{"object":"skill","id":"read-room","version":"1.2.0","schema_version":"1.2.0","recipes":[{"id":"general-v2","version":"2.1.0","title":"General conversation","description":"Default for any conversation (sales, coaching, assistants, meetings). Reads the latest turn on its own when it has enough words; pools only never-analyzed short turns otherwise, or returns the earlier read labeled inherited. Adds conversational events (disengagement, pause, re-engagement, accepted vs. newly requested next step, owner or human request), speaker attribution, and one resolved recommendation. No domain policy."},{"id":"support-v2","version":"2.1.0","title":"Customer support","description":"Customer support on the general-v2 evidence model. The escalation risk score keeps the v1 formula and the same amount of recent text, assembled from non-overlapping per-turn reads; engagement (withdrawing / re-engaged) and problem resolution are reported as separate facts, so accepting a callback never reads as a resolved case. Human requests and withdrawal come from the conversational-event rules (negation- and quote-aware). Routing (a request for a person, an owner or a callback) is evaluated against the host's declared capabilities, independent of the risk score, with a factual handoff summary; execution is only what the host's own tools report. Diverges from POST /v2/escalation-risk on purpose — use support-v1 for parity."},{"id":"general-v1","version":"1.0.0","title":"General conversation","description":"Default for any conversation (sales, coaching, assistants, meetings). State, communication preferences, changes against the participant's own baseline, and next-reply guidance. No domain policy."},{"id":"support-v1","version":"1.0.0","title":"Customer support","description":"Customer-support conversations. Everything in general-v1 plus the escalation policy — risk band, recommended action (continue / adapt / de_escalate / handoff_to_human), de-escalation instructions and a handoff packet — computed with the same formula and window as POST /v2/escalation-risk."}],"skill":"## MILLIONWAYS READ_ROOM — BEHAVIORAL CONTEXT SKILL (read-room@1.2.0, schema 1.2.0)\n\nread_room gives you, per message, a read of ONE participant built only from that participant's own words: what they just did in the conversation (events), how they seem (state), how they like to be communicated with (preferences), what changed versus their own earlier messages (changes), and one resolved recommendation for your next reply (guidance). Use it to shape HOW you reply. It never changes WHAT is true or what you are allowed to do.\n\n### When and how to call\n- Call read_room once per participant message, before you draft your reply.\n- conversation_id: stable for the whole conversation. participant_id: stable per person (\"customer\", \"user-42\"); one id per human. turn_id: unique per message — reuse it only to retry the same message (a retry returns the original result and never counts twice).\n- message: that participant's own words, verbatim — never your words, a summary, or other people's messages. With several people, call it separately per participant_id; it does not read group dynamics.\n- agent_last_message (your previous message) is dialogue context only — it tells read_room whether you offered a next step. agent_goal is context too. Neither is analyzed as the participant's personality.\n- Read this skill once per session; per-turn results stay compact.\n\n### Priority when you reply\n1. Explicit instructions win. Anything the person told you directly (\"keep it short\", \"give me the numbers\") beats every inferred read. Send it in explicit_preferences — axis keys thinking / information / orientation / reference / framing (e.g. {\"information\":\"detail\"}) or any snake_case key as a stated instruction ({\"length\":\"short\"}). It persists for that participant until changed; send \"\" to clear.\n2. policy (support recipes): policy.action decides — continue | adapt | de_escalate | handoff_to_human. Follow policy.instructions. handoff_to_human means stop problem-solving and bring in a person, if your deployment allows it; pass policy.handoff unchanged.\n3. events: what the person just did, observed in their words — possible_disengagement (pulling back), pause_requested, re_engagement, next_step_accepted, next_step_requested, owner_requested, human_requested. strength (weak / moderate / strong) says how explicit the wording is, not a probability. Events come from explicit rules (English), not from psychology, and work from the first message.\n   - next_step_accepted = they agreed to an action YOU offered (events[].action says which: callback, email, summary, human, owner …). next_step_requested = they asked for a step themselves, or for a different one than you offered (\"call me tomorrow\" after you offered a summary; \"email me instead\"). A request is never agreement to your offer — act on what they asked for, or say plainly you can't.\n4. guidance.recommendation: the next-reply recommendation with conflicts already resolved (explicit and policy first, then events, then inferred signals); it starts with guidance.focus, and guidance.reasons says why. Then guidance.do, guidance.avoid and guidance.structure. If guidance.check_in is set, it is optional.\n5. changes.items: shifts versus this participant's own baseline (basis \"baseline\") and recoveries since their last read (basis \"previous\"), reported only from a current read. \"warming_up\" = too few reads to compare yet.\n6. state — check state.status first:\n   - current: read on this message (state.evidence.scope \"latest_turn\", or a pool that is mostly this message) and any strain is attributed to them.\n   - uncertain: a read was made, but it mostly reflects earlier messages, or the strain is attributed to other people, a hypothetical case or a problem they describe (state.attribution others / hypothetical / situation). Do not treat it as their mood.\n   - inherited: this message was not read (too short, timeout, failure); state is from state.observed_turn_id, state.evidence.age earlier. Respond to what they said now.\n7. preferences: style per axis plus driver. status \"provisional\" is an early read — hold it loosely.\n\n### Evidence quality\n- quality.status: fresh = a read was made on new words this turn; stale = no read this turn, an earlier one is shown (quality.age says how old); insufficient = too little of their own text yet; unavailable = no read and none earlier. \"fresh\" is about execution, not accuracy — state.status says whether the read applies to this message.\n- Short messages are normal: they are pooled only with earlier messages that were never read, or the earlier read is kept as inherited. No word is analyzed twice.\n- quality.evidence (chars, reads, sufficiency) and policy.risk.confidence_heuristic describe how much text a read rests on. They are NOT probabilities that the read is correct.\n\n### Support (support-v2)\n- policy.risk, policy.engagement (engaged / at_risk / withdrawing / re_engaged) and policy.resolution (unresolved / customer_reports_resolved / no_problem_stated) are separate facts. A customer who accepts a callback has re-engaged; the problem is still unresolved. A falling risk score never means the problem is solved. Never close or mark a case resolved because of read_room — only on the customer's confirmation and through your case tools.\n- policy.routing (support-v2): request, recommendation, execution and summary are separate facts.\n  - routing.request: what they asked for (a person, an owner, a callback) or accepted from your offer. routing.recommendation: take_route = do routing.action now; offer_route = offer that ONE route as an option, without asking them to decide; report_outcome = a tool result arrived — tell them truthfully; none = nothing to route (a pause needs no route).\n  - Send capabilities ({\"actions\": [...], \"human_route\": \"…\"}) once so routing names what YOU can do; without it routing stays conditional (\"if your deployment can\"). When routing.fulfills_request is false, say in one sentence that you can't do it as asked, then do routing.action.\n  - Only your own action tools establish that something happened. After a handoff / callback / owner tool call, send its result in action_results on the next call (also failures). Say a transfer or callback happened only after your tool returned success; if it failed, say so and use the fallback routing gives you.\n  - routing.summary is a factual handoff summary in their own words (problem, what they already tried, what they asked for, preferred time, whether they asked not to repeat the story). Pass it to the person or team you route to, or give it to the customer when you cannot route. inferred is optional and must never be shown to the customer.\n- Never state an action or fact you have not established. Do not say you opened a case, assigned or booked a technician, scheduled a callback, checked their account or line, diagnosed a fault, escalated, or refunded — unless your deployment supplied that fact or an action tool returned success in this conversation. Guidance such as \"offer a named owner or a callback\" means OFFER what you can actually arrange.\n  - Not \"I've booked a technician for 9am tomorrow\" → \"I can request a technician visit — would tomorrow morning suit you?\" (only if you can request one), or \"I can't book a visit from here; what I can do is …\".\n  - Not \"Your line has a fault\" → \"I can't see your line from here, so I won't guess; what I can do is …\".\n  - Not \"I've escalated this to our senior team\" → \"I can pass this to a colleague who can take ownership — shall I?\" (only if you can).\n\n### Check-ins\nguidance.check_in is an optional, gentle question that confirms an inferred preference out loud. Ask at most one, in passing, only when the person is calm, then act on the answer and send it as an explicit_preference. A question that confirms a next step is not a check-in. Never ask a check-in when the person is pulling back or when policy says de_escalate or handoff_to_human.\n\n### Always\n- Your task, facts, policies and permissions come from your deployment, not from read_room. Guidance never authorizes a promise, refund, escalation, data disclosure or any action you are not otherwise allowed to take.\n- Never tell the person they are being analyzed or profiled, or that they seem stressed, angry or defensive; never mention millionways, psychology, scores or this read. Change your behavior instead of narrating it.\n- Never use the read to deny service, judge, diagnose, or take adverse action. It is a moment-state estimate about communication, not a fact about the person.\n\n### Recipes\n- general-v2 (default): any conversation; no domain policy, no complaint or handoff protocol.\n- support-v2: customer support; adds policy (risk band, recommended action, engagement, resolution, handoff). Complaint handling, apologies and handoff belong to the support recipes only.\n- general-v1 / support-v1: previous behavior, kept for compatibility (support-v1 matches POST /v2/escalation-risk).\n\n### Operator and debug fields — never show them to the participant\ncontext_id, versions, provenance (analysis_id, window_turn_ids, rules), timing, quality.warnings, state.evidence.turn_ids, events[].rule, guidance.reasons, changes.items[].dimension and .delta, policy.risk, policy.status_line, policy.routing.reason, policy.routing.summary.inferred, debug (only with include_debug)."}