twincat-scope

An agent skill for TwinCAT 3 Scope: build scope projects that actually record, and read the recordings they produce without drowning in samples. Built and tested with Claude Code; the same files follow the skill layout GitHub Copilot in VS Code reads, which is untested.

Companion skill: twincat-st writes and reviews the Structured Text. This one measures what that code does on the machine. Install either alone.

Reference material only. Not for safety functions. Scope configurations, recording plans and diagnoses produced with this skill must be reviewed and validated by a qualified engineer against the drive and platform documentation before anyone acts on them.

Not affiliated with or endorsed by Beckhoff Automation GmbH & Co. KG. "TwinCAT" and "Beckhoff" are trademarks of Beckhoff Automation GmbH & Co. KG, used here nominatively to describe what this skill works with. "EtherCAT" is a registered trademark and patented technology, licensed by Beckhoff Automation GmbH, Germany.

Why

Two things go wrong when you hand a model a scope problem without help.

The configuration looks right and records nothing. NC axis symbols live on port 501, not 851. Scope wants REAL64, not LREAL, and reads the IEC name as VOID. A display channel wired to a stale GUID opens perfectly and plots an empty chart. None of this is guessable, and a model without the skill confidently calls such a file "sound".

The recording is too big to read. Ten minutes of twenty channels at 1 kHz is twelve million samples, and the fault is usually three of them. Decimate it for a plot and a 3-sample spike shows up about one time in seven. Read a multi-rate export as one table and cause and effect can come out reversed.

This skill gives the agent the Scope-specific facts and a small tool that answers questions about a recording in a few hundred bytes or one picture, instead of returning samples.

What it has shown so far. In the latest blinded evals (evals/results-iteration-6.md: seven evals, three runs each, the same model with and without the skill) answers with the skill scored 35.0 of 38 against 24.7 without, and an independent judge ranked the skill ahead on six of seven. On the seventh, a scaled CSV, the skill's own wording produced a confident wrong answer; after that wording was fixed, a re-run of that eval was ranked ahead too. On a planted-fault recording both scored full marks. The skill does not save tokens; it used about 18% more. On real machines, files it generates open and record unedited, and checkscope --tmc caught every planted symbol error against a real compiled program (see Status).

Features

Build and check a recording

• newscope — writes a .tcscopex from a template for the symbols you name: fresh GUIDs with every display channel still wired to its acquisition, Scope's own data types, NC axis symbols routed to port 501, sample and record time set.
• Readable layout by default — one chart tab per device, one stacked band per quantity, at most eight traces per band, so a following error of microns is not flattened by a position of a metre. Dark or light charts, following the TwinCAT IDE's theme by default.
• checkscope — catches the files that open fine and record nothing: unwired or dangling channels, duplicate GUIDs, wrong ports, IEC type names, size/type mismatches, placeholder names and NetIDs. Also reads the project stored inside a .svdx.
• Recording-plan warnings — sample load against the target, a fixed window that will probably miss an intermittent fault, ring-buffer and trigger behaviour.
• --tmc — checks every PLC symbol against the compiled program: typos, renamed variables, types read at the wrong width.
• No dependencies — doctor, newscope and checkscope run on plain Python, so they work on a locked-down engineering PC.

Read a recording

• ingest — converts a .svdx (through Beckhoff's TC3ScopeExportTool.exe) or a Scope CSV to Parquet once, so every later question takes seconds instead of minutes.
• manifest — channels, units, sample rates, duration, gaps, and whether timing across acquisition groups can be trusted at all.
• stats — per-channel health: time pinned at a rail, flat stretches, quantisation, outlier-heavy distributions.
• events — steps, spikes, ramps, flatlines, holds, wraps, clipping, standing following errors, digital transitions and threshold crossings, ranked worst-first and spread across the recording. Tells a setpoint at rest from a frozen sensor.
• plot — a PNG drawn as a min/max envelope per pixel, so a 3-sample spike is always visible.
• correlate — which channel moved first, with the lag and its sign spelled out; refuses channels on different clocks unless you ask it to resample.
• window — the actual numbers for a short time range, capped so a broad question cannot flood the conversation.
• Real export formats — TAB, European ;/, and ,/. dialects, multi-rate groups, repeat-padded or truncated slow groups, channels exported several times. Layouts it cannot read correctly are refused by name rather than misread.

Guidance for the agent

• A diagnosis method that walks from summary to raw rows, with how to read each result.
• How to size a recording so it does not disturb the machine it is measuring.
• Four hard rules: no safety logic, no writes to a live machine, no unverified claims, and a human starts every recording.

Install

# Claude Code
git submodule add https://github.com/SionVerhoef/twincat-scope .claude/skills/twincat-scope

# GitHub Copilot in VS Code (untested)
git submodule add https://github.com/SionVerhoef/twincat-scope .github/skills/twincat-scope

Pin a release rather than following main, and commit the pin:

git -C .claude/skills/twincat-scope checkout v1.1.2
git add .claude/skills/twincat-scope

A plain git clone of your repository leaves a submodule folder empty. Everyone who clones afterwards needs git clone --recurse-submodules <your-repo>, or git submodule update --init in an existing clone.

If your team would rather not use submodules, tools/update-skill.ps1 downloads a release zip instead (-Version v1.1.2, or the latest release by default). Its default target is the Copilot folder, so for Claude Code pass -Target .claude\skills\twincat-scope. A copy of the script saved from a browser is blocked by the usual RemoteSigned policy until you unblock it:

Unblock-File .\update-skill.ps1
.\update-skill.ps1 -Target .claude\skills\twincat-scope

Requirements

uv. It fetches Python and the analysis dependencies on first run, needs no administrator rights, and installs into your user profile:

winget install --id=astral-sh.uv -e

Behind a corporate proxy:

$env:UV_NATIVE_TLS = "true"                        # use the Windows certificate store
$env:UV_DEFAULT_INDEX = "https://nexus.example/repository/pypi/simple"

Without UV_NATIVE_TLS, an intercepting proxy breaks TLS with an opaque error.

doctor tells you what is missing and how to fix it:

py -3 scripts/tcscope.py doctor        # python3 on Linux or macOS

doctor, newscope and checkscope need no packages, but they do need Python 3.11 or newer itself. A fresh Windows PC may have none: python.exe there can be a 0-byte Microsoft Store stub, and py missing. Install it for your user only, no administrator rights needed:

winget install --id=Python.Python.3.12 -e --scope user

or run the python.org installer with Install for all users unticked.

Layout

twincat-scope/
├── SKILL.md                        entry point — rules, workflow, routing
├── references/
│   ├── data-triage.md              the method for reading a recording
│   ├── recording-load.md           sizing a recording safely
│   ├── scope-configuration.md      .tcscopex schema, ports, types, layout
│   └── export-tool.md              TC3ScopeExportTool.exe and the CSV traps
├── scripts/tcscope.py              every verb
├── templates/                      known-good .tcscopex
├── examples/                       real recordings — empty by design
├── tests/                          synthetic fixtures with planted defects
├── evals/                          skill evals and field-test write-ups
└── tools/update-skill.ps1          zip install, for teams avoiding submodules

Status

There is no TwinCAT installation where this skill is developed. Everything below was checked on real machines in field sessions, written up in evals/field-review-*.md.

Verified

• Files generated by newscope, unedited, open and record in Scope View: NC axis channels on 501 and PLC BIT/INT16/REAL64 channels on 851, several tabs, both themes.
• The whole path once end to end: generate, record with a trigger, convert the .svdx with the real export tool, ingest, manifest.
• The CSV reader against 19 genuine export-tool CSVs and every CSV option in Scope View's export dialog. Each layout either reads correctly or is refused by name.
• checkscope against 25 real project files, and eight of Scope View's eleven trigger actions. A ring-buffer recording keeps exactly its record time, ending at the stop.
• Both templates, loaded as shipped: they open with the documented layout, and refuse to record until the NetId and symbols are set, as intended.
• --tmc against a real 862-symbol .tmc: known-good symbols resolve, a misspelled symbol and a wrongly-sized type are both caught.
• NC channels by every path newscope types — direct axis fields and ToPlc/FromPlc members, as UINT32, UINT16, INT32 and REAL64 — record on a moving axis and keep their types.
• Timelines All on a 10-channel group reads correctly.
• events on a real 600 s, 33-channel recording, over several rounds of fixes.
• An axis parked at a software limit or on an end stop gives no clipping; on the end stop events reports one standing following error and leaves a real settling tail out.
• The test suite on Windows and Linux.

Not yet verified

• The analysis verbs across a wide variety of real recordings — most real data so far comes from one machine.
• A genuine saturation, and marker tables in an export.
• What a Subsave trigger records: it needs a Scope View Professional licence, which no test machine had.
• A Scope View CSV exported with Scale values on cannot be told from a raw one: manifest warns that its values may be scaled, and with a Name-only header cannot see the scaling at all. ingest on the .svdx reads raw values.
• Installation through GitHub Copilot in VS Code.

SKILL.md rule 3 tells the agent never to claim something is verified when it is not. The same applies to this list.

The most useful contribution is a redacted recording of a known fault — see examples/README.md — or a run of evals/field-test-brief.md on a machine with TwinCAT.

Tests

py -3 tests/make_fixture.py && py -3 tests/make_real_fixtures.py
py -3 tests/test_verbs.py              # end-to-end checks of every verb
py -3 evals/test_grader.py             # the eval grader against known answers
py -3 evals/test_stage_runs.py         # the eval staging harness

These are plain scripts, not pytest suites. evals/README.md explains how the skill itself is evaluated against a no-skill baseline.

Licence

MIT — see LICENSE.