@topy-ai/maggie 0.8.14 → 0.8.16

package/README-zh-TW.md

@@ -143,7 +143,7 @@ structure drift 或重複 treatment product 時 fail closed。
 deployment、memory、feedback 和 MaggieDash。內容先 draft/review，外部寫入、
 publish 與 production deployment 需要明確確認。
 
-目前 npm 版本為 `0.8.14`,新增 `maggie-ai-media`(去背/放大/arm's-length ComfyUI 的命令生成器,自身不下載模型、不用 GPU),讓 Q 類達到同級;重型執行仍受成本把關。先前 0.8.13 新增 AI 媒體生成規劃 reference(Q 類:自託管模型概覽、程式碼授權≠權重授權規則、商用安全子集 rembg/Real-ESRGAN、成本/BYOK 守則;實際建置待明確許可)。先前 0.8.12 收尾其餘設計類:新增 `maggie-image`(影像處理 CLI:尺寸/OG/favicon/WebP-AVIF/svg/montage)與 web 動效、生成式視覺、圖像配方、字幕四個 reference。先前 0.8.11 收尾 N 類影片功能:`maggie_video.py compile` 再補上視頻 clip ken-burns、sidechain ducking、以及非透明度 overlay 效果(move/slide、圖片 zoom),並新增 `autocut` 子命令(silencedetect 去靜音)。先前 0.8.10 新增圖片 ken-burns + 文字透明度關鍵幀與 `captions`(whisper)子命令。先前 0.8.9 新增 `compile`(宣告式 timeline JSON→單一 ffmpeg 命令)。先前 0.8.8 新增 `maggie-video`(製作與處理行銷影片:剪輯、字幕、裁切成各渠道比例、浮水印、轉碼、響度正規化、clip 組裝、程式化動畫),附一個實測可用的 ffmpeg 配方 CLI(`maggie_video.py`),從參考庫的影片專案(moviepy/editly/ffmpeg 命令層、Revideo/Motion Canvas/CCapture 程式化渲染、MLT/OpenShot/Olive 的 timeline 資料形狀)提煉而來;另加 `ffmpeg-recipes`、`programmatic-video`、`video-timeline-model` 三個 reference。先前 0.8.7 把 redesign 稽核併入 `maggie-design`(改編自 MIT 的 `elayadesign/redesign-skill`):掃描 + 診斷 AI 通病/缺失狀態/雜亂程式碼,再經 in-place/verify 流程就地修正,因此「稽核/打磨/升級既有站點」會路由到 maggie-design。先前 0.8.6 補齊兩個設計技能:加入完整的生產 reference,並把設計來源明確為 Figma 非必需(實時 URL、既有程式碼、Figma、或純截圖)。先前 0.8.5 採納兩個設計技能(改編自 MIT 的 `elayadesign/ai-design-skills`):maggie-landing-page(轉化策略 + 視覺系統紀律,具體數值為可被專案 DESIGN.md 覆蓋的預設值;由 maggie-design 生產與驗證)與 maggie-design-motion(靜態設計→像素級動畫 HTML 複刻→匯出 MP4 + GIF)。先前 0.8.4 將技能庫擴展成完整的「一人用 AI 跑一家公司」C-suite,新增八條職能線:maggie-sales(CRM 中立的銷售管線)、maggie-finance(只解釋的 CFO 指標,含確定性 maggie_finance.py 計算器)、maggie-support(面向客戶的客戶成功)、maggie-product(路線圖 + 預註冊實驗)、maggie-fundraising(pitch deck、data room、投資人更新)、maggie-legal(ToS/隱私/DPA/SOC2 readiness)、maggie-monetization(定價/paywall/dunning,金錢操作需人工批准)、maggie-orchestrator(CEO 編排層:將目標拆成帶審批門的 mission 分派到各職能線,含 maggie_orchestrator.py 規劃器)。先前 0.8.3 將 cmo.so/www 開發經驗沉澱進技能庫:新增兩個技能 maggie-skill-authoring(依 catalog 與 MCP 工具白名單撰寫/校驗社區營銷 playbook 技能)與 maggie-feedback-triage(反饋的團隊接收端:決策/定級/分派、PRD + issue 流程、24/7 隊列守護);新增多個 reference(OAuth PKCE 登入、Expo 約定與 Router 骨架、MCP evals 測試框架、conversation cards、WordPress→Firebase 認證遷移手冊);並增補 maggie-assistant-tools(tool≤user 每輪工具過濾、agent loop 安全不變量、preview-only 橋接代理、MCP resources/prompts、注解加固、CLI 與 ChatGPT 插件打包)、maggie-auth-reference(自清理登入測試套件、自簽單次令牌)、maggie-mobile-app(OTA 開機拉取、live==built 部署守衛)、maggie-deployment(pm2 單包發布、live-hash 校驗)與 maggie-ops(template-as-code 郵件同步)。先前 0.8.2 新增 maggie-spirit(為 AI 品牌生成平面像素 spirit/sprite 系統)與 goal-led AI agent 設計參考;0.8.0 為安全優化 roadmap(scan、image-patch、標頭評分、SSVC、GDPR 日誌規則、新 MCP 工具)。先前 0.8.0 為安全優化 roadmap(scan、image-patch、標頭評分、SSVC、GDPR 日誌規則、新 MCP 工具),含 fix-plan 修補風險分級 T1–T4(判斷補丁是否安全套用)與 host-impact 補丁影響面量測(哪些服務需重啟、是否需重開機)、public-check、Maggie Security、batch 31 排程 agent routines、報告收件匣與無人值守 agent 限制、PHP runtime 盤點與主機鎖、pull 網站改寫、query 點子與社群佇列檢查，batch 30 MaggieDash 先裝核心、release privacy gate、feedback 備援、社群連結與圖片檢查、大量頁面產生與 app 檢查，batch 29 預約規則、部署、app QA、助理與社群功能，以及 batch 28 產品 app、assistant tools、mobile、product release、skills catalog 與 bridge host 功能、batch 26、27 feedback 修正與 idempotent `maggie update`與 Booking email-template 合約/CLI validator 同步。本版完成 2026-09-24 booking feedback batch：Calendar grid/free-slot entry 與 host modal focus-restoration evidence。本版新增 Starter marketplace provisioning plan；本版完成 feedback batches 22、23：加入 declarative browser interaction evidence、design markup/style/content-key gate、compiler source anchoring、entity double-escape protection、`update --help` read-only、noindex hypothetical indexability、可過期 baseline acknowledgement、DOM-only link extraction regression，以及 MaggieDash 0.2.10 的 content override、短期 browser session、CDN-safe public-session cache contracts。authentication 和 isolated dev server 仍由 host 負責。`maggie-booking` workflow、傳統 email/password、Stripe、Resend、public booking flow 及既有安全邊界維持不變；provider-bound video generation 與 transactional delivery 仍記錄在 issue #39，前一批延後功能仍在 issue #38，不會直接成為 active memory。0.7.33 記錄並審查一筆更正回饋；確認 0.7.32 已正確處理 batch index，沒有修改已驗證的 shared behavior。0.7.32 加入 feedback 一基索引支援、MaggieDash panel 的 source/freshness/error evidence、Consent Mode 與實際 write/no-op reconciliation、SEO baseline recapture、sitemap origin rebasing、content-diff scope、migration ledger reconciliation、deployment credential preflight、release provenance/runtime preflight、icon release gate、opt-in blog auto-publish gate，以及 least-privilege VPS deployer。也修復 top-level `maggie feedback` dispatcher parity。0.7.31 加入 update 差異摘要、atomic installer copy、subset install manifest merge，以及 inventory drift 時 doctor non-zero。0.7.30 補上 feedback tracking issue closeout gate 與 batch 14/15 audit ledger。0.7.29 修正 npm package landing page 使用英文 README；0.7.28 加入 versioned Gemini model policy、明確 fallback 與 provenance、bundled-first CLI
+目前 npm 版本為 `0.8.16`,把先前僅 README 的 repo 解稀疏為完整源碼並逐一代碼級研究(驗證所有 reference 並補 net-new:Remotion delayRender gate、opentoonz BSD 可移植的 xsheet 時間線、綠幕/LUT/卡拉OK 濾鏡、whisper turbo + 字幕旗標、Babylon NullEngine 無頭),並新增 maggie_image.py trim/auto-orient 與更豐富的 captions 旗標。先前 0.8.15 補齊除 Q 外所有設計類的覆蓋:maggie_image.py 新增 svg-export(Inkscape)/composite/look,新增 design-editor-patterns reference,並擴充 web-animation/generative-visuals/video-timeline;docs/design-coverage-audit.md 逐 repo 確認。先前 0.8.14 新增 `maggie-ai-media`(去背/放大/arm's-length ComfyUI 的命令生成器,自身不下載模型、不用 GPU),讓 Q 類達到同級;重型執行仍受成本把關。先前 0.8.13 新增 AI 媒體生成規劃 reference(Q 類:自託管模型概覽、程式碼授權≠權重授權規則、商用安全子集 rembg/Real-ESRGAN、成本/BYOK 守則;實際建置待明確許可)。先前 0.8.12 收尾其餘設計類:新增 `maggie-image`(影像處理 CLI:尺寸/OG/favicon/WebP-AVIF/svg/montage)與 web 動效、生成式視覺、圖像配方、字幕四個 reference。先前 0.8.11 收尾 N 類影片功能:`maggie_video.py compile` 再補上視頻 clip ken-burns、sidechain ducking、以及非透明度 overlay 效果(move/slide、圖片 zoom),並新增 `autocut` 子命令(silencedetect 去靜音)。先前 0.8.10 新增圖片 ken-burns + 文字透明度關鍵幀與 `captions`(whisper)子命令。先前 0.8.9 新增 `compile`(宣告式 timeline JSON→單一 ffmpeg 命令)。先前 0.8.8 新增 `maggie-video`(製作與處理行銷影片:剪輯、字幕、裁切成各渠道比例、浮水印、轉碼、響度正規化、clip 組裝、程式化動畫),附一個實測可用的 ffmpeg 配方 CLI(`maggie_video.py`),從參考庫的影片專案(moviepy/editly/ffmpeg 命令層、Revideo/Motion Canvas/CCapture 程式化渲染、MLT/OpenShot/Olive 的 timeline 資料形狀)提煉而來;另加 `ffmpeg-recipes`、`programmatic-video`、`video-timeline-model` 三個 reference。先前 0.8.7 把 redesign 稽核併入 `maggie-design`(改編自 MIT 的 `elayadesign/redesign-skill`):掃描 + 診斷 AI 通病/缺失狀態/雜亂程式碼,再經 in-place/verify 流程就地修正,因此「稽核/打磨/升級既有站點」會路由到 maggie-design。先前 0.8.6 補齊兩個設計技能:加入完整的生產 reference,並把設計來源明確為 Figma 非必需(實時 URL、既有程式碼、Figma、或純截圖)。先前 0.8.5 採納兩個設計技能(改編自 MIT 的 `elayadesign/ai-design-skills`):maggie-landing-page(轉化策略 + 視覺系統紀律,具體數值為可被專案 DESIGN.md 覆蓋的預設值;由 maggie-design 生產與驗證)與 maggie-design-motion(靜態設計→像素級動畫 HTML 複刻→匯出 MP4 + GIF)。先前 0.8.4 將技能庫擴展成完整的「一人用 AI 跑一家公司」C-suite,新增八條職能線:maggie-sales(CRM 中立的銷售管線)、maggie-finance(只解釋的 CFO 指標,含確定性 maggie_finance.py 計算器)、maggie-support(面向客戶的客戶成功)、maggie-product(路線圖 + 預註冊實驗)、maggie-fundraising(pitch deck、data room、投資人更新)、maggie-legal(ToS/隱私/DPA/SOC2 readiness)、maggie-monetization(定價/paywall/dunning,金錢操作需人工批准)、maggie-orchestrator(CEO 編排層:將目標拆成帶審批門的 mission 分派到各職能線,含 maggie_orchestrator.py 規劃器)。先前 0.8.3 將 cmo.so/www 開發經驗沉澱進技能庫:新增兩個技能 maggie-skill-authoring(依 catalog 與 MCP 工具白名單撰寫/校驗社區營銷 playbook 技能)與 maggie-feedback-triage(反饋的團隊接收端:決策/定級/分派、PRD + issue 流程、24/7 隊列守護);新增多個 reference(OAuth PKCE 登入、Expo 約定與 Router 骨架、MCP evals 測試框架、conversation cards、WordPress→Firebase 認證遷移手冊);並增補 maggie-assistant-tools(tool≤user 每輪工具過濾、agent loop 安全不變量、preview-only 橋接代理、MCP resources/prompts、注解加固、CLI 與 ChatGPT 插件打包)、maggie-auth-reference(自清理登入測試套件、自簽單次令牌)、maggie-mobile-app(OTA 開機拉取、live==built 部署守衛)、maggie-deployment(pm2 單包發布、live-hash 校驗)與 maggie-ops(template-as-code 郵件同步)。先前 0.8.2 新增 maggie-spirit(為 AI 品牌生成平面像素 spirit/sprite 系統)與 goal-led AI agent 設計參考;0.8.0 為安全優化 roadmap(scan、image-patch、標頭評分、SSVC、GDPR 日誌規則、新 MCP 工具)。先前 0.8.0 為安全優化 roadmap(scan、image-patch、標頭評分、SSVC、GDPR 日誌規則、新 MCP 工具),含 fix-plan 修補風險分級 T1–T4(判斷補丁是否安全套用)與 host-impact 補丁影響面量測(哪些服務需重啟、是否需重開機)、public-check、Maggie Security、batch 31 排程 agent routines、報告收件匣與無人值守 agent 限制、PHP runtime 盤點與主機鎖、pull 網站改寫、query 點子與社群佇列檢查，batch 30 MaggieDash 先裝核心、release privacy gate、feedback 備援、社群連結與圖片檢查、大量頁面產生與 app 檢查，batch 29 預約規則、部署、app QA、助理與社群功能，以及 batch 28 產品 app、assistant tools、mobile、product release、skills catalog 與 bridge host 功能、batch 26、27 feedback 修正與 idempotent `maggie update`與 Booking email-template 合約/CLI validator 同步。本版完成 2026-09-24 booking feedback batch：Calendar grid/free-slot entry 與 host modal focus-restoration evidence。本版新增 Starter marketplace provisioning plan；本版完成 feedback batches 22、23：加入 declarative browser interaction evidence、design markup/style/content-key gate、compiler source anchoring、entity double-escape protection、`update --help` read-only、noindex hypothetical indexability、可過期 baseline acknowledgement、DOM-only link extraction regression，以及 MaggieDash 0.2.10 的 content override、短期 browser session、CDN-safe public-session cache contracts。authentication 和 isolated dev server 仍由 host 負責。`maggie-booking` workflow、傳統 email/password、Stripe、Resend、public booking flow 及既有安全邊界維持不變；provider-bound video generation 與 transactional delivery 仍記錄在 issue #39，前一批延後功能仍在 issue #38，不會直接成為 active memory。0.7.33 記錄並審查一筆更正回饋；確認 0.7.32 已正確處理 batch index，沒有修改已驗證的 shared behavior。0.7.32 加入 feedback 一基索引支援、MaggieDash panel 的 source/freshness/error evidence、Consent Mode 與實際 write/no-op reconciliation、SEO baseline recapture、sitemap origin rebasing、content-diff scope、migration ledger reconciliation、deployment credential preflight、release provenance/runtime preflight、icon release gate、opt-in blog auto-publish gate，以及 least-privilege VPS deployer。也修復 top-level `maggie feedback` dispatcher parity。0.7.31 加入 update 差異摘要、atomic installer copy、subset install manifest merge，以及 inventory drift 時 doctor non-zero。0.7.30 補上 feedback tracking issue closeout gate 與 batch 14/15 audit ledger。0.7.29 修正 npm package landing page 使用英文 README；0.7.28 加入 versioned Gemini model policy、明確 fallback 與 provenance、bundled-first CLI
 dispatch、`maggie --version`、feedback batch review 聚合與重複偵測、marketplace enrichment evidence、booking
 worker/resolver evidence，以及 read-only migration preflight。它也包含 host-owned mobile app surface contract、signed-in
 camera-state QA、直接 Astro route resolution、correlated feedback batch，以及

package/README.md

@@ -89,7 +89,7 @@ Maggie keeps the existing project foundation and asks for decisions before
 shared routes, analytics, or publishing boundaries change. The current
 package ships 39 installable skills and a local-first MaggieDash foundation.
 
-This package is `0.8.14` — adds `maggie-ai-media` (bg removal / upscaling / arm's-length ComfyUI as a command generator; no model downloads or GPU itself). `0.8.13` adds the AI-media-generation planning reference (category Q: licensing + safe subset + cost guardrails; build gated on go-ahead). `0.8.12` adds `maggie-image` (process images via an ImageMagick recipe CLI) + references for web animation, generative visuals, image recipes, and captions. `0.8.11` - `maggie_video.py compile` now also does video-clip ken-burns, sidechain ducking, and non-opacity overlay effects (move/slide, image zoom), and a new `autocut` subcommand removes silent gaps. `0.8.10` added image ken-burns + text opacity keyframes and the `captions` (whisper) subcommand. `0.8.9` added `compile` (declarative timeline JSON -> one ffmpeg command). `0.8.8` added `maggie-video` (produce and process marketing video: cut, caption, resize to channel aspects, watermark, transcode, normalize loudness, assemble clips, render code-drawn animation) with a tested ffmpeg-recipe CLI, distilled from the reference library's video projects; plus references `ffmpeg-recipes`, `programmatic-video`, `video-timeline-model`. `0.8.7` folded a redesign audit into `maggie-design` (from the MIT `elayadesign/redesign-skill`): scan + diagnose generic AI-design patterns/weak states/sloppy code, then fix in place via the in-place/verify flow, so "audit/polish/upgrade an existing site" routes to maggie-design. `0.8.6` completed the two adopted design skills with full production references and made the design source Figma-optional (live URL, existing code, Figma, or a screenshot). `0.8.5` adopted the two design skills (adapted, MIT, from `elayadesign/ai-design-skills`): `maggie-landing-page` (conversion strategy + a visual-system discipline whose values defer to the project's DESIGN.md) and `maggie-design-motion` (static design → pixel-exact animated HTML replica → MP4 + GIF export). `0.8.4` expanded the library into a full "one person runs a company with AI" C-suite with eight new function lanes: `maggie-sales` (CRM-neutral pipeline), `maggie-finance` (explain-only CFO metrics + a deterministic `maggie_finance.py` calculator), `maggie-support` (customer success), `maggie-product` (roadmap + pre-registered experiments), `maggie-fundraising` (deck, data room, investor updates), `maggie-legal` (ToS/privacy/DPA/SOC2 readiness), `maggie-monetization` (pricing/paywalls/dunning; money actions human-approved), and `maggie-orchestrator` (the CEO layer that decomposes a goal into gated missions across the lanes, with a `maggie_orchestrator.py` planner). `0.8.3` folded cmo.so/www development learnings in (maggie-skill-authoring, maggie-feedback-triage, and mobile/MCP/auth references); `0.8.2` added `maggie-spirit`; `0.8.0` was the security optimization roadmap.
+This package is `0.8.16` — de-sparses the earlier README-only repos to full source and studies them at code level (confirm + net-new: Remotion delayRender gate, opentoonz BSD port target, green-screen/LUT/karaoke filters, whisper turbo + caption flags, Babylon NullEngine), plus maggie_image.py trim/auto-orient and richer captions flags. `0.8.15` closed the non-Q design coverage gap: `maggie_image.py` gains svg-export/composite/look, a new design-editor-patterns reference, and growth in web-animation/generative-visuals/video-timeline references. `0.8.14` adds `maggie-ai-media` (bg removal / upscaling / arm's-length ComfyUI as a command generator; no model downloads or GPU itself). `0.8.13` adds the AI-media-generation planning reference (category Q: licensing + safe subset + cost guardrails; build gated on go-ahead). `0.8.12` adds `maggie-image` (process images via an ImageMagick recipe CLI) + references for web animation, generative visuals, image recipes, and captions. `0.8.11` - `maggie_video.py compile` now also does video-clip ken-burns, sidechain ducking, and non-opacity overlay effects (move/slide, image zoom), and a new `autocut` subcommand removes silent gaps. `0.8.10` added image ken-burns + text opacity keyframes and the `captions` (whisper) subcommand. `0.8.9` added `compile` (declarative timeline JSON -> one ffmpeg command). `0.8.8` added `maggie-video` (produce and process marketing video: cut, caption, resize to channel aspects, watermark, transcode, normalize loudness, assemble clips, render code-drawn animation) with a tested ffmpeg-recipe CLI, distilled from the reference library's video projects; plus references `ffmpeg-recipes`, `programmatic-video`, `video-timeline-model`. `0.8.7` folded a redesign audit into `maggie-design` (from the MIT `elayadesign/redesign-skill`): scan + diagnose generic AI-design patterns/weak states/sloppy code, then fix in place via the in-place/verify flow, so "audit/polish/upgrade an existing site" routes to maggie-design. `0.8.6` completed the two adopted design skills with full production references and made the design source Figma-optional (live URL, existing code, Figma, or a screenshot). `0.8.5` adopted the two design skills (adapted, MIT, from `elayadesign/ai-design-skills`): `maggie-landing-page` (conversion strategy + a visual-system discipline whose values defer to the project's DESIGN.md) and `maggie-design-motion` (static design → pixel-exact animated HTML replica → MP4 + GIF export). `0.8.4` expanded the library into a full "one person runs a company with AI" C-suite with eight new function lanes: `maggie-sales` (CRM-neutral pipeline), `maggie-finance` (explain-only CFO metrics + a deterministic `maggie_finance.py` calculator), `maggie-support` (customer success), `maggie-product` (roadmap + pre-registered experiments), `maggie-fundraising` (deck, data room, investor updates), `maggie-legal` (ToS/privacy/DPA/SOC2 readiness), `maggie-monetization` (pricing/paywalls/dunning; money actions human-approved), and `maggie-orchestrator` (the CEO layer that decomposes a goal into gated missions across the lanes, with a `maggie_orchestrator.py` planner). `0.8.3` folded cmo.so/www development learnings in (maggie-skill-authoring, maggie-feedback-triage, and mobile/MCP/auth references); `0.8.2` added `maggie-spirit`; `0.8.0` was the security optimization roadmap.
 
 The 0.7.59 release made `maggie security host-audit` (the deep
 lynis audit over SSH) robust on hosts where a storage test can stall: lynis runs

package/bundled-references/captions-transcription.md

@@ -9,7 +9,7 @@ Distilled from the reference library's R-category projects.
 
 | Engine | What / how | License | Reach for it when |
 |---|---|---|---|
-| openai-whisper | The reference Python CLI: `whisper IN --model base --output_format srt --output_dir DIR` | MIT | the default; simplest install (`pip install openai-whisper`) |
+| openai-whisper | `whisper IN --model turbo --output_format srt --output_dir DIR` (default model is now **`turbo`** = large-v3-turbo; formats `txt/vtt/srt/tsv/json`) | MIT | the default; simplest install (`pip install openai-whisper`) |
 | whisper.cpp | C/C++ port, fast on CPU, GGML models: `whisper-cli -m model.bin -f IN -osrt -of OUTBASE` | MIT | no GPU / fastest CPU / embedding in a pipeline |
 | whisperX | whisper + forced alignment for **word-level timestamps** and diarization: `whisperx IN --model large-v3 --output_format srt --output_dir DIR` | BSD-4 | word-accurate caption timing or speaker labels |
 
@@ -38,4 +38,24 @@ Model size trades speed for accuracy: `tiny` / `base` (fast, rough) -> `small` /
   API is ever wired in instead, the global spend-first rule applies - state the cost.
 - Keep caption text inside the safe area (see `maggie-video` channel presets).
 
+## Caption-shaping flags (built into openai-whisper)
+
+The core CLI already does what we previously routed to whisperX or post-processing
+(`maggie_video.py captions` exposes these):
+
+- `--word_timestamps True` - word-level timings (whisperX still wins on alignment +
+  diarization, but core whisper now has them).
+- `--highlight_words True` - **karaoke-style** per-word highlight in SRT/VTT (the
+  social-caption effect); render the ASS with ffmpeg `subtitles`.
+- `--max_line_width` / `--max_line_count` / `--max_words_per_line` - native line
+  wrapping so captions fit the channel safe-area, no post-processing.
+- `--initial_prompt "<brand, product names>"` - bias the model toward correct brand/
+  jargon spelling (the concrete mitigation for the mishears-brand-terms risk; still
+  human-review).
+- Anti-hallucination over music/silence: `--hallucination_silence_threshold`,
+  `--compression_ratio_threshold`, `--logprob_threshold`, `--no_speech_threshold` -
+  worth setting for music-bed marketing clips.
+- `--task translate` (foreign audio -> English subs), `--clip_timestamps` (only given
+  ranges).
+
 See `maggie-video` for the subcommands and the burn-in / soft-mux / compile options.

package/bundled-references/design-editor-patterns.md

@@ -0,0 +1,104 @@
+# Design-editor patterns
+
+Borrowable logic and creativity from open-source design, whiteboard, vector and raster
+**editors** (reference library category P), for making and handling marketing visuals
+programmatically. We adopt **concepts and data shapes, not code** (most are GPL/AGPL/
+MPL or custom-licensed; the few MIT ones are reimplemented, not bulk-copied). Pairs
+with [image recipes](image-recipes.md), [generative visuals](generative-visuals.md),
+and [AI media generation](ai-media-generation.md).
+
+## Non-destructive edit recipe (darktable)
+
+darktable (GPL) edits by a **replayable history stack of named module invocations**,
+stored in an **XMP sidecar** beside the untouched original - never mutating the source;
+the image is reconstructed by replaying the stack. Borrow this as a maggie "image edit
+recipe": a small JSON `[ {op, params}, ... ]` stored next to the original and replayed
+by `maggie_image.py` (each op -> an ImageMagick/sharp command). It is **reproducible,
+auditable, re-editable, and batchable** - apply one "look" recipe to every photo in a
+campaign for consistent brand grading. `darktable-cli <input> [sidecar.xmp]
+<output-dir>` renders from such a stack (XMP optional/auto-discovered, final arg is a
+**directory**; GPL binary, arm's-length). Two source details worth adopting: **params
+versioning** (each module keeps `legacy_params` for forward-compatible migration -
+version your recipe ops) and **"Styles" = saved reusable history** (name a look once,
+apply to many).
+
+## LLM-generatable scene / diagram formats
+
+- **Excalidraw** (`.excalidraw` JSON, MIT) - a flat list of elements (`{type:
+  rectangle|ellipse|diamond|text|line|arrow|freedraw|image|frame, x, y, width, height,
+  strokeColor, backgroundColor, fillStyle, roughness, seed, roundness, groupIds,
+  boundElements}`) with a separate `files` asset map and ratio-based arrow/label
+  bindings that survive resize. **Emit it from a prompt** for diagrams, flowcharts,
+  annotated screenshots -> open in-app or render to SVG/PNG, with an on-brand hand-drawn
+  (rough.js) look.
+- **drawio / mxGraph** (`mxGraphModel` XML, Apache-2.0 code; diagram output
+  unrestricted) - `<mxCell>` nodes with `value` (label), a compact `style` DSL
+  (`rounded=1;fillColor=#dae8fc;strokeColor=#6c8ebf;`), `vertex="1"` or `edge="1"` with
+  `source`/`target`, and `<mxGeometry>`. Trivially generatable for org charts, funnels,
+  architecture diagrams; render headless to SVG/PNG (don't redistribute their icon/
+  stencil sets).
+
+## SVG as the document (Method-Draw) + safe ingest
+
+Method-Draw (MIT, from SVG-Edit) treats the **live SVG DOM as the single source of
+truth** - `XMLSerializer` is the save format; no separate scene JSON. Two borrows:
+(a) the simplest possible model for programmatic marketing SVGs (build the DOM,
+serialize); (b) a **sanitize whitelist** (`svgWhiteList_`) of allowed SVG elements/
+attributes - essential for safely ingesting user- or AI-supplied SVG before rendering
+(strips scripts/unknowns); pair with svgo in [image recipes](image-recipes.md). Its
+command `apply/unapply` history is the generic edit-history pattern (and the backbone
+of the edit-recipe above).
+
+## Design-doc patterns (penpot, tldraw)
+
+- **penpot** (MPL, concept) - a design doc in **open standards (SVG/CSS/HTML/JSON)**
+  with first-class **Design Tokens as the single source of truth** between brand and
+  code, **Components + Variants** (one card, many variants = reusable marketing
+  templates), CSS Grid/Flex layout on the doc, and an **MCP server** making the doc
+  agent-readable. Strong alignment with Maggie's brand-kit + MCP posture.
+- **tldraw** (custom license, concept) - an infinite-canvas **reactive record store**
+  of typed shape records (geometry + props + `parentId` + fractional index) with
+  bindings/assets/pages, and the explicit framing that **an LLM agent operates on the
+  canvas document as data** - reinforcing the "generate the scene JSON" approach.
+
+## AI-edit vocabulary (krita-ai-diffusion)
+
+krita-ai-diffusion (GPL, concept) bolts generative AI onto a layer/selection editor
+over a ComfyUI backend. The borrowable vocabulary for precise AI image edits (beyond
+full-frame generation): **mask/selection-scoped generative fill** (inpaint, expand,
+add/remove inside a region), **layer-as-region prompting** (a text prompt per image
+area defined by a layer), and **conditioning** via ControlNet (scribble/lineart/canny/
+pose/depth/segmentation) and IP-Adapter (reference image, style/composition transfer).
+The plugin-over-backend boundary (editor model stays; AI is a swappable backend)
+mirrors the arm's-length posture in [AI media generation](ai-media-generation.md).
+
+## Licensing
+
+Invoke binaries at arm's length (never vendor): inkscape, darktable (GPL). Concepts
+only: penpot (MPL), tldraw (custom), krita/krita-ai-diffusion (GPL), drawio icon sets.
+MIT and freely borrowable as ideas (still reimplement, don't bulk-copy): excalidraw,
+Method-Draw, rough.js, Pinta. drawio code Apache-2.0; its diagram output is
+unrestricted.
+
+## From full-source study (net-new)
+
+- **tldraw** - the LLM mutation surface is `editor.createShapes/updateShapes/
+  deleteShapes` (batch partial mutations); the store ships schema **migration
+  sequences**, a composable **validation layer**, and a **side-effects manager**
+  (before/after create/change/delete hooks). `parentId` can be a page *or* a shape;
+  pages carry a fractional `index`. Adopt the batch-mutation + validate + migrate
+  shape as the model for an agent-written canvas doc.
+- **penpot** - **design tokens are W3C DTCG** (import/export + aliasing), components
+  have **variants**, shapes carry **flex/grid layout**; the agent boundary is an
+  **official MCP server talking to a sandboxed in-app plugin over WebSocket** - a
+  strong arm's-length model for making a design doc agent-writable (mirrors Maggie's
+  own MCP posture).
+- **drawio** - the `mxConstants` STYLE_* key table is a **validated allow-list** for
+  LLM-emitted styles; `mxEdgeStyle` routing + hierarchical/tree/organic **auto-layout**
+  mean a generated node+edge list can self-position; `.drawio` pages are often
+  **deflate+base64** compressed.
+- **krita** - an **adjustment/filter layer holds a serialized filter config** = a
+  non-destructive recipe living in the layer stack; ops are **mask/selection-scoped**;
+  `.kra` is `zip(manifest XML + per-layer blobs)`; a **Python scripting API** and a
+  JSON-manifest plugin boundary let a generative backend register like any filter -
+  the "AI as a swappable backend" posture.

package/bundled-references/ffmpeg-recipes.md

@@ -128,3 +128,24 @@ maggie tool maggie_video.py compile --spec timeline.json   # declarative timelin
 maggie tool maggie_video.py --run encode --input in.mov --output out.mp4   # actually runs
 maggie tool maggie_video.py presets
 ```
+
+## More filters worth reaching for (confirmed in libavfilter)
+
+- **Green-screen / keying** (UGC, product on a brand background):
+  `chromakey=0x00FF00:0.1:0.0` (or `colorkey`), clean fringes with `despill`.
+- **Fade to/from black** (distinct from clip crossfade): `fade=t=in:st=0:d=0.5`,
+  `fade=t=out:st=9.5:d=0.5`.
+- **Brand color grade / LUT**: `lut3d=brand.cube`, or `curves`/`eq`/`colorbalance`/`hue`.
+- **Karaoke captions**: `subtitles=in.ass` with a full ASS file (word highlight) -
+  pairs with whisper `--highlight_words` ([captions](captions-transcription.md)).
+- **Focus/credits**: `gblur` + `vignette`; `scroll` for scrolling outros.
+- **GPU fast path** when hardware is present: `overlay_cuda`/`overlay_vulkan`,
+  `xfade_opencl` for long renders.
+
+## Target-filesize social presets (HandBrake idea)
+
+Beyond aspect, encode to **fit under a platform's upload cap**: pick a two-pass video
+bitrate from the cap and duration - `bitrate_bits = size_bytes * 8 / duration - audio`
+- e.g. "25 MB / 30 s / 1080p". A named preset bundles `(upload cap, max duration) ->
+resolution + bitrate`. (HandBrake ships exactly these as "Social 25 MB 30 Seconds"
+presets.)

package/bundled-references/generative-visuals.md

@@ -92,3 +92,53 @@ zdog, pts, hydra, regl, Babylon.js, react-three-fiber, fogleman/primitive, svg.j
 nannou - licenses in the table above. Capture pipeline: see
 [programmatic video](programmatic-video.md), [design-motion production](design-motion-production.md),
 [ffmpeg recipes](ffmpeg-recipes.md).
+
+## More patterns (from the desktop/editor study)
+
+- **The shared sketch model** (processing / openFrameworks / nannou / p5): a `setup()`
+  once + per-frame `draw()` loop with seeded `random()/noise()`. openFrameworks (MIT)
+  splits it into `setup()/update()/draw()` (sim vs render); nannou (MIT/Apache)
+  formalizes Model-Update-View. The reusable discipline: **keep state separate from the
+  render fn, index by frame, seed all randomness** - that is what makes a loop
+  reproducible and headless-capturable. Default to p5/three/regl in the browser;
+  reach for a desktop framework only for heavy GPU throughput or offline high-res
+  render the browser can't sustain. (Processing non-core is GPL - reimplement.)
+- **Hand-drawn scene JSON (excalidraw, MIT).** `.excalidraw` is a flat list of
+  elements (`{type, x, y, w, h, strokeColor, backgroundColor, fillStyle:
+  hachure|cross-hatch|solid|zigzag, roughness, seed, roundness}`) rendered with
+  rough.js. It is **LLM-generatable**: emit the JSON for a diagram/flowchart/annotated
+  screenshot, then render to SVG/PNG. The rough.js mapping for an on-brand hand-made
+  look: `seed` (determinism), `fillStyle`, `fillWeight = strokeWidth/2`,
+  `hachureGap = strokeWidth*4`, `roughness`, `disableMultiStroke` for non-solid strokes.
+- **Vector shape morphing** (synfig concept): tween between two paths by interpolating
+  each control point's vertex + in/out tangents + width - smooth logo/blob/path morphs.
+- **Repeater / radial stamp** (Lottie concept): one shape -> N transformed copies with
+  an incremental transform + opacity ramp (bursts, confetti, radial brand patterns) -
+  cheap, on-brand, deterministic.
+- For AI-assisted editing vocabulary (mask/selection-scoped generation, layer-as-region
+  prompting, ControlNet/IP-Adapter conditioning), see
+  [AI media generation](ai-media-generation.md) and
+  [design editor patterns](design-editor-patterns.md).
+
+### Headless capture + 3D export (from full-source study)
+
+How to render these engines **without a screen**, and extra output formats worth
+knowing:
+
+- **Babylon.js `NullEngine`** (Apache-2.0) - first-class **server-side headless**
+  rendering (no GPU/canvas); the cleanest headless story for render-automation jobs,
+  plus a built-in **procedural-texture library** (brick/cloud/fire/marble/noise/wood),
+  `CreateScreenshotUsingRenderTarget` for 4K+ stills from a small viewport, and
+  **NodeMaterial** (shaders serialize to JSON - an LLM-emittable look). This is the
+  concrete reason to pick Babylon over three for automation.
+- **three.js** (MIT) - `readRenderTargetPixels()`/`...Async` read pixels from an
+  offscreen `WebGLRenderTarget` (true headless high-res, not on-screen canvas);
+  **SVGRenderer** emits a 3D scene as editable SVG; **USDZExporter** turns a 3D product
+  into a free **iOS AR Quick Look** asset; **GPUComputationRenderer** (ping-pong FBO
+  sim: flocking/fluid) = a living, seeded generative hero background.
+- **openFrameworks** (MIT) - `ofFbo` offscreen framebuffer for arbitrary-res stills;
+  **ofxEmscripten** compiles a desktop sketch to WebAssembly for the browser;
+  **ofxSvg** parses + emits structured SVG.
+- **Processing** - `PSurfaceNone` (headless) and `createGraphics(big)` (render larger
+  than the screen -> high-res still). License at feature granularity: core raster is
+  LGPL, but **SVG/PDF export is in non-core (GPL)** - reimplement, don't vendor.

package/bundled-references/image-recipes.md

@@ -87,3 +87,50 @@ sharp (Apache-2.0), squoosh codecs (Apache-2.0), imagemin (MIT), cwebp/avifenc
 (BSD). **Avoid vendoring code from** Upscayl (AGPL-3.0) - for AI upscaling invoke the
 separate `realesrgan-ncnn-vulkan` binary instead. sharp/squoosh quality defaults are
 adopted as the preset numbers.
+
+## SVG beyond resvg/ImageMagick (Inkscape, arm's-length)
+
+Inkscape's headless CLI (GPL app, invoked like ImageMagick) fills gaps resvg and IM
+can't: **font outlining, PDF/EPS, and per-id slicing**. `maggie_image.py svg-export`
+generates these.
+
+```bash
+inkscape in.svg --export-type=png --export-filename=out.png --export-width=1200 \
+  --export-area-drawing --export-background=#ffffff --export-background-opacity=1
+inkscape in.svg --export-type=pdf --export-filename=out.pdf                      # print collateral
+inkscape in.svg --export-text-to-path --export-type=svg --export-filename=flat.svg  # outline fonts for handoff
+inkscape in.svg --export-id=badge --export-id-only --export-type=png --export-filename=badge.png  # slice one element
+inkscape in.svg --export-type=png --export-dpi=300 --export-area-page --export-filename=print.png
+```
+
+## Layered composite + named looks (Pinta vocabulary)
+
+Blend modes map from a raster editor's layer stack onto ImageMagick `-compose`:
+
+```bash
+convert base.png overlay.png -gravity center -compose Multiply -composite out.png
+# modes: Over, Multiply, Screen, Overlay, Darken, Lighten, ColorDodge, ColorBurn,
+#        HardLight, SoftLight, Difference, Exclusion
+```
+
+Named "looks" (each a one-line style preset - a lightweight take on darktable's
+non-destructive edit recipe): `grayscale` (`-colorspace Gray`), `sepia`
+(`-sepia-tone 80%`), `posterize` (`-posterize 4`), `duotone` (`-colorspace Gray
+-level-colors '#1a1a2e','#e94560'`), `vignette` (`-background black -vignette 0x20`),
+`oilpaint` (`-paint 4`), `sketch` (`-colorspace Gray -charcoal 1`), `blur`/`sharpen`.
+`maggie_image.py composite` and `maggie_image.py look` generate these. For a
+reusable, batchable edit *recipe* (a JSON stack of named ops applied to many photos),
+see [design editor patterns](design-editor-patterns.md).
+
+## More ImageMagick ops worth baking in (confirmed in source)
+
+- **`-auto-orient`** first in every pipeline - fixes sideways phone photos by EXIF
+  (`maggie_image.py` resize/social/trim already prepend it).
+- **`-thumbnail WxH`** - metadata-stripping fast resize; a better default than
+  `-resize` for web/OG thumbnails.
+- **`-fuzz N% -trim +repage`** - trim a uniform border or cut a product/logo out of
+  its background (`maggie_image.py trim`).
+- **`-clut` / `-hald-clut`** - repeatable LUT brand grading (raster equivalent of
+  ffmpeg `lut3d`).
+- **`-distort perspective '<src coords> <dst coords>'`** - place a screenshot into a
+  device mockup.

package/bundled-references/programmatic-video.md

@@ -78,3 +78,25 @@ apply.
 `libx264 -crf 14-20 -preset slow/medium -pix_fmt yuv420p -movflags +faststart`,
 even dimensions, plus 960px / 800x500 GIF deliverables. For transparent output,
 Revideo uses per-format pixel formats (`yuva420p` WebM, `yuva444p10le` ProRes).
+
+## Remotion's real renderer (from source) and the ideas to borrow
+
+Source study (Remotion License - concept only, no vendoring) shows Remotion is **not**
+the base64-PNG pipe we lump with Revideo/Motion Canvas: it drives headless Chromium
+over CDP, captures frames via **shared memory**, and encodes with a **per-platform
+native Rust sidecar that statically links FFmpeg** (a long-running streaming process),
+plus a bundled `ffmpeg` for the final stitch. Borrowable ideas (improve our Part C):
+
+- **A readiness gate, not just a fake clock.** `delayRender()`/`continueRender()`: a
+  component declares async work pending and the renderer blocks the screenshot until
+  all handles clear, reporting **labeled** open handles on timeout. Add an explicit
+  readiness handshake (plus a per-frame `await document.fonts.ready`) to Part C so no
+  frame is captured before fonts/data/images are ready - prevents FOUT/half-loaded
+  frames.
+- **Tab-pool concurrency**: open N browser pages, each seeking a different frame in
+  parallel (frame range / N).
+- **Chunked parallel encode + seamless concat**: split the frame range, encode chunks
+  independently, concat without re-encode - throughput for long renders.
+- **Kept-open, LRU-cached source-video decoders**: for embedded video assets, keep the
+  decoder open and cache frames so repeated seeks to a timestamp are cheap (our flat
+  timeline currently punts this to raw ffmpeg `-ss`).

package/bundled-references/video-timeline-model.md

@@ -130,3 +130,59 @@ arm's length, never linked or vendored.
   keyframe/ease vocabulary, the licensing note - stable and engine-neutral.
 - **`maggie-video` skill:** channel presets, caption/CTA safe-areas, the operational
   ffmpeg command layer (via `maggie_video.py`), and tool wiring.
+
+## Extensions from the animation-tool study (O-category)
+
+Concepts to grow the model (adopt the data shape / vocabulary; code only from the
+MIT/BSD sources - rlottie MIT, opentoonz/tahoma2d BSD; synfig/friction/wick/Natron
+GPL = concept only):
+
+- **Richer keyframe easing** (synfig). Extend `{t, v, ease}` with: **per-side ease**
+  (`easeIn` != `easeOut`), **`clamped`** (auto-smooth that won't overshoot the keyed
+  value), **`hold`/`constant`**, and optional **TCB** `{tension, continuity, bias}`.
+  Add **`fullRotations`** on a rotation keyframe (wick) to control spin count/direction
+  rather than shortest-path. These precompute to per-frame values for ffmpeg and map to
+  CSS/Lottie beziers. (compile currently interpolates linearly; this is the target
+  vocabulary.)
+- **Lottie JSON as the companion vector-motion-graphics format** (rlottie, MIT). For
+  logo stings, animated icons and kinetic type, use Lottie (After-Effects/Figma
+  exportable, in-browser renderable via lottie-web) alongside the MLT/NLE shape. Borrow
+  its primitives: **Trim-path** (`start%/end%/offset` - the formal version of the
+  "draw a line on" trick in [design-motion production](design-motion-production.md)),
+  **Repeater** (N copies with a transform + opacity ramp), **track mattes**
+  (`Alpha/Luma`), **transform parenting**, and **time remap** (`tm`).
+- **Nested clips / precomps** (wick/lottie). A clip may hold its own timeline with a
+  playback mode (`loop | playOnce | single`) and be instanced - reusable lower-thirds,
+  intro/outro stingers; a composable alternative to the flat track list.
+- **Decomposed transform + animatable pivot** (friction). Split transform into
+  `pos / scale / rot / shear / pivot / opacity`, each its own keyframe curve, with a
+  **movable pivot** for correct scale/rotate around a moving anchor.
+- **Blend-mode / compositing vocabulary** (Natron `MergingFunctionEnum`). The allowed
+  `blendMode` values on overlay clips: Porter-Duff (`Over, Under, In, Out, ATop, Copy,
+  Matte`) + blends (`Multiply, Screen, Overlay, SoftLight, HardLight, ColorBurn,
+  ColorDodge, Difference, Exclusion, Hue, Saturation, Color, Luminosity`).
+- **Xsheet tabular timing** (opentoonz/tahoma2d, **BSD - code reusable with
+  attribution, a direct port target**). An optional rows=frames / columns=assets
+  exposure grid (cells = a drawing exposed for N frames; a *hold* is the same cell
+  repeated). The source implements this (and more of our model than we first credited)
+  in `TXsheet` / `TXshCellColumn` (`std::vector<TXshCell> m_cells`) / `TXshCell`
+  (`{level, TFrameId}`) / `TFrameId` (`{frame, letter}`), plus the cell-op vocabulary
+  (`reverseCells, swingCells, stepCells, reframeCells, timeStretch`). It also backs, as
+  BSD code: the **decomposed transform channels** (`TStageObject::Channel` T_X/T_Y/
+  T_Angle/T_ScaleX/... each a keyframe curve, with parenting + `T_Path` follow-spline),
+  a **richer easing enum** (`TDoubleKeyframe::Type` = Constant / Linear / SpeedInOut /
+  EaseInOut / EaseInOutPercentage / Exponential / Expression / File), **nested
+  sub-xsheets** (`TXshChildLevel` = precomp), and a **node-graph fx schematic**
+  (`FxDag`/`TFx`). So the easing/transform/nested-clip/effect-graph extensions above
+  have a BSD implementation to port, not only GPL concept sources.
+- **Strip/effect model + full blend set** (blender VSE `Strip`, concept). The Video
+  Sequence Editor's `Strip` (input1/input2, effect strips CROSS/WIPE/GLOW/SPEED/
+  ADJUSTMENT/TEXT, a Photoshop-complete `StripBlendMode`) and **F-curve modifier stack**
+  (noise/cycles/limits/stepped) + **drivers** (property-driven channels) are strong
+  concept borrows for overlay/effect clips.
+- **OTIO interchange** (kdenlive, concept). Document a mapping to OpenTimelineIO (the
+  ASWF-standard timeline interchange): gap=blank, stack=track, externalReference=asset
+  src - so the model can round-trip with real NLEs.
+- **Optional node-graph "effect graph"** (Natron, concept) for branched layered effects
+  (mask -> blur -> merge) that a linear track can't express - heavy; a documented
+  escape hatch, not the default.

package/bundled-references/web-animation-patterns.md

@@ -362,3 +362,37 @@ Most are MIT and safe to reference/learn from freely. The four to flag: **GSAP**
 (proprietary no-charge), **typed.js** (GPL), **animate.css** (Hippocratic 2.1), and
 **synfig/friction** (GPL desktop tools). For the flagged ones we adopt the *pattern*
 and author our own code.
+
+## Authoring motion visually, playing it in code (Theatre.js)
+
+[`theatre-js/theatre`](https://github.com/theatre-js/theatre) (Apache-2.0) is a
+motion-design *editor* (`@theatre/studio`) that writes a plain serializable **project
+state**, replayed by a tiny headless runtime (`@theatre/core`) against three.js,
+HTML/SVG, or any JS variable. The borrowable pattern is the **declarative keyframe
+document**, not the editor: store motion as JSON - a `sequence` with `length` +
+`subUnitsPerUnit` (an fps snap-grid) -> **tracks keyed by property path** ->
+**keyframes** each carrying `{ position, value, handles:[lx,ly,rx,ry],
+type:'bezier'|'hold' }` (per-keyframe bezier handles = per-segment easing baked into
+data; `value` may be non-numeric). Playback mirrors CSS: `play({ iterationCount,
+range, rate, direction:'alternateReverse' })`. Author a brand motion once, commit the
+JSON as a reusable asset, replay it deterministically with no editor shipped - the
+authoring story for `maggie-design-motion` (keep the motion *spec* as data, not
+imperative GSAP calls). Reach for it when motion must be re-edited by non-coders or
+reused across pages; a one-shot reveal stays plain CSS/GSAP.
+
+## Interactive, input-driven motion (Rive state machines)
+
+[`rive-app/rive-runtime`](https://github.com/rive-app/rive-runtime) (MIT) drives
+vector animation with a **state machine** instead of fixed playback - the pattern for
+interactive hero elements, product configurators and mascots. Model: a small **input**
+surface (`bool`, `number`, `trigger`) -> **transitions** guarded by
+**conditions/comparators** (`number > 0.5`, trigger fired) -> **states** (play an
+animation, or a **BlendState1D** that blends along a number input). A **listener**
+layer maps pointer/scroll/keyboard events to input changes; states run in parallel
+**layers**; transitions carry exit-time, ease interpolators, and **random weights**
+for non-repetitive idles. Borrow it in the browser without Rive: model an interactive
+hero as `inputs -> conditions -> states` - declare inputs (`hovered:bool`,
+`scroll:0..1`, `clicked:trigger`), bind DOM events to them, and switch CSS/WAAPI/GSAP
+timelines by state rather than scripting each handler; a number input + interpolation
+gives scroll-linked response, random-picked idle transitions give a "living" element.
+(Rive's own MIT player is also a legitimate option.)

package/bundled-skills/maggie-image/SKILL.md

@@ -2,7 +2,7 @@
 name: maggie-image
 description: Process marketing images - resize and fit to channel/OG sizes, convert to WebP/AVIF, make favicon and OG-image sets, strip metadata, compress, optimize and render SVG, and build contact-sheet montages. A recipe generator over ImageMagick (primary) with dedicated tools (svgo/resvg/cwebp/avifenc/pngquant) used when installed. Use when the user wants to resize, convert, optimize, favicon, watermark, or batch-process images for a site or social post.
 metadata:
-  version: 0.1.1
+  version: 0.1.3
 ---
 
 # Maggie Image
@@ -58,8 +58,14 @@ maggie tool maggie_image.py presets
 ```
 
 Subcommands: `resize`, `social`, `webp`, `avif`, `favicon`, `strip`, `compress`,
-`svgo`, `svg2png`, `montage`, `presets`. `--run` (before the subcommand) executes
-when the tool is present; otherwise the command is printed for review.
+`svgo`, `svg2png`, `svg-export` (Inkscape: PDF/EPS, text-to-path, per-id slicing),
+`composite` (blend-mode layering), `look` (named style presets: sepia/duotone/
+posterize/vignette/...), `montage`, `presets`. `--run` (before the subcommand)
+executes when the tool is present; otherwise the command is printed for review.
+
+For reusable, batchable edit *recipes* and generatable scene/diagram formats
+(excalidraw/drawio), the non-destructive edit-stack idea, and SVG-ingest safety, see
+[design editor patterns](../../references/design-editor-patterns.md).
 
 ## Defaults and quality
 

package/bundled-skills/maggie-video/SKILL.md

@@ -2,7 +2,7 @@
 name: maggie-video
 description: Produce and process marketing video - render a composition to MP4/GIF/WebM, assemble clips/images/text/captions/audio on a timeline, cut to channel aspects (reel/square/youtube), burn in captions, normalize loudness, watermark, transcode, and make social cuts. Engine-neutral over the renderer (ffmpeg for assembly/transcode, a programmatic engine for code-drawn animation). Use when the user wants to make, edit, cut, caption, resize, or export a video for marketing.
 metadata:
-  version: 0.1.4
+  version: 0.1.5
 ---
 
 # Maggie Video

package/bundled-tools/clis/maggie_image.py

@@ -10,7 +10,7 @@ and are generate-only until it is installed. Web-delivery quality defaults follo
 sharp/squoosh (JPEG/WebP ~80, AVIF ~50); OG/social are cover-cropped, logos pad.
 
 Subcommands: resize, social, webp, avif, favicon, strip, compress, svgo, svg2png,
-montage, presets.
+svg-export, composite, look, montage, presets.
 """
 import argparse
 import shlex
@@ -59,18 +59,25 @@ def _resize_ops(w, h, mode, pad):
 
 def c_resize(a):
     w, h = a.size.lower().split("x")
-    cmd = _im() + [a.input, "-strip"] + _resize_ops(w, h, a.mode, a.pad) + \
+    cmd = _im() + [a.input, "-auto-orient", "-strip"] + _resize_ops(w, h, a.mode, a.pad) + \
         ["-quality", str(a.quality), a.output]
     return emit(cmd, a.run, _im_tool())
 
 
 def c_social(a):
     w, h = SOCIAL[a.preset]
-    cmd = _im() + [a.input, "-strip"] + _resize_ops(w, h, a.mode, a.pad) + \
+    cmd = _im() + [a.input, "-auto-orient", "-strip"] + _resize_ops(w, h, a.mode, a.pad) + \
         ["-quality", str(a.quality), a.output]
     return emit(cmd, a.run, _im_tool())
 
 
+def c_trim(a):
+    """Trim uniform border / cut a product or logo out of its background
+    (ImageMagick -fuzz -trim; +repage resets the virtual canvas)."""
+    cmd = _im() + [a.input, "-auto-orient", "-fuzz", f"{a.fuzz}%", "-trim", "+repage", a.output]
+    return emit(cmd, a.run, _im_tool())
+
+
 def c_webp(a):
     if shutil.which("cwebp") and not a.force_im:
         return emit(["cwebp", "-q", str(a.quality), "-m", "6", a.input, "-o", a.output], a.run, "cwebp")
@@ -152,6 +159,63 @@ def c_montage(a):
     return emit(cmd, a.run, _montage()[0])
 
 
+def c_svg_export(a):
+    """Inkscape headless export - fills the gaps resvg/ImageMagick can't: text-to-path
+    font outlining, PDF/EPS for print, and per-id slicing. GPL app, invoked at arm's
+    length (like ImageMagick)."""
+    args = ["inkscape", a.input, f"--export-type={a.type}", f"--export-filename={a.output}"]
+    if a.width:
+        args.append(f"--export-width={a.width}")
+    if a.dpi:
+        args.append(f"--export-dpi={a.dpi}")
+    if a.area == "drawing":
+        args.append("--export-area-drawing")
+    elif a.area == "page":
+        args.append("--export-area-page")
+    if a.id:
+        args += [f"--export-id={a.id}", "--export-id-only"]
+    if a.text_to_path:
+        args.append("--export-text-to-path")
+    if a.background:
+        args += [f"--export-background={a.background}", "--export-background-opacity=1"]
+    print("# Inkscape: text-to-path outlining (font-independent handoff), PDF/EPS, per-id slicing.")
+    return emit(args, a.run, "inkscape")
+
+
+def c_composite(a):
+    """Layer an overlay onto a base with a blend mode (Pinta/Natron vocabulary ->
+    ImageMagick -compose). Modes: Over, Multiply, Screen, Overlay, Darken, Lighten,
+    ColorDodge, ColorBurn, HardLight, SoftLight, Difference, Exclusion."""
+    cmd = _im() + [a.base, a.overlay]
+    if a.gravity:
+        cmd += ["-gravity", a.gravity]
+    cmd += ["-compose", a.mode, "-composite", a.output]
+    return emit(cmd, a.run, _im_tool())
+
+
+LOOKS = {
+    "grayscale": ["-colorspace", "Gray"],
+    "sepia": ["-sepia-tone", "80%"],
+    "posterize": ["-posterize", "4"],
+    "oilpaint": ["-paint", "4"],
+    "sketch": ["-colorspace", "Gray", "-charcoal", "1"],
+    "vignette": ["-background", "black", "-vignette", "0x20"],
+    "blur": ["-blur", "0x8"],
+    "sharpen": ["-unsharp", "0x1.5"],
+}
+
+
+def c_look(a):
+    """Apply a named 'look' (a reusable, non-destructive style preset; the darktable
+    edit-recipe idea in a single op). `duotone` maps shadows->--dark, highlights->--light."""
+    if a.look == "duotone":
+        ops = ["-colorspace", "Gray", "-level-colors", f"{a.dark},{a.light}"]
+    else:
+        ops = LOOKS[a.look]
+    cmd = _im() + [a.input, "-strip"] + ops + [a.output]
+    return emit(cmd, a.run, _im_tool())
+
+
 def c_presets(a):
     print("Social/OG size presets (use with `social --preset <name>`; cover-cropped):")
     for name, (w, h) in SOCIAL.items():
@@ -196,6 +260,9 @@ def main(argv=None):
     fv.add_argument("--input", required=True); fv.add_argument("--out-dir", dest="out_dir", required=True)
     fv.set_defaults(fn=c_favicon)
 
+    tr = sub.add_parser("trim", help="trim border / cut product out of its background")
+    io(tr); tr.add_argument("--fuzz", type=int, default=5, help="color tolerance %%"); tr.set_defaults(fn=c_trim)
+
     st = sub.add_parser("strip", help="strip metadata (EXIF/GPS/ICC)"); io(st); st.set_defaults(fn=c_strip)
     co = sub.add_parser("compress", help="compress JPEG/PNG"); io(co)
     co.add_argument("--quality", type=int, default=82); co.set_defaults(fn=c_compress)
@@ -211,6 +278,26 @@ def main(argv=None):
     mo.add_argument("--pad", type=int, default=10); mo.add_argument("--background", default="#ffffff")
     mo.set_defaults(fn=c_montage)
 
+    se = sub.add_parser("svg-export", help="Inkscape headless export (pdf/eps/png/svg, text-to-path, by-id)")
+    se.add_argument("--input", required=True); se.add_argument("--output", required=True)
+    se.add_argument("--type", default="png", help="png,pdf,ps,eps,svg")
+    se.add_argument("--width", type=int); se.add_argument("--dpi", type=int)
+    se.add_argument("--area", choices=["drawing", "page"], default="drawing")
+    se.add_argument("--id", help="export only this element/layer id")
+    se.add_argument("--text-to-path", dest="text_to_path", action="store_true", help="outline fonts")
+    se.add_argument("--background"); se.set_defaults(fn=c_svg_export)
+
+    cp = sub.add_parser("composite", help="layer an overlay on a base with a blend mode")
+    cp.add_argument("--base", required=True); cp.add_argument("--overlay", required=True)
+    cp.add_argument("--output", required=True); cp.add_argument("--mode", default="Over")
+    cp.add_argument("--gravity"); cp.set_defaults(fn=c_composite)
+
+    lk = sub.add_parser("look", help="apply a named look (sepia/duotone/posterize/vignette/...)")
+    lk.add_argument("--input", required=True); lk.add_argument("--output", required=True)
+    lk.add_argument("--look", choices=list(LOOKS) + ["duotone"], required=True)
+    lk.add_argument("--dark", default="#1a1a2e"); lk.add_argument("--light", default="#e94560")
+    lk.set_defaults(fn=c_look)
+
     sub.add_parser("presets", help="list size presets").set_defaults(fn=c_presets)
 
     a = ap.parse_args(argv)

package/bundled-tools/clis/maggie_video.py

@@ -511,17 +511,20 @@ def c_captions(a):
         cmd = [tool, "-m", a.model, "-f", a.input, "-osrt", "-of",
                os.path.join(out_dir, base)]
         print(f"# whisper.cpp writes {os.path.join(out_dir, base)}.srt; --model is the .bin path")
-    elif a.engine == "whisperx":
-        tool = "whisperx"
-        cmd = [tool, a.input, "--model", a.model, "--output_format", "srt",
+    else:  # openai-whisper (default) / whisperx - same flag surface
+        tool = "whisperx" if a.engine == "whisperx" else "whisper"
+        cmd = [tool, a.input, "--model", a.model, "--output_format", a.format,
                "--output_dir", out_dir] + lang
-        print(f"# whisperx writes {os.path.join(out_dir, base)}.srt in --output_dir")
-    else:  # openai-whisper (local, pip install openai-whisper)
-        tool = "whisper"
-        cmd = [tool, a.input, "--model", a.model, "--output_format", "srt",
-               "--output_dir", out_dir] + lang
-        print(f"# openai-whisper writes {os.path.join(out_dir, base)}.srt in --output_dir")
-    print("# transcription is local/offline (no paid API); a large model is slower but better.")
+        if a.word_timestamps:
+            cmd += ["--word_timestamps", "True"]
+        if a.highlight_words:
+            cmd += ["--highlight_words", "True"]   # karaoke-style per-word highlight
+        if a.initial_prompt:
+            cmd += ["--initial_prompt", a.initial_prompt]  # bias brand/product spellings
+        print(f"# {tool} writes {os.path.join(out_dir, base)}.{a.format} in --output_dir; "
+              f"openai-whisper's default model is now 'turbo' (large-v3-turbo).")
+    print("# transcription is local/offline (no paid API); a larger model is slower but better.")
+    print("# --highlight-words gives karaoke captions; --initial-prompt fixes brand-term spelling.")
     print("# then: maggie_video.py subtitle --input video.mp4 --srt <that>.srt --output out.mp4  (burn in)")
     return emit(cmd, a.run, tool=tool)
 
@@ -602,11 +605,16 @@ def main(argv=None):
     ac.add_argument("--margin", type=float, default=0.1, help="padding kept around speech (s)")
     ac.set_defaults(fn=c_autocut)
 
-    ca = sub.add_parser("captions", help="transcribe speech to SRT (local whisper)")
-    ca.add_argument("--input", required=True); ca.add_argument("--output", required=True, help="target .srt")
+    ca = sub.add_parser("captions", help="transcribe speech to SRT/VTT (local whisper)")
+    ca.add_argument("--input", required=True); ca.add_argument("--output", required=True, help="target caption file")
     ca.add_argument("--engine", choices=["whisper", "whispercpp", "whisperx"], default="whisper")
-    ca.add_argument("--model", default="base", help="model name, or the .bin path for whispercpp")
-    ca.add_argument("--lang", default="auto"); ca.set_defaults(fn=c_captions)
+    ca.add_argument("--model", default="turbo", help="model (openai-whisper default 'turbo'); .bin path for whispercpp")
+    ca.add_argument("--format", default="srt", help="srt|vtt|txt|tsv|json (openai-whisper/whisperx)")
+    ca.add_argument("--lang", default="auto")
+    ca.add_argument("--word-timestamps", dest="word_timestamps", action="store_true")
+    ca.add_argument("--highlight-words", dest="highlight_words", action="store_true", help="karaoke per-word highlight")
+    ca.add_argument("--initial-prompt", dest="initial_prompt", help="bias brand/product term spelling")
+    ca.set_defaults(fn=c_captions)
 
     sub.add_parser("presets", help="list channel aspect presets").set_defaults(fn=c_presets)
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@topy-ai/maggie",
-  "version": "0.8.14",
+  "version": "0.8.16",
   "description": "Install and manage Maggie Skills for AI coding agents",
   "license": "MIT",
   "type": "module",

package/references/captions-transcription.md

@@ -9,7 +9,7 @@ Distilled from the reference library's R-category projects.
 
 | Engine | What / how | License | Reach for it when |
 |---|---|---|---|
-| openai-whisper | The reference Python CLI: `whisper IN --model base --output_format srt --output_dir DIR` | MIT | the default; simplest install (`pip install openai-whisper`) |
+| openai-whisper | `whisper IN --model turbo --output_format srt --output_dir DIR` (default model is now **`turbo`** = large-v3-turbo; formats `txt/vtt/srt/tsv/json`) | MIT | the default; simplest install (`pip install openai-whisper`) |
 | whisper.cpp | C/C++ port, fast on CPU, GGML models: `whisper-cli -m model.bin -f IN -osrt -of OUTBASE` | MIT | no GPU / fastest CPU / embedding in a pipeline |
 | whisperX | whisper + forced alignment for **word-level timestamps** and diarization: `whisperx IN --model large-v3 --output_format srt --output_dir DIR` | BSD-4 | word-accurate caption timing or speaker labels |
 
@@ -38,4 +38,24 @@ Model size trades speed for accuracy: `tiny` / `base` (fast, rough) -> `small` /
   API is ever wired in instead, the global spend-first rule applies - state the cost.
 - Keep caption text inside the safe area (see `maggie-video` channel presets).
 
+## Caption-shaping flags (built into openai-whisper)
+
+The core CLI already does what we previously routed to whisperX or post-processing
+(`maggie_video.py captions` exposes these):
+
+- `--word_timestamps True` - word-level timings (whisperX still wins on alignment +
+  diarization, but core whisper now has them).
+- `--highlight_words True` - **karaoke-style** per-word highlight in SRT/VTT (the
+  social-caption effect); render the ASS with ffmpeg `subtitles`.
+- `--max_line_width` / `--max_line_count` / `--max_words_per_line` - native line
+  wrapping so captions fit the channel safe-area, no post-processing.
+- `--initial_prompt "<brand, product names>"` - bias the model toward correct brand/
+  jargon spelling (the concrete mitigation for the mishears-brand-terms risk; still
+  human-review).
+- Anti-hallucination over music/silence: `--hallucination_silence_threshold`,
+  `--compression_ratio_threshold`, `--logprob_threshold`, `--no_speech_threshold` -
+  worth setting for music-bed marketing clips.
+- `--task translate` (foreign audio -> English subs), `--clip_timestamps` (only given
+  ranges).
+
 See `maggie-video` for the subcommands and the burn-in / soft-mux / compile options.

package/references/design-editor-patterns.md

@@ -0,0 +1,104 @@
+# Design-editor patterns
+
+Borrowable logic and creativity from open-source design, whiteboard, vector and raster
+**editors** (reference library category P), for making and handling marketing visuals
+programmatically. We adopt **concepts and data shapes, not code** (most are GPL/AGPL/
+MPL or custom-licensed; the few MIT ones are reimplemented, not bulk-copied). Pairs
+with [image recipes](image-recipes.md), [generative visuals](generative-visuals.md),
+and [AI media generation](ai-media-generation.md).
+
+## Non-destructive edit recipe (darktable)
+
+darktable (GPL) edits by a **replayable history stack of named module invocations**,
+stored in an **XMP sidecar** beside the untouched original - never mutating the source;
+the image is reconstructed by replaying the stack. Borrow this as a maggie "image edit
+recipe": a small JSON `[ {op, params}, ... ]` stored next to the original and replayed
+by `maggie_image.py` (each op -> an ImageMagick/sharp command). It is **reproducible,
+auditable, re-editable, and batchable** - apply one "look" recipe to every photo in a
+campaign for consistent brand grading. `darktable-cli <input> [sidecar.xmp]
+<output-dir>` renders from such a stack (XMP optional/auto-discovered, final arg is a
+**directory**; GPL binary, arm's-length). Two source details worth adopting: **params
+versioning** (each module keeps `legacy_params` for forward-compatible migration -
+version your recipe ops) and **"Styles" = saved reusable history** (name a look once,
+apply to many).
+
+## LLM-generatable scene / diagram formats
+
+- **Excalidraw** (`.excalidraw` JSON, MIT) - a flat list of elements (`{type:
+  rectangle|ellipse|diamond|text|line|arrow|freedraw|image|frame, x, y, width, height,
+  strokeColor, backgroundColor, fillStyle, roughness, seed, roundness, groupIds,
+  boundElements}`) with a separate `files` asset map and ratio-based arrow/label
+  bindings that survive resize. **Emit it from a prompt** for diagrams, flowcharts,
+  annotated screenshots -> open in-app or render to SVG/PNG, with an on-brand hand-drawn
+  (rough.js) look.
+- **drawio / mxGraph** (`mxGraphModel` XML, Apache-2.0 code; diagram output
+  unrestricted) - `<mxCell>` nodes with `value` (label), a compact `style` DSL
+  (`rounded=1;fillColor=#dae8fc;strokeColor=#6c8ebf;`), `vertex="1"` or `edge="1"` with
+  `source`/`target`, and `<mxGeometry>`. Trivially generatable for org charts, funnels,
+  architecture diagrams; render headless to SVG/PNG (don't redistribute their icon/
+  stencil sets).
+
+## SVG as the document (Method-Draw) + safe ingest
+
+Method-Draw (MIT, from SVG-Edit) treats the **live SVG DOM as the single source of
+truth** - `XMLSerializer` is the save format; no separate scene JSON. Two borrows:
+(a) the simplest possible model for programmatic marketing SVGs (build the DOM,
+serialize); (b) a **sanitize whitelist** (`svgWhiteList_`) of allowed SVG elements/
+attributes - essential for safely ingesting user- or AI-supplied SVG before rendering
+(strips scripts/unknowns); pair with svgo in [image recipes](image-recipes.md). Its
+command `apply/unapply` history is the generic edit-history pattern (and the backbone
+of the edit-recipe above).
+
+## Design-doc patterns (penpot, tldraw)
+
+- **penpot** (MPL, concept) - a design doc in **open standards (SVG/CSS/HTML/JSON)**
+  with first-class **Design Tokens as the single source of truth** between brand and
+  code, **Components + Variants** (one card, many variants = reusable marketing
+  templates), CSS Grid/Flex layout on the doc, and an **MCP server** making the doc
+  agent-readable. Strong alignment with Maggie's brand-kit + MCP posture.
+- **tldraw** (custom license, concept) - an infinite-canvas **reactive record store**
+  of typed shape records (geometry + props + `parentId` + fractional index) with
+  bindings/assets/pages, and the explicit framing that **an LLM agent operates on the
+  canvas document as data** - reinforcing the "generate the scene JSON" approach.
+
+## AI-edit vocabulary (krita-ai-diffusion)
+
+krita-ai-diffusion (GPL, concept) bolts generative AI onto a layer/selection editor
+over a ComfyUI backend. The borrowable vocabulary for precise AI image edits (beyond
+full-frame generation): **mask/selection-scoped generative fill** (inpaint, expand,
+add/remove inside a region), **layer-as-region prompting** (a text prompt per image
+area defined by a layer), and **conditioning** via ControlNet (scribble/lineart/canny/
+pose/depth/segmentation) and IP-Adapter (reference image, style/composition transfer).
+The plugin-over-backend boundary (editor model stays; AI is a swappable backend)
+mirrors the arm's-length posture in [AI media generation](ai-media-generation.md).
+
+## Licensing
+
+Invoke binaries at arm's length (never vendor): inkscape, darktable (GPL). Concepts
+only: penpot (MPL), tldraw (custom), krita/krita-ai-diffusion (GPL), drawio icon sets.
+MIT and freely borrowable as ideas (still reimplement, don't bulk-copy): excalidraw,
+Method-Draw, rough.js, Pinta. drawio code Apache-2.0; its diagram output is
+unrestricted.
+
+## From full-source study (net-new)
+
+- **tldraw** - the LLM mutation surface is `editor.createShapes/updateShapes/
+  deleteShapes` (batch partial mutations); the store ships schema **migration
+  sequences**, a composable **validation layer**, and a **side-effects manager**
+  (before/after create/change/delete hooks). `parentId` can be a page *or* a shape;
+  pages carry a fractional `index`. Adopt the batch-mutation + validate + migrate
+  shape as the model for an agent-written canvas doc.
+- **penpot** - **design tokens are W3C DTCG** (import/export + aliasing), components
+  have **variants**, shapes carry **flex/grid layout**; the agent boundary is an
+  **official MCP server talking to a sandboxed in-app plugin over WebSocket** - a
+  strong arm's-length model for making a design doc agent-writable (mirrors Maggie's
+  own MCP posture).
+- **drawio** - the `mxConstants` STYLE_* key table is a **validated allow-list** for
+  LLM-emitted styles; `mxEdgeStyle` routing + hierarchical/tree/organic **auto-layout**
+  mean a generated node+edge list can self-position; `.drawio` pages are often
+  **deflate+base64** compressed.
+- **krita** - an **adjustment/filter layer holds a serialized filter config** = a
+  non-destructive recipe living in the layer stack; ops are **mask/selection-scoped**;
+  `.kra` is `zip(manifest XML + per-layer blobs)`; a **Python scripting API** and a
+  JSON-manifest plugin boundary let a generative backend register like any filter -
+  the "AI as a swappable backend" posture.

package/references/ffmpeg-recipes.md

@@ -128,3 +128,24 @@ maggie tool maggie_video.py compile --spec timeline.json   # declarative timelin
 maggie tool maggie_video.py --run encode --input in.mov --output out.mp4   # actually runs
 maggie tool maggie_video.py presets
 ```
+
+## More filters worth reaching for (confirmed in libavfilter)
+
+- **Green-screen / keying** (UGC, product on a brand background):
+  `chromakey=0x00FF00:0.1:0.0` (or `colorkey`), clean fringes with `despill`.
+- **Fade to/from black** (distinct from clip crossfade): `fade=t=in:st=0:d=0.5`,
+  `fade=t=out:st=9.5:d=0.5`.
+- **Brand color grade / LUT**: `lut3d=brand.cube`, or `curves`/`eq`/`colorbalance`/`hue`.
+- **Karaoke captions**: `subtitles=in.ass` with a full ASS file (word highlight) -
+  pairs with whisper `--highlight_words` ([captions](captions-transcription.md)).
+- **Focus/credits**: `gblur` + `vignette`; `scroll` for scrolling outros.
+- **GPU fast path** when hardware is present: `overlay_cuda`/`overlay_vulkan`,
+  `xfade_opencl` for long renders.
+
+## Target-filesize social presets (HandBrake idea)
+
+Beyond aspect, encode to **fit under a platform's upload cap**: pick a two-pass video
+bitrate from the cap and duration - `bitrate_bits = size_bytes * 8 / duration - audio`
+- e.g. "25 MB / 30 s / 1080p". A named preset bundles `(upload cap, max duration) ->
+resolution + bitrate`. (HandBrake ships exactly these as "Social 25 MB 30 Seconds"
+presets.)

package/references/generative-visuals.md

@@ -92,3 +92,53 @@ zdog, pts, hydra, regl, Babylon.js, react-three-fiber, fogleman/primitive, svg.j
 nannou - licenses in the table above. Capture pipeline: see
 [programmatic video](programmatic-video.md), [design-motion production](design-motion-production.md),
 [ffmpeg recipes](ffmpeg-recipes.md).
+
+## More patterns (from the desktop/editor study)
+
+- **The shared sketch model** (processing / openFrameworks / nannou / p5): a `setup()`
+  once + per-frame `draw()` loop with seeded `random()/noise()`. openFrameworks (MIT)
+  splits it into `setup()/update()/draw()` (sim vs render); nannou (MIT/Apache)
+  formalizes Model-Update-View. The reusable discipline: **keep state separate from the
+  render fn, index by frame, seed all randomness** - that is what makes a loop
+  reproducible and headless-capturable. Default to p5/three/regl in the browser;
+  reach for a desktop framework only for heavy GPU throughput or offline high-res
+  render the browser can't sustain. (Processing non-core is GPL - reimplement.)
+- **Hand-drawn scene JSON (excalidraw, MIT).** `.excalidraw` is a flat list of
+  elements (`{type, x, y, w, h, strokeColor, backgroundColor, fillStyle:
+  hachure|cross-hatch|solid|zigzag, roughness, seed, roundness}`) rendered with
+  rough.js. It is **LLM-generatable**: emit the JSON for a diagram/flowchart/annotated
+  screenshot, then render to SVG/PNG. The rough.js mapping for an on-brand hand-made
+  look: `seed` (determinism), `fillStyle`, `fillWeight = strokeWidth/2`,
+  `hachureGap = strokeWidth*4`, `roughness`, `disableMultiStroke` for non-solid strokes.
+- **Vector shape morphing** (synfig concept): tween between two paths by interpolating
+  each control point's vertex + in/out tangents + width - smooth logo/blob/path morphs.
+- **Repeater / radial stamp** (Lottie concept): one shape -> N transformed copies with
+  an incremental transform + opacity ramp (bursts, confetti, radial brand patterns) -
+  cheap, on-brand, deterministic.
+- For AI-assisted editing vocabulary (mask/selection-scoped generation, layer-as-region
+  prompting, ControlNet/IP-Adapter conditioning), see
+  [AI media generation](ai-media-generation.md) and
+  [design editor patterns](design-editor-patterns.md).
+
+### Headless capture + 3D export (from full-source study)
+
+How to render these engines **without a screen**, and extra output formats worth
+knowing:
+
+- **Babylon.js `NullEngine`** (Apache-2.0) - first-class **server-side headless**
+  rendering (no GPU/canvas); the cleanest headless story for render-automation jobs,
+  plus a built-in **procedural-texture library** (brick/cloud/fire/marble/noise/wood),
+  `CreateScreenshotUsingRenderTarget` for 4K+ stills from a small viewport, and
+  **NodeMaterial** (shaders serialize to JSON - an LLM-emittable look). This is the
+  concrete reason to pick Babylon over three for automation.
+- **three.js** (MIT) - `readRenderTargetPixels()`/`...Async` read pixels from an
+  offscreen `WebGLRenderTarget` (true headless high-res, not on-screen canvas);
+  **SVGRenderer** emits a 3D scene as editable SVG; **USDZExporter** turns a 3D product
+  into a free **iOS AR Quick Look** asset; **GPUComputationRenderer** (ping-pong FBO
+  sim: flocking/fluid) = a living, seeded generative hero background.
+- **openFrameworks** (MIT) - `ofFbo` offscreen framebuffer for arbitrary-res stills;
+  **ofxEmscripten** compiles a desktop sketch to WebAssembly for the browser;
+  **ofxSvg** parses + emits structured SVG.
+- **Processing** - `PSurfaceNone` (headless) and `createGraphics(big)` (render larger
+  than the screen -> high-res still). License at feature granularity: core raster is
+  LGPL, but **SVG/PDF export is in non-core (GPL)** - reimplement, don't vendor.

package/references/image-recipes.md

@@ -87,3 +87,50 @@ sharp (Apache-2.0), squoosh codecs (Apache-2.0), imagemin (MIT), cwebp/avifenc
 (BSD). **Avoid vendoring code from** Upscayl (AGPL-3.0) - for AI upscaling invoke the
 separate `realesrgan-ncnn-vulkan` binary instead. sharp/squoosh quality defaults are
 adopted as the preset numbers.
+
+## SVG beyond resvg/ImageMagick (Inkscape, arm's-length)
+
+Inkscape's headless CLI (GPL app, invoked like ImageMagick) fills gaps resvg and IM
+can't: **font outlining, PDF/EPS, and per-id slicing**. `maggie_image.py svg-export`
+generates these.
+
+```bash
+inkscape in.svg --export-type=png --export-filename=out.png --export-width=1200 \
+  --export-area-drawing --export-background=#ffffff --export-background-opacity=1
+inkscape in.svg --export-type=pdf --export-filename=out.pdf                      # print collateral
+inkscape in.svg --export-text-to-path --export-type=svg --export-filename=flat.svg  # outline fonts for handoff
+inkscape in.svg --export-id=badge --export-id-only --export-type=png --export-filename=badge.png  # slice one element
+inkscape in.svg --export-type=png --export-dpi=300 --export-area-page --export-filename=print.png
+```
+
+## Layered composite + named looks (Pinta vocabulary)
+
+Blend modes map from a raster editor's layer stack onto ImageMagick `-compose`:
+
+```bash
+convert base.png overlay.png -gravity center -compose Multiply -composite out.png
+# modes: Over, Multiply, Screen, Overlay, Darken, Lighten, ColorDodge, ColorBurn,
+#        HardLight, SoftLight, Difference, Exclusion
+```
+
+Named "looks" (each a one-line style preset - a lightweight take on darktable's
+non-destructive edit recipe): `grayscale` (`-colorspace Gray`), `sepia`
+(`-sepia-tone 80%`), `posterize` (`-posterize 4`), `duotone` (`-colorspace Gray
+-level-colors '#1a1a2e','#e94560'`), `vignette` (`-background black -vignette 0x20`),
+`oilpaint` (`-paint 4`), `sketch` (`-colorspace Gray -charcoal 1`), `blur`/`sharpen`.
+`maggie_image.py composite` and `maggie_image.py look` generate these. For a
+reusable, batchable edit *recipe* (a JSON stack of named ops applied to many photos),
+see [design editor patterns](design-editor-patterns.md).
+
+## More ImageMagick ops worth baking in (confirmed in source)
+
+- **`-auto-orient`** first in every pipeline - fixes sideways phone photos by EXIF
+  (`maggie_image.py` resize/social/trim already prepend it).
+- **`-thumbnail WxH`** - metadata-stripping fast resize; a better default than
+  `-resize` for web/OG thumbnails.
+- **`-fuzz N% -trim +repage`** - trim a uniform border or cut a product/logo out of
+  its background (`maggie_image.py trim`).
+- **`-clut` / `-hald-clut`** - repeatable LUT brand grading (raster equivalent of
+  ffmpeg `lut3d`).
+- **`-distort perspective '<src coords> <dst coords>'`** - place a screenshot into a
+  device mockup.

package/references/programmatic-video.md

@@ -78,3 +78,25 @@ apply.
 `libx264 -crf 14-20 -preset slow/medium -pix_fmt yuv420p -movflags +faststart`,
 even dimensions, plus 960px / 800x500 GIF deliverables. For transparent output,
 Revideo uses per-format pixel formats (`yuva420p` WebM, `yuva444p10le` ProRes).
+
+## Remotion's real renderer (from source) and the ideas to borrow
+
+Source study (Remotion License - concept only, no vendoring) shows Remotion is **not**
+the base64-PNG pipe we lump with Revideo/Motion Canvas: it drives headless Chromium
+over CDP, captures frames via **shared memory**, and encodes with a **per-platform
+native Rust sidecar that statically links FFmpeg** (a long-running streaming process),
+plus a bundled `ffmpeg` for the final stitch. Borrowable ideas (improve our Part C):
+
+- **A readiness gate, not just a fake clock.** `delayRender()`/`continueRender()`: a
+  component declares async work pending and the renderer blocks the screenshot until
+  all handles clear, reporting **labeled** open handles on timeout. Add an explicit
+  readiness handshake (plus a per-frame `await document.fonts.ready`) to Part C so no
+  frame is captured before fonts/data/images are ready - prevents FOUT/half-loaded
+  frames.
+- **Tab-pool concurrency**: open N browser pages, each seeking a different frame in
+  parallel (frame range / N).
+- **Chunked parallel encode + seamless concat**: split the frame range, encode chunks
+  independently, concat without re-encode - throughput for long renders.
+- **Kept-open, LRU-cached source-video decoders**: for embedded video assets, keep the
+  decoder open and cache frames so repeated seeks to a timestamp are cheap (our flat
+  timeline currently punts this to raw ffmpeg `-ss`).

package/references/video-timeline-model.md

@@ -130,3 +130,59 @@ arm's length, never linked or vendored.
   keyframe/ease vocabulary, the licensing note - stable and engine-neutral.
 - **`maggie-video` skill:** channel presets, caption/CTA safe-areas, the operational
   ffmpeg command layer (via `maggie_video.py`), and tool wiring.
+
+## Extensions from the animation-tool study (O-category)
+
+Concepts to grow the model (adopt the data shape / vocabulary; code only from the
+MIT/BSD sources - rlottie MIT, opentoonz/tahoma2d BSD; synfig/friction/wick/Natron
+GPL = concept only):
+
+- **Richer keyframe easing** (synfig). Extend `{t, v, ease}` with: **per-side ease**
+  (`easeIn` != `easeOut`), **`clamped`** (auto-smooth that won't overshoot the keyed
+  value), **`hold`/`constant`**, and optional **TCB** `{tension, continuity, bias}`.
+  Add **`fullRotations`** on a rotation keyframe (wick) to control spin count/direction
+  rather than shortest-path. These precompute to per-frame values for ffmpeg and map to
+  CSS/Lottie beziers. (compile currently interpolates linearly; this is the target
+  vocabulary.)
+- **Lottie JSON as the companion vector-motion-graphics format** (rlottie, MIT). For
+  logo stings, animated icons and kinetic type, use Lottie (After-Effects/Figma
+  exportable, in-browser renderable via lottie-web) alongside the MLT/NLE shape. Borrow
+  its primitives: **Trim-path** (`start%/end%/offset` - the formal version of the
+  "draw a line on" trick in [design-motion production](design-motion-production.md)),
+  **Repeater** (N copies with a transform + opacity ramp), **track mattes**
+  (`Alpha/Luma`), **transform parenting**, and **time remap** (`tm`).
+- **Nested clips / precomps** (wick/lottie). A clip may hold its own timeline with a
+  playback mode (`loop | playOnce | single`) and be instanced - reusable lower-thirds,
+  intro/outro stingers; a composable alternative to the flat track list.
+- **Decomposed transform + animatable pivot** (friction). Split transform into
+  `pos / scale / rot / shear / pivot / opacity`, each its own keyframe curve, with a
+  **movable pivot** for correct scale/rotate around a moving anchor.
+- **Blend-mode / compositing vocabulary** (Natron `MergingFunctionEnum`). The allowed
+  `blendMode` values on overlay clips: Porter-Duff (`Over, Under, In, Out, ATop, Copy,
+  Matte`) + blends (`Multiply, Screen, Overlay, SoftLight, HardLight, ColorBurn,
+  ColorDodge, Difference, Exclusion, Hue, Saturation, Color, Luminosity`).
+- **Xsheet tabular timing** (opentoonz/tahoma2d, **BSD - code reusable with
+  attribution, a direct port target**). An optional rows=frames / columns=assets
+  exposure grid (cells = a drawing exposed for N frames; a *hold* is the same cell
+  repeated). The source implements this (and more of our model than we first credited)
+  in `TXsheet` / `TXshCellColumn` (`std::vector<TXshCell> m_cells`) / `TXshCell`
+  (`{level, TFrameId}`) / `TFrameId` (`{frame, letter}`), plus the cell-op vocabulary
+  (`reverseCells, swingCells, stepCells, reframeCells, timeStretch`). It also backs, as
+  BSD code: the **decomposed transform channels** (`TStageObject::Channel` T_X/T_Y/
+  T_Angle/T_ScaleX/... each a keyframe curve, with parenting + `T_Path` follow-spline),
+  a **richer easing enum** (`TDoubleKeyframe::Type` = Constant / Linear / SpeedInOut /
+  EaseInOut / EaseInOutPercentage / Exponential / Expression / File), **nested
+  sub-xsheets** (`TXshChildLevel` = precomp), and a **node-graph fx schematic**
+  (`FxDag`/`TFx`). So the easing/transform/nested-clip/effect-graph extensions above
+  have a BSD implementation to port, not only GPL concept sources.
+- **Strip/effect model + full blend set** (blender VSE `Strip`, concept). The Video
+  Sequence Editor's `Strip` (input1/input2, effect strips CROSS/WIPE/GLOW/SPEED/
+  ADJUSTMENT/TEXT, a Photoshop-complete `StripBlendMode`) and **F-curve modifier stack**
+  (noise/cycles/limits/stepped) + **drivers** (property-driven channels) are strong
+  concept borrows for overlay/effect clips.
+- **OTIO interchange** (kdenlive, concept). Document a mapping to OpenTimelineIO (the
+  ASWF-standard timeline interchange): gap=blank, stack=track, externalReference=asset
+  src - so the model can round-trip with real NLEs.
+- **Optional node-graph "effect graph"** (Natron, concept) for branched layered effects
+  (mask -> blur -> merge) that a linear track can't express - heavy; a documented
+  escape hatch, not the default.

package/references/web-animation-patterns.md

@@ -362,3 +362,37 @@ Most are MIT and safe to reference/learn from freely. The four to flag: **GSAP**
 (proprietary no-charge), **typed.js** (GPL), **animate.css** (Hippocratic 2.1), and
 **synfig/friction** (GPL desktop tools). For the flagged ones we adopt the *pattern*
 and author our own code.
+
+## Authoring motion visually, playing it in code (Theatre.js)
+
+[`theatre-js/theatre`](https://github.com/theatre-js/theatre) (Apache-2.0) is a
+motion-design *editor* (`@theatre/studio`) that writes a plain serializable **project
+state**, replayed by a tiny headless runtime (`@theatre/core`) against three.js,
+HTML/SVG, or any JS variable. The borrowable pattern is the **declarative keyframe
+document**, not the editor: store motion as JSON - a `sequence` with `length` +
+`subUnitsPerUnit` (an fps snap-grid) -> **tracks keyed by property path** ->
+**keyframes** each carrying `{ position, value, handles:[lx,ly,rx,ry],
+type:'bezier'|'hold' }` (per-keyframe bezier handles = per-segment easing baked into
+data; `value` may be non-numeric). Playback mirrors CSS: `play({ iterationCount,
+range, rate, direction:'alternateReverse' })`. Author a brand motion once, commit the
+JSON as a reusable asset, replay it deterministically with no editor shipped - the
+authoring story for `maggie-design-motion` (keep the motion *spec* as data, not
+imperative GSAP calls). Reach for it when motion must be re-edited by non-coders or
+reused across pages; a one-shot reveal stays plain CSS/GSAP.
+
+## Interactive, input-driven motion (Rive state machines)
+
+[`rive-app/rive-runtime`](https://github.com/rive-app/rive-runtime) (MIT) drives
+vector animation with a **state machine** instead of fixed playback - the pattern for
+interactive hero elements, product configurators and mascots. Model: a small **input**
+surface (`bool`, `number`, `trigger`) -> **transitions** guarded by
+**conditions/comparators** (`number > 0.5`, trigger fired) -> **states** (play an
+animation, or a **BlendState1D** that blends along a number input). A **listener**
+layer maps pointer/scroll/keyboard events to input changes; states run in parallel
+**layers**; transitions carry exit-time, ease interpolators, and **random weights**
+for non-repetitive idles. Borrow it in the browser without Rive: model an interactive
+hero as `inputs -> conditions -> states` - declare inputs (`hovered:bool`,
+`scroll:0..1`, `clicked:trigger`), bind DOM events to them, and switch CSS/WAAPI/GSAP
+timelines by state rather than scripting each handler; a number input + interpolation
+gives scroll-linked response, random-picked idle transitions give a "living" element.
+(Rive's own MIT player is also a legitimate option.)