@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/hbg-life-simulation/references/local-account-api.md

@@ -1,112 +0,0 @@
-# Local account API with CLIProxyAPI
-
-Use this as the preferred HBG image channel only after a per-project explicit consent gate. The Skill must check readiness first, ask before consuming account capacity and before installation, startup, repair, login, or OAuth, and use built-in ImageGen when the user declines or the account lacks image entitlement.
-
-CLIProxyAPI is a third-party MIT-licensed project: <https://github.com/router-for-me/CLIProxyAPI>. It supports Codex, Claude, Antigravity, Kimi, and xAI OAuth, plus OpenAI-compatible Responses and Images endpoints. It is not an official provider client and does not bypass quotas, entitlements, rate limits, or provider terms.
-
-## Security contract
-
-- Treat Skill invocation and a request to make a video as insufficient consent to use account capacity, set up the proxy, or start OAuth.
-- Require a fresh confirmation before using the local account channel for each new video project, even when the proxy is already ready.
-- Require explicit OAuth-risk acceptance before login.
-- Ask before starting a configured but stopped proxy.
-- Use only accounts owned or administered by the user.
-- Bind to `127.0.0.1`; never expose the proxy on `0.0.0.0`, a LAN address, a tunnel, or a public server.
-- Disable remote management and the management panel.
-- Store the binary, config, local client key, OAuth records, PID, and logs outside every project and Git worktree.
-- Use permissions `0700` for credential directories and `0600` for config and environment files.
-- Never print, paste, summarize, upload, or commit provider OAuth records or the generated local client key.
-- Never inspect browser cookies, browser storage, account pages, `auths/`, `client.env`, or provider OAuth files. The user completes browser/device authentication personally.
-- Do not patch the running Codex application's own transport configuration. Use the local endpoint only from explicit compatible clients and batch scripts.
-
-## Bootstrap
-
-First run the read-only readiness check:
-
-```bash
-scripts/setup_local_cliproxyapi.sh ready
-```
-
-It returns exactly one non-secret state: `ready`, `not-configured`, `not-running`, or `unhealthy`. Do not install, start, repair, or log in until the user has agreed after seeing the third-party, account-capacity, and OAuth disclosure.
-
-After consent, from the installed skill directory:
-
-```bash
-scripts/setup_local_cliproxyapi.sh bootstrap \
-  --provider codex \
-  --accept-oauth-risk
-```
-
-For a terminal without a browser callback, use the Codex device flow:
-
-```bash
-scripts/setup_local_cliproxyapi.sh bootstrap \
-  --provider codex-device \
-  --accept-oauth-risk \
-  --no-browser
-```
-
-The script downloads a pinned upstream release, verifies its published SHA-256 checksum, creates a localhost-only configuration, runs OAuth, starts the service, and checks the authenticated `/v1/models` endpoint.
-
-OAuth is an interactive user action. The Agent may launch the flow or display the device code only after consent, but must not operate the provider login page, read the account, retrieve cookies, or inspect the resulting OAuth files.
-
-Default local state:
-
-```text
-${XDG_DATA_HOME:-$HOME/.local/share}/hbg-life-simulation/cliproxyapi/
-├── bin/cli-proxy-api
-├── config.yaml
-├── client.env
-├── auths/
-├── logs/
-├── VERSION
-└── UPSTREAM
-```
-
-`client.env` exports `OPENAI_BASE_URL`, a generated localhost-only `OPENAI_API_KEY`, and `HBG_IMAGE_MODEL`. Source it only in the process that needs the API:
-
-```bash
-source "$(scripts/setup_local_cliproxyapi.sh env-path)"
-```
-
-## Lifecycle
-
-```bash
-scripts/setup_local_cliproxyapi.sh status
-scripts/setup_local_cliproxyapi.sh ready
-scripts/setup_local_cliproxyapi.sh smoke-test
-scripts/setup_local_cliproxyapi.sh stop
-scripts/setup_local_cliproxyapi.sh start
-scripts/setup_local_cliproxyapi.sh login --provider codex --accept-oauth-risk
-```
-
-Do not delete or replace an unknown process. The stop command acts only on the PID whose command line matches this integration's binary and config paths.
-
-## HBG image batches
-
-CLIProxyAPI currently provides `/v1/images/generations` and `/v1/images/edits` for supported image models, including `gpt-image-2`. Confirm the local models endpoint and one deliberate single-image test before starting a large batch. A successful models request does not prove that the user's account has image entitlement.
-
-```bash
-source "$(scripts/setup_local_cliproxyapi.sh env-path)"
-
-IMAGE_GEN="${CODEX_HOME:-$HOME/.codex}/skills/.system/imagegen/scripts/image_gen.py"
-python "$IMAGE_GEN" generate \
-  --model "${HBG_IMAGE_MODEL:-gpt-image-2}" \
-  --prompt "A simple comic-style empty room, no text" \
-  --size 1536x1024 \
-  --out /tmp/hbg-local-api-smoke.png \
-  --no-augment
-```
-
-The image test consumes account capacity and must be intentional. Display and inspect the returned image in chat before approving this channel for project batches.
-
-After it passes, use the existing `generate-batch` path with concurrency 5. Do not raise concurrency above 10 or use the proxy to evade account limits. Keep the same `SHEET_MAP.json`, staging, dimension normalization, anatomy review, and accounting requirements as any other compatible API.
-
-## Failure handling
-
-- `401` from the local endpoint: load the correct `client.env`; do not replace it with the provider OAuth token.
-- No models or upstream authorization failure: rerun the selected provider login and inspect only non-secret status output.
-- Image endpoint unsupported or entitlement missing: fall back to built-in ImageGen or another user-authorized Images API.
-- OAuth callback port conflict: use `codex-device` or stop the process occupying the provider callback port; do not weaken the loopback binding.
-- Release download or checksum failure: stop. Never install an unverified archive.
-- Port 8317 conflict: choose another local port during first install; keep `OPENAI_BASE_URL` and the config synchronized through the setup script.

package/skills/hbg-life-simulation/references/opening-system.md

@@ -1,45 +0,0 @@
-# HBG opening and audio system
-
-## Opening timing
-
-- Start lead narration almost immediately.
-- Place the flash/SFX no more than 0.05 seconds after the lead finishes.
-- Use about nine alternate lives across 1.4–1.8 seconds.
-- Hold the selected life for 0.1–0.2 seconds before the reveal voice starts.
-- Keep the selected frame visible until the reveal finishes.
-- Start the first body image only after reveal narration ends.
-
-Keep the opening clean. Show centered life labels during the flash and the full selected-life title on the final card. Do not show progress bars, counters, `副本载入完成`, or numeric HUD badges.
-
-Use orientation-native flash art. In landscape, each final candidate image must be 1920×1080 and composed as a wide shot before zoom is applied. A 5×2 sheet on a landscape canvas produces portrait-shaped cells and is therefore forbidden for landscape flash extraction. Use 2×2 sheets of landscape panels or standalone landscape images instead. Portrait projects follow the inverse rule at 1080×1920.
-
-Render each candidate preview to a revisioned filename. After approval, bind that exact file through `audio_meta.opening.previewVideo` or the canonical approved path in `HBG_STYLE.json`. The long renderer must not search old `v2`/`v3` previews or substitute the raw flash/SFX asset.
-
-## HBG camera language
-
-- Alternate aggressive short zoom-in and zoom-out during every flash image. For a 0.14–0.20 second flash frame, use about `scale 1.00 → 1.18–1.22` and `1.18–1.22 → 1.00`, with a small opposing x/y drift. A sequence of static hard cuts does not satisfy the opening motion requirement.
-- On the selected card, use a clearly readable slow `scale 1.00 → 1.13–1.15` plus slight horizontal drift. Verify the first and last encoded hold frames; if the crop change is not immediately visible side by side, strengthen it before the full render.
-- In the body, alternate zoom and pan; do not repeat the same direction across consecutive images.
-
-Run `scripts/validate_style_system.mjs PROJECT_DIR` after the preview is approved. It samples two frames inside the first flash and two frames across the selected hold; an unmeasurable crop change is a hard failure.
-
-## Fixed BGM workflow
-
-Default to one static BGM baseline. Do not duck under narration unless the user explicitly requests it.
-
-1. Probe source integrated loudness, true peak, and the first 20 seconds.
-2. Add BGM as a separate audio element with one fixed `data-volume`.
-3. Render a 15–20 second opening preview covering lead, flash, reveal, and the first body sentence.
-4. If the BGM is inaudible, raise it by 3–4 dB and rerender; do not make tiny adjustments.
-5. Check final true peak and retain at least 3 dB headroom.
-
-Validated project example:
-
-- source BGM integrated loudness: approximately `-15.3 LUFS`;
-- `data-volume=0.055`: too quiet;
-- `data-volume=0.14`: audible but still conservative;
-- `data-volume=0.22`: accepted louder review level;
-- no fades, no ducking, no volume keyframes;
-- encoded true peak remained approximately `-3.6 dBTP`.
-
-Do not copy `0.22` blindly to a different track. Match perceived level using the source loudness and the 3–4 dB iteration rule.

package/skills/hbg-life-simulation/references/project-spec.md

@@ -1,62 +0,0 @@
-# Project specification
-
-Use `PROJECT_SPEC.json` for every value that changes with the story. Builders must never contain a story title, fixed chapter count, character identity, correction list, voice choice, BGM filename, flash-life list, or selected-life image path.
-
-Required shape:
-
-```json
-{
-  "version": 1,
-  "title": "完整的人生副本名称",
-  "titleLines": ["标题第一行", "标题第二行"],
-  "source": {
-    "openingSentence": "今天体验的人生副本是完整的人生副本名称。",
-    "corrections": [{ "from": "明显错词", "to": "修正后" }],
-    "chapters": [{ "title": "章节名", "cue": "本章在正文中唯一的开头短语" }]
-  },
-  "narration": {
-    "voice": "zh-CN-YunjianNeural",
-    "bodyRate": "+0%",
-    "leadRate": "+0%",
-    "revealRate": "+0%",
-    "pitch": "+0Hz",
-    "flashDuration": 1.667,
-    "captionMaxChars": 18,
-    "captionMinChars": 6,
-    "captionMinDuration": 0.55
-  },
-  "imageGeneration": {
-    "channel": "local-account-api",
-    "consent": "explicit",
-    "provider": "codex",
-    "model": "gpt-image-2",
-    "concurrency": 5
-  },
-  "opening": {
-    "leadDisplayText": "今天体验的人生副本是……",
-    "flashMedia": "assets/opening/flash.mp4",
-    "finalImage": "assets/opening/final.jpg",
-    "flashLives": [{ "asset": "assets/opening/life-01.png", "label": "另一种人生" }]
-  },
-  "audio": {
-    "bgmSource": "assets/audio/bgm/source.mp3",
-    "bgmLooped": "assets/audio/bgm/looped.m4a",
-    "narrationOutput": "assets/audio/narration.m4a"
-  }
-}
-```
-
-Rules:
-
-- Keep `SCRIPT_SOURCE.md` verbatim. Put only unmistakable transcription repairs in `source.corrections`.
-- Use any reasonable chapter count. Builders derive headings, act labels, sub-compositions, and timing from the array length.
-- Make each chapter cue unique and ordered after corrections.
-- Keep character identity in `CHARACTERS.md`, not in this shared specification or a builder.
-- Set `imageGeneration.channel` to `local-account-api`, `built-in-imagegen`, or `external-compatible-api`. Record `consent`, provider label, model, and concurrency only; never record an endpoint, key, token, Cookie, OAuth path, account identity, or contents of protected proxy files.
-- Use `local-account-api` only after the Skill consent/readiness gate. Use `built-in-imagegen` after decline or unavailable entitlement. Keep API concurrency at 5 by default and never above 10.
-- Keep orientation-dependent pixels in `HBG_STYLE.json`, never in `PROJECT_SPEC.json`.
-- Keep semantic scene selection in `STORYBOARD_BASE.json`; code may validate it but must not invent example-specific beats.
-- Use `captionMaxChars=18` as the usual landscape ceiling and `14` as the usual portrait ceiling. Generated splits must meet `captionMinChars`; only a complete original utterance may set the short-caption exception.
-- Record missing punctuation before a spoken list or a broken source phrase in `source.corrections`. Do not use source line breaks as synthetic TTS pauses.
-- Resolve every `opening.flashLives[].asset` to the exact orientation canvas before validation.
-- Resolve or copy every configured asset into the project before composition building.

package/skills/hbg-life-simulation/references/render-and-qa.md

@@ -1,107 +0,0 @@
-# Long render and encoded-media QA
-
-Use this process for videos longer than four minutes or roughly 7,200 frames at 30 fps.
-
-## Render preflight
-
-Before disk preflight, run the style gate:
-
-```bash
-# Default landscape project
-node scripts/init_project_style.mjs PROJECT_DIR
-
-# Use only for an explicit portrait request
-node scripts/init_project_style.mjs PROJECT_DIR --orientation portrait
-
-node scripts/validate_style_system.mjs PROJECT_DIR
-```
-
-The default initialization is landscape 1920×1080. Portrait 1080×1920 requires the explicit flag. This gate rejects a declared orientation that contradicts the canvas, an opening with the wrong dimensions/orientation, an opening shorter than `bodyStart`, unsafe orientation-specific caption ratios, zero ASS box padding, and opening sequences without measurable flash/selected-life motion.
-
-Run:
-
-```bash
-scripts/preflight_long_render.sh PROJECT_DIR 12
-```
-
-The 12 GiB default is a conservative floor for a high-quality 1080p, 8–12 minute image-heavy render. Treat it as a gate, not an estimate. Increase it for 4K, 60 fps, unusually detailed frames, or non-streaming capture.
-
-The preflight script reports existing `renders/work-*` directories but never deletes them. Before removing one, match its render job ID to a completed, interrupted, or failed process and confirm no matching render process is alive. Remove only the exact abandoned directory. Never broadly clear `renders/`, a workspace root, source assets, or completed MP4 files.
-
-## Long-form render
-
-Run:
-
-```bash
-scripts/render_long_video.sh PROJECT_DIR renders/story-v1-complete.mp4
-```
-
-Defaults:
-
-- high quality;
-- `PRODUCER_ENABLE_CHUNKED_ENCODE=true`;
-- `FFMPEG_ENCODE_TIMEOUT_MS=7200000`;
-- hardware FFmpeg encoding on macOS;
-- a new revisioned output path supplied by the caller.
-
-Override only when needed:
-
-```bash
-HBG_RENDER_QUALITY=standard HBG_USE_GPU=0 HBG_MIN_FREE_GIB=16 \
-  scripts/render_long_video.sh PROJECT_DIR renders/story-v2-complete.mp4
-```
-
-If capture succeeds but encoding times out, do not repeat the same command blindly. Enable chunked encoding, raise the timeout, and use a supported hardware encoder. If the renderer has already deleted its captured frames, the next run must recapture; make the corrected settings before restarting.
-
-## Low-disk streaming fallback
-
-Use the bundled streaming renderer when HyperFrames frame capture would exceed safe free disk space or when a measured test shows that the renderer materializes the entire long timeline as image files:
-
-```bash
-HBG_VALIDATE_ONLY=1 node scripts/render_streaming_ffmpeg.mjs PROJECT_DIR renders/story-v1.mp4
-node scripts/render_streaming_ffmpeg.mjs PROJECT_DIR renders/story-v1.mp4
-```
-
-The project must already contain final `PROJECT_SPEC.json`, `STORYBOARD.json`, `HBG_STYLE.json`, `audio_meta.json`, all scene assets, an explicitly approved opening video, narration, BGM, ratchet SFX, and reveal audio. Story-specific paths and chapter metadata must come from the spec/audio metadata rather than renderer fallbacks. The script renders one short MP4 per still with deterministic zoom/pan, concatenates them, burns synchronized ASS captions and chapter labels, mixes the fixed-level audio bed, and streams the final H.264/AAC output without materializing every timeline frame.
-
-Defaults come from the selected orientation template and the project's locked `HBG_STYLE.json`. Override paths or mix settings only when the project differs:
-
-```bash
-HBG_OPENING_VIDEO=qa/opening-final.mp4 \
-HBG_BGM_AUDIO=assets/audio/bgm/story.m4a \
-HBG_BGM_VOLUME=0.22 \
-node scripts/render_streaming_ffmpeg.mjs PROJECT_DIR renders/story-v2.mp4
-```
-
-The streaming renderer removes its exact `ffmpeg-work-<output-name>` directory only after a non-empty final MP4 exists. Set `HBG_KEEP_STREAM_WORK=1` to preserve intermediates for diagnosis. On failure, preserve the work directory until the cause is known; remove only that exact inactive directory.
-
-The streaming ASS style reads `HBG_STYLE.json`. Landscape defaults to 48 px white bold text, a semi-opaque warm-black box, 12 px box padding, and `MarginV=58` at 1920×1080. The validated portrait treatment uses the same 48 px type and box treatment but `MarginV=280` at 1080×1920. `BorderStyle=3` with zero outline/padding is forbidden because the background box disappears. Environment overrides are diagnostic escape hatches, not the normal authoring path.
-
-## Encoded MP4 verification
-
-Run with meaningful semantic timestamps:
-
-```bash
-scripts/verify_final_video.sh OUTPUT.mp4 QA_DIR \
-  0.5:opening-lead \
-  3.2:opening-flash \
-  5.3:selected-title \
-  449.5:hands-start \
-  455:hands-mid \
-  462:hands-end \
-  572.2:phone \
-  648.4:keyboard \
-  663:ending
-```
-
-The script writes:
-
-- `ffprobe.json` for resolution, codecs, frame rate, duration, frame count, and audio/video streams;
-- `blackdetect.txt` for black intervals;
-- `silencedetect.txt` for long silence despite the intended BGM bed;
-- `ebur128.txt` for integrated loudness, loudness range, and true peak;
-- named encoded PNG frames plus `contact-sheet.jpg`.
-
-Open every suspicious contact-sheet panel at full resolution. For hand or prop corrections, inspect at least the start, midpoint, and end of the encoded motion. Passing source-image review is insufficient because zoom/pan can reveal cropped defects.
-
-For the validated HBG mix, preserve at least 3 dB true-peak headroom. A constant BGM track should span the full composition and should not create one-second-or-longer digital silence in the final mix.

package/skills/hbg-life-simulation/references/storyboarding.md

@@ -1,69 +0,0 @@
-# Storyboarding rules
-
-## Beat selection
-
-Choose a new still when at least one changes:
-
-- location or time of day;
-- dominant action;
-- emotional power balance;
-- focal object or visual metaphor;
-- narration moves from observation to memory or decision.
-
-Do not create a new image for connective wording alone. Combine consecutive sentences when they describe the same action and emotional state.
-
-## Duration-density audit
-
-Run this audit only after real narration timing exists:
-
-- 4–8 seconds: normal still;
-- 8–12 seconds: acceptable only when the frame carries enough action, detail, or emotion;
-- over 12 seconds: split unless the hold is explicitly intentional;
-- over 16 seconds: always treat as missing visual coverage and add at least one beat.
-
-For an 8–12 minute HBG story, target roughly 70–110 images. Prioritize extra images in emotional peaks, dialogue exchanges, object memories, and ending reflections. A stronger zoom does not compensate for a 25–55 second still.
-
-## Shot rotation
-
-Avoid repeating the same face-centric composition. Rotate among:
-
-1. wide establishing frame;
-2. medium character action;
-3. two-shot with visible distance;
-4. over-shoulder observation;
-5. hand/object close-up;
-6. environmental or symbolic insert.
-
-At least one in every four scenes should work without showing either face.
-
-## Chapter selection
-
-Choose chapter boundaries from the current story's irreversible changes: a new age or life stage, location, relationship phase, power shift, decision, gain, loss, departure, or final reinterpretation. Use as many chapters as the story needs; do not force every project into ten acts.
-
-## Storyboard JSON shape
-
-```json
-{
-  "id": "s001",
-  "chapter": 1,
-  "narration": "旁白文本",
-  "visual": "one concrete visible moment",
-  "characters": ["protagonist", "supporting-character-id"],
-  "participants": { "count": 2, "allowed": ["protagonist", "hr-manager"] },
-  "shot": "wide | medium | close | insert | two-shot",
-  "motion": "zoom-in | zoom-out | pan-left | pan-right | hold",
-  "duration_hint": 4.2,
-  "asset": "assets/generated/scenes/s001.png"
-}
-```
-
-`participants` is required for API generation. Author it semantically while reading the story; never derive it later only from surface keywords. `count` is the exact visible-person count for this frame, and `allowed` lists only the people permitted to appear. A crowd frame may use a fixed count such as seven with a bounded description such as `protagonist plus six ordinary commuters`.
-
-## Image prompt checklist
-
-- Repeat immutable identity traits from `visual-system.md`.
-- Describe one visible moment, not a paragraph of plot.
-- State camera distance and who is in focus.
-- State the current story's environment continuity.
-- Prohibit text and extra lookalike protagonists.
-- Keep hands anatomically plausible when soldering, smoking, eating, or applying bandages.

package/skills/hbg-life-simulation/references/style-system.md

@@ -1,80 +0,0 @@
-# HBG orientation and style system
-
-Use one project-level `HBG_STYLE.json` as the single source of truth for canvas, captions, opening motion/layout, chapter-label layout, and the accepted BGM baseline. Initialize it before writing composition CSS or rendering:
-
-```bash
-# Default: landscape 16:9
-node scripts/init_project_style.mjs PROJECT_DIR
-
-# Explicit portrait request only
-node scripts/init_project_style.mjs PROJECT_DIR --orientation portrait
-```
-
-Do not independently hard-code caption or opening values in HTML and FFmpeg. Composition builders must read `HBG_STYLE.json`; the bundled streaming renderer already does.
-
-## Orientation selection rule
-
-- If the user does not specify orientation, use landscape 16:9 at 1920×1080.
-- Use portrait 9:16 at 1080×1920 only when the user explicitly asks for portrait, vertical, 9:16, or an equivalent format.
-- Do not infer portrait from the platform, story genre, short-video duration, or a prior project.
-- Treat an existing project's `HBG_STYLE.json` as locked. Do not overwrite or rotate it implicitly.
-- Landscape and portrait share visual identity and motion language, but canvas, caption placement, maximum text width, crop framing, and safe areas are orientation-specific. Never copy raw coordinates between them.
-- Keep story title, chapter names, voice, BGM paths, flash-life labels, and image paths in `PROJECT_SPEC.json`; keep only orientation-dependent pixels and colors here.
-
-## Accepted landscape baseline (default)
-
-For 1920×1080:
-
-- caption font: PingFang SC semibold/bold, 48 px;
-- caption box: warm black at about 78% opacity;
-- HTML padding: `13px 28px 16px`, 12 px radius;
-- ASS `BorderStyle=3` with `Outline=12`;
-- caption bottom margin: 58 px, or about 5.4% of frame height;
-- caption maximum width: 1540 px;
-- semantic caption length: normally 8–18 Chinese characters; prefer a complete phrase over a shorter arbitrary cut;
-- flash motion: alternate `1.00 → 1.20` and `1.20 → 1.00` inside each short frame;
-- selected-life hold: `1.00 → 1.15` with slight drift.
-- opening and chapter-label widths, positions, and font sizes come from the landscape template.
-
-## Accepted portrait baseline
-
-For 1080×1920:
-
-- caption font: PingFang SC semibold/bold, 48 px;
-- caption box: warm black at about 78% opacity;
-- HTML padding: `13px 28px 16px`, 12 px radius;
-- ASS `BorderStyle=3` with `Outline=12`; never use zero outline because the background box effectively disappears;
-- caption bottom margin: 280 px, or 14.6% of frame height;
-- caption maximum width: 940 px;
-- semantic caption length: normally 8–14 Chinese characters;
-- flash motion: alternate `1.00 → 1.20` and `1.20 → 1.00` inside each short frame;
-- selected-life hold: `1.00 → 1.15` with slight drift.
-- opening and chapter-label widths, positions, and font sizes come from the portrait template.
-
-When a reference video has a different aspect ratio, transfer visual appearance, not raw coordinates. Measure font size, box opacity, and padding from the reference, but derive placement and cropping from the target orientation's safe-zone ratios. Never copy `bottom`, `MarginV`, `maxWidth`, pan distance, or crop coordinates directly across orientations.
-
-Every opening flash source must already match the selected canvas. Do not treat `object-fit: cover` as an orientation converter: a portrait source inside a landscape frame discards most of its vertical composition and creates an unintended close-up. The style validator probes every configured flash image and rejects dimensions other than the project's exact canvas.
-
-## HTML and ASS parity gate
-
-The short HyperFrames preview and the long streaming render use different caption engines. A correct HTML preview does not prove the final ASS captions are correct.
-
-Before a long render:
-
-1. Render the 15–20 second HyperFrames preview.
-2. Set `audio_meta.opening.previewVideo` to the exact approved preview file, or adopt it to the canonical path in `HBG_STYLE.json`.
-3. Run `node scripts/validate_style_system.mjs PROJECT_DIR`.
-4. Confirm the validator reports the requested orientation and dimensions, safe orientation-specific caption ratios, measurable flash motion, and measurable selected-life motion.
-5. Run the streaming renderer with `HBG_VALIDATE_ONLY=1`; confirm it resolves the same approved opening path.
-
-The full renderer must fail if the approved opening is missing or has the wrong orientation. Do not fall back to an older preview, the raw ratchet video, or an asset made for the other orientation.
-
-## Reference extraction rule
-
-When the user names an accepted prior video, extract encoded frames from that final MP4 rather than relying only on its source CSS. Compare at least:
-
-- two caption frames on light and dark backgrounds;
-- two frames from within one flash image;
-- selected-life hold near its start and end.
-
-Only freeze parameters after the encoded reference and encoded target have both been inspected.

package/skills/hbg-life-simulation/references/visual-system.md

@@ -1,44 +0,0 @@
-# Visual and character system
-
-## Shared HBG look
-
-- Default canvas: 1920×1080, landscape 16:9, 30 fps. Use 1080×1920 portrait 9:16 only when explicitly requested; frame and crop for the target orientation rather than stretching the other orientation.
-- Medium: polished Chinese webtoon / grounded seinen manga.
-- Rendering: bold black ink contour, realistic anatomy, detailed lived-in environments, subtle halftone and paper grain.
-- Base palette: deep brown-black `#1C1410`, tomato red `#D8000F`, warm off-white `#F5F2EF`, plus restrained colors appropriate to the story setting.
-- Generated images contain no text. Add all copy in HTML.
-
-Do not carry occupations, ages, relationships, clothing, locations, or mood from a previous story. Derive them from the current text and record them in the project's `CHARACTERS.md`.
-
-## Project identity lock
-
-For every recurring character, record:
-
-- role and relationship;
-- age or age stages;
-- face shape, eyes, eyebrows, nose, mouth, ears, hair, and one or two subtle recognition marks;
-- body type and posture;
-- clothing by time period or location;
-- emotional register and behaviors;
-- traits that may change and traits that must remain invariant.
-
-For a protagonist shown across ages, preserve at least three identity bridges such as eye shape, eyebrow direction, ear shape, mole placement, hair growth pattern, or smile asymmetry. Do not rely on a name alone.
-
-Generate one anchor for a single recurring protagonist, a dual/group anchor only when multiple characters recur together, and separate age anchors when one face changes substantially over time.
-
-## Environment continuity
-
-Record recurring locations, architecture, lighting, weather, season, period details, and important props in the project files. Reuse those facts in every related prompt. Change them only when the narration changes time or place.
-
-## Physical-reality gate
-
-- Treat real-world plausibility as a hard requirement, not an optional polish pass.
-- Use standalone generation for anatomy- or orientation-sensitive frames: overlapping hands, close hand-object contact, phones, keyboards, cigarettes, chopsticks, tools, sports equipment, mirrors, and identity-critical portraits. Reserve 2×2 sheets for lower-risk continuity beats. If one grid panel fails, regenerate that beat separately instead of cropping to conceal the problem.
-- Count anatomy before evaluating aesthetics: trace each shoulder to one arm and each wrist to one hand; count visible palms and fingers; distinguish intentional occlusion from duplication. Open every hand close-up at 100% and trace every fingertip back through exactly one palm, wrist, and forearm.
-- A two-hand contact image must contain exactly two wrist-to-palm chains, with no third palm-shaped mass or duplicated finger fan. Reject extra, fused, detached, mirrored, ownership-ambiguous, or structurally impossible hands and limbs.
-- Check the visible front/back, top/bottom, and functional direction of every prop before accepting a frame.
-- For phones, the character's eyes must face the screen side; the screen, rear cameras, case, buttons, and charging edge must not contradict one another. Fingers must grip the edges naturally and must not cover or emerge through the device.
-- For tools and food props, verify grip, contact point, gravity, heat source, working end, and intended use.
-- For mirrors and reflective surfaces, verify that pose, handedness, device direction, gaze, and surrounding geometry agree with the real subject.
-- Check doors, chairs, beds, tables, clothing, shadows, and body support for usable geometry and believable contact.
-- Reject and regenerate any frame with reversed, impossible, floating, intersecting, or functionally unusable objects, even if character identity and style are correct.