Node Red

Node-RED on vulcan

Where things live

How to edit flows — Admin API first, always

Preferred: Admin API (PUT /flow/<tab-id>). Live reload, no restart, no editor disconnect, surgical (only the named tab changes). This is the default path for any edit John asks for.

# Always read the token via shell substitution; never echo it.
TOKEN=$(cat /run/secrets/node-red-admin-token)
NR=http://localhost:1880

# Read all flows (returns array of nodes including tab/subflow definitions)
curl -sS -H "Authorization: Bearer $TOKEN" $NR/flows

# Read one tab (returns {id, label, nodes:[...], configs:[...], info, env})
curl -sS -H "Authorization: Bearer $TOKEN" $NR/flow/<tab-id>

# Replace one tab — surgical, leaves all other tabs untouched
curl -sS -X PUT -H "Authorization: Bearer $TOKEN" \
     -H "Content-Type: application/json" \
     -d @updated-tab.json $NR/flow/<tab-id>

# Create a new tab (returns the new id)
curl -sS -X POST -H "Authorization: Bearer $TOKEN" \
     -H "Content-Type: application/json" \
     -d @new-tab.json $NR/flow

# Delete a tab
curl -sS -X DELETE -H "Authorization: Bearer $TOKEN" $NR/flow/<tab-id>

# Replace the entire flows array (DESTRUCTIVE — use only when unavoidable)
curl -sS -X POST -H "Authorization: Bearer $TOKEN" \
     -H "Content-Type: application/json" \
     -H "Node-RED-Deployment-Type: full" \
     -d @flows.json $NR/flows

Layout-preserving edits: GET /flow/<id> → modify only the field that's wrong on the specific node(s) by id → PUT /flow/<id> with the same node array. Node coordinates, wires, and IDs are preserved because you sent them back unchanged.

Token rules:

• Always read via $(cat /run/secrets/node-red-admin-token) or assign to a shell variable. Never echo, log, or pass the token to a tool that surfaces output in this conversation. If a curl command would dump headers, redirect them to a tempfile.
• The token is owned by johnw:users mode 0400 — no sudo needed.

Fallback paths (in order, only when the API can't help):

1. Editor UI — for one-off edits John wants to make himself, or when a generated flow benefits from human review before publish. Output the JSON to ~/*.json, suggest Menu → Import → Select a file → Deploy.
2. Direct flows.json edit + restart — only if Node-RED is down or auth is broken. Procedure: backup → sudo read/edit → validate_flow.py → sudo chown node-red:node-red → sudo systemctl restart node-red. ~6 s editor disconnect.

Don't write to /var/lib/node-red/flows.json directly when the API is reachable. Don't ask the user to re-import a tab when a PUT /flow/<id> would do the same job without losing the layout.

House style — match this

Read every "House style" point below before producing a flow. Most of John's prior corrections trace to one of these.

Time triggers

• Use chronos-scheduler (not stock inject with cron).
• Crontab values are 6-field CronosJS: 0 0 23 * * 2,4,6 = sec min hour dom mon dow. Day-of-week list goes in the last field.
• Sun-relative: type:"sun", value:"sunsetStart"|"goldenHour"|"night"|..., plus a random offset (15–240 min) to spread fires.
• Configs share f1c80506d19d3de2 (location node).

HA service calls (api-call-service)

• entityId is ALWAYS the array field. Single: ["switch.x"]. Multi: ["climate.a","climate.b"]. Script/scene: [].
• dataType is ALWAYS "jsonata". Never "json".
• data is either "" (no extra payload) or compact JSONata: {"preset_mode": "eco"}, {"temperature": $env("Temperature")}, TTS like {"cache": true, "media_player_entity_id": "media_player.vlc_telnet", "message": '...' & $string(...) & '...'}.

State gates (api-current-state with halt_if)

• Default has 2 outputs. Output 0 fires when state matches halt_if; output 1 fires when it does NOT match.
• John writes gates as questions: anyone home?, office door closed?, john home?. The question's "yes" answer routes to output 0; "no" to output 1.
• Wire ONE output to the continuation; leave the other empty. Pick which output based on plain-English intent.
• For comparisons, JSONata halt is supported: halt_if_type:"jsonata", halt_if:"3*24*60*60", halt_if_compare:"gt".
• DO NOT GUESS the direction. Always read the existing wires for context. The same halt_if string is used both ways in this codebase.

Naming

• Triggers carry their for: duration in the name: mac inactive 15min, TV on 2min, Nasim leaves 15min, out of office 15min.
• Gates are lowercase questions ending in ?: anyone home?, office door closed?, rain delay?, vacuum cleaning?.
• Actions are imperatives or device-verb-param: Turn off HVAC, purifier on, upstairs heat_cool 78-82, bedroom heat off, tv_room set 78 heat.
• Inject buttons: time-shaped (06:00 daily, Shut-off 23:15) or state-shaped (Lockdown, Turn on).
• Schedulers: descriptive — 12:00-15:00, ~Golden Hour till ~11 PM, Program A 23:00, Pool ON 09:00.

Layout

• Vertical bands per logical section, stacked top-to-bottom with ~100–220 px gaps.
• Comment-as-header anchors each band at x ≈ 150–200, y = first row of the band.
• Flow goes left-to-right within each band; comment uses sentence-headline style with em-dashes/ellipses: When I leave the computer…, Pre-cool upstairs for Institute Nights, B-Hyve Program A — Sac County Odd Addr (Tu/Th/Sa).

Subflow status output

Wire your "success" branch through a small function that emits msg.payload = {fill, shape, text} to the subflow's status port:

const stamp = new Date().toLocaleString('en-US', {
  month: 'short', day: 'numeric', hour: 'numeric', minute: '2-digit', hour12: true
});
msg.payload = { fill: 'green', shape: 'dot',
                text: `${env.get('Action')} called : ${stamp}` };
return msg;

The runtime TZ is local, so no offset to hardcode. Pattern in production: subflow Act until observed.

Top pitfalls (each has bitten us)

1. api-current-state output direction. Output 0 = match, output 1 = no-match. Same halt_if is wired both ways in different parts of the codebase. Always check existing wiring, never assume from the name alone. (See Office HVAC misfire — office door closed? halt_if="off" wired to output 0 means "fire when door IS closed".)
2. chronos-repeat JSONata = milliseconds, not seconds. Returning 5 is 5 ms. Use $number($env("Repeat")) * 1000 for seconds.
3. chronos-repeat "env" type doesn't read subflow env vars. In a subflow context, switch the interval type to jsonata and use $env("VarName").
4. server-state-changed for: N is HA-side: entity must STAY in matching state. Flickery sensors reset the dwell timer continuously. binary_sensor.johns_mac_studio_active is unreliable for presence — use sensor.johns_mac_studio_active_camera or _audio_output ≠ Inactive instead.
5. join-wait reset semantics. msg.reset = true silently drains the queue; msg.complete drains to the expired output. Don't conflate.
6. Manage Palette can delete Nix-overlay plugins during a Node-RED package version bump. If node-red-event-logger disappears, sudo systemctl restart node-red-event-logger-install re-installs it.
7. node-red postgres role is INSERT-only on msg_events/audit_events. Reads require sudo -u postgres psql -d nodered_events. Grafana queries fine.
8. msg.payload truncation in event log — 4096 UTF-8 bytes max. Large payloads stored as {"_truncated": true, "preview": "..."} with payload_size recording the original byte count.
9. CronosJS cron is 6-field, not 5-field. First field is seconds. 0 0 23 * * 2,4,6 not 0 23 * * 2,4,6.
10. server-state-changed v6 uses entities: {entity: [...], substring: [...], regex: [...]}, NOT the flat entityId/entityIdType from older versions. Wrong schema → TypeError: Cannot read properties of undefined (reading 'entity') on startup, six errors for six nodes, etc. Always use the nested form when emitting JSON for v6.
11. api-call-service v7 needs action: "<domain>.<service>" in addition to the legacy domain/service fields, plus floorId: [], labelId: [], and blockInputOverrides. Omitting any of these makes the editor flag the node as invalid (red triangle) even though the runtime might still execute it. Reference example: the user's working 09238a6ff00540ec node.
12. api-current-state outputProperties valueTypes that are actually valid: entityState, entityId, jsonata, str, num, bool, flow, global, msg, env, date, bin, eventData. The string entity is NOT a valid valueType — use jsonata with $entity().attributes.<key> to get attributes. Also include override_topic: false (working nodes always have it).
13. Never echo or log the Admin API token. When using /run/secrets/node-red-admin-token, wrap it in $(cat …) or assign to a shell variable that's only consumed by curl. If you need to see whether the token works, check the curl HTTP code (-w "%{http_code}") and response length — never the request headers.
14. Palette/API installs need bash in the service PATH. Many npm packages (e.g. core-js) have postinstall scripts that spawn sh. The default node-red.service PATH on this host (nodejs, gcc-wrapper, coreutils, findutils, grep, sed, systemd) has no shell — installs ENOENT with npm error syscall spawn sh. Fixed by systemd.services.node-red.path = [ pkgs.bash ]; in modules/services/node-red.nix. Anytime an install fails with "spawn sh ENOENT", verify the service path still has bash.

Debugging workflow

Event log captures onSend and onComplete for every node into Postgres. Primary UI: Grafana → Node-RED Events dashboard. SQL backup if Grafana is offline.

"X didn't fire":

SELECT ts, msgid, topic, payload FROM msg_events
WHERE node_id = '<trigger-uuid>' AND hook = 'onSend'
  AND ts > now() - INTERVAL '24 hours'
ORDER BY ts;

Zero rows → upstream issue. Rows present → drill in via msgid.

"X fired when it shouldn't":

1. Find the actuator's onSend in Grafana panel "All events" (filter node_name, hook=onSend).
2. Copy the msgid → dashboard variable $msgid.
3. Read the trace panel top-to-bottom — first row is the trigger, each subsequent onSend is a hop. Find where a predicate wrongly evaluated true and inspect payload at that hop.

Full schema, retention rules, and more queries: references/event_logging.md.

Plugin field guide

20+ contrib plugins installed. Used heavily:

• node-red-contrib-home-assistant-websocket — main driver.
• node-red-contrib-chronos — every timer / sun trigger / "act until observed" loop.
• node-red-contrib-join-wait — multi-input debounce (canonical: Office confirmed absent).
• node-red-contrib-postgresql — used internally by the event logger.
• node-red-contrib-actionflows — only in the Act until observed subflow.
• node-red-debugger — plugin (sidebar), not nodes; off by default.

Lesser-used: collector, bool-gate, boolean-logic-ultimate, pid-controller-isa, prometheus-exporter, simple-gate, threshold-control, openai-api, email, ping, prowl, introspection.

Per-plugin pitfalls + idiomatic usage: references/plugins.md.

Domain entities (HA)

Quick recall list — full catalog and tab UUIDs are in references/patterns.md.

• Climates (Nest): climate.{upstairs,guest_bedroom,home_office,living_room,tv_room,master_bedroom}
• Pool (IntelliCenter): switch.{pool,spa_waterfall,spa,jets}, water_heater.{pool,spa}, sensor.{water_sensor_1,solar_sensor_1,vsf_rpm,vsf_gpm}, binary_sensor.{pool_schedule,spa_waterfall_schedule}
• Sprinklers (B-Hyve): zones via switch.sprinkler_control_<zone>_smart_watering (call bhyve.start_watering with minutes), rain delay switch.sprinkler_control_rain_delay
• Presence: person.john_wiegley (home/not_home), binary_sensor.office_door_sensor_p2_office_door (Matter: on=open, off=closed)
• Mac activity: prefer sensor.johns_mac_studio_active_camera/_audio_output over binary_sensor.johns_mac_studio_active.

When to load references

• Plugin gotcha or unsure of node config → references/plugins.md
• Event-log query or Grafana panel → references/event_logging.md
• Reproducing John's wiring style on a new tab → references/patterns.md
• Function node code patterns → references/function_snippets.md
• Admin API or generic node schema lookup → references/api_reference.md, references/node_schemas.md

Available scripts

• scripts/generate_uuid.py [count] — Node-RED 16-char hex UUIDs
• scripts/validate_flow.py <file> — JSON + wire integrity
• scripts/wire_nodes.py <file> <src> <tgt> [output] — programmatic wiring
• scripts/create_flow_template.py <type> [out] — generic boilerplate (mqtt/http-api/data-pipeline/error-handler). These are not in John's style — use as scaffolding only.

Things to avoid offering

• Don't use Manage Palette to install a new plugin permanently — Nix overlay is the right vehicle.
• Don't suggest ~/.node-red/ paths; those don't exist on this host.
• Don't write to flows.json directly when the API is reachable — use PUT /flow/<id> for surgical, layout-preserving updates.
• Don't ask the user to re-import a tab to apply a small fix — fetch with GET /flow/<id>, patch, PUT /flow/<id>. Same end state, no manual work.
• Don't propose mocking the event-logger DB in tests — use real Postgres (CLAUDE.md rule).
• Don't fabricate entity IDs — verify against /var/lib/hass/.storage/core.entity_registry (jq filtered by platform).