# CLAUDE.md — wrps-demo-kit

Instructions for Claude Code working in this repository. Human-facing orientation
is in `README.md`; read that first, then this.

---

> [!IMPORTANT]
> ## 🪜 Working protocol — one small step at a time
>
> This project is built **incrementally, with every step tested and confirmed
> before the next one starts.** Do not batch phases or run ahead.
>
> 1. **State the plan** — what you are about to change, and the **acceptance test**
>    that will prove it works.
> 2. **Make the minimal change** for that step alone.
> 3. **Test it and show the evidence** — command output, a Modbus read, a
>    container log. Not a claim.
> 4. **Wait for the user to confirm** before starting the next step. Never proceed
>    on an untested or failing step; fix and re-test.
> 5. After each confirmed step, **commit** with a clear message.
>
> If a step is too big for one testable change, split it. When unsure whether
> something counts as done, ask.

---

## 1. What this is

A reusable method for building demo automation use cases on **Yokogawa CI Server**
(SCADA) and **OpenPLC Runtime v4** (soft PLC, in Docker), with the Waterloo Road
Pump Station as its worked example.

**The demo is live and in use.** A simulated three-pump station runs on the PLC and
is polled by CI Server. Treat everything on those machines as production.

## 2. The two machines

| Role | Host | Address |
|---|---|---|
| SCADA — CI Server R1.05 | `yau-poc-cicore1` | `10.0.0.21` |
| PLC — OpenPLC Runtime v4 container | `yau-sls-poc-lin001` | `10.0.0.17`, public `20.211.144.151` |

SSH alias `lin001` → `azureuser@20.211.144.151`. `10.0.0.17` is only routable from
inside the VNet or over the WireGuard VPN.

> [!WARNING]
> **`yau-sls-poc-lin001` is a shared, live host** running ~28 containers for
> several unrelated projects behind Caddy and Authelia, serving customer-facing
> demos. **This project owns exactly one container, `openplc-runtime`.**
>
> Never restart Caddy or Authelia. Never publish a port on `0.0.0.0` — the host has
> a public IP and no firewall (`ufw` inactive; the Azure NSG governs). Never put
> growing data on the 62 GB root disk.

Full detail: `02-environment/README.md`. The host owner's own brief, including
rules that exist because breaking them caused outages:
`02-environment/YAU_Linux_Host_Onboarding.md`.

## 3. Architecture invariant — do not violate

**OpenPLC Runtime is the Modbus TCP _server_ (slave) on port 502. CI Server is the
_client_ (master) and polls it.** Never the reverse. The PLC must never be
configured as a Modbus master.

Modbus unit / slave id: `1`.

## 4. The generation chain — never hand-edit downstream

```
03-plc/src/*.st  ─build.py─►  03-plc/register-map.csv  ─gen_scada_points.py─►
      04-scada/modbus_points/ci-server-points.csv  ─gen_ciserver_qli.py─►  *.qli
```

To change an address, type or scaling: edit the **`REGISTERS` table in
`03-plc/build.py`**, rebuild, then re-run both SCADA generators. Propagate all the
way or the two sides disagree silently — the reads still succeed, they are just
wrong.

Every generated file reproduces byte-identically. **If a regeneration produces a
diff you did not intend, stop and find out why** before committing it.

Hand-maintained exceptions, both documented in place:
`04-scada/modbus_points/historian/` and `03-plc/editor-devices/`.

## 5. Things that are true and non-obvious

Getting any of these wrong produces plausible-looking wrong answers rather than
errors.

- **`%MW` is not holding register 0.** `%QW` occupies HR 0–1023; `%MW` starts at
  **HR 1024**, so `%MW3` is HR 1027. A wrong-by-1024 read succeeds and returns
  nonsense.
- **`%QW17` (alarm bitmask) is unsigned.** Bit 15 does not fit a signed INT.
  `%QW7` is genuinely signed.
- **The deployed build is the SIMULATION build.** All `%IW`/`%IX` read 0 — correct,
  the mux takes simulated values. Do not poll FC04/FC02 expecting live data.
- **Port 502 answers only while a program is running** *and* the Editor project
  defines a Modbus Server. A refusal looks like a firewall drop and is not.
- **STruC++ ships only inside the OpenPLC Editor GUI.** There is no CLI. Any change
  to PLC logic goes through the Editor — and **the Editor is not currently
  installed anywhere** (`03-plc/DEPLOY.md` §0).
- **The runtime image was made with `docker commit`** and exists in no registry.
  Deleting it loses the compiled program.
- **Every CI Server item import renumbers every item**, killing every display. The
  fix is to **validate the display in CI Server's Editor Module** — one action,
  before any other diagnosis.
- **`PS_*` names the four Modbus poll groups and nothing else.** Placeholder `PS_`
  *tag* names were removed; they named nothing that exists and caused a real
  defect. CI Server items are `AID.WRPS.<SECTION>.<LEAF>`.
- **Live setpoint tuning is not in this repo.** Operators retune over Modbus; those
  values live only in the container volume and revert on restart.

## 6. Guardrails

**MAY do freely:** edit repo files, SSH to `lin001` for **read-only** inspection
(`docker ps`, `docker logs`, `docker inspect`, reading files), run the generators,
run `05-tests/` against a fake PLC or a live read.

**MUST ask first:**
- anything that stops, restarts, recreates or updates the `openplc-runtime`
  container — **it is live control and CI Server is polling it**
- anything touching another container on that host, or any host-level config
- changing the register map — it is a contract with a system you cannot see
- writing to `%MW` on the live PLC (setpoints and the command word are real
  commands; `%MW1` executes on a rising edge)
- pushing to `origin`, or any history rewrite

**MUST NOT:**
- advance past an untested or unconfirmed step
- commit secrets — `*.local.md`, `*.pem`, tokens
- edit `99-reference/`, or `03-plc/as-built/`
- hand-edit any generated file
- change CI Server configuration — the user does that; you prepare the files
- widen a port bind to `0.0.0.0`

## 7. Verification, not assertion

The user is an automation engineer auditing this work. **Show evidence.**

- Read the live PLC rather than inferring its state.
- After changing a generator, prove the output still reproduces.
- Say plainly when something is unverified. `yau-poc-cicore1` is **not reachable**
  from here — anything about CI Server's own state is belief, not fact, and must be
  labelled as such.
- Do not describe a step as done until its acceptance test has passed.

## 8. Key commands

```bash
# the PLC container (read-only)
ssh lin001 "docker ps --filter name=openplc-runtime"
ssh lin001 "docker logs --tail 50 openplc-runtime"

# poll the live PLC from inside its own network (10.0.0.17 is not routable here)
ssh lin001 "docker run --rm --network openplc-net python:3.12-alpine \
  sh -c 'pip install -q pymodbus && python -c \"...\"'"

# rebuild the chain
cd 03-plc              && python build.py --mode sim
cd 04-scada/modbus_points && python gen_scada_points.py && python gen_ciserver_qli.py
cd 04-scada/hmi        && python build_display.py

# verify the map (needs a route to 10.0.0.17, or use fake_plc.py locally)
cd 05-tests && python verify_modbus.py --host 10.0.0.17 --port 502 --unit 1
```

## 9. Conventions

- **ST style:** IEC keywords uppercase, 4-space indent, one action per line, a
  `(* ... *)` header on every POU. No vendor extensions.
- **Control and simulation never interleave.** Separate POUs; only `10_globals.st`
  and `50_prog_io_mux.st` may touch located variables, and `build.py` enforces it.
- **CI Server items:** `AID.WRPS.<SECTION>.<LEAF>`.
- **Commits:** one per confirmed step. Say what was proven, not just what changed.

## 10. Open items

Do not present these as solved. `README.md` carries the full list; the two that
most often catch people out:

1. **None of the 20 acceptance tests in `01-design/WRPS-CTL-003` have ever been
   run.** The program compiles and publishes correct registers; its control
   behaviour is unverified.
2. **There is no route to deploy a new program** until the OpenPLC Editor is
   installed somewhere that can reach `10.0.0.17:8443`.