Repository navigation
Releases
PlantCraft 0.6.1
PlantCraft 0.6.1
Released 2026-11-09. The setup guide (Heron) for new installations, a demo plant of 100 loops with loop health that matches the known faults, app packages and licences for closed apps, PlantCraft PID, the Historian's data lake feed, and fixes from the public demo and the engineer sweeps.
[0.6.1] - 2026-11-09
Added
- Simulator and demo: 100 loops, most of them healthy. Beside the 50 loops of the demo
story, the Simulator serves 50 healthy, well-tuned loops over every unit, so the demo reads as
a real plant where a minority of loops need attention. - Platform: the setup guide. A small robot walks an administrator through a new
installation: a language model, the basics (time zone, language), the plant's data (OPC UA,
PI Web API or a file), documents, the asset tree, the apps' first steps, the team and email.
Each step knows by itself when it is done. "Do it with me" opens the assistant with that
step's work and proposes each change as a card the person applies; the model never writes
(setup-guide.md). - Platform: asset trees drafted from more tag conventions. DeltaV paths, Experion names,
area prefixes, Portuguese and Spanish role words, and tags with no ISA tag (Pump1_Speed)
all draft a tree; a CSV whose times carry an offset reads. - Platform and SDK: app packages and licences.
app.tomlsays which PlantCraft versions an
app installs on and was tested on, its languages and whether it is closed; PlantCraft checks
them before an install or an upgrade. Closed apps run with a signed licence for one
installation, added in Apps › Licences (app-package.md). - PlantCraft PID, the first closed app: the P&ID reader and plant map, moved out of
Documents. - Historian: data lake feed. A saved export goes to the data lake every hour or every day.
- Docs: integrators' guide for PlantCraft next to MES, a data lake and an on-premises
language model (integrators.md). - Development:
make previewandmake preview-lightstart the whole suite locally with the
demo story, apart from any other installation.
Changed
- SDK: written once. Analytics runs block code in the SDK's one sandbox (as Alarms and
Documents do); the asset picker, the apps' command line and the PI Web API client are one copy
in the SDK. - The apps build against the repository's own SDK and ui, and pull requests run the tests of
every folder they change. - Tours are shorter, and Operations has its own.
- Translations: the pt-BR and es wording follows the words industry tools already use.
Fixed
- Public demo: a visit no longer ends while someone reads (scrolling, pointer moves and page
loads count as use); the clock no longer covers pages or the assistant's box; Documents opens
inside the suite's frame; the assistant retries a model that limits the rate, follows answers
without stalling, and charts the right hours. - Platform: an asset tree drafted from a source now routes its tags to that source, so the
plant shows values after the setup; a text signal's last value (a loop's mode) shows; Desk's
Coming days leaves out what already happened; notifications of a row of changes arrive as one;
long references wrap on a phone; Customize no longer covers an app's notice on its Overview. - Analytics: loop health matches the known faults. Checked against the Simulator's 100
loops, the likely cause matches in all 100 at live sampling (about 22 of 50 before): a move to
a new setpoint is no longer counted as error or oscillation, stiction, noise and valve travel
read right, and a new diagnostic names a slow output drift. Each loop's row shows its likely
cause beside the health score. - Analytics: a database gives disk space back by itself (older ones are compacted once, in
a quiet hour); the engine hands back the memory a round used; the Overview is built once at a
time; phone layouts fit. - Alarms, Documents, Historian, Models, View: overview cards translated; clearer messages
and labels (a rule without a place, "found by words", compression figures, asset types in
View's picker, readable coloured table cells). - Install: a bundle over 2 GiB is written in parts GitHub accepts; the machine profile sizes
Analytics' read cache; the public demo's limits reach PlantCraft from.env.
Security
- Historian: a PI Web API source only ever reads (its session refuses anything but GET, HEAD
and OPTIONS).
Downloads
| Bundle | For | Size | SHA-256 |
|---|---|---|---|
plantcraft-0.6.1-linux-amd64-bundle.tar.gz |
most servers | 2,253,065,117 bytes (2.1 GiB) | fb412706b4891c6da44238bdd3de77ee98f92427714a3a27abd9a263c0d2aee1 |
plantcraft-0.6.1-linux-arm64-bundle.tar.gz |
ARM servers and Apple silicon | 2,183,450,902 bytes (2.0 GiB) | 678b095b3061584fbb10c21bf82b3fe6586752f31e5f9efe775d55c4387e4ff1 |
Each bundle comes with its .sha256 and .sha256.sig. docs/verify-a-release.md explains how to check them with ssh-keygen as release@plantcraft (OpenSSH fingerprint SHA256:9l/yDgR8/Pzel1pncxrXz2IGCIhFirKpDT78tO49fG8).
Joining the parts
GitHub limits each file to 2 GiB, so a larger bundle comes in parts. Download every part of your processor, join them, then check the hash and signature as usual:
cat plantcraft-0.6.1-linux-amd64-bundle.tar.gz.part-* > plantcraft-0.6.1-linux-amd64-bundle.tar.gz
Upgrading
From 0.6.0: unpack the new bundle and run ./install.sh. It checks the bundle's signature, backs up the data folder before updating, and rolls back automatically if the update fails (./install.sh --rollback also returns to 0.6.0 by hand). Data, settings and accounts are kept. The only schema change is one new table, app_licenses; nothing is migrated by hand.
Each app's own changes are in its CHANGELOG.md under apps/.
PlantCraft 0.6.0
PlantCraft 0.6.0
The suite in three languages (English, Brazilian Portuguese and Spanish), an installation file and requirements check for whoever installs the server, Operations as a built-in app, per-visit public demo sessions, and security hardening across the platform and apps.
[0.6.0] - 2026-11-02
Added
The suite in three languages
- Brazilian Portuguese and Spanish for the whole suite. Every screen, message, report and
help page of the platform and the six apps, Analytics' block catalogue included, with one
language choice for the whole suite. The assistant answers in the language chosen. A check in
make cikeeps the three languages complete (translations.md).
Installing and running the suite
-
Platform: the installation file. Choices that belong to whoever installs the server live
in/etc/plantcraft/installation.conf. The file is owned by root and never mounted in a
container. It sets the apps, demo, machine profile, CPU and memory ceilings, network and
language-model switch, and its values win over the command line and.env.
./install.sh --show-configsays whether it is applied (installation-file.md). -
Platform: requirements check before installing.
./install.sh --check(orpreflight.sh,
--json) checks these and changes nothing:- system and processor;
- Docker, Compose and Docker Desktop's licence;
- control groups, CPUs, memory, swap and free disk;
- the HTTPS port, internal networks and clock.
install.shruns it first and stops on a FAIL. It now also accepts Rancher Desktop'sdocker
in WSL. -
Platform: one machine profile sizes the whole suite.
./install.sh --machine small|medium|large|xlarge|auto|offsets every container's CPU and memory limit, processes
and pools from one choice. The defaultautopicks by the CPUs and memory Docker sees, and
values edited by hand in.envare kept. Operations shows the machine and its profile, and
suggests the profile that fits after the machine changes. -
Platform: Docker Engine in the bundle. On a Linux server with no Docker,
./install.sh --install-dockerinstalls the Docker Engine 29.9.2 and Compose 5.6.1 the bundle
carries, offline. -
Platform: setup wizard (first version). The installer's screens open in a browser on the
server (127.0.0.1, a random key per start). They ask for the apps, check the requirements for
those apps, write the installation file and run the install. -
Platform:
./install.sh --rotate-admin-password. A new emergency administrator password,
shown once. -
Platform: uptime by app and maintenance windows in Operations › Status, and an official
Zabbix template in the bundle (monitoring/plantcraft_zabbix_agent.yaml). -
Platform: health that tells the truth.
/healthzanswers 503 when an app's engine is in a
crash loop, so Docker, System health, Operations and Zabbix all see it. -
Platform:
plantcraft llm set|show. The default language model and its key are set from
the server's shell, the key read from standard input, with no browser. -
Operations: the plant's holidays stop the clocks of queues with working hours.
-
Compute: a revoked worker enrolls again with a new one-time code, under the same name.
The public demo
- Each visit is its own. A visible 30-minute clock, the visit ends after 5 idle minutes, 30
places with a waiting page, and a notice before an update. With the switch on, each visit gets
its own copy of every app's data, so what a visitor saves in Alarms, Analytics and View goes
with the visit; writes to the platform stay refused. The demo says which parts are locked for
security, and how long the visit lasts. - The assistant can be on in the demo. Ask-only, with per-visit chats, 10 questions per
visit, limits per address, minute and day, and the tokens of every answer counted
(PLANTCRAFT_PUBLIC_DEMO_LLM, off by default). A provider without credit reads as a pause, not
an error, and the question does not count.
Operations, a built-in app for the people who keep the suite running
- Operations sits beside Desk for administrators and the new Operations role.
- Requests: anyone asks IT or an integrator for help from any app ("Ask for help"). Each
request has first-answer and resolution targets per priority, and a queue's clocks run only
in its working hours. - Partners: the companies working at the plant, each with its coordinator, teams, a sponsor
at the plant and a last day of access. A sponsor approves each month's hours. - Temporary support access: a person gets Operations for up to 72 hours, for a reason. The
access ends by itself.
Apps
- Analytics: a plant map on the Overview, with every loop coloured by the KPI you pick. Also:
- effective control and loop-hours lost by cause;
- days until a falling health reaches its limit;
- oscillations several loops share, grouped, with the likely origin;
- overshoot, rise and settling times and IAE of each setpoint step found in the history;
- a read-only PI tuning suggestion (Lambda/IMC);
- a loop compared with a recorded reference;
- Horch's cross-correlation in the stiction screen;
- signal checks for rate of change, NE 43 fault levels and the source's own quality;
- a Virtual sensor algorithm (Kalman filter and drift);
- algorithms that run on every asset of an equipment type;
- third-party APC health and the cost of each active limit: % on, MVs in service, CVs at
a limit, and limits ranked by shadow price, also in the weekly report; - virtual sensors and models without a physical model: lab fusion (the online value
corrected by the lab at each sample's time), a regression block (OLS, Ridge, PLS) trained by
Calibrate with its validity range, an empirical model block (PCA, T² and SPE), mass balance
reconciliation with gross error detection, and drift of a redundant pair; - "New action" on every loop result, with the loop's charts attached;
- the diagram of an algorithm with lit, clickable wires and full screen.
- Alarms:
- import the DCS alarm list (CSV or Excel) as a draft;
- consoles (operator positions) with their own alarm-rate targets;
- limits by product grade;
- state-based alarms, enabled only while a permissive holds;
- an alarm performance section (ISA-18.2) in View's reports;
- the EEMUA 191 performance level, floods with hysteresis, and what the journal tells.
- Historian:
- store and forward at the gateway (MQTT, Modbus and SQL survive a collector outage);
- sources with owners, a contact and an expected interval, whose owners hear when a source
fails; - audited edits of archived values;
- OPC UA status codes kept with quality;
- notes on a signal;
- shortest interval between archived values (CompMin);
- laboratory results as signals, and a LIMS read as a SQL source;
- a limit on the points it records (
MAX_SIGNALS); - daily and monthly rollups, each with its own retention, and the disk each one takes;
- interpolated reads with a maximum gap, and summaries (count, % good, totals, time-weighted
statistics) in API v1; - Sparkplug B sources, read only, with points added from births;
- OPC UA subscriptions, opt-in per source (still read only).
- View:
- a multivariable controller (APC) panel;
- reports made when a batch or a start-up ends;
- assistant report sections that cite their facts as numbered footnotes;
- reports over whole calendar periods;
- trends compared with earlier periods;
- laboratory results in trends;
- ISA-101 analog indicators and alarm marks by priority;
- a generated Level 1 overview of the whole plant;
- Back between diagrams.
- Models:
- tracking: the model checked against the plant on a schedule, with model-minus-plant signals
and a "Model off the plant" event; - an example optimization that tunes a PI for the least IAE;
- OMChemSim in the runner, with a methanol and water flash drum that runs;
- Compile a model and see whether it is compiled.
- tracking: the model checked against the plant on a schedule, with model-minus-plant signals
- Simulator: operator training with scheduled faults and a separate answer key, snapshots,
freeze and speed.
For app developers (plantcraft-sdk 0.80.0 → 0.105.0)
- Images in an app's language model calls (
POST /api/v1/llm/{use}, 0.86.0). PlantCraftClient.lab_samplesreads laboratory results with their status and times.GET /api/v1/signals/{id}/valuetakesend, for the value of a window that closed earlier.plantcraft_sdk.files.write_atomicandwrite_json(0.91.0).- A binary series answer,
plantcraft_sdk.wire(0.85.0). plantcraft_sdk.server.workers(0.80.0).- Starter asset types for a third-party multivariable controller:
apc,apc_mvandapc_cv
(0.93.0). fuse_lab(0.94.0);series(..., sampling="interpolated", max_gap=)andsummary()
(0.95.0).- A copy of every app's database for each public demo visit (0.97.0, off by default).
askTextandshowToastin@plantcraft/uireplacewindow.promptandwindow.alert.- One translation mechanism:
createI18nin@plantcraft/ui,plantcraft_sdk.i18nwith
LanguageMiddleware, andpython -m plantcraft_sdk.i18n_check --strict(0.102.0 to 0.105.0). - The app template (StarterCraft 0.2.0) keeps only the basics in its code, with full guides and
recipes in itsdocs/.
Changed
-
Platform and apps: a new look for the whole suite (ui 0.102.0): page, card and text on one
vertical rhythm, the new PlantCraft mark as the favicon in every app, and a display size in
five steps, from Larger text to More space. -
Platform and apps: pages load less script.
- Plotly and the fonts download once for the whole suite, not once per app (ui 0.96.0).
- Pages fetch Plotly onl...
PlantCraft 0.5.1
PlantCraft 0.5.1 (2026-10-30)
Maintenance release after the 0.5.0 launch: new look, the plantcraft name everywhere, a clear contribution path, faster Historian reads, bounded View memory and a working install on macOS with Docker Desktop. No new apps.
Changed
- New look: the PC valve logo and Steel Blue. The mark is now the
letters P and C drawn as a control valve on a process line; the brand colour moves from
#0A66D0to Steel Blue#2E5A88(#3D78B6on dark), across buttons, links, selection, focus
and the sidebar accent. App tiles keep their hues, toned toward the new blue, with dark values
that hold on black. New app marks: Analytics (two connected blocks), Models (a distillation
column), Desk (console with the day's checklist), Alarms (a bell without the limit line),
Historian (the data cylinder without the time axis), Documents (a page without the
magnifier). The family image no longer shows the plant simulator, which is not an app. New
brand/folder with PNG logos and the LinkedIn avatar and banner.
Added
-
"Talk to us" in the public demo. A visit shows a small button at the bottom right on every
page, the apps included, that opens the site's contact form in a new tab
(PLANTCRAFT_PUBLIC_DEMO_CONTACT_URL, an https address;deploy/cloud/vm.shpoints it at
plantcraft.io/#talk, which opens the form at once). The page after a visit ends offers it
too. Without the setting nothing shows. -
The public demo starts the same for every visitor. A visit starts
with the newest 20 notifications unread and the rest read, whatever the live plant added, and
the seed leaves the newest message of 4 conversations unread (the Messages badge shows 4).
Opening a conversation is now the visit's own read mark, like every other mark since the
check and favorites: before, it marked the shared demo account, so the next visitor found
everything read. -
Tables and pages on a phone (fine pass 2026-10-28, ui 0.86.0). A table column can be left
out on a phone (DataColumn.phone: "hide"), cells are tighter there, a number and its unit
never split, and a header actions cell keeps the header's rule. Overview widgets stacked on a
phone are as tall as what they show (no empty desktop row under a short list;CardGrid
phoneHeight). Page tabs that scroll sideways fade at the edge with more tabs and keep the
active one in view. System health and Backups leave their secondary columns out on a phone; a
state, a size and a backup's name stay on one line. -
Check, seen or not, favorites. A notification gets a check (done) in its
row, in the quick-access panel, when it is open and on the page it opens (a small bar on the
action or the alarm); one button turns it seen and not seen again. A conversation and an
assistant chat can be marked unread until opened; the three take a star, favorites first.
PUT /api/me/marks. On the public demo each visit marks for itself only.
Changed
-
The platform's Python package is
plantcraft. The folder at the root of
the repository, every import andpython -m plantcraft.<module>follow the product's name, and
the distribution isplantcraft.processcraftstill runs: the console script, and inside the
imagespython -m processcraft.<module>(plantcraft/legacy.py), so an olderinstall.sh, a
rollback to 0.5.0 and the Backups page keep driving every version's installer image. Image and
service names, data folders, the app id, cookies and stored ids are unchanged. -
Contributing is one clear path. The README leads with how the suite fits together and how
to contribute (commands moved todocs/development.md). Issue forms for a bug, an idea, a
question and a change to the core;.github/labels.ymlis the label set a workflow keeps on the
repository;docs/maintainers.mdis how issues and pull requests are answered, and a Claude
Code skill (.claude/skills/review-contribution) checks whether a pull request fits. The
agent rules and module pages are written for any contributor: no dated notes, no internal
codes, and STATUS and devlog pages start fresh (the CHANGELOG is the record). -
Analytics says what it is: scheduled monitors, diagnostics and KPIs for signals and control
loops, in its card, the guides and the site. The name stays. -
Turn the assistant on in four steps.
docs/assistant.mdopens with them (it is off until an
administrator connects a model), linked from the README and the install guide; the public
demo's assistant says so too. -
Architecture in pictures and the zones of a contribution.
docs/architecture.mdopens with
the suite in one diagram, how the apps talk (one sticky valve) and what is core and what is open;
CONTRIBUTING.mdsays when an issue comes first. The Zones check (dev/pr-zones) shows a pull
request's zones on its run page and holds a fork's core change to an agreed issue.
Fixed
-
The Historian answers raw reads about four times faster. Writing the time of every value
one by one was 79 % of a raw read of millisecond samples (a day of the demo's 167 plant
signals: 3.1 s, now 0.85 s).plantcraft-sdkwrites sub-second times for a whole series at
once, with the same text as before. With ten methods reading at once, Analytics' round went
from 30 to 16 s. -
A new install on macOS or Windows starts. On a Docker Desktop bind mount several
PlantCraft processes wrote a new database at once and corrupted it ("database disk image is
malformed"): file locks do not hold between processes there. PlantCraft now runs one process
wherever SQLite cannot use WAL, whateverPLANTCRAFT_WORKERSsays. Building a release works
withpython3only (macOS) in every app's Makefile too. -
View no longer runs out of memory with many viewers. Its short cache kept expired
diagram data until it held 4,096 entries; on the public demo View reached its 2 GiB limit and
was stopped (2026-10-29). Expired data now leaves at once, and viewers asking for the same
data together share one read. -
The old name is gone from what people type and download. The command
isplantcraft(processcraftstill works), release files are
plantcraft-<version>-linux-<arch>-bundle.tar.gz,ssh-keygenchecks them as
release@plantcraft(same key; a signature does not carry the name), the update page names
every app with its current name (Analytics, View, Historian, Simulator) and the user guides say
Analytics and View where they still said Link and Dashboards. The site's chat knows the former
names. The Python package, images, services and data ids keepprocesscraft, so installations
and their data do not change. -
Running from source says which Python to use (#10). The demo from source is tested on
Python 3.12 and 3.13, and on 3.15 asyncua, the OPC UA library, fails to read, so signals show
no values. The README and the install guide now say 3.12 or 3.13 (the Simulator needs 3.12),
andrun_opcua_demo.pywarns on a newer Python. Thanks to the reporter of #10. -
Analytics rounds no longer fail with "PlantCraft could not be reached: timed out" (demo,
2026-10-29, SDK 0.78.0). A loop round reads a day of every loop's signals; while the
Historian served other long reads, the round's read waited its turn in the Historian's read
budget (up to 30 s) longer than the SDK client's 15 s, so the client asked again while its
first read still ran, three times, and the round failed (22 times in 8 h on the demo, "failed"
in the Queue). The SDK's series reads now wait 90 s (SERIES_TIMEOUT, other calls keep 15 s)
and PlantCraft waits 75 s for the Historian (HISTCRAFT_TIMEOUT_SECONDS, was 40 s), longer
than the budget wait plus the Historian's own 30 s read limit. -
Public demo, found by testing it online (2026-10-27). The maintenance and manager guides
opened "Report not found": the demo's visitor joins the Instrumentation and control group,
which View's Control loops folder is shown to. The demo's four calculated signals are made
again (their formulas named the Simulator's tags withoutCPM.Loops.). The visit's clock sits
at the bottom (it covered Desk's greeting, and on a phone the tour's buttons); the guide's
card no longer overflows with a long role name. An IPv6 visitor's places are counted by its
/64, not by each device. Desk's Unread messages shows a reference by its name, not
#[name](action:9). -
The site's chat gives OpenAI's reasoning models room to think (empty answers came back as
"cannot answer right now"), asks once more when the model fails, logs why, and answers
Portuguese in Brazilian Portuguese.
Downloads
| Bundle | For | Size | SHA-256 |
|---|---|---|---|
plantcraft-0.5.1-linux-amd64-bundle.tar.gz |
most servers | 2,028,246,573 bytes (1.9 GiB) | 23a020279a8fdbed35441626e412750663405f23bdee0ea955ed800b52586ba4 |
plantcraft-0.5.1-linux-arm64-bundle.tar.gz |
ARM servers and Apple silicon | 1,977,122,823 bytes (1.8 GiB) | 63c5eab031cce434ada7d69826c183678c8abe02dd51822161e26b788ce587ca |
Each bundle comes with its .sha256 and .sha256.sig. docs/install.md explains how to check them with ssh-keygen as release@plantcraft (OpenSSH fingerprint SHA256:9l/yDgR8/Pzel1pncxrXz2IGCIhFirKpDT78tO49fG8).
Upgrading
Run ./install.sh from the new bundle. Data, settings and accounts are kept, and the previous version stays available for rollback (./install.sh --rollback).
Each app's own changes are in its CHANGELOG.md under apps/.