demo1-meta-display-simple-caption

Imported from UnlimitedAbandonWare/AbandonWareAi (.agents/skills/demo1-meta-display-simple-caption/SKILL.md). Install upstream with npx skills add UnlimitedAbandonWare/AbandonWareAi --skill demo1-meta-display-simple-caption. Copyright stays with the author.

Meta Display · 대화·힌트만 단순 표시

안경(Meta Ray-Ban Display)에 그릴 내용은 대화 전사/요약과 짧은 힌트만이다. Fold6·웹은 수음·제어·설정을 맡고, 렌즈는 읽기 전용 캡션 표면으로 둔다.

기준 구현 (SSOT 행동)

1. 정상 표시가 확인된 단순 페이지를 기준으로 삼는다. 예: display-test-01 계열(어두운 배경, 밝은 테두리, 큰 글씨 1~2덩어리, 스크롤·장식·진단 오버레이 없음).
2. 기존 복잡한 Display 페이지를 여러 방식(DAT / 공식 Web App / 커스텀 릴레이)으로 섞어 추적하지 않는다. 경로가 커스텀 companion·정적 HTML이면 그 경로만 최소 수정한다.
3. 렌즈 본문 허용 필드만 사용한다:
   • conversation — 최신 전사 또는 짧은 대화 덩어리
   • hint — 짧은 힌트 카드(기본 생성 계약: 목표 약 480–540자, hard cap 612자, 렌즈에는 측정된 페이지(기본 약 8줄/페이지)로 표시; 목표·상한·페이지 줄 수는 개발자 설정으로 조절 가능하며 기본값이 아래의 FIELD_TESTED 수치)
   • hintId — Card.requestId; 힌트 세대 식별자. 같은 본문의 재생성(새 id)과 같은 힌트의 재수신(같은 id)을 구분한다
4. 렌즈에 넣지 않는다: 디버그 배너, ACK/CONNECTED/owner 원문, QR, 설정 패널, 버튼 다수, 긴 RAG 덤프, 다중 카드, 자동 스크롤 피드.
5. 한 화면에 한 덩어리만 보여 준다. 길면 다음 페이지/다음 힌트로 넘긴다(힌트 생성 계약 기본 최대 612자; 표시 페이징이지 임의 truncate가 아니다). 600×600·무스크롤·고대비를 유지한다. 각 힌트는 최초 표시 기준으로만 만료된다(기본 수명 20초; Fold 안경 표시 설정에서 전사 유지·힌트 유지·페이지 간격을 서로 독립적으로 1–100초 정수 입력 — 서버·클라이언트가 같은 값을 쓰게 한다) — 동일 힌트 반복 수신·전사 갱신·페이지 전환은 수명을 연장하지 않고, 만료/교체된 힌트는 다시 수신돼도 재등장하지 않는다. 한 페이지 기본 줄 수(기본 11줄)를 넘는 힌트는 측정된 줄 높이로 페이지 분할해 전체를 보여 준다(자동 넘김 기본 5초 — 안경에는 키 입력 경로가 없어 유일한 페이지 전환 수단이다 — + 수동 이동 병행, 간격은 1–100초 설정 가능하고 0=끄기; 수명보다 간격이 길면 다음 페이지 전에 힌트가 사라져도 간격을 당기거나 수명을 늘리지 않는다; 모든 페이지가 같은 수명을 공유하고, 수동 이동 경로는 항상 남겨 둔다). 전사와 힌트가 함께 보일 때 두 영역은 합산 줄 예산(기본 약 13개 실측 줄, 설정 가능)을 공유하며, 전사는 최신 단위(마지막 문장·질문 포함)를 우선 유지하고 오래된 단위부터 줄인다. 이번 작업이 표시시간만 바꿀 때는 글자 크기·줄 수·힌트 본문을 줄여 맞추지 않는다.
6. ACK 증가·서버 CONNECTED만으로 “렌즈에 보였다”고 단정하지 않는다. 단순 텍스트가 안경에서 보인 뒤에만 기능을 한 단계씩 붙인다.
7. 이 스킬은 표시 규칙을 정의한다. 장시간 수음·전사 파이프라인 구현 세부는 별도 작업/스킬에 두고, 여기서는 안경 출력 계약만 강제한다.

Codex가 할 일 / 하지 말 일

최소 수용 기준

• 안경에서 밝은 테두리 안 텍스트가 읽힌다.
• 보이는 문자열은 대화/힌트 계열뿐이다.
• Fold6(또는 웹 컨트롤) 없이 렌즈만으로 설정을 바꾸려 하지 않는다.

롤백

이 스킬 폴더와 AGENTS.md의 simple-caption 포인터만 제거한다. display-test-01 및 기존 Display 역할 스킬은 보존한다.

Soft-evidence mode (hints > silence)

When RAG citations/evidence are empty, the chat/display pipeline should still emit a non-empty conversation/hint. Local/dev softens gates via:

• gate.evidence.allow-empty=true — EvidenceGate soft-allows totalEvidence==0
• gate.citation.log-only=true, guard.citation.require_official=false, low min counts
• gate.finalSigmoid.mode=log-only, jammini.guard.mode=soft
• AnswerExpander must not instruct [NO_EVIDENCE]; keep the original draft if a marker still appears

Do not show only "근거 없다" on the lens. Prefer a short conversational hint. Re-tighten for production by flipping those properties off.

Runtime policy (FIELD_TESTED)

Canonical runtime details moved here from AGENTS.md. Numbers below are factory defaults. Live Fold #lens-display prefs win; persist via lensSettings / saveSetting('lensDisplay').

• OFFICIAL (Meta-verified): 600×600 additive display; black background is a transparent canvas. D-pad/Enter focus; interactive controls follow .focusable. Scrolling containers, Neural Band, on-glasses text composer, and opt-in drag exist — do not absolutize "no scroll" / "no text input".
• Length policy (configurable; FIELD_TESTED defaults): generation target default ~480–540 Hangul chars; hard cap default 612 only when context needs it. 4–8 complete lines stays the default hint-generation shape (ConversateCardPrompt), not a rendering cap. These numbers are adjustable defaults, not immutable constants: the user may raise/lower the generation target·cap via dev settings or application-meta-display.yml knobs when measured evidence (under-generation vs finishReason=length) supports it — never blanket-double tokens/calls/hint counts, and keep other cue routes' default behavior unchanged. On-lens display uses a measured shared line budget (transcript+hint), default ~13 rendered lines but configurable: #transcript keeps newest complete units first, default transcript max ~4 lines (configurable) beside a hint; #hint renders a configurable number of lines per page (default ~8). Font default 26–30px; the user may adjust transcript/hint font size via dev settings — adjust layout only by measured fit (scrollHeight/clientHeight), never by silently overriding a user-chosen size. Transcript size/color/weight must not switch when a hint appears or disappears: keep one configured style from first show, and avoid layout jumps on hint show/hide. Goal = a complete hint fast/reliably — not shortest possible.
• Lens lifecycle: a hint owns its TTL from first show (default 20 s). Fold 안경 표시 설정 exposes three independent integer-second fields in range 1–100 (step 1, not presets): transcript hold, hint hold, page interval. Defaults stay 20 s / 20 s / 5 s until the user changes them. Server LensDisplayPrefs (transcriptTtlMs sibling + existing hintTtlMs + autoPageMs) and client receiver.js state.display must use the same applied values — never a leftover HINT_TTL_MS constant or a hidden 20 s / 2 s floor. Identical re-polls, transcript updates and page turns never extend the hint; expired/superseded hints stay hidden (judge expiry by expiry time, not by timer callback firing alone). Changing settings must not restart an already-running hint clock or revive an expired hint. /api/assist/display/lens/text = {conversation,hint,hintId,hintExpiresAt,conversationExpiresAt,display} (hintId = Card.requestId; text is fallback identity only). #hint caps a configurable number of lines per page (default ~11); longer hints are paged client-side (auto-advance on by default, 5 s — the glasses have no key input, so it is the only page-turn path — interval 1–100 s, 0 = off). Do not shrink/stretch the page interval so the last page still appears before expiry — if hint TTL is 5 s and page interval is 10 s, page 2 may never show; that is intended. Manual next/prev must always work, 힌트 n/m in #status. #transcript fills the remaining share of the shared line budget (default ~13) and drops oldest complete units first — never the latest sentence. A settings round trip (UI → server validation → lens/text / relay display → lens render) must apply without a rebuild/restart after this code is live; localStorage alone never proves on-glasses application. Out-of-range input is rejected at the input/invalid_lens_settings boundary — never silently clamped to another value. Show both the typed seconds and the applied seconds in the Fold menu. Bump ?v= in meta/index.html when receiver.js changes.
• Force-hint / rolling trigger (generation cycle — settings-driven, not frozen): YAML factory defaults are force-hint 180 s / Δ≥50 chars, rolling quiet 2.5 s (trigger-quiet-ms), cooldown 10 s, card hold hintHoldUntil from hint TTL / display-ttl-ms 20 s. The time knobs (quiet, cooldown, force-after, display/hint hold, page interval) follow persisted Fold 안경 표시 설정 — after save and after refresh/reconnect the running cycle must change. Do not keep a leftover YAML floor when the user saved a different value (including a saved 180 s force-hint). Char thresholds (Δ50, 120) stay count gates, not seconds. Rolling path stays deterministic: final utterance with normalized transcript delta ≥ 120 chars fires utterance_end; accumulated ≥120 plus quiet ≥ the set trigger-quiet-ms fires transcript_delta; both gated by the set cooldown-ms and hintHoldUntil. sample-probability was removed — do not reintroduce.
• Tokens vs characters: max-output-tokens is an API output budget, not characters. For 480–612 Hangul hints, review/raise per-model output budget so replies are not cut; change only with cut evidence (finishReason=length) or measured under-generation. A mid-layer Card/DTO/render truncating early is a hidden cap — fix that path.
• Routing / JSON / logs: adaptive routing from latency EWMA, success rate, timeout/validation failures, and cost — no hardcoded permanent #1 model. JSON: provider native structured output → strict validation → limited recovery → fallback; lenient parsers only for real format drift. Log on existing structured logs; no parallel log system. Do not mark a model stale from its name alone — only official deprecation/retirement, real probe failure, or credential issues.
• Ops: Fold6/web = mic/ASR/controls; lens = quiet text. Keep cueHint/Conversate + soft-evidence; no parallel hint stack. Live Java/YAML: compile then ForceRestart/DevWatch; never claim live success from a stale bootRun. Refresh build/desktop-meta-display*/... static copies via processResources when live loads from build.

Authorized source patch: Fold display + generation timing

When the user wants Fold 개발자 메뉴에서 전사 유지·힌트 유지·페이지 간격 또는 힌트 생성 주기(quiet / cooldown / force-after / display hold)를 직접 입력하게 하면, this is an authorized application-source patch, not a report, not E0 resume, not a new settings page. Layout/font/output volume stay as they are. $demo1-meta-display-resume sourceMutationAllowed=false does not apply. Skip triad / three-way preflight. Still use work-ledger + source_edit_session lease/preimage.

Menu (reuse, do not duplicate)

Fold main/resources/static/assets/display/index.html → 연결·입력 설정 → 안경 표시 설정 (#lens-display). Reuse ld-cap-ttl / ld-hint-ttl / ld-auto-page + app.js lensSettings / saveSetting('lensDisplay'). Add sibling integer-second inputs for cue quiet / cooldown / force-after if missing. Do not reuse #segment-preset (ASR restart). Keep 자동 넘김 0=끔. Inputs: type=number step=1; hold/page 1–100 s; force-after range must still accept the factory 180 s. Show typed seconds and server-echoed applied seconds in #ld-status. Persist so refresh/reconnect restores the last saved cycle.

Classify leftover time floors (connect, do not freeze)

Entering 1 for a time field must yield 1 s, not a hidden 2 s. A saved 180 s force-hint must not be silently clamped down.

Independent clocks

transcriptTtlMs, hintTtlMs, autoPageMs, and the cue time prefs are independent. Do not max/min them together. Hint lifetime starts at first show; re-receive / page turn / settings edit must not reset a running clock or revive an expired hint. Rolling-caption start (ROLLING_EXPIRY) stays as-is — only the hold value becomes the setting. Overflow still uses existing page split; do not shrink fonts or truncate hint text to avoid paging.

Wire-through files (minimal)

LensDisplayPrefs.java (prefs + describe/defaults/Patch), ConversateSessionService.java (caption vs hint TTL split; quiet/cooldown/force read prefs), assets/display/index.html + app.js, assets/display/meta/receiver.js, tests. Bump receiver.js?v= in meta/index.html. First Java deploy: wear Close-Meta-Display.bat then Start-Meta-Display.bat. After that, changing the numbers must apply without rebuild/restart. Log setting apply / page turn / expiry / cue trigger with applied ms + timestamp only — no transcript, no secrets.

Verify

1 s and max values, mid values such as 7/23/46, transcript vs hint vs page vs quiet/cooldown/force set independently, 1/2 → 2/2 follows page interval, TTL 5 s + page 10 s with no auto-correction, refresh/reconnect keeps last saved prefs and the generation cycle.

Focus display (Nova Focus) — separate from hints

Nova Focus answers travel as a separate additive focus projection, never inside hint (its char cap, TTL, and hints-OFF would hide them). Their display contract is sequential: grapheme-by-grapheme reveal, oldest completed line rolls off when full, last char → tail hold → fade — no page numbers or next/prev. The followup-idle clock starts at presentation_done (final answer received + queue drained + last char in DOM), not at first render or answer-generated. Contract, phases, and gates: .agents/skills/demo1-nova-focus/SKILL.md. Every rule above (hint/conversation TTL, paging, shared line budget, settings) keeps applying to hint/conversation only — do not apply them to focus and do not route focus answers through hint.