste it into.
  • A knowledge bit your account cannot open is named as such, with the one way in: register, when a free account
    opens it; ask for more details, when the set is opened on request. Signed in, the row says the set is not in your
    account and names it, and never offers Register; asking re-reads your account, so a set opened for you meanwhile is
    there for the next task. The answer continues from what it has read and never fills the missing steps from general
    knowledge.
  • The LEW840x request form asks for one thing: the full source with the cellular rung.
  • The BLG20x request form can ask for the demo firmware, which the card's Demo line has always offered to ask
    about. The card no longer lists Wi-Fi, which this firmware does not use, and Start begins the guided first run.
  • The BLG20x image pair now carries each half's bootloader, so a unit straight from the box comes up, and its
    programming steps write the protection back before the reset, so the part stays unlocked.
  • Each provider's model list is read live at startup and when the picker opens; retired models are no longer offered,
    and DeepSeek's current model shows its own context size, cache prices and thinking setting.
  • A log or monitor capture reads the device and never resets it unless you ask, on nRF and on ESP.
  • Anonymous usage counts now cover the new gateway and demo cards, the locked-bit row, the request form and the
    sign-in link: which card and which action, never what you typed. See TELEMETRY.md.

Fixed

  • A task no longer slows to a step a minute while the editor window is hidden or behind another app.
  • A reply that stops arriving is ended, retried once, and then reported with a sentence you can act on, instead of
    leaving the task waiting.
  • The account icon draws correctly in remote windows.
  • The request form says when an option cannot be granted for that board, and that nothing was sent, instead of
    asking you to try again.
  • On the free tier, a task could fail to write a file and retry with less each time; the model is now asked for tool
    calls in the format it makes them.
  • One downloaded tool with a bad descriptor no longer hides every tool after it, and signing in reaches the tool list
    without a reload.
  • An account with BLG20x access sees the BLG20x card even before the board is detected.

v0.4.0

Choose a tag to compare

@sssadsum sssadsum released this 09 Sep 14:50

[0.4.0] - 2026-09-09

0.3.2 and 0.3.3 were bench pre-releases and were never published; their work ships here.

Added

  • A redesigned entry surface. The panel opens on where you are and what is on your desk (folder,
    detected boards, toolchains), then the runs worth starting, ranked by what was actually detected,
    each saying why. Typing in the box is the new session. One named resume is always offered when
    there is a past session for the folder you are in.
  • The environment in two densities. A header row carries the folder and one verdict per platform
    from the toolchain check itself (the board by name; no toolchain yet, what to install; a serial
    device that will not answer, in the warning colour, with why), and one click opens the full view
    in place. The row can never tick a platform the check calls missing.
  • Past sessions have one home, the editor's History: search by a session's given name, rename,
    delete. The entry surface offers one named resume and one door to every other run.
  • Guided builds for partner hardware are offered from a cold start, alongside three sample runs that
    need no hardware and nothing installed.
  • A free registered account (GitHub or email, no card) unlocks cellular, satellite NB-NTN and
    nRF54 edge-AI basics, and the LEW840x gateway firmware as signed images (BLE scanner, ESP32 uplink,
    nRF9160 bearer), free to flash and evaluate, with licensed source per module available on request
    from Settings. BLE, Wi-Fi and Ethernet are
    unchanged and need nothing. A card that needs an account says so, says what registering unlocks,
    and takes you there in your browser; the extension never sees a password, and your projects and
    logs stay on your machine.
  • An Account section in Settings: what your account has unlocked, in words; a way to ask for the
    prebuilt gateway template source; sign out; and delete your account, which removes everything keyed
    on you and leaves this machine's projects, logs and free-tier allowance untouched.
  • A task that needs a bit your account cannot open keeps going and says which bit it went without,
    and still credits whoever curated it.
  • Send a message to a session that is already working. It is queued, shown at the end of the
    conversation with a way to take it back, and delivered at the agent's next step, so a run can be
    steered without cancelling it. Stop is still its own button, and returns anything undelivered to
    the chat box.
  • Your own model prices. adsum-iot-coder.modelPricing takes a price per model in USD per million
    tokens and overrides everything else. Built-in prices can also update without an extension
    release, the way advisory data already does.
  • The board on your desk is named from the connected hardware, including the nRF9151 SMA DK; on ESP
    the USB-UART bridge is named and esptool's own reason for a silent chip is reported.
  • A pre-flash check on nRF reads the reset vector first, so an image that would flash clean and never
    execute is caught before it is written.
  • Contributor profiles come from the registry, so a new co-author is credited with a link without
    waiting for a release.

Changed

  • Session handover to an external coding agent is not offered in this release.
  • Knowledge is loaded, not "skilled". The agent's own vocabulary now says bit everywhere it used
    to say skill: the instruction it follows is MANDATORY BIT LOAD, governed by rules/bit-loading.md.
    Your editor's own Skills feature is untouched; the two were colliding in one place.
  • One notice at a time. The CRA finding, the registered receipt, the layout tip, the upgrade card and
    the review nudge no longer stack; the review nudge is last, because a favour must never interrupt
    work. The free-tier strip is dismissible and the account chip lives in the runs drawer.
  • The environment line reports the exception when there is one (a serial device that will not
    answer, nrfutil missing beside the nRF extension) instead of the same line every day.
  • Guided builds load the measurement doctrine instead of a summary of it, so gates report evidence
    where they owe you evidence.
  • A tool's own limit is reported as a limit, not as a fact about the device: the board shell waits
    for a cellular round trip, and the loggers fail on a zero capture instead of summarising it.
  • The About page names what Adsum runs on today: Nordic nRF52/53/54L and nRF91 with NB-IoT, LTE-M
    and GNSS, Espressif ESP32/S3/C6, and the supported partner gateways.
  • The message shown when a run is stuck says what unsticks it (narrow the step, name the file, give
    the command) rather than recommending a different vendor's model.
  • Bench guidance no longer tells the agent to kill every J-Link process on a shared bench; it
    identifies its own probe first.
  • "Bring your own coding agent" is marked Coming soon and cannot be selected while it is unproven.

Fixed

  • DeepSeek and GLM ignored the thinking depth you chose. Picking Low stored and displayed the
    choice, but neither the setting nor the depth reached the request, so the provider applied
    full-depth thinking on every call. A depth chosen now counts as thinking on, and an existing
    configuration starts working without being touched.
  • The OpenAI reasoning-effort setting did nothing. It was read, stored and counted, but the
    request never carried it.
  • DeepSeek costs were understated four to five times. Built-in prices are re-read from the
    vendor, and DeepSeek's off-peak half-price window is applied from the clock.
  • A provider stream that went silent without erroring left the task running forever with a
    half-written sentence (11% of bench runs). A stalled stream is now aborted and retried like any
    other stream error.
  • "Free tier quota exhausted" was shown for any empty reply, including provider hiccups, while the
    strip above it still showed tokens left. Exhaustion is now reported only on the backend's word.
  • Sign-in from a Remote-SSH window, a codespace or a dev container opened the browser on the remote
    machine, so the button looked dead. The editor now opens the link where you are.
  • Sign-in with more than one editor window open could land in a window that refused it. The
    round trip now completes in whichever window receives it, and says which window to return to.
  • The four cellular cards stayed behind the "Register to unlock" panel after a successful sign-in.
  • The registered receipt appeared below the suggested runs; it is now at the top of the panel.
  • A request for template source never showed as sent, so people asked twice.
  • Two board-name rows were wrong: PCA10112 is the nRF21540 DK, not the nRF9160 DK (PCA10090), which
    was missing. Every row is now read from Nordic's own tables, and the name has one home, the
    board-identity bit.
  • A Node-based tool bit could not run on Windows (the advertised command opened a cache file in the
    editor); tools now ship a launcher, and product knowledge paths resolve as bits.
  • The composer could be left permanently disabled, with no way to send, after a task finished.
  • Text the agent had already finished saying could be re-rendered when a tool call followed it, and
    a stalled reply that the host had already retried was left on screen as a stray fragment.
  • A message typed while the agent was working could be taken as the answer to a tool approval that
    had not been shown yet. Nothing typed mid-run can approve a tool.
  • A guided build's step banner was hard-coded to five steps, so a seven-step product build rendered
    "Step 4/7" as a plain heading.
  • A workspace whose applications sit in subfolders (a gateway with its ESP app in esp32/ and its
    Zephyr apps beside it) loaded no platform or product knowledge at all. Each platform's knowledge is
    now built from the application folder the classifier found, and a folder that was empty when the
    extension activated is classified again once files land in it.
  • A compound terminal command (a; b; c) waited out the four-minute silence backstop instead of
    completing when its last sub-command did.
  • GNSS knowledge was gated on NTN, so a GNSS task on an nRF9161 could never load it.
  • The agent could invent a load_skill call against an iot-knowledge MCP server that does not
    exist; both rule bits now refuse the whole class and say the paths are files.
  • The terminal tab was named for the upstream project in one of the two places that create it.
  • The bits roster scrolls and says how many bits there are.
  • 93 reference screenshots used only by the visual tests were shipping inside the package; it is
    10 MB smaller.
  • The update card and toast describe this release, and the "New" badges two sample runs had carried
    since 0.1.7 are gone. Release messages now live in one place, so a version bump without a copy
    review fails the build.

v0.3.1

Choose a tag to compare

@sssadsum sssadsum released this 26 Aug 11:30

Everything since v0.2.1, the last version anyone received.

0.3.0 was written up as a release and never published, so its cellular work reaches you here for the first time. Both entries are included below rather than only the newest, because both are new to you.


[0.3.1] - 2026-08-26

Added

  • Every device tool is a Tool bit: versioned, credited, and hash-verified before it runs.
  • CRA advisory hints and package mappings update from the registry instead of a release.
  • Crash-address decoder that picks addr2line by chip architecture and reports where it looked.
  • Board and chip names come from knowledge, including a catalogue of ten shipping ESP32 targets.
  • Fanstel LEW840X boards route to their own product knowledge.

Changed

  • A newer registry copy of a bit or tool wins over the bundled one; bundled remains the fallback.
  • An update cannot widen what a tool may do on its own; the stricter rule stands.
  • Credit reports the copy that actually served, not always the bundled one.
  • About names both platforms and the CRA check, and links the docs rather than a chip vendor.
  • Building the extension no longer depends on the build machine's Node version.

Fixed

  • A CVE scan that could not run no longer reads as a clean one.
  • Large captures no longer return truncated with a success status.
  • A tool improved in the registry could still be read back from an older copy on disk.
  • Third-party code bundled into a tool is now named in its notice file.
  • Two links pointed at a repository that had been renamed.
  • Setting ADSUM_AUTHOR_TOKEN empty now means "published only", so drafts can be compared against it.
  • A multi-chip product had nowhere to keep its project memory.
  • The ESP monitor put the chip into the bootloader when opening the port.
  • A silent ESP capture now retries the reset once, and a retry never leaves less evidence than the first attempt.
  • ESP commands source the toolchain environment in the terminal that actually runs them.
  • idf.py is not recognized is reported as a broken virtual environment, not a missing source step.
  • The native-module allow-list never matched on Windows.

Known issues

  • Modem trace needs firmware built with tracing enabled, and a SIM, or the capture is empty.

[0.3.0] - 2026-08-22

0.2.2 was never released; its work ships here.

Added

  • nRF91 cellular: NB-IoT, LTE-M, NTN and DECT NR+, each loaded only for the project that needs it.
  • GNSS positioning knowledge for nRF91.
  • Board knowledge for the nRF9161 DK, nRF9151 DK and nRF9160 DK.
  • Device tools are Tool bits, each with its own descriptor and author credit.
  • Tool bits can arrive from the registry, verified against a published hash before running.
  • Board shell: send AT and Zephyr shell commands and read until the board finishes, on all platforms.
  • Modem trace: capture, decode and explain registration state, refusal reason, radio mode and search events.
  • log-shape: describe a log's structure in one call, then read any message kind in context.
  • Knowledge is named when it loads, not only when the agent reads a file.

Fixed

  • A tool that could not run no longer reports as though it had.
  • Device tools no longer try to install Python packages; the wrappers no longer hide the install location.
  • An empty answer from a device is no longer reported as a finding.
  • The "Open project folder" button did nothing, could point at the wrong project, and never appeared for copied samples.
  • A scaffolded project you never opened stopped following you into unrelated tasks.
  • Refusing to write project memory now says plainly that it is a blocker.
  • Log and code search: the search program was not found, so nearly every search returned no results.
  • Mermaid diagrams render more reliably.

Known issues

  • NTN needs an nRF9151 with the LACA A1A variant and its own modem firmware.
  • DECT NR+ needs a separate modem firmware image, available from Nordic sales rather than as a download.
  • Two sessions using one development kit at the same time can interfere with each other.

v0.2.1

Choose a tag to compare

@sssadsum sssadsum released this 17 Aug 20:06

[0.2.1] - 2026-08-17

Added

  • Project memory: an .adsum/ folder per project holding the board, transport, goal, open defects, and researched notes, written when something is learned and read at the start of every task. Multi-app repositories get one memory per app plus a shared one above them.
  • Log search: captured RTT, UART, HCI, and sniffer logs are searched by pattern and read by line range instead of loaded whole. A real capture that cost 333,000 tokens now costs a few thousand. A search with no matches falls back to the log's structure and the firmware's own vocabulary, or asks, instead of reading thousands of lines.
  • nRF54 board knowledge: the nRF54L15 DK, the nRF54LM20 DK, and the nRF54LM20A, plus a migration guide from nRF52840 to nRF54L.
  • The board is identified from connected hardware and the project's own board settings before the first build, instead of guessed from build output that does not exist yet.
  • Board knowledge can state the nRF Connect SDK version it needs; you are warned when yours is older, at the point the knowledge is used.
  • DeepSeek as a native provider: correct context length, pricing, and cache rates. Extended thinking can be turned off for routine steps, or set to Low, High, or Max.
  • Export a session as a single redacted file, and hand a running task to your own coding agent (beta).

Fixed

  • Three context-window accounting bugs: cached tokens were counted twice; the window ignored the space a model reserves for its reply, which caused Prompt exceeds max length; and an oversized-request refusal went unrecognised, so the existing recovery never ran.
  • Compaction is announced before it happens, reports what it kept, and preserves the goal, the board, the bug being chased, and the file in hand.
  • Tool results are bounded: file edits return a diff with surrounding lines instead of echoing the file, and long command output is folded with every error and warning kept.
  • Typing while the agent works no longer loses your draft when an approval request arrives.
  • Espressif boards are no longer re-probed in the terminal when the chip, revision, and port are already detected.
  • Checkpoint messages say what is happening: a slow first snapshot on a large repo reads as work in progress, says so once, and takes itself down; nothing is shown where checkpoints simply do not apply.
  • A prototype run ends on an "Open project folder" button instead of a paragraph of instructions; the conversation is saved and returns from History after the reload.
  • GLM 4.7 and 5.x are recognised, so they are no longer scored as unreliable models.

Known issues

  • Opening a scaffolded project is still a manual step; until the folder is open, the new project has no memory or checkpoints.
  • nRF54L support is still maturing. Building and flashing works, with rough edges around board-root setup and runner selection. nRF52 and nRF53 paths are unaffected.

v0.2.0

Choose a tag to compare

@sssadsum sssadsum released this 21 Jul 23:58

[0.2.0] - 2026-07-21

A full end-to-end gateway build, your model or your key, credited expertise, and a second marketplace.

Build a whole product, not just a fix

Expanded curated firmware knowledge now carries a full, long-horizon build across two chips and two toolchains. The launch build is a complete two-chip BLE-to-WiFi gateway, an nRF52840 BLE scanner on Zephyr and an ESP32 Wi-Fi and MQTT uplink on ESP-IDF, for the Fanstel board, from one spec in one working session with the developer approving each step. See the walkthrough.

Choose your model

A curated model picker: the free tier, the GLM Coding Plan (glm-5.2 with 1M context, plus glm-5-turbo and glm-4.7), Claude (Sonnet 5, now the default, Opus 4.8, Haiku 4.5), DeepSeek V4 (flash and pro), and any OpenAI or Anthropic-compatible endpoint, cloud or local. Switch on a live task with no restart.

Credited expertise

When the agent uses a piece of curated knowledge, it names the engineer who wrote it, in the conversation and linked, with a provenance popover (author, maintainer, version, license, source). Human-curated, never anonymous AI output.

A next step, everywhere

The forward-handoff behavior CRA runs got in 0.1.8 now applies to every workflow: a finished task offers concrete next actions instead of a dead-end "done".

Longer sessions hold up

Token-aware context handling replaces blunt transcript truncation, so a long session (the gateway build above, a big BLE capture, a full CRA sweep) is less likely to lose the data it needs mid-session.

Now on Open VSX

Adsum is published to Open VSX as well as the VS Code Marketplace, so Cursor, Windsurf, VSCodium, and other VS Code-compatible editors can install it.

Reliability

  • ESP-IDF version parsing fixed for more install layouts.
  • The Stop button now interrupts a hanging terminal command.
  • The CRA nudge no longer fires on Adsum's own repository.

Preview

  • Hand a running task to your own coding agent (beta): Adsum supplies the embedded knowledge and drives the toolchain while your agent does the work. Rolling out gradually.

v0.1.8 — hardening & bug fixes

Choose a tag to compare

@sssadsum sssadsum released this 08 Jul 18:42

[0.1.8] - 2026-07-07

3-layer debug lands (app, HCI, and radio), alongside a hardening pass for the CRA Readiness Check driven by real field runs on Windows and macOS.

3-layer debug

Adsum now debugs across all three layers of a BLE connection: the application log, the HCI host↔controller bus, and the over-the-air radio, correlated to show where a flow actually broke. It builds on the HCI decoding shipped in 0.1.7 (parsing Host Controller Interface event streams into human-readable BLE protocol events). The guided HCI + Sniffer sample run promised in 0.1.7 is now the on-ramp: a walkthrough of a real one-directional BLE bug that lands on the one-line fix and bridges into the CRA readiness check. It runs on a bundled sample with no hardware needed, and the same debugging works on your own nRF boards.

CRA runs don't dead-end anymore

  • A CRA run now rests on an open question, never a "task complete" box. No more dead-end endings or "I'll continue later" traps, the run always offers concrete forward actions (triage this CVE, start closing this gap), and you leave simply by moving on. The old completion scorecards (with pass/fail glyphs) are blocked at the source.
  • Knowledge-loading is self-healing. A transient network blip on a knowledge fetch now retries silently; a mistyped knowledge path auto-corrects when the catalog has exactly one match (a real run dead-ended on cra/rules/core.md vs cra/core.md, that class of failure is gone). Error messages now say precisely what failed: transient fetch vs not-in-catalog vs registry unreachable.
  • The readiness-report integrity guard is fairer and clearer. It no longer misreads honest phrasing like "62 total packages: 10 queryable" as a wrong count (a correct report was rejected 3× for this), and every rejection now quotes the exact line it objected to, so a rewrite lands in one attempt.

Honesty & safety hardening (from real run reviews)

  • Never weakens your project to make a scan work. New hard rules: the agent must never disable your security features (secure boot, flash encryption, signed OTA) to force a build, must never edit your SDK/toolchain installation (a run had patched a script inside C:\ncs\, now banned), and if an earlier scan run left your config modified, the posture check detects it and offers a restore instead of counting those as your gaps.
  • ESP SBOM generation fixed for IDF 5.x. The documented --output-file flag is used, and idf.py sbom-create --spdx-file (absent on IDF 5.5.4) is version-checked before use instead of failing.
  • Silent commands aren't "failures" anymore. Commands that legitimately produce no output (mkdir, cp, …) no longer report a scary "technical issue", the agent is told plainly: silent success, verify state directly if it matters.

Windows reliability

  • The terminal just works on a fresh Windows install. New Windows machines ship PowerShell locked down (Restricted execution policy), which silently blocks VS Code's shell integration so the agent can't read command output. The extension now detects and repairs this in the background at startup, sets the execution policy to RemoteSigned (current-user scope), selects a working default terminal profile, and restarts stale terminals, with a dismissible note of what changed. Group Policy-managed machines are left untouched.

Compact input area

  • The input stack is dramatically slimmer: one-line input that grows as you type, auto-approve as a compact ⚡ chip next to @, and the wide Cancel/Resume buttons replaced by a Claude-style send ↔ stop morphing icon (brand cyan, instant tooltips). The chat input glows cyan on focus.
  • The "What's new" card icon is now theme-consistent (no more OS-style emoji).

v0.1.7 — CRA readiness

Choose a tag to compare

@sssadsum sssadsum released this 08 Jul 19:09

[0.1.7] - 2026-06-24

CRA Readiness Check: get CRA-ready as you build

A new one-click CRA Readiness Check helps you prepare for the EU Cyber Resilience Act, on both nRF and ESP. It's a readiness snapshot to help you prepare, not a conformity assessment, and not legal advice.

  • SBOM from your real build. A machine-readable software bill of materials (SPDX), the CRA's named artifact, generated from the actual build, not a guess.
  • Secure-by-design posture. A checklist against your build's real configuration (secure boot, signed updates, debug-port lock, secure pairing, secure storage, and more), each item evidence-grounded (the literal config fact, not a generic assertion), ordered so prerequisites come first.
  • Advisories for your SDK version. Surfaced with links to review, never an automatic verdict.
  • Help you start. The agent offers to begin closing the top gap (e.g. add a secure bootloader), routing into the normal add-feature flow.
  • Writes a compliance/ folder: a human-readable report, a machine-readable JSON companion, and the SBOM. Run it on your project, or try it on a bundled sample with nothing open.

It also tells you which CRA date applies to you up front, whether you're getting a head start on the Dec 2027 essential requirements, or already in scope for the Sep 2026 reporting duty.

HCI deep-debug: three layers for BLE projects

When a BLE project is open (CONFIG_BT=y detected), the welcome screen now surfaces a one-line shortcut to the full debug stack:

app logs · HCI · radio sniffer (soon)

The HCI layer is new: the extension can now parse Host Controller Interface (HCI) event streams and decode them to human-readable BLE protocol events, helping you see exactly what's happening at the controller level, not just the application log.

A guided HCI + Sniffer sample run, a three-layer walkthrough (app log → HCI bus → over-the-air) that lands on the one-line fix and bridges to the CRA readiness check, is coming in a follow-up release.

nRF terminal: self-contained, version-aware

The nRF terminal now detects which NCS version your build uses and executes commands against that SDK automatically, no manual path wrangling. When you have multiple NCS installs (e.g. v3.2.1 and v3.3.1), the agent follows the build directory's version, not whichever was compiled most recently.

ESP reliability

  • Multi-board builds. When multiple ESP-IDF versions are installed, the extension now shows all of them, not just the first, removing the "ambiguous forever" state where a project pinned to a different version than the global install.
  • Version detection on git-clone installs. ESP-IDF git checkouts have no version.txt; the extension now reads tools/cmake/version.cmake as a reliable fallback, so the platform strip always shows the correct IDF version regardless of install method.

Welcome screen: adaptive, not static

  • "Try it on a sample" picker. A new inline picker lists available demos (BLE bug, CRA readiness check, and more). Appears when no project is open; becomes a re-run link once you've tried one.
  • CRA nudge for BLE projects. When a BLE project is open but has no compliance/ folder yet, a dismissible note surfaces: "A connected product likely falls under the EU CRA, preview your secure-by-design posture." Evidence-grounded, never a verdict.
  • Compact platform status. The nRF/ESP detection panel collapses to a one-line summary per detected platform (e.g. nRF · NCS 3.2.1 · nRF5340 DK). Click to expand for full detail.

Diagrams that match your theme

Mermaid diagrams in chat now follow your VS Code light/dark theme and the Adsum palette, previously they used a fixed dark theme that was hard to read in light mode.

Honesty hardening

The agent's CRA output is now verified against a 108-fixture scanner before it leaves the model. New rules added this release: the agent can no longer produce numeric readiness scores, "non-compliant" verdicts, or citations to CRA articles that don't exist. All posture items are evidence-mode only, literal config facts, with "verify" always the next step.

Legal

  • NOTICE file added (Apache 2.0 §4c attribution).
  • iot-knowledge/LICENSE makes the open k-bits' license explicit.
  • README updated: open-core model, AI limitations, trademark notices.