stillmade-sdk 0.1.37 → 0.1.39

package/MODULES.md

@@ -77,7 +77,11 @@ const voice = await stillmade.hosted.run('audio.speech', { text: plan.script, se
 ```
 
 `image.describe`, `media.analyze` and `audio.transcribe` take saved references
-(`image`/`images`, `media`, `audio`); `image.generate` takes `prompt` and optional
+(`image`/`images`, `media`, `audio`). `audio.transcribe` takes a saved audio or a
+saved video, whose own sound it transcribes, so there is no need to extract the
+audio first. `media.analyze` returns `{schemaVersion: 1, items}`, one item per
+picture or video, and for one video takes an optional `range: {start, end}` in
+seconds, at most 12 apart, to look closely at that part. `image.generate` takes `prompt` and optional
 `references`, one to eight saved images the new picture keeps. Look at pictures with `media.analyze`:
 StillMade credits pay for it, while `image.describe` runs only on the person's
 own OpenRouter key. `web.fetch` takes `url` and `instruction`;
@@ -103,20 +107,29 @@ to the next transform.
 | `{operation: 'video.concat', sources, settings?}` | A saved MP4 of 2 to 50 videos in order. |
 | `{operation: 'video.resize', source, settings}` | A saved MP4 re-encoded with new settings. |
 | `{operation: 'image.thumbnail', source, at?, width?}` | A saved PNG of one frame (default the first frame, 1280 px wide). |
-| `{operation: 'document.text', source}` | `{text, pages}`: the words of a PDF or text file the person gave, page by page (up to 500 pages). |
+| `{operation: 'document.text', source}` | `{text, pages}`: all the words of a PDF or text file the person gave, and `pages`, a list holding each page's words as one string (up to 500 pages). A file with no words (a scanned PDF) is refused. |
 | `{operation: 'image.compose', width, height, layers, background?, format?, text?}` | A saved PNG (or JPEG) of saved images joined into one picture: side by side, a grid, a contact sheet, a poster with a title. |
 | `{operation: 'timeline.render', clips, kind?, aspect?, audio?, captions?, settings?}` | A saved MP4 (or, with `kind: 'audio'`, an MP3) of your own timeline, rendered the way the Editor exports. |
 
 `timeline.render` is how a Block exports: there is no separate Export, Captions,
 Transition or Color grade Block to call. Each clip is `{source, start, duration,
-from?, transition?: {type, duration?}, look?, mute?}` (a saved image or video; a
-video keeps its own sound unless `mute`), each audio item `{source, start,
-duration, from?, volume?}`, and each caption `{text, start, end, words?,
-position?, highlight?, animation?}`, where `words` (`[{text, start, end}]`)
-highlights each word as it is spoken. Times are seconds on the timeline: each
+from?, transition?: {type, duration?}, look?, mute?, motion?, fit?, focus?}` (a saved
+image or video; a video keeps its own sound unless `mute`; `motion` is a slow
+camera move over the clip: `zoom-in`, `zoom-out`, `pan-left`, `pan-right`,
+`pan-up` or `pan-down`, which always keeps the picture filling the frame), each
+audio item `{source, start, duration, from?, volume?}`, and each caption
+`{text, start, end, words?, position?, highlight?, animation?, size?}`, where
+`words` (`[{text, start, end}]`) highlights each word as it is spoken and `size`
+is the text height in pixels of the rendered frame (16 to 200; default 44, so
+give a 1080 by 1920 vertical video about 80). Times are seconds on the timeline: each
 clip starts where the one before it ends, less the transition on that earlier
 clip, which blends it into the next one; clips that overlap more are refused.
-`aspect` is `16:9`, `9:16`, `1:1` or `4:5`. Looks, transitions and caption
+`aspect` is `16:9`, `9:16`, `1:1` or `4:5`. A clip of another shape than the
+frame (a wide sermon in a `9:16` clip) is cropped to fill it, around its middle
+or around `focus: {x, y}`, the point to keep in view as fractions 0 to 1 across
+the picture (`{x: 0.3}` keeps a speaker standing left of centre), and
+`focus.scale` (1 to 2.5, instead of a motion) holds the clip that much closer on
+that point; `fit: 'contain'` fits it whole with bars instead. Looks, transitions and caption
 animations use the Editor's names (`TIMELINE_LOOKS`, `TIMELINE_TRANSITIONS`,
 `CAPTION_ANIMATIONS`). The render follows the person's export plan, like any
 export from StillMade.
@@ -139,8 +152,9 @@ for 9 photos on a 1800 by 1800 sheet, three columns of 600 pixels.
 
 `source`/`sources` are saved references (`{kind, assetId, versionId, url}`).
 Render `settings` are `{fps: 24|30|60, resolution: 720|1080|2160, quality:
-'low'|'medium'|'high'}`; the default is 30 fps, 1080, high. Video keeps the
-source's aspect ratio. Rendering follows the person's export plan, like any
+'low'|'medium'|'high'}`; the default is 30 fps, 1080, high. `video.cut`,
+`video.concat` and `video.resize` keep the source's aspect ratio;
+`timeline.render` makes the `aspect` it is given. Rendering follows the person's export plan, like any
 export.
 
 Pure helpers ship as `@stillmade/block-sdk/media` for you to bundle (the file is
@@ -276,9 +290,13 @@ never hold a NUL character (`\u0000`): StillMade cannot store it, and import ref
 A fixture's photo, recording, clip or file can be one of StillMade's example files,
 named by address: `/api/media/example/photo-1.jpg` to `photo-4.jpg`, `sheet.png`,
 `room.png`, `kitchen.wav`, `hall.wav`, `click-120.wav`, `audio.mp3`, `clip.mp4`,
-`notes.txt` and `sales.csv` (sizes in PASTE_TO_LLM.md). `test` reads their real bytes
+`notes.txt`, `sales.csv` and `lesson.pdf`, a two-page PDF of words (sizes in
+PASTE_TO_LLM.md). `test` reads their real bytes
 when the fixture gives no mock for them, and StillMade, its preview and its phone and
 browser check serve the same files, so the example a person sees is the one you tested.
+Point what a fixture's mocks answer at them too (`/api/media/example/clip.mp4` for a
+rendered clip, `kitchen.wav` for speech): `dev` then shows and plays the result, where a
+made-up address shows an empty player.
 
 A fixture can record what the host answers, so a test of a Block that calls
 hosted operations costs nothing and gives the same result every run:

package/PASTE_TO_LLM.md

@@ -166,7 +166,7 @@ Every Block lists its separate tools in manifest.features, one entry per tool or
 
 ### Port meanings
 
-For each newly created Block, declare the meaning of every input and output in its manifest contract, not just a physical type. Each input needs required:boolean, a short plain-language description and a semantic (import refuses an input without one, and gate checks it). An input accepts values declared with its meaning or a narrower one: an input for any photo a Step hands on uses source_image, as StillMade's Image editor does (generated_image and processed_image are kinds of source_image), and a photo editor's result is processed_image. Each output needs semantic and a short description. Use lowercase snake_case or dotted semantic names, such as company_profile or email_subject; different meanings are not interchangeable merely because both are text. If an existing role is present, semantic must equal that role. Reuse an exact existing meaning only when the information actually has that meaning. Describe structured fields with schema when their shape matters, and always on a json or object output (import refuses one without it): for example {name:"string",website:"string",industry:"string?"}. Supported schema fields are string/text, number, integer, boolean, object, json, arrays of those types, and optional fields with ?. A value of a structured type (transcript, script, brief, shot-plan and the like) is an object with schemaVersion:1, in code and fixtures; a transcript as the Transcription Block hands it on is {schemaVersion:1,text:"...",words:[{text:"Today",start:0,end:0.4}]}, times in seconds. Declare sources only when a requirement restricts where its answer may come from: upstream, project, adapter, chat, default, user or external. Do not mark a credential as ordinary text or put secrets in defaults. Contracts describe information; they never authorize network access, model calls, project edits, credit spending or new runtime APIs. Keep existing imported, installed and base Block contracts unchanged; do not silently relabel old published contracts to make a connection appear compatible. New contracts must match actual code and fixtures.
+For each newly created Block, declare the meaning of every input and output in its manifest contract, not just a physical type. Each input needs required:boolean, a short plain-language description and a semantic (import refuses an input without one, and gate checks it). An input accepts values declared with its meaning or a narrower one: an input for any photo a Step hands on uses source_image, as StillMade's Image editor does (generated_image and processed_image are kinds of source_image), and a photo editor's result is processed_image. Videos and sounds work the same way: an input for any video uses source_video (generated_video, edited_video and rendered_video are kinds of it) and one for any sound source_audio (narration_audio and rendered_audio are kinds of it); a video the Block renders is rendered_video, a sound it renders rendered_audio. Each output needs semantic and a short description. Use lowercase snake_case or dotted semantic names, such as company_profile or email_subject; different meanings are not interchangeable merely because both are text. If an existing role is present, semantic must equal that role. Reuse an exact existing meaning only when the information actually has that meaning. Describe structured fields with schema when their shape matters, and always on a json or object output (import refuses one without it): for example {name:"string",website:"string",industry:"string?"}. Supported schema fields are string/text, number, integer, boolean, object, json, arrays of those types, and optional fields with ?. A value of a structured type (transcript, script, brief, shot-plan and the like) is an object with schemaVersion:1, in code and fixtures; a transcript as the Transcription Block hands it on is {schemaVersion:1,text:"...",words:[{text:"Today",start:0,end:0.4}]}, times in seconds. Declare sources only when a requirement restricts where its answer may come from: upstream, project, adapter, chat, default, user or external. Do not mark a credential as ordinary text or put secrets in defaults. Contracts describe information; they never authorize network access, model calls, project edits, credit spending or new runtime APIs. Keep existing imported, installed and base Block contracts unchanged; do not silently relabel old published contracts to make a connection appear compatible. New contracts must match actual code and fixtures.
 Custom inheritance is persisted in manifest.semantics, an array of {name,description?,parents?}. Use namespaced custom names, at most 64 definitions per Block, eight parents per definition, and 500 description characters. Never redefine unnamespaced standard meanings, create cycles, or combine conflicting parent sets. Preserve effective inherited meanings of all existing base inputs, outputs and accepts declarations; adding or changing a parent is a contract change even when port labels stay identical. New independent ports may declare new meanings.
 In isolated computation code, StillMade.emit(outputName,value) collects declared final outputs without a return object; emit each name once and never combine it with a returned output object or job.pending result. StillMade.resolveRequirement(inputName) reads the fixed host-validated declared input snapshot as {resolved:true,requirement,value}, or {resolved:false,requirement} for an absent optional input. StillMade.getProjectContext() returns only context fields explicitly bound in manifest.inputs with their matching context.FIELD.read permissions; getProjectContext(field) reads one declared field and rejects undeclared fields. These snapshot readers do not run the resolver again, request Chat/user answers, read extra project data, expose credentials, write projects or grant permissions. StillMade.validate(value,schema) checks JSON values against SDK semantic schemas and returns a boolean; it does not replace host output validation, prove semantic meaning or authorize execution. These methods are not available in view.javascript.
 
@@ -186,7 +186,7 @@ A capability package performs exactly one supported operation. A job that needs
 
 ### Module Blocks
 
-A job that needs several hosted operations, or code around one (reading a transcript's words and times, shaping a list), or media work on the person's device even with no hosted operation (rendering clips and photos into one video, joining pictures into one), is a module Block: {manifest,moduleSource,tests}, runtime:"module", moduleSource:{"main.js":"<the code>"}. StillMade names main.js by its digest and adds module. StillMade's in-app builder takes one file: give no other files, and main.js is plain JavaScript that imports nothing and uses no browser APIs (no fetch, TextEncoder, timers or WebCodecs). That single-file rule is the in-app builder's; a module Block built with the SDK may use timers, WebCodecs, bundled npm packages and .wasm files as MODULES.md describes. main.js: export default async function run(input, stillmade) returns the declared outputs. It calls each hosted operation with await stillmade.hosted.run(operation, request), listed in manifest.permissions.capabilities: text.generate, image.generate, video.generate, audio.speech, audio.music, audio.sfx, audio.transcribe, image.describe, media.analyze, web.fetch, web.research. Requests, and what each call returns (a test's mocks record exactly that value): text.generate {prompt, system, maxTokens} returns the text as a string, and with schema (a JSON schema object) returns that object, so ask for JSON with schema; one text.generate prompt holds at most 32,000 characters, so a longer file or transcript is read in parts (one call per part of about 25,000 characters, then one call that joins what they found), never cut off, and when the result is the whole text again (a cleaned transcript, a translation) each part's answer is kept and joined in order in main.js, since one answer holds at most 4,096 tokens; audio.transcribe {audio: the input's saved reference, language:"auto"} returns {schemaVersion:1,text,words:[{text,start,end}]}; image.generate {prompt, references?: one to eight saved images the new picture keeps, such as the product, person or place in the person's photos} and audio.speech {text} return a saved image or audio reference; web.fetch {url, instruction} returns a string; media.analyze {media: one to four saved images or one saved video, question} returns {items:[{kind,summary,observations,detectedText}]}, one item per image, and is how a Block looks at pictures (picking the photos that show different angles, reading a menu off a photo): StillMade credits pay for it; image.describe {image, instruction} returns a string but runs only on the person's own OpenRouter key, so never use it; web.research {topic} returns {summary,sources}. A file the person gives the Block (a text, CSV, Markdown or JSON file on a file input) is read with await stillmade.files.read(input.<port>), which returns its bytes; turn them into text in main.js as the worked file input below does, and give the test's input file an assetId whose text mocks.files records as {files:{"<assetId>":{text:"..."}}}. A Block that reads a file the person gives it is a module Block, never a text.generate capability package. A PDF (lesson slides, a lease, a contract) is read with await stillmade.media.transform({operation:"document.text",source:input.<port>}), which returns {text, pages}, the words of each page, on the person's device and without credits (a text file reads the same way); a scanned PDF has no words to read. A call that fails throws: let it fail the run rather than catching it and returning an empty result. Media work on the person's device (cut, join or resize video, decode or extract audio, thumbnails, joining images) is await stillmade.media.transform(request). Joining saved images into one picture (two photos side by side with Before and After labels, a grid or collage, a contact sheet, a storyboard sheet, a poster with a title) is {operation:"image.compose",width,height,background:"#ffffff",layers:[{source,x,y,width,height,fit:"cover" or "contain"}],text:[{text,x,y,size,color,background,align:"left" or "center" or "right",width}]}: a canvas up to 4096 pixels a side, positions in pixels worked out in main.js, layers drawn in order and text on top (y is the top of the line, background puts a box behind it, colours are #rrggbb or, see-through, #rrggbbaa such as #00000099 for a caption box, width wraps the words onto lines no wider than it: give a caption under a frame the frame's width so it never runs into the next one); it returns the saved image, and main.js never decodes pixels itself. Exporting is a transform too: {operation:"timeline.render",kind:"video" or "audio",aspect:"9:16",clips:[{source,start,duration,from,transition:{type:"Fade"},look:"Filmic"}],audio:[{source,start,duration}],captions:[{text,start,end,words:[{text,start,end}],highlight:"#ffd400"}]} renders the Block's own timeline with the Editor's transitions, looks and word-by-word captions under the person's export plan (start is a clip's place on that timeline, so each clip starts where the one before it ends, less the transition on the clip before it, which blends that clip into the next, and from is where it begins in its own file; overlapping clips are refused), so never ask for a separate export, captions or color Block; stillmade.log and stillmade.progress(fraction, label) report. One press of the Step's button approves each hosted operation once at the price shown beside it, and media.analyze for every picture or video the Step is given, across as many calls as that takes; any further call asks the person first. A module Block runs in the person's browser while StillMade is open, never on a schedule or event. Tests are {name,input,expected,mocks} and compare outputs, so every test is a case that runs to its outputs; an input the Block refuses (it throws a plain Error with a message for the person, such as a clip with no duration) is never a test. mocks records what each call answers, as {hosted:{"audio.transcribe":{schemaVersion:1,text:"...",words:[...]},"text.generate":{...}},transforms:{...}} (a list answers successive calls in order), and expected is the exact output main.js returns for those answers: exact values of each output's type, never capability bounds such as minDuration or maxBytes. In a module test, mock each media call with its complete saved reference (kind, assetId, versionId, url, mimeType and its measured fields: duration, bytes, sampleRate, channels for audio; width, height, bytes for images) and repeat those references unchanged in expected; names or labels go in their own output. Worked outputs, each a passing test. A document output is the document object on a document port (not text): {"outputs":{"sheet":{"type":"document"}},"main.js":"export default async function run(input, stillmade) { const text = await stillmade.hosted.run('text.generate', {prompt: 'A Markdown care sheet for ' + input.plant, maxTokens: 800}); return {sheet: {schemaVersion: 1, format: 'markdown', text, title: input.plant + ' care'}}; }","test":{"name":"Fern","input":{"plant":"Fern"},"mocks":{"hosted":{"text.generate":"## Light\nBright, indirect light."}},"expected":{"sheet":{"schemaVersion":1,"format":"markdown","text":"## Light\nBright, indirect light.","title":"Fern care"}}}}. A file output is the reference await stillmade.files.write(data, {mediaType, name}) returns, handed on unchanged; mediaType is one of application/pdf, text/plain, text/csv, text/markdown, application/json, application/zip: {"outputs":{"report":{"type":"file"}},"main.js":"export default async function run(input, stillmade) { const csv = 'Room\\n' + input.rooms.join('\\n'); return {report: await stillmade.files.write(csv, {mediaType: 'text/csv', name: 'rooms.csv'})}; }","test":{"name":"Two rooms","input":{"rooms":["Kitchen","Hall"]},"expected":{"report":{"kind":"file","mimeType":"text/csv","name":"rooms.csv"}}}} (in a test, files.write answers with fields the host makes: assetId local-written-file-N, versionId, url /api/media/stillmade-mock/written-file-N.csv with the extension of its mediaType, and bytes; expected gives that reference as {kind:"file", mimeType, name} and it matches when those three match). An audio[] output is the list of references audio.speech returned, unchanged; its mock is a list answering successive calls, and the labels go in a separate text[] output: {"outputs":{"tracks":{"type":"audio[]"},"names":{"type":"text[]"}},"main.js":"export default async function run(input, stillmade) { const tracks = []; for (const room of input.rooms) tracks.push(await stillmade.hosted.run('audio.speech', {text: 'Welcome to the ' + room + '.'})); return {tracks, names: input.rooms}; }","test":{"name":"Two rooms","input":{"rooms":["Kitchen","Hall"]},"mocks":{"hosted":{"audio.speech":[{"kind":"audio","assetId":"speech-1","versionId":"v1","url":"/api/media/example/kitchen.wav","mimeType":"audio/wav","duration":2,"bytes":96044,"sampleRate":24000,"channels":1},{"kind":"audio","assetId":"speech-2","versionId":"v1","url":"/api/media/example/hall.wav","mimeType":"audio/wav","duration":2,"bytes":96044,"sampleRate":24000,"channels":1}]}},"expected":{"tracks":[{"kind":"audio","assetId":"speech-1","versionId":"v1","url":"/api/media/example/kitchen.wav","mimeType":"audio/wav","duration":2,"bytes":96044,"sampleRate":24000,"channels":1},{"kind":"audio","assetId":"speech-2","versionId":"v1","url":"/api/media/example/hall.wav","mimeType":"audio/wav","duration":2,"bytes":96044,"sampleRate":24000,"channels":1}],"names":["Kitchen","Hall"]}}}. A file input read as text: {"outputs":{"guide":{"type":"text"}},"main.js":"function textOf(buffer) { const bytes = new Uint8Array(buffer); let binary = ''; for (let i = 0; i < bytes.length; i += 8192) binary += String.fromCharCode.apply(null, bytes.subarray(i, i + 8192)); try { return decodeURIComponent(escape(binary)); } catch (error) { return binary; } }\nexport default async function run(input, stillmade) { const notes = textOf(await stillmade.files.read(input.notes)); const guide = await stillmade.hosted.run('text.generate', {prompt: 'Rewrite these notes as numbered steps, keeping every number as written:\\n' + notes, maxTokens: 1200}); return {guide}; }","test":{"name":"Two notes","input":{"notes":{"kind":"file","assetId":"notes-1","versionId":"v1","url":"/api/media/example/notes.txt","mimeType":"text/plain","name":"notes.txt","bytes":31}},"mocks":{"files":{"notes-1":{"text":"pull the probe\nwipe the tines"}},"hosted":{"text.generate":"1. Pull the probe.\n2. Wipe the tines."}},"expected":{"guide":"1. Pull the probe.\n2. Wipe the tines."}}}. Photos joined into one image, its transform mocked with the saved image it returns: {"outputs":{"sheet":{"type":"image"}},"main.js":"export default async function run(input, stillmade) { const photos = input.photos; if (!photos.length) throw new Error('Add at least one photo.'); const columns = Math.ceil(Math.sqrt(photos.length)), rows = Math.ceil(photos.length / columns), cell = 400, gap = 20, top = 160; const width = columns * cell + (columns + 1) * gap, height = top + rows * cell + (rows + 1) * gap; const layers = photos.map((source, i) => ({source, x: gap + (i % columns) * (cell + gap), y: top + gap + Math.floor(i / columns) * (cell + gap), width: cell, height: cell, fit: 'cover'})); const sheet = await stillmade.media.transform({operation: 'image.compose', width, height, background: '#ffffff', layers, text: [{text: input.title, x: width / 2, y: 50, size: 64, align: 'center', color: '#111111'}]}); return {sheet}; }","test":{"name":"Four photos","input":{"photos":[{"kind":"image","assetId":"photo-1","versionId":"v1","url":"/api/media/example/photo-1.jpg","mimeType":"image/jpeg","width":1600,"height":1200,"bytes":412000},{"kind":"image","assetId":"photo-2","versionId":"v1","url":"/api/media/example/photo-2.jpg","mimeType":"image/jpeg","width":1600,"height":1200,"bytes":412000},{"kind":"image","assetId":"photo-3","versionId":"v1","url":"/api/media/example/photo-3.jpg","mimeType":"image/jpeg","width":1600,"height":1200,"bytes":412000},{"kind":"image","assetId":"photo-4","versionId":"v1","url":"/api/media/example/photo-4.jpg","mimeType":"image/jpeg","width":1600,"height":1200,"bytes":412000}],"title":"Saturday games"},"mocks":{"transforms":{"image.compose":{"kind":"image","assetId":"sheet-1","versionId":"v1","url":"/api/media/example/sheet.png","mimeType":"image/png","width":1300,"height":1440,"bytes":1820000}}},"expected":{"sheet":{"kind":"image","assetId":"sheet-1","versionId":"v1","url":"/api/media/example/sheet.png","mimeType":"image/png","width":1300,"height":1440,"bytes":1820000}}}}. StillMade runs every test before anyone sees the Block, without spending credits. With the SDK instead, the same code is module/main.js in a Block folder and the same tests are tests/fixtures.json; stillmade-block pack writes the file to import (MODULES.md).
+A job that needs several hosted operations, or code around one (reading a transcript's words and times, shaping a list), or media work on the person's device even with no hosted operation (rendering clips and photos into one video, joining pictures into one), is a module Block: {manifest,moduleSource,tests}, runtime:"module", moduleSource:{"main.js":"<the code>"}. StillMade names main.js by its digest and adds module. StillMade's in-app builder takes one file: give no other files, and main.js is plain JavaScript that imports nothing and uses no browser APIs (no fetch, TextEncoder, timers or WebCodecs). That single-file rule is the in-app builder's; a module Block built with the SDK may use timers, WebCodecs, bundled npm packages and .wasm files as MODULES.md describes. main.js: export default async function run(input, stillmade) returns the declared outputs. It calls each hosted operation with await stillmade.hosted.run(operation, request), listed in manifest.permissions.capabilities: text.generate, image.generate, video.generate, audio.speech, audio.music, audio.sfx, audio.transcribe, image.describe, media.analyze, web.fetch, web.research. Requests, and what each call returns (a test's mocks record exactly that value): text.generate {prompt, system, maxTokens} returns the text as a string, and with schema (a JSON schema object) returns that object, so ask for JSON with schema; one text.generate prompt holds at most 32,000 characters, so a longer file or transcript is read in parts (one call per part of about 25,000 characters, then one call that joins what they found), never cut off, and when the result is the whole text again (a cleaned transcript, a translation) each part's answer is kept and joined in order in main.js, since one answer holds at most 4,096 tokens; audio.transcribe {audio: the input's saved audio or video reference (a video's own sound is transcribed, so never extract it first), language:"auto"} returns {schemaVersion:1,text,words:[{text,start,end}]}; image.generate {prompt, references?: one to eight saved images the new picture keeps, such as the product, person or place in the person's photos} and audio.speech {text} return a saved image or audio reference (an audio reference always carries its measured duration); web.fetch {url, instruction} returns a string; media.analyze {media: one to four saved images or one saved video, question, range?: {start, end} in seconds of that video, at most 12 apart, to look closely at one part of it} returns {schemaVersion:1,items:[{kind,summary,observations,detectedText}]}, one item per image, and is how a Block looks at pictures (picking the photos that show different angles, reading a menu off a photo): StillMade credits pay for it; image.describe {image, instruction} returns a string but runs only on the person's own OpenRouter key, so never use it; web.research {topic} returns {summary,sources}. A file the person gives the Block (a text, CSV, Markdown or JSON file on a file input) is read with await stillmade.files.read(input.<port>), which returns its bytes; turn them into text in main.js as the worked file input below does, and give the test's input file an assetId whose text mocks.files records as {files:{"<assetId>":{text:"..."}}}. A Block that reads a file the person gives it is a module Block, never a text.generate capability package. A PDF (lesson slides, a lease, a contract) is read with await stillmade.media.transform({operation:"document.text",source:input.<port>}), which returns {text, pages}, all the words and pages, a list holding each page's words as one string, on the person's device and without credits (a text file reads the same way); a scanned PDF has no words to read. A call that fails throws: let it fail the run rather than catching it and returning an empty result. Media work on the person's device (cut, join or resize video, decode or extract audio, thumbnails, joining images) is await stillmade.media.transform(request). Joining saved images into one picture (two photos side by side with Before and After labels, a grid or collage, a contact sheet, a storyboard sheet, a poster with a title) is {operation:"image.compose",width,height,background:"#ffffff",layers:[{source,x,y,width,height,fit:"cover" or "contain"}],text:[{text,x,y,size,color,background,align:"left" or "center" or "right",width}]}: a canvas up to 4096 pixels a side, positions in pixels worked out in main.js, layers drawn in order and text on top (y is the top of the line, background puts a box behind it, colours are #rrggbb or, see-through, #rrggbbaa such as #00000099 for a caption box, width wraps the words onto lines no wider than it: give a caption under a frame the frame's width so it never runs into the next one); it returns the saved image, and main.js never decodes pixels itself. Exporting is a transform too: {operation:"timeline.render",kind:"video" or "audio",aspect:"9:16",clips:[{source,start,duration,from,transition:{type:"Fade"},look:"Filmic",motion:"zoom-in"}],audio:[{source,start,duration}],captions:[{text,start,end,words:[{text,start,end}],highlight:"#ffd400",size:80}]} (motion, a slow camera move over a photo, is zoom-in, zoom-out, pan-left, pan-right, pan-up or pan-down; size is the caption height in pixels, 80 for a vertical video) renders the Block's own timeline with the Editor's transitions, looks and word-by-word captions under the person's export plan (start is a clip's place on that timeline, so each clip starts where the one before it ends, less the transition on the clip before it, which blends that clip into the next, and from is where it begins in its own file; overlapping clips are refused; a clip of another shape than the frame is cropped to fill it around its middle, or around focus:{x,y} on the clip, the point to keep in view as fractions 0 to 1 across the picture, such as where the speaker stands (focus.scale, 1 to 2.5 and instead of a motion, holds the clip that much closer on that point), or is fitted whole with fit:"contain"), so never ask for a separate export, captions or color Block; stillmade.log and stillmade.progress(fraction, label) report. One press of the Step's button approves each hosted operation once at the price shown beside it, and media.analyze for every picture or video the Step is given, across as many calls as that takes; any further call asks the person first. A module Block runs in the person's browser while StillMade is open, never on a schedule or event. Tests are {name,input,expected,mocks} and compare outputs, so every test is a case that runs to its outputs; an input the Block refuses (it throws a plain Error with a message for the person, such as a clip with no duration) is never a test. mocks records what each call answers, as {hosted:{"audio.transcribe":{schemaVersion:1,text:"...",words:[...]},"text.generate":{...}},transforms:{...}} (a list answers successive calls in order), and expected is the exact output main.js returns for those answers: exact values of each output's type, never capability bounds such as minDuration or maxBytes. In a module test, mock each media call with its complete saved reference (kind, assetId, versionId, url, mimeType and its measured fields: duration, bytes, sampleRate, channels for audio; width, height, bytes for images), its url an example file such as /api/media/example/clip.mp4 so the preview can play it, and repeat those references unchanged in expected; names or labels go in their own output. Worked outputs, each a passing test. A document output is the document object on a document port (not text): {"outputs":{"sheet":{"type":"document"}},"main.js":"export default async function run(input, stillmade) { const text = await stillmade.hosted.run('text.generate', {prompt: 'A Markdown care sheet for ' + input.plant, maxTokens: 800}); return {sheet: {schemaVersion: 1, format: 'markdown', text, title: input.plant + ' care'}}; }","test":{"name":"Fern","input":{"plant":"Fern"},"mocks":{"hosted":{"text.generate":"## Light\nBright, indirect light."}},"expected":{"sheet":{"schemaVersion":1,"format":"markdown","text":"## Light\nBright, indirect light.","title":"Fern care"}}}}. A file output is the reference await stillmade.files.write(data, {mediaType, name}) returns, handed on unchanged; mediaType is one of application/pdf, text/plain, text/csv, text/markdown, application/json, application/zip: {"outputs":{"report":{"type":"file"}},"main.js":"export default async function run(input, stillmade) { const csv = 'Room\\n' + input.rooms.join('\\n'); return {report: await stillmade.files.write(csv, {mediaType: 'text/csv', name: 'rooms.csv'})}; }","test":{"name":"Two rooms","input":{"rooms":["Kitchen","Hall"]},"expected":{"report":{"kind":"file","mimeType":"text/csv","name":"rooms.csv"}}}} (in a test, files.write answers with fields the host makes: assetId local-written-file-N, versionId, url /api/media/stillmade-mock/written-file-N.csv with the extension of its mediaType, and bytes; expected gives that reference as {kind:"file", mimeType, name} and it matches when those three match; passed on in an asset output as kind "asset", it is given as {kind:"asset", mimeType, name} and matches the same way). An audio[] output is the list of references audio.speech returned, unchanged; its mock is a list answering successive calls, and the labels go in a separate text[] output: {"outputs":{"tracks":{"type":"audio[]"},"names":{"type":"text[]"}},"main.js":"export default async function run(input, stillmade) { const tracks = []; for (const room of input.rooms) tracks.push(await stillmade.hosted.run('audio.speech', {text: 'Welcome to the ' + room + '.'})); return {tracks, names: input.rooms}; }","test":{"name":"Two rooms","input":{"rooms":["Kitchen","Hall"]},"mocks":{"hosted":{"audio.speech":[{"kind":"audio","assetId":"speech-1","versionId":"v1","url":"/api/media/example/kitchen.wav","mimeType":"audio/wav","duration":2,"bytes":96044,"sampleRate":24000,"channels":1},{"kind":"audio","assetId":"speech-2","versionId":"v1","url":"/api/media/example/hall.wav","mimeType":"audio/wav","duration":2,"bytes":96044,"sampleRate":24000,"channels":1}]}},"expected":{"tracks":[{"kind":"audio","assetId":"speech-1","versionId":"v1","url":"/api/media/example/kitchen.wav","mimeType":"audio/wav","duration":2,"bytes":96044,"sampleRate":24000,"channels":1},{"kind":"audio","assetId":"speech-2","versionId":"v1","url":"/api/media/example/hall.wav","mimeType":"audio/wav","duration":2,"bytes":96044,"sampleRate":24000,"channels":1}],"names":["Kitchen","Hall"]}}}. A file input read as text: {"outputs":{"guide":{"type":"text"}},"main.js":"function textOf(buffer) { const bytes = new Uint8Array(buffer); let binary = ''; for (let i = 0; i < bytes.length; i += 8192) binary += String.fromCharCode.apply(null, bytes.subarray(i, i + 8192)); try { return decodeURIComponent(escape(binary)); } catch (error) { return binary; } }\nexport default async function run(input, stillmade) { const notes = textOf(await stillmade.files.read(input.notes)); const guide = await stillmade.hosted.run('text.generate', {prompt: 'Rewrite these notes as numbered steps, keeping every number as written:\\n' + notes, maxTokens: 1200}); return {guide}; }","test":{"name":"Two notes","input":{"notes":{"kind":"file","assetId":"notes-1","versionId":"v1","url":"/api/media/example/notes.txt","mimeType":"text/plain","name":"notes.txt","bytes":31}},"mocks":{"files":{"notes-1":{"text":"pull the probe\nwipe the tines"}},"hosted":{"text.generate":"1. Pull the probe.\n2. Wipe the tines."}},"expected":{"guide":"1. Pull the probe.\n2. Wipe the tines."}}}. Photos joined into one image, its transform mocked with the saved image it returns: {"outputs":{"sheet":{"type":"image"}},"main.js":"export default async function run(input, stillmade) { const photos = input.photos; if (!photos.length) throw new Error('Add at least one photo.'); const columns = Math.ceil(Math.sqrt(photos.length)), rows = Math.ceil(photos.length / columns), cell = 400, gap = 20, top = 160; const width = columns * cell + (columns + 1) * gap, height = top + rows * cell + (rows + 1) * gap; const layers = photos.map((source, i) => ({source, x: gap + (i % columns) * (cell + gap), y: top + gap + Math.floor(i / columns) * (cell + gap), width: cell, height: cell, fit: 'cover'})); const sheet = await stillmade.media.transform({operation: 'image.compose', width, height, background: '#ffffff', layers, text: [{text: input.title, x: width / 2, y: 50, size: 64, align: 'center', color: '#111111'}]}); return {sheet}; }","test":{"name":"Four photos","input":{"photos":[{"kind":"image","assetId":"photo-1","versionId":"v1","url":"/api/media/example/photo-1.jpg","mimeType":"image/jpeg","width":1600,"height":1200,"bytes":412000},{"kind":"image","assetId":"photo-2","versionId":"v1","url":"/api/media/example/photo-2.jpg","mimeType":"image/jpeg","width":1600,"height":1200,"bytes":412000},{"kind":"image","assetId":"photo-3","versionId":"v1","url":"/api/media/example/photo-3.jpg","mimeType":"image/jpeg","width":1600,"height":1200,"bytes":412000},{"kind":"image","assetId":"photo-4","versionId":"v1","url":"/api/media/example/photo-4.jpg","mimeType":"image/jpeg","width":1600,"height":1200,"bytes":412000}],"title":"Saturday games"},"mocks":{"transforms":{"image.compose":{"kind":"image","assetId":"sheet-1","versionId":"v1","url":"/api/media/example/sheet.png","mimeType":"image/png","width":1300,"height":1440,"bytes":1820000}}},"expected":{"sheet":{"kind":"image","assetId":"sheet-1","versionId":"v1","url":"/api/media/example/sheet.png","mimeType":"image/png","width":1300,"height":1440,"bytes":1820000}}}}. StillMade runs every test before anyone sees the Block, without spending credits. With the SDK instead, the same code is module/main.js in a Block folder and the same tests are tests/fixtures.json; stillmade-block pack writes the file to import (MODULES.md).
 
 ### ComfyUI workflows
 
@@ -234,7 +234,7 @@ Optional custom view interfaces are supported, but generated manifest.ui control
 
 *When:* StillMade's builder adds this section when the person asks for an interface, preview or controls, or refines a Block that has one.
 
-A custom interface is the optional package-level view:{html,css?,themedCss?,javascript?}; when refining, preserve and refine base.view. Design it by StillMade's interface principles, in order: Function first: unless the person asks for a design direction, build from the Block kit (`sm-block`, `sm-stage`, `sm-choices`, `sm-actions`, `sm-more`) and ship little or no CSS of your own. A direction the person names wins over the kit. Show the thing being made: the preview, media or result gets the largest area (`sm-stage`). Never a JSON dump, source editor, manifest or diagnostic panel unless the task is about source. One obvious next action (`sm-primary`); peers are quiet buttons or named icons. Seven or fewer choices in view. The action button names what it does ("Write", "Generate image", "Make video"), never "Run", "Run step", "Submit" or "Review cost". Show no credit numbers: StillMade puts the price beside the verb on its own run button ("Write · 1 credit"), with a chip for the model, quality or count when they can change. Pressing it runs at that price; a confirmation appears only when the server price is higher. Pictures over words: choices people see or hear are thumbnails, swatches, samples or glyphs (`sm-choices`), not word lists. Icons carry an accessible name and tooltip; never emoji. Words only when needed. Drop what nobody needs to finish the task (intros, "how it works", restating the Block name). Words that might help start collapsed (`sm-hint`, `sm-more`). Errors, spending approvals, stale inputs and destructive steps stay visible, in few words. Strong defaults; secondary settings go in `sm-more`, closed. Compact: the whole Block fits one laptop screen without scrolling. A short setting is one row, its label beside the control; no card around a single control. Status is a dot or a few words ("No narration yet"), never a paragraph; the reason sits behind an `sm-hint`. Files are tiles: a saved file is a thumbnail or chip and adding one is a + tile. Never the bare browser file picker, and no size limits in view until a file is too big. Rare tools (update, schedule, history, reload) are named icon buttons beside the action, never sections under the work. An (i) beside anything a newcomer could misread: a setting, status or number whose effect is not obvious from its label gets a closed `sm-hint` with one or two sentences. Give every input that needs one a `description`; StillMade shows it as that (i). Obvious controls get none. One workspace: no cards inside cards, no second title, step counter or Back/Next bar; the host owns navigation. Status sits beside the action it describes (`sm-status`), in a few words. Keep the page in normal scroll flow; no sticky media or control strips unless the editing task needs them. Use the width: from 860px wide the kit puts the controls on the left at a reading width (300–380px) and the result (`sm-stage`) fills the right, in its real shape; under that everything stacks with the result first and the action within reach. Extra width goes to the result, never to wider text fields. Before there is a result, the stage previews what will be made from the inputs. `data-layout="stack"` opts out. Settings sit side by side (`sm-settings`), next to what they change. Three or four short options are one segmented row (`data-segmented`), not a row of big tiles. Lists fill the width at a small size: input files are small tiles (`sm-gallery data-size="small"`), many items are one line each (`sm-list`): a thumbnail, one line of text, a short fact. Tapping an item opens all of its fields in place. Extra width means more items per row, never bigger cards. On a phone chip and segment rows scroll sideways instead of wrapping; long lists show their first three with "N more" (`data-peek`) and long text five lines until tapped (`sm-clamp`). Nothing is lost when a Block gets smaller: every control stays on the screen, behind an (i), in `sm-more`, or inside the item it belongs to. Compacting moves controls; it never deletes them. Text scene descriptions are planning notes, never a storyboard or an approved frame. Its DOM controller is separate from manifest runtime code. The view may also declare appearance:"original" or "host". For an imported design, preserve its colors with appearance:"original" and view.css; no themed replacement is required. Preserve base.view.appearance when refining unless the user requests a change. Omitted/host appearance retains the existing Light, Dark and White requirement: supply semantic colors in view.css or a complete replacement view.themedCss. An original-style view may optionally supply valid themedCss for a StillMade appearance preview. Use StillMade variables (--bg,--fg,--fg-muted,--border-tok,--r-3) only when opting into host styling; do not redefine host tokens in the themed variant. Appearance never relaxes sandbox constraints: no CSS escapes/filters/blending, imports/url(), or element.style mutations in JS. Use classes for UI state; media pixels retain their colors. Both CSS fields have the same 32 KiB limit; all existing shared-state, security and permission checks still apply. Declare elements in html, bind with document.getElementById/querySelector and addEventListener; use literal property access, array methods, textContent, input.value, button.disabled and canvas. Initial values: StillMade.onInput(values=>{...}); await StillMade.run(inputBindings) returns validated output bindings from the isolated package runtime; catch errors and show results. No computation in the DOM controller; put it in code/recipe. Disable the action button while pending.
+A custom interface is the optional package-level view:{html,css?,themedCss?,javascript?}; when refining, preserve and refine base.view. Design it by StillMade's interface principles, in order: Function first: unless the person asks for a design direction, build from the Block kit (`sm-block`, `sm-stage`, `sm-choices`, `sm-actions`, `sm-more`) and ship little or no CSS of your own. A direction the person names wins over the kit. Show the thing being made: the preview, media or result gets the largest area (`sm-stage`). A picture or clip the Block makes fills it in its own shape, as large as fits; several fill it as a grid, each big enough to watch. Small tiles are for the files a person gives the Block, never for what it makes. Never a JSON dump, source editor, manifest or diagnostic panel unless the task is about source. One obvious next action (`sm-primary`); peers are quiet buttons or named icons. Seven or fewer choices in view. The action button names what it does ("Write", "Generate image", "Make video"), never "Run", "Run step", "Submit" or "Review cost". Show no credit numbers: StillMade puts the price beside the verb on its own run button ("Write · 1 credit"), with a chip for the model, quality or count when they can change. Pressing it runs at that price; a confirmation appears only when the server price is higher. Pictures over words: choices people see or hear are thumbnails, swatches, samples or glyphs (`sm-choices`), not word lists. Icons carry an accessible name and tooltip; never emoji. Words only when needed. Drop what nobody needs to finish the task (intros, "how it works", restating the Block name). Words that might help start collapsed (`sm-hint`, `sm-more`). Errors, spending approvals, stale inputs and destructive steps stay visible, in few words. Strong defaults; secondary settings go in `sm-more`, closed. Compact: the whole Block fits one laptop screen without scrolling. A short setting is one row, its label beside the control; no card around a single control. Status is a dot or a few words ("No narration yet"), never a paragraph; the reason sits behind an `sm-hint`. Files are tiles: a saved file is a thumbnail or chip and adding one is a + tile. Never the bare browser file picker, and no size limits in view until a file is too big. Rare tools (update, schedule, history, reload) are named icon buttons beside the action, never sections under the work. An (i) beside anything a newcomer could misread: a setting, status or number whose effect is not obvious from its label gets a closed `sm-hint` with one or two sentences. Give every input that needs one a `description`; StillMade shows it as that (i). Obvious controls get none. One workspace: no cards inside cards, no second title, step counter or Back/Next bar; the host owns navigation. Status sits beside the action it describes (`sm-status`), in a few words. Keep the page in normal scroll flow; no sticky media or control strips unless the editing task needs them. Use the width: from 860px wide the kit puts the controls on the left at a reading width (300–380px) and the result (`sm-stage`) fills the right, in its real shape; under that everything stacks with the result first and the action within reach. Extra width goes to the result, never to wider text fields. Before there is a result, the stage previews what will be made from the inputs. `data-layout="stack"` opts out. Size with the kit, never fixed sizes: no fixed pixel widths or heights for the workspace or stage, no 100vh, no position:fixed panels. StillMade zooms a Block's screen to fill the space it has (up to 150%, never down to a phone layout on a computer), so a Block laid out with the kit's spacing fills its Step without empty space. Settings sit side by side (`sm-settings`), next to what they change. Three or four short options are one segmented row (`data-segmented`), not a row of big tiles. Lists fill the width at a small size: input files are small tiles (`sm-gallery data-size="small"`), many items are one line each (`sm-list`): a thumbnail, one line of text, a short fact. Tapping an item opens all of its fields in place. Extra width means more items per row, never bigger cards. On a phone chip and segment rows scroll sideways instead of wrapping; long lists show their first three with "N more" (`data-peek`) and long text five lines until tapped (`sm-clamp`). Nothing is lost when a Block gets smaller: every control stays on the screen, behind an (i), in `sm-more`, or inside the item it belongs to. Compacting moves controls; it never deletes them. Text scene descriptions are planning notes, never a storyboard or an approved frame. Its DOM controller is separate from manifest runtime code. The view may also declare appearance:"original" or "host". For an imported design, preserve its colors with appearance:"original" and view.css; no themed replacement is required. Preserve base.view.appearance when refining unless the user requests a change. Omitted/host appearance retains the existing Light, Dark and White requirement: supply semantic colors in view.css or a complete replacement view.themedCss. An original-style view may optionally supply valid themedCss for a StillMade appearance preview. Use StillMade variables (--bg,--fg,--fg-muted,--border-tok,--r-3) only when opting into host styling; do not redefine host tokens in the themed variant. Appearance never relaxes sandbox constraints: no CSS escapes/filters/blending, imports/url(), or element.style mutations in JS. Use classes for UI state; media pixels retain their colors. Both CSS fields have the same 32 KiB limit; all existing shared-state, security and permission checks still apply. Declare elements in html, bind with document.getElementById/querySelector and addEventListener; use literal property access, array methods, textContent, input.value, button.disabled and canvas. Initial values: StillMade.onInput(values=>{...}); await StillMade.run(inputBindings) returns validated output bindings from the isolated package runtime; catch errors and show results. No computation in the DOM controller; put it in code/recipe. Disable the action button while pending.
 Standard custom view code (no `view.config.json` frame runtime) is checked before it runs. These rules come straight from the validator; frame views run in their own sandbox and may build the page with ordinary DOM calls such as `createElement`:
 
 - **HTML elements:** only `div`, `section`, `article`, `header`, `footer`, `main`, `aside`, `nav`, `h1`, `h2`, `h3`, `h4`, `p`, `span`, `strong`, `em`, `i`, `small`, `br`, `hr`, `label`, `input`, `textarea`, `button`, `select`, `option`, `optgroup`, `fieldset`, `legend`, `output`, `progress`, `meter`, `ul`, `ol`, `li`, `dl`, `dt`, `dd`, `figure`, `figcaption`, `img`, `video`, `audio`, `source`, `canvas`, `table`, `thead`, `tbody`, `tfoot`, `tr`, `th`, `td`, `pre`, `code`, `details`, `summary`. No scripts, iframes, SVG, forms, comments or inline event attributes; media URLs must be `data:` or `blob:`.
@@ -329,7 +329,8 @@ A test or example that needs a real photo, recording, clip or file can name one
 StillMade's example files instead of carrying its own: `/api/media/example/photo-1.jpg`
 to `photo-4.jpg` (1600×1200, `photo-2.jpg` 1200×1600), `sheet.png` (1300×1440),
 `room.png` (1200×900), `kitchen.wav`, `hall.wav` and `click-120.wav` (2 s, 24 kHz mono),
-`audio.mp3` (2 s), `clip.mp4` (640×360, 2 s), `notes.txt` and `sales.csv`, for example
+`audio.mp3` (2 s), `clip.mp4` (640×360, 2 s), `notes.txt`, `sales.csv` and `lesson.pdf`
+(two pages of words a PDF Block reads with `document.text`), for example
 `{"kind":"image","assetId":"photo-1","versionId":"v1","url":"/api/media/example/photo-1.jpg","mimeType":"image/jpeg","width":1600,"height":1200}`.
 StillMade serves them, its preview and phone and browser check open them, and `test`
 reads them when the test gives no mock for them. A Block whose interface needs a photo
@@ -1277,8 +1278,8 @@ inside the iframe; larger media should use the existing native Editor workspace.
 
 ## Tests and interconnection
 
-Include 2–6 small deterministic fixtures covering normal behavior and edge cases
-(a capability Block takes 1–3). The fields of each fixture depend on the runtime:
+Include 2 to 6 small deterministic fixtures covering normal behavior and edge cases
+(at most 50 are accepted; a capability Block takes 1–3). The fields of each fixture depend on the runtime:
 
 | Runtime | Each fixture |
 | --- | --- |