Releases: DrOlu/neuralosd
Release list
v1.4.0
Full Changelog: v1.3.5...v1.4.0
v1.3.5
v1.3.4
v1.3.2
v1.3.1
v1.2.4
neuralosd v1.2.3 — group-by breakdowns
Standalone binaries for the neuralOS deterministic-first runtime.
New: group-by breakdowns
neuralosd init now generates a breakdown probe for every data source —
a real GROUP BY with caged arguments, so plain English fills them:
neuralosd ask --instance-dir ./store "revenue breakdown by year"
neuralosd ask --instance-dir ./store "sales by market"
neuralosd ask --instance-dir ./store "profit by category"
2012 $2,259,450.90
2013 $2,677,438.69
2014 $3,405,746.45
2015 $4,299,865.87
TOTAL $12,642,501.91
Dimensions are the low-cardinality categorical columns plus year / quarter /
month derived from a detected date column. Measures are the non-identifier
numeric columns, and the measure is optional with a default — which is why
"revenue" resolves to the Sales column without naming it.
Fixed: a confident wrong answer
The router used to take the first probe whose args were extractable.
"sales by quarter" therefore matched a flat total_sales and returned a
single grand total — a confident answer to a question nobody asked. It now
prefers the probe that accounts for the most of the question through its
extracted arguments. Verified with a regression test that fails when the fix
is reverted.
Verified
Numbers cross-checked against an independent openpyxl pass on the real
workbook — identical to the cent. 168 unit tests · 12-check CLI smoke ·
strict backend assertion · three sidecar proofs, on every platform.
Variants
| Variant | File | Adds |
|---|---|---|
| base | neuralosd-<platform> |
framework + 4 doc volumes + 5 skills + model + Excel |
| boxlite | neuralosd-boxlite-<platform> |
+ the BoxLite microVM engine |
| msb | neuralosd-msb-<platform> |
+ the Microsandbox msb runtime |
Full Changelog: v1.2.2...v1.2.3
neuralosd v1.2.2 — the model fallback actually works now
Standalone binaries for the neuralOS deterministic-first runtime.
The headline fix: --model did nothing
Found by installing from PyPI and using it. --model loaded the model and
then never consulted it, for two independent reasons:
- The fallback imported a module that did not exist (
_cmd/model_bridge),
and a bareexceptswallowed the error — so it returned nothing on every
question, silently. The real bridge now exists. - The router refused before offering anything to the model when lexical
retrieval found no overlapping tokens — exactly the paraphrase case the
model is for.
Now:
| Retrieval | With --model |
|---|---|
| a probe matches and its args extract | deterministic — the model is never loaded |
| a probe matches but args are missing | model picks among the retrieved candidates |
| nothing matches | model is shown the whole menu |
nothing matches, no --model |
honest refusal: no probe matched this question |
Also fixed: a segfault under load
Six parallel /ask calls against serve --model killed the server
(segfault). The on-device engine is not thread-safe at construction either,
not just at inference, so serialising run() alone was insufficient. All
model work now sits behind one lock; the deterministic path stays fully
concurrent. A regression test asserts maximum overlap == 1 and was verified to
fail with the lock removed.
Verified on this host, from PyPI
pip install 'neuralosd[all]' → deterministic ask, model fallback, honest
refusal, both sandbox backends available, bundled docs including the new
model-fallback sections.
Variants
| Variant | File | Adds |
|---|---|---|
| base | neuralosd-<platform> |
framework + 4 doc volumes + 5 skills + model + Excel |
| boxlite | neuralosd-boxlite-<platform> |
+ the BoxLite microVM engine |
| msb | neuralosd-msb-<platform> |
+ the Microsandbox msb runtime |
12-check CLI smoke · strict backend assertion · three sidecar proofs ·
151 unit tests.
Full Changelog: v1.2.1...v1.2.2
neuralosd v1.2.1 — docs now shipped inside the binary
Standalone binaries for the neuralOS deterministic-first runtime.
What changed
The binary bundles four documentation volumes, readable without a network:
neuralosd docs usage # operations guide
neuralosd docs architecture # design manual
neuralosd docs cookbook # recipes
neuralosd docs reference # full API reference
1.2.0 shipped the sidecar and uv provisioning but the bundled volumes did
not describe them — that material had only been written into the README, which
is not part of the binary. Now all four volumes cover it:
- usage — how to provision with
neuralosd sidecar --setup, what uv
actually runs, where the environment lives - architecture — why bootstrapping exists and why it must stay explicit
- cookbook — a recipe: "Use a library the binary doesn't have"
- reference — the sidecar protocol,
sidecar_client,provisionAPI,
and every environment variable
The five bundled skills are readable too (neuralosd.skills()).
Everything else from 1.2.0
uv install paths · sidecar delegation (in-process first, host Python for what
the binary lacks) · explicit neuralosd sidecar --setup provisioning with
zero-config discovery · Excel support baked in.
Variants
| Variant | File | Adds |
|---|---|---|
| base | neuralosd-<platform> |
framework + 4 doc volumes + 5 skills + model + Excel |
| boxlite | neuralosd-boxlite-<platform> |
+ the BoxLite microVM engine |
| msb | neuralosd-msb-<platform> |
+ the Microsandbox msb runtime |
Verified on every platform: 12-check CLI smoke, strict backend assertion, and
the three sidecar proofs (clean failure / delegation / uv-provisioned with
zero configuration).
Full Changelog: v1.2.0...v1.2.1
neuralosd v1.2.0 — uv provisioning (sidecar with zero configuration)
Standalone Nuitka binaries for the neuralOS deterministic-first runtime.
New in 1.2.0 — uv is a first-class path
Install / run with uv:
uv pip install "neuralosd[all]"
uv tool install neuralosd
uvx neuralosd ask --instance-dir . "how many rows"
Provision the sidecar — no Python needed on the machine:
neuralosd sidecar --status
neuralosd sidecar --setup --with pypdf,pywinrm
neuralosd sidecar --setup --force
uv supplies its own CPython, so this works on a host with no usable
interpreter — which was the sidecar's one remaining limitation. The binary
then finds the environment automatically, with no environment variables.
Provisioning is always explicit: it touches the network, so it never happens
silently while answering a question.
How the sidecar decides
| Situation | Behaviour |
|---|---|
| all of a probe's imports resolve in the binary | runs in-process (fast) |
| an import is missing, even lazily inside the function | retried in the sidecar |
probe declared tier="sidecar" |
always runs in the sidecar |
| missing import, no sidecar | exits with pip install <lib> + the setup hint |
NEURALOSD_SIDECAR=off disables delegation.
Verified on every platform
12-check CLI smoke · strict backend assertion · and three sidecar proofs:
without a helper the binary fails with pip install pypdf; with one it
returns a real PDF page count from a library it does not contain; and with a
uv-provisioned environment it does so with zero configuration.
Variants
| Variant | File | Adds |
|---|---|---|
| base | neuralosd-<platform> |
framework + docs + skills + model + Excel |
| boxlite | neuralosd-boxlite-<platform> |
+ the BoxLite microVM engine |
| msb | neuralosd-msb-<platform> |
+ the Microsandbox msb runtime |
Full Changelog: v1.1.0...v1.2.0