Host moves — the build plan

Host-moves build plan · created 2026-07-20 · eleven owner-triggered steps; the furniture session runs parallel (furniture.html); the study is host-moves.html. Step 1 implemented + self-tested 2026-07-20 (v536): the roster line teaches @{u:N}, lint_mentions guards output (mention_ok/mention_bad in parse health), chips render in host bubbles. Owner-tested same day — worked first turn (both hosts @-mentioned both humans on request); v537 polish from the test: the self-mention pill (a mention of the viewer fills coral-soft — your own name jumps out) and a roster-line line stopping the "@Dan Dan" name-repeat. Box deploy batched. Found in passing: the panel ALREADY receives a per-turn human roster in multi-human rooms — the study §1's "cannot enumerate" claim holds only for solo-human rooms. Step 3 implemented + self-tested 2026-07-20 (v539); owner's first live room (MrBeast, deepseek-v4-pro) surfaced the step's real lesson: never strip working grammar from the panel's own memory — v539 stripped the <roll> tags from hist, so the model saw results "just happening," learned "I say rolling and dice appear," and by turn 4 narrated intent with no tag. Fixed same day: hist keeps the tags (clients still never see them), roll events land after the turn's events (replay order = live order), and the pending note re-teaches the grammar. In-context evidence beats any SP instruction — the model's own transcript is its strongest teacher. Owner's Plato-room test (five clean rolls, correct parity narration) surfaced the timing critique: the card was welded to the reply that asked — ask and answer as one staged beat, sometimes a bare card with no speech at all. v540 polish: the tumble beat (a live card lands mid-tumble, resolves in place ~1.4s later; replay resolves instantly) + the SP rule never roll silently. The full "who rolls, when" answer is step 4 — the handed roll, where the player's tap decides the moment. Step 4 implemented + self-tested 2026-07-20 (v541), owner-directed widening: THE TABLE — the room's die became live state in a fixed dock (always visible while dice are in play; pending「点一下掷出 / 等 X 掷」→ tumble in place → the latest face), the stream demoted to compact 🎲 record capsules, placement rule now three surfaces (stream=records · table=state · composer=your action). Mechanics: roll_offer/roll_result event pair + dice_state() projection (rides replay, cold-open safe, v539 rolls read as idle faces); <roll for="uN"/> hands the die — the addressee's tap resolves it (auth'd, current-die-only) and the tap is the human input: it queues the narration turn and wakes the worker; the host's own rolls auto-resolve behind a 1.6s beat. Found: only speak events minted lids — the dice pair joined the sequence. v542 (owner test): the dock painted under the bubbles and wasn't tappable — content-under-bars slides the stream's box over anything unpositioned in that band; the table joined the z-ladder (relative, z 20 — topbar 36 · composer 30 · table 20). Step 4.5 (v543, owner-directed redesign after live play): armed single-cast — the die is armed for exactly ONE throw (for="uN" exclusive · for="any" the room, first throw wins · no for = the host's own) so re-roll cheating is structurally impossible; the table went read-only (state + attribution:「2d6 · 5+6 · Dan 掷」) and the hand became the only cast control — the die in your hand, fixed above the composer at thumb reach, tap → tumble → fly up to the table; the patient tap — a human's cast no longer summons the host: the narration cue waits (~8s, extends while typing) so roll-then-speak batches into ONE turn; continuation turns give the host speak-roll-speak on its own casts (loop-guarded: a roll-cue turn cannot cast again); <roll away/> clears the table, bare <roll/> re-casts. Trap hit: the measured --composer-h var ran stale (86px vs a real 138px) — the v488 lesson again; the hand anchors on the composer's live rect instead. Owner's Mark Twain session (v543): mechanics held ("when it works it feels great"); v544 polish — the table's attribution wears the roller's avatar + colour (by_slug rides the events). The remaining wobble — "who has the dice" wrong at times — is the model's game-state tracking (whose-turn + parity slips, self-aware: 「我这人霸占骰子的毛病」), i.e. exactly what steps 5–6 externalize (rules in the note, whose-turn on the board). Patience window (8s) noted as tight; owner defers. Step 5 implemented + self-tested 2026-07-21 (v545): THE HINGE is real — extract_panel_blocks is the routing table (rolls · reacts · notes · reveals in one pass, hist-frozen first; a future move = one row + its apply), and the envelope works end to end: <note>…</note> seals into the panel's persisted notebook (rides every turn's content; hist keeps tag+content — double memory) with a visible timestamped capsule; <reveal/> opens the oldest seal as the card — content · seal time · 「封存后未曾改动 ✓」. The acceptance battery ran the full 20-Questions chain: the reveal carries the original seal timestamp, the notebook survives a restart, and the secret never touched a bubble. SP gains SEALED NOTES (seal BEFORE play · answer FROM the note — "an unsealed secret is an improvised one, and rooms can tell"). Step 6 implemented + self-tested same day (v546), one router row as promised: <board title="比分">…</board> pins/replaces the public state strip (board_set events + projection — update-in-place, restart-safe, rides every turn's content so the panel remembers what it pinned); <board away/> takes it down; zero stream records — the acceptance ran a trivia round with a live scoreboard and no scoreboard re-posts. The dock became one zone with two slots (board pin above the die, collapsible, hides only when both empty). SP: THE BOARD — state lives on the board, never re-posted as chat; whose turn belongs there (the Twain whose-turn fix, structural at last). The Holmes room (owner test) found the router's blind spot — the model played a flawless quiz (sealed every answer, boarded every score, revealed every round) but placed its action tags outside the <speak> blocks, and the router only scanned decoded speech: every board and reveal silently vanished, and the visible wreckage (no whiteboard · unrevealed seals · cascading prose leaks) looked like model failure. Fixed same day: extract_outside_blocks salvages tags above/below the speak blocks (paired spans stripped first, unclosed speak swallows the rest — no double-collect; guarded off the json-content/quarantine paths where raw IS the replies). Actions are actions wherever they sit. SP also gains the anti-leak line: a re-ask or hint must never contain the answer — the reveal card is the only door. The Trivia room (10 clean rounds — board · seals · reveals all flowing) exposed the envelope’s last gap: on composed logic puzzles, rulings contradicted the seals (sealed 一小时, ruled 45分钟 correct; sealed one riddle’s answer, judged another) — facts converge across turns, compositions don’t: the model re-derives instead of consulting. Countered same day (fbf3a74): notes must be labeled（第5题答案：…）, the notebook’s riding note now binds rulings, and SP gains THE JUDGING RULE — if the note seems wrong, rule by it and admit the flaw; a ruling that contradicts a seal is the one unforgivable move. Step 2 implemented + self-tested 2026-07-20 (v538): hosts react via <react emoji="…"/> (stripped, applied as p:<slug> events on the batch's guest line; a react-only reply = the nod without the bubble); humans' reactions reach the panel as delta notes; the sheet shows host reactors — attribution public. Owner-tested same day: first direct ask landed a real 👍 chip (old room, post-restart). Finding worth keeping: compliance is character-gated by design — Jack Ma taps 👍 cheerfully, Bob Dylan declines in voice ("I'm a song and dance man, not a telegraph operator"). The tool works; the taste belongs to the persona — the SP's persona-first rule extending to furniture, as intended. Organic (unprompted) react rate = watch the parse-health reacts counter before any SP warming. Step 7a implemented + self-tested 2026-07-21 (v547): THE CIRCLE, sealed collect end-to-end — <gate>问题</gate> opens a go-around over the room's humans (one router row, as promised); while open every human message is HELD in the gate object (state.json only) — no event, no transcript line, no panel wake: the read-path law enforced by construction, so neither replay path can show a held answer (Ben's replay literally does not contain Amy's text). A second message revises; the sender's bubble seals (dashed + 🔒), everyone else sees only the dock's roster card (question · who's in · n/m — never what). Release = all-in auto · any member's two-tap Release · <gate away/> (posts what's in — typed words are never destroyed); the answers become real events AT ONCE (original nonces stamp the sealed bubbles into ordinary ones) and ONE cue-carrying turn reads the whole round. Live probe (Twain, deepseek, two users over HTTP): four rounds — brainwriting round released all-in and the model read the answers as a set（「"复苏"与"泥土"——两位各抓住了一半的春天」）, early release via the route posted the lone answer, and the release cue batched with a queued human line into one turn (the batching seam holding). Bug caught at paint level: the `gate` close SSE outruns the release's `post` events — clearing my-held-nonce there left the sender's bubble sealed forever; the post with my nonce is what unseals it. Remaining for step 7: the collect-roll (7b) — for="u3,u5" arming through the gate, ties re-rolled server-side, the 定先手 acceptance. The Cameron-ish room (owner test, v548) caught the one-brain hole live: solo-human RPS, the host as referee-and-player — its own move was never sealed anywhere, and it "revealed" the perfect counter after reading the human's answer four rounds straight (4:0; honest odds 1-in-81). Nothing had taught the two tools to compose. Fixed same day: SP gains the host-as-player rule (seal YOUR move via <note> WITH the gate; <reveal/> FIRST at release — "announcing a move you never sealed, after reading everyone's answers, is cheating, and the timestamps prove it either way"), the roster card grows the opener's ✉ chip (host_sealed = an unrevealed seal by the opener — the human can SEE the commitment), and the release cue demands reveal-before-ruling. Re-probed live: Cameron sealed（石头）, revealed on the answer, and ruled against himself（「你的布把我包了——你赢了」）. Also from the owner test: all circle copy left internal jargon（"opened the circle"）for plain 密封收集 language — the open capsule may be the host's entire move (a pure-gate reply drops its bubble), so it must explain itself. Two open observations: first-turn tool compliance is variable (the model sometimes mimes the furniture in prose before operating it next turn — the step-8 card's opener will pin the first-move discipline), and for a solo human the envelope alone already makes the game fair — the gate earns its keep at two-plus humans. 7c (v549, owner redesign, same day): SEALED BUBBLES — one commitment channel for everyone. The owner's call superseded the Cameron patch with structure: every circle submission — human or persona — lands in the stream immediately as a sealed bubble (author + 🔒; content server-held; yours shows your text), and all open together at release. The persona submits via <seal>出招</seal> riding the gate-opening reply (one router row; subs keyed p:<slug>) — its move is committed before any human answer exists, which is the whole fairness; at release persona moves unseal first as real speak events (chronology on record), and the cue lists them verbatim — or tells a playing host that sealed nothing you have no move; judge only. Live-probed through four RPS rounds: the no-seal backstop held（「这把只判，不出招——布赢了空气」）, the sealed round ruled honestly from the committed move（sealed 石头 → 「石头砸剪刀——我的」）, a three-player round released Cameron-剪刀 / Amy-剪刀 / Ben-石头 → 「Ben赢！」, and both viewer seats verified at paint level (the observer sees placeholders land live; the completer's own bubble unseals with its lid — the release runs inside the completing /say, so its post SSE outruns the response: the route hands the lid back and the client unseals directly). The one-brain caveat stands for persona-vs-persona simultaneity — and the owner's stated future architecture (one independent LLM per persona) dissolves it when it arrives. 7c′ (v550, owner feedback from the Ang Lee room): two fixes. The seal leak — the model narrated its move in the visible prose AND sealed it（「手指蜷了蜷…剪刀 ✌️」+ <seal>剪刀 ✌️</seal>, three rounds; the router was fine, the theater leaked）; countered twice: a server guard strips any standalone visible line that IS the sealed text verbatim (the exact observed pattern — safe because the gate question lists options inline, never as a bare matching line), and the SP now forbids the move appearing in visible words in any form（named · emoji · theater — "perform the gesture, never the content"）. True send order — sealed bubbles now render in submission order（gate_payload.order）, your own held bubble sits in its slot, and at release every bubble reveals in place (posts and persona unseals replace their placeholders; the server interleaves release events by the same order, so replay = the order the room lived it). Re-probed live: order ticked [ang-lee] → [ang-lee, u8] → reveal 布·石头·布 in sequence, opening bubble clean. Owner-tested ✓ same day（"It worked"）— the sealed-bubble/unsealing visuals deferred by owner for a later pass. 7b, the COLLECT-ROLL (v551) — step 7's last piece: <roll dice="1d20" for="all"/>（or for="u3,u7"）arms one die per person — roll_cast events land publicly as they come (server facts have no anchoring problem; the table's meta grows per-thrower chips), each member's hand die retires with their throw, and when the last is in the server settles it: top ties re-roll automatically (round-stamped, on record), then ONE multi roll_result（the 定先手 capsule「Ben 8 · Amy 5 — Ben 先」）and ONE patient narration cue — partial casts wake nothing, so the host speaks exactly once per round: the step's acceptance, passed live first try（Twain armed for="all" unprompted; "the dice have spoken — the floor is yours, Ben"）. Grammar note: all now means everyone rolls; any/anyone stay first tap wins. Owner's Newton-ish + Ang Lee tests (v552) sharpened the boundary law: the collect record had "goes first" baked in — wrong for a compare-rolls game — so the furniture now states FACTS only (the numbers, winner bold; the cue hands the host「highest: X」and says the meaning — turn order, the round's point — is the game's rule, declared once). The Ang Lee RPS meltdown (a two-布 round scored to nobody, an improvised tie house-rule, +2 self-award, then「这算谁赢？」asked at the players) plus Newton's initial flat prop-denial（「这里没有实物的骰子和白板」— zero tags four turns, cured only by an explicit-syntax opener no real user should ever type）both point the same way: per-round rule-judging, scoring, termination, and put-the-tools-away are SCENARIO discipline, and scenario discipline is the CARD (step 8) — pinned rules the host re-reads, an opener that makes the first move mandatory, a closing move that clears the table. Also overdue: the one-game sonnet A/B on game-state discipline (third data point now: Trivia seals · Lee's judging · Lee's termination). v553 — the owner's three pre-step-8 sanity rulings, all shipped. ① The abstract die: a roll is a draw from a distribution — numeric（NdM）or categorical（<roll pick="红, 蓝, 绿" for="all"/> now collects too）. ② Winning is the room's contract: the server never ranks, never names a winner, and the auto tie re-roll is gone（it assumed highest-wins-unique; in 比大小 a tie means no points, and re-rolling would change the game）— results post as facts in throw order, ties stand, the cue says meaning is the game's agreed rule（default conventions allowed, said aloud）and re-arm-the-tied is the break. ③ Tool invocation: the SP gains ONE WORKED EXAMPLE — a complete tagged opening reply（board + collect in one <speak>）plus the two anti-patterns（never narrate props into existence · never claim the room has none）. A/B'd immediately: the same Newton + the same natural-language prompt that drew flat denial in the morning → first reply armed the collect AND pinned the board, judged round 1 right, board ticked 1:0. In-context shape beats description — the v539 law applied forward. (Escalation ladder if compliance sags again: hist-seeded exemplar exchange → keyword-triggered riding nudge → the card's pinned opener.)