Skip to main content

SerenGanglion

Seren's reflexes on the M5Stack LLM module.

A ganglion is a cluster of nerve cells outside the brain that runs reflexes without asking it, and a wake word is a reflex. This one lives on the LLM module inside a StackChan: a StackFlow unit beside M5's own, on the module's own Linux. It does the module's work there, and the robot's core hears only what it needs. It sits in the Seren family beside Sinew, Meninges, Soma and Psycheion.

What's in it (today)

Early. The unit joins StackFlow and says hello; listening, sounds and LEDs come next.

  • seren-ganglion check: what this module is, read-only. Its Python and C library, the versions it carries, StackFlow's sockets and who may connect to them, how M5 starts its units, and whether a microSD card shows up. Every part answers on its own, so one that can't run says why and the rest still report. It writes nothing and opens no socket.
  • The StackFlow wire (seren_ganglion.wire): the framing M5's units use between themselves, read from M5's own C++ (pzmq.hpp, StackFlow.cpp), standard library only.
  • The sys client (seren_ganglion.sys_client): registering with StackFlow's sys unit and calling any unit over ZeroMQ. It's the one part that needs pyzmq.
  • seren-ganglion run, the unit (seren_ganglion.unit): it answers StackFlow's seven actions. setup registers a task with sys and pushes a ganglion.hello down the serial line to the core: its version, what it carries, and whether its box holds. taskinfo and exit work. pause, work, link and unlink answer -18, "not have unit action!", the way M5's own units do. Nothing in it opens a mic.
  • The .deb, built on every release: one file with everything inside, for a module with no network, and seren-ganglion.service, enabled when it installs.

Install

On a module

Download seren-ganglion_<version>_arm64.deb from the releases and check it against that release's SHA256SUMS. Get it onto the module however you have to hand (below), then:

dpkg -i seren-ganglion_<version>_arm64.deb
seren-ganglion check

The module needs no pip, no venv and no network. Everything is under /opt/seren/ganglion/, and dpkg -r seren-ganglion takes it all away again.

For working on it

pip install seren-ganglion

That gives you the Python package on a desktop, for development and tests. To fetch the wheels for a module by hand: Ganglion and Sinew are pure Python, but pyzmq has to be fetched for the module's platform. It comes as one file, pyzmq-26.4.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl, with ZeroMQ's library inside:

pip download seren-ganglion seren-sinew --no-deps -d ganglion-wheels
pip download pyzmq --only-binary=:all: --platform manylinux2014_aarch64 --python-version 3.10 --no-deps -d ganglion-wheels

--no-deps because Sinew's declared dependencies (starlette and its family) are never imported here. On the module a wheel is a zip, and python3 -m zipfile -e <wheel> <folder> needs nothing that isn't already there.

Where it runs

The LLM module (AX630C) is its own computer inside the StackChan: Ubuntu 22.04 (arm64), kernel 4.19, about 1 GB of RAM, 32 GB of eMMC, M5's StackFlow units, and no network of its own. It talks to the CoreS3 over one UART, as JSON lines at up to 1.5 Mbaud. What's known about the module, and how it was found, is in SerenSoma's firmware/docs/appendages.md, under "The LLM module as ears".

A package, not a distro. Forking M5's image would mean their Windows-only reflash, with the module out of the stack, for every mistake, and owning a kernel 4.19 BSP with closed NPU blobs through every M5 update. Everything we want lives above the kernel, so it's one .deb beside M5's packages.

Python 3.10 is the floor because it's the module's. Raise it and the module falls off.

What it will do

It joins StackFlow as the unit ganglion, so the core reaches it as ganglion.<n> through the same locked route it uses for M5's units: one more name on that route's allow list. bashexec, push, upgrade and rmmode stay off the robot's API.

  • Listening, wired inside the module. Wake word, VAD and Whisper chained on the module. The core gets one finished event, heard, with its text and timing. Only text crosses the UART.
  • Face sounds. It plays the MP3s the core forwards. M5's audio.play takes wav, mp3 or pcm, not Opus, so Psycheion's Face Studio makes the MP3s; at most about 64 KB a line.
  • The three LEDs, its health, its versions.
  • Its own upkeep: updating M5's StackFlow as a script anyone can run, telnet on port 23 switched off at install, and a refusal to run on a StackFlow older than the one tested.

Its shapes live in Sinew, beside the robot's, with the same contract test, so the two ends can't drift. Sinew's body shapes import with the standard library alone from 5.10 on, and tests/test_sinew_on_the_module.py holds us to it: it fails on 5.9, which imports starlette at its front door.

It can be Python. StackFlow units talk ZeroMQ over socket files, and the protocol doesn't care what a unit is written in (Psycheion read pzmq.hpp and StackFlow.cpp, 20 Nov; checked the same night):

  • Each unit answers requests on a REP socket at ipc:///tmp/rpc.<unit>. A call is two frames, the action then its data, and the answer is one frame.
  • Two values travel as one string: a byte giving the first one's length, the first, then the second.
  • A unit joins by calling sys's register_unit with its name. The answer is nested: its work id number, paired with a second pair of its output URL and its inference URL.

The mic light is the core's, never the ganglion's

Psycheion, 20 Nov; Chad: "yes yes yes". A unit that crashes or hangs could leave a mic open and never say so, which is the hidden microphone again. So:

  1. The ganglion opens a mic only when the core tells it to. Never on its own, and never by itself after a restart.
  2. The core doesn't take its word for it. Every few seconds it asks sys.lstask. If a task that captures audio is set up (ours, or M5's audio, kws, vad, asr or whisper), the light is on. Tasks, not services (Psycheion, 21 Nov): all five of M5's listening services run all day as daemons waiting for a setup, so "the service is up" would keep the light on forever and mean nothing. Never build this against systemctl.
  3. No answer counts as listening. A module that stops replying keeps the light on until it answers.

The ganglion's own "mic open" event is still wanted, for the face's listening look, because it's fast. It never decides the light.

The bench test for the rule is the kernel's own word. ALSA shows whether a capture stream is actually open, whoever opened it. The module has one capture device, and with nothing set up it reads closed (Chad's module, 21 Nov):

cat /proc/asound/card0/pcm0c/sub0/status

On the bench the light must be on exactly while that says RUNNING.

Connections: however the person brings it

Chad, 20 Nov: the module has a microSD slot and a USB-C port, and the Ethernet board "doesn't really go on for the real stuff". So nothing assumes a network.

  • Talking to the robot is always the UART, through the core. It's the one link every StackChan with a module has. The ganglion listens on no network port.

  • Getting the .deb and models on takes whichever of these is to hand:

    Way How
    microSD Copy it onto the card. Unproven: on 21 Nov no card was in, so whether the slot shows up in its Linux is still open
    USB-C adb push from a computer
    Ethernet, if someone has it Download it on the module
    The UART Later, through Soma (below). About ten seconds for the .deb

Who it runs as

Psycheion, 21 Nov: the ganglion parses whatever the core sends, so a parsing bug as root is a root bug, on the box we locked bashexec away from. M5's units run as root, and their sockets may admit only root. Connecting to a unix socket needs write permission on its file, which check reports.

Answered by check on Chad's module, 21 Nov (v0.3.0): root-only. Every StackFlow socket (/tmp/rpc.sys and one per unit, plus /tmp/llm/5556.sock) is root:root, srwxr-xr-x: no group or other may connect. M5 starts each unit as its own systemd service, llm-<unit>.service.

  • Had the sockets let a group write, the .deb would make a ganglion system user in that group. They don't, so:

  • It runs as root, boxed by systemd, and this section says so until it's fixed:

    NoNewPrivileges=yes
    CapabilityBoundingSet=
    ProtectSystem=strict
    ReadWritePaths=/var/lib/seren-ganglion /tmp
    ProtectHome=yes
    PrivateNetwork=yes
    RestrictAddressFamilies=AF_UNIX
    

    That means no capabilities, a read-only system apart from its own state and the sockets, and no network at all. Path-based unix sockets still work across a private network namespace, so StackFlow is reachable and nothing else is. PrivateTmp stays off, because the sockets live in /tmp. These need checking against the module's systemd before they're trusted.

  • We don't change M5's units to loosen their sockets. A drop-in that changed how their root processes create files would reach well past ours.

The service follows M5's own (systemctl cat llm-sys llm-skel, 21 Nov): theirs run as root from /opt/m5stack, Restart=always, RestartSec=1, StartLimitInterval=0, and every unit but sys has After= and Requires=llm-sys.service. Ours, seren-ganglion.service, does the same, plus the box above. The module's systemd is 249, built with +SECCOMP, and the kernel has seccomp (the Seccomp: line in /proc/self/status only exists when it does).

The unit tests its own box when it starts. Whether the kernel has seccomp filters, which RestrictAddressFamilies needs, doesn't show from outside. So the first thing it does is try to open a network socket: under the box that fails. If it succeeds, the box has a hole, and the unit says so in its log and in taskinfo rather than carrying on quietly. PrivateNetwork=yes still holds the line by itself: a network socket in a private namespace reaches nothing but its own loopback.

The kernel has everything the box needs (/proc/config.gz, 21 Nov): CONFIG_SECCOMP_FILTER=y and CONFIG_NET_NS=y. The self-test stays anyway, for the day a module ships a kernel without them. On the NUC's systemd, a process under RestrictAddressFamilies=AF_UNIX got "Address family not supported by protocol" for a network socket and opened a unix one, which is what the self-test reads.

The service ships in the .deb and is enabled on install: lib/systemd/system/seren-ganglion.service. StartLimitIntervalSec=0 sits in [Unit], where systemd reads it; in [Service] it's ignored (systemd-analyze verify caught it). M5 writes the older StartLimitInterval, which [Service] still takes.

Later: Soma installs it for you

Soma already fetches firmware from GitHub releases and checks checksums before staging it. It could fetch this .deb the same way and send it to the module through the core. Psycheion, 21 Nov: installing a .deb is dpkg -i as root, so while that install mode is on, whoever holds the robot's token has root on the module. A checksum on the NUC proves the right file left the NUC, not who's sending it now. So, before anyone builds the convenient version:

  • The switch is hands on the robot, not the token: for example, holding the head for five seconds in the first two minutes after a boot. Soma's page says "touch the robot to allow the install".
  • It closes after one install or ten minutes, whichever comes first.
  • It takes one file to one fixed path and runs one fixed command, dpkg -i on that path. The core builds the command itself; nothing the caller sends ends up in it. It never takes shell lines.

Releasing

The family's shape (.github/workflows/release.yml):

  1. Every push: the tests, on Python 3.10 to 3.14. The first set is the leanest, Sinew and pytest only, as close to the module as a runner gets; [dev] adds pyzmq for the socket tests.
  2. A vX.Y.Z tag on main: setuptools_scm takes the version from the tag, so the wheel and the .deb always carry the same number.
  3. The wheel and sdist are built, then the .deb by tools/build_deb.py. That's pure Python, so it builds the same on a desktop as in CI, with no dpkg needed. Everything inside is pure Python or an already-built wheel, so there's nothing to compile for ARM. It carries the wheel just built, the newest Sinew the pyproject allows, and pyzmq for the module, and records each with its checksum in /opt/seren/ganglion/BUNDLED. File times come from the tagged commit, so a rebuild is byte-identical.
  4. The GitHub release gets the wheel, the sdist, the .deb and SHA256SUMS. PyPI gets the wheel and sdist, by trusted publishing, so no token is stored.

A dev version like 0.2.0.dev3 becomes 0.2.0~dev3 in the .deb, so dpkg sorts it before 0.2.0.

Layout

Where What
seren_ganglion/check.py seren-ganglion check, read-only
seren_ganglion/unit.py seren-ganglion run: the StackFlow unit
seren_ganglion/wire.py StackFlow's framing. Standard library only
seren_ganglion/sys_client.py RPC to sys or any unit over ZeroMQ
tools/build_deb.py The .deb, from wheels
Sinew, seren_sinew.ganglion (to come) The shapes the ganglion and the core exchange
SerenSoma The core's half: forwarding to ganglion, the lstask check behind the mic light, the allow list

Tests

cd SerenGanglion
pip install -e ".[test]"
python -m pytest

The socket tests run against a fake sys on tcp, because Windows has no ipc sockets. The .deb is read back the way dpkg reads it, and a real one was checked with dpkg-deb on 21 Nov.

What's coming

  1. check on a real module done 21 Nov: Python 3.10.12, glibc 2.35, libzmq 4.3.5 loads, sockets root-only, units as llm-<unit>.service, systemd 249, the mic closed at rest.
  2. The unit joins StackFlow built for 0.4: registers, says hello, taskinfo, exit. Next is ganglion.hello reaching Soma through the core for real, then listening.
  3. A systemd unit ships in 0.4. Still to do: telnet off on install, once we know what serves port 23.
  4. Its shapes in Sinew, and the core's half in Soma.

AGPL-3.0-or-later.

Metadata

Release files for seren-ganglion 0.4.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for seren-ganglion 0.4.0
File Size Uploaded Provenance Actions
seren_ganglion-0.4.0.tar.gz 32.0 kB With provenance Details for seren_ganglion-0.4.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for seren-ganglion 0.4.0
File Interpreter ABI Platform Actions
seren_ganglion-0.4.0-py3-none-any.whl Python 3 none any Details for seren_ganglion-0.4.0-py3-none-any.whl

Total release size: 49.3 kB

Release files / seren_ganglion-0.4.0.tar.gz

Download URL seren_ganglion-0.4.0.tar.gz
Size 32.0 kB
Tags Source
SHA-256 checksum
How to use checksums
153f4087fc7931e31f304374d41e7388baa1028fd4df25b5c66ae26ea32148fd
BLAKE2b-256 checksum
How to use checksums
cb7dd39886cfd0a465c062f03de2deda011438a88c543f3fd5a64ae4bd65ef02
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 22, 2026.

Transparency log

Release files / seren_ganglion-0.4.0-py3-none-any.whl

Download URL seren_ganglion-0.4.0-py3-none-any.whl
Size 17.3 kB
Tags Python 3
SHA-256 checksum
How to use checksums
933e0b9c76992c7bc1cc53c0db347e058ab81b851f3b9447dae1448c2b63ed6f
BLAKE2b-256 checksum
How to use checksums
b5579e79addf6db5699b9ee73b552a121fb99f202cf5c396f82bc4739a19e4fc
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 22, 2026.

Transparency log

Security for seren-ganglion 0.4.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for seren_ganglion-0.4.0.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 22, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page