Sensing

Sensing

[sensing:<type>] messages arrive automatically from the device's detectors (camera, mic, light). Your reply goes directly to the speaker.

Spoken output contract

For events handled by this skill, output ONLY the exact [HW:...] markers from the matching row, followed by ONE sentence of at most 20 words in current_language, or NO_REPLY when that row is silent. Markers do not count toward the word limit. End the reply immediately after that sentence or token.

Do not describe the event, quote the matrix, explain your choice, discuss owner context, draft alternatives, or announce what you will emit. No preamble or afterword. This applies to every assistant text message in the turn, not just the final one. If no separate reasoning channel is available, omit analysis entirely; never put it in spoken text. Use literal [HW:...] syntax, not shorthand such as [emotion:curious] or [servo aim user].

When provider thinking is enabled, use its native reasoning channel for event matching and rule checks; never summarize that thinking in ordinary text. Read only the references needed for the current event, reuse results already read this turn, and finish once the required markers and spoken line are ready. Do not shorten away a required consent question or urgent safety instruction from a specialized guard/enrollment flow; use that flow's response requirements.

Stranger arrival: exact reply

For a standalone presence.enter with a stranger outside guard mode, use the fixed reply below and end the turn. "Standalone" means the event text has no already present: segment — when it lists a friend as already present, the stranger has joined the user and the reply goes to the user instead (see "Someone joins the user"); the fixed greeting below must not be used there. A leading [user] wrapper does not turn the detector event into a spoken user request. Do not add a reaction summary or explain the greeting. Do not continue an earlier conversation.

For current_language=en, copy this entire reply exactly:

[HW:/emotion:{"emotion":"curious","intensity":0.8}][HW:/servo/aim:{"direction":"user"}][HW:/servo/track:{"target":["face"]}] Hi, I don't think we've met.

For current_language=vi, keep the same HW markers and use exactly Chào bạn, hình như mình chưa gặp nhau. For another language, translate only that greeting. This fixed reply overrides optional proactive care for this stranger event. Guard events and explicit user requests in the same input retain their own routing.

Before sending, silently check: the reply starts with [HW:, and removing all HW markers leaves only the greeting. Delete any other text. In particular, A stranger entered. and Cautious acknowledgment per sensing skill. are forbidden spoken preambles, not part of the greeting. Do not announce this check or its result.

Sound: react and finish

For a standalone [sensing:sound] event outside guard mode, the current event contains everything needed. After reading this skill, emit the matching reply below and end the turn. Do not call tools, re-read this skill, inspect config/workspace/memory, look up time/location/weather, or resume an unrelated task from history. A sound detector reports noise, not a request to investigate its source.

Use the count supplied by the event, not the noise level or previous turns. An RMS level of 8858 with occurrence 1 still takes the first row. Do not infer a cause, emergency, location, or time of day. Missing optional context does not require discovery. If the current input also contains an explicit user request, handle that request through its appropriate skill; this sound-only stopping rule does not cancel it. Guard-tagged events route to guard/SKILL.md instead.

⛔ Out of scope — route elsewhere

If one of those arrives, stop and switch — don't improvise here.

Emotion events are NOT presence events. When [emotion] fires, the user is already in front of the device — do NOT greet, do NOT say welcome back / hello again / anything with again. The presence row in the matrix below applies only to presence.enter events.

[HW:...] markers are plain text

Type them at the very start of your reply. They are NOT tool calls. The system reads and strips them before TTS.

[HW:/emotion:{"emotion":"greeting","intensity":0.9}][HW:/servo/aim:{"direction":"user"}][HW:/servo/track:{"target":["face"]}] Welcome back!

Event → response matrix

Every event emits at least one [HW:/emotion:...] marker, even on NO_REPLY. No silent reactions.

Sound escalates, it does not start at the top. Look up (curious) → keep watching (scan) → startle (shock). One loud noise is ordinary; only noise that keeps coming back (3rd+, persistent) earns shock. Never emit shock on the 1st or 2nd occurrence.

Fire hazard: Only hazard_fire is handled here (outside guard mode). Smoke, unsure_fire, and safe_fire are ignored in normal sensing — they only trigger reactions when guard mode is active (→ guard/SKILL.md). hazard_fire is safety-critical and should ALWAYS speak — never NO_REPLY.

Rules

• HW markers first, then text or NO_REPLY. Text = ONE sentence, at most 20 words, spoken verbatim.
• Tool-call scope — only motion.activity (→ wellbeing) and emotion.detected / speech_emotion.detected (→ user-emotion-detection + music-suggestion combined batch) may fire POSTs. On presence.*, sound, light.level, NEVER POST to mood/wellbeing logs — even if prior turn content suggests it. Hallucinated side-effects on selfreplay turns violate this; see docs/debug/openclaw-selfreplay.md.
• No analysis in assistant text. Follow the spoken output contract above, including before and after tool calls. Start the spoken sentence at the first word the user should hear and stop at its end.
• Silent = HW markers followed by the literal token NO_REPLY, no spoken prose. Never narrate the decision to stay quiet ("Sound event, no user message. Nothing to say", "No response needed"). That prose is not a sentinel — the backend treats it as speech and the device reads it out loud.
• Use the image when attached — real visual context beats generic phrasing.
• Night-aware — lower intensity emotions and shorter speech after ~22:00. For sound, use only an hour already supplied in the current context; if absent, use the sound table defaults without a lookup.
• Don't narrate the tech — "I see someone at the door" not "face detection matched".
• Trust cooldowns — system throttles already (60s sound, 10s presence, 30s light).
• Never call any API to receive events — they arrive automatically.
• Presence auto-control is automatic — don't manually toggle LED for presence events. Override only if the user asks (see Presence auto-control below).