CueFrame

Atomically apply an array of CompositionOps to a project composition. All ops succeed or none do (transactional). Call describe_composition_ops without type for the compact operation list, then with a type for its exact JSON Schema. For negotiated component execution, read get_composition with componentExecutionVersions=["3","4"] and includeDependencies="true", then send componentExecutionVersions:[3,4] and that exact expectedBrandIdentity. A changed brand refuses the command; never refresh the witness to conceal a conflict. Explicit builtinSources selectors attach only the approved exact source pins referenced by the final composition. The additive `concurrency` envelope requires an endpoint confirmed to enforce guardedCompositionV1; unsupported hosted endpoints refuse it with unsupported_concurrency. Obtain dependency expectations from the observed composition; never rebuild a stale basis from current state. Pass `if_match` for optimistic-concurrency control (409 stale_etag on mismatch) and `dry_run: true` to validate the batch without persisting. Ops land on the project's ACTIVE composition unless you pass `composition_id` (mutually exclusive with `session_id`), which targets a specific one — that is how you author a second composition minted with new_composition and run an A/B. Use this to seed a composition (add media clips) or make targeted edits (set a crop intent, reorder clips, set the format). CAPTIONS FOR EDITED TIMELINES: after saving all clip edits, call again with {type:"captions.fromComposition", trackId} to build captions from every audible clip on that track, honoring trims, exclusions and placement. Clips must have duration matching playable source duration divided by playbackRate; unsupported timing fails explicitly. Re-run it after later timeline edits; captions do not update automatically. {type:"captions.fromTranscript", mediaId, window?} instead replaces all captions from one media transcript, rebased to zero; it does not follow a montage. Style with {type:"setCaptionStyle", style}; or set explicit words with {type:"setCaptions", segments, style?} (e.g. translations or your own VO timings). Brand is NOT set here — it lives on the project (create_project / update_project with brandKitId — or PATCH /projects/:id over REST). CARDS: a {kind:"card", html, tokens} clip projects each `tokens` KEY `x` to the CSS custom property `--cf-x`, so the html must read it as `var(--cf-x)` — write `tokens:{"plate":"#0af"}` + `background:var(--cf-plate)`. Do NOT put dashes in the key: `"--plate"` becomes `--cf---plate` and your `var(--plate)` then resolves to NOTHING, which CSS silently degrades to a transparent background and black text. Responses carry `warnings[]` when a card reads a var nothing defines. SCENE SHOTS: for a 3D product shot — a tilted page screenshot, a lit laptop/phone showing a real screenshot, two-colour type, a colour-field backdrop, focus/grain/vignette and a camera move — use {type:"scene.add", clipId, newTrack|trackId, startTime, duration, scene, assets} and revise it with {type:"scene.update", clipId, scene?, assets?, startTime?, duration?}. You author a scene SPEC, never wrapper TSX; the server registers the component for you. There is NO built-in scene primitive in list_catalog — READ cueframe://scene-shot first for the JSON schema, the canonical V2 world-objects recipe with nested groups, slabs, pages, phones, laptops, world/screen text, admitted surfaces, camera motion, lighting, and the current resource limits.

Insert a resolved builtin component package into an existing project. Pass template as the selected builtin id and body with projectId, the resolved documentSha256, insert:{startTime,idPrefix}, parameter values and explicit media bindings. Read get_composition first and retain expectedRevision, commandId, componentExecutionVersions:[3,4] and expectedBrandIdentity as required by the contract. Insertion preserves existing timeline content and package source offsets. Select pinned samples explicitly with media:{slot:{sample:true}}, or bind real media IDs. Missing required media is refused; samples are never selected implicitly. A stale revision, pin or brand refuses; do not silently refresh the witness. For ordinary templates, this operation retains its existing apply behavior.

Cancel an in-flight compose job (stop a runaway/mis-priced ensemble). Needs projectId + composeJobId (from compose). The director cancel is confirmed before the row flips; any linked charge is released. No-op-reject (422) if already terminal.

Cancel an in-flight NLE export (fcpxml/premiere). Needs projectId + exportId (from the export enqueue). No-op-reject (422) if already terminal.

Cancel a running async job (verify / preview / brand-kit-extract) returned by an enqueue call — abort a mis-priced job instead of waiting out its TTL. For box-backed jobs the warm box is reaped best-effort; cancel never meters. No-op-reject (422) if already terminal.

Cancel an in-flight render (stop a runaway/queued job). Needs projectId + renderId (from create_render). No-op if already terminal.

Start an async Director compose job: it authors an ensemble of full composition candidates, verifies each, and saves the winner to the project. Body must set EXACTLY ONE source (mutually exclusive): `briefId` — author FROM a persisted Brief (the PLANNED path: create_brief first; every Brief field is honored or the compose is refused loudly naming the field); `fromComposition: true` — author from the project's CURRENT composition (the GENERAL path: seed it with apply_composition first — your media clips are the intent, and any graphics you authored are CARRIED VERBATIM as locked clips the Director authors around, never re-authors or drops); or `suggestionId` — author from an AI clip suggestion (the CLIPPING path, paired with suggest_briefs). Returns { composeJobId, status, streamUrl }. Then call wait_job(kind="compose", id=composeJobId, projectId=...) to block until the winner is selected.

Ask the Director for editing advice on the project's CURRENT composition — it NEVER writes your composition (you stay the decider). Async: returns a consultJobId; then wait_job(kind="consult", id=consultJobId, projectId=...) surfaces the advisory envelope (assessment / proposals / principles / plan). `body.mode`: 'advise' = the planner agent (fast, ~30s-2min) returns a `plan` (framing/cuts/captions/graphics intents) as GUIDANCE you re-author from — no grounded scores, nothing to adopt; 'amend' (default) = the full author→judge→refine loop (minutes) returns a graded `assessment` + a fix candidate you ADOPT via select_candidate with its candidateIndex. ('explore' is coming soon → use amend.) `body.intent` = your goal in plain English; optional `body.focus` (editorial|spatial|brand|caption); `body.ifMatch` binds to your current etag. Use advise for a fast directional nudge, amend for a grounded, adoptable fix.

Create or UPSERT a brand kit (re-POSTing the same kitId updates it in place — this is also the brand-kit UPDATE path). Sets fonts/colors/logo. This is brand step 0 of the workflow: author the kit, then bind it to a PROJECT so colors/fonts/motion theme the output AUTOMATICALLY — pass its kitId as brandKitId to create_project, or attach it to an existing project via update_project (PATCH /projects/:id over REST). Without a bound kit, create_render produces UNBRANDED output. Tenant-scoped (no project arg).

Persist the video's INTENT as a Brief — the Director's rich input contract (spec: the storyboard skill's output lands here). Fields: goal (required), audience, platform, format, durationSec, tone, cta, locale (output language — all on-screen copy + captions), hook, caption, beats[] (each with an editorial intent and optional MOMENT PINNING {mediaId, startSec, endSec} — honored to-the-frame or refused), mustIncludes, captions {styleId, emphasisPhrases} (the big-serif emphasis treatment), motionStyle, seededGraphics (your authored overlay/component/card clips, carried VERBATIM as locked clips — never re-authored), exclusions (hard negatives: graphic kinds the Director must NEVER author + 'no-transitions'), gates {holdAt, autoApprove}. Then compose it: compose with {briefId}. The response carries the QUOTE (compose + estimated render) — relay it to your human BEFORE composing.

Top up the CURRENT org's credit wallet. Body { plan: "credit_pack_10"|"credit_pack_30"|"credit_pack_100" } → returns { url }: a hosted Stripe checkout page to open in a browser and confirm payment. There are no subscription tiers — a pack is a one-off top-up, and it is also the only thing that raises the generation (premium_credits) allowance. Always a URL — an existing saved card is NEVER charged silently; the human confirms on Stripe. The org is resolved from your key, never the body. Requires the billing:write scope. Only send a user here when get_account shows the relevant `balance` at or near zero. Discover the packs with list_plans.

Create and store a custom component for the tenant: a native scene that arranges its declared images and videos as tiles in 3D space. Returns the ComponentResponse (id + metadata) you can then reference from a composition. No project arg — components are tenant-scoped. SOURCE CONTRACT (version 4; the server instructions' Component Contract has the full rules and an example): send the complete source object {version:4, componentId, template, manifest, definition}. `template` is one sub-composition: a root <div id="root" data-composition-id="<componentId>" data-width data-height data-no-timeline data-composition-variables='…'> holding <div data-cf-scene></div>, and one inline script ending in __cfScene.mount({ root: document.getElementById('root'), component, variables: __hyperframes.getVariables() }). data-composition-variables is the definition's projection (scalar inputSchema properties, then manifest slots). The manifest is {version:4, runtime:'hyperframes', hfContract:1, kind:'scene', assets, graphics:{api:'webgl2',…}, alphaPolicy}; slots are image or video. The source is admitted before it is stored: template lint, determinism (no timers, Date, network or CSS animation) and the declaration are checked, and a refusal names the rule. A version-3 source is retired and refused as component_source_retired. Place it with execution {version:4, componentId, sourceVersionId, params, seed, assets:{ slot: mediaId }} and placement {version:3, kind:'clip'}; slots never travel in params.

Create a new project — the unit you compose and render into. compose / apply_composition / create_render / the preview family all require a projectId; this mints one. Pass name (required) plus optionally a format (aspectRatio/fps/resolution) and brandKitId. brandKitId binds a brand kit to the PROJECT (pass the kitId from create_brand_kit): colors/fonts/motion theme every render of this project automatically (omit it and the output is unbranded). Bind or change it later via update_project (PATCH /projects/:id over REST). Returns the ProjectResponse (id + metadata).

Queue an MP4 render of a project's composition. Returns { id (renderId), status, warnings? } — `warnings[]` carries non-fatal advisories (omitted when none). The render themes from the project's brand kit (bind one via create_project / update_project with brandKitId — or PATCH /projects/:id over REST; without one the output is unbranded). Then call wait_job(kind="render", id=renderId, projectId=...) to block until the file is ready and get the download URL, or register a webhook (create_webhook on render.completed — status rides the payload).

Ingest a LOCAL file — bytes you hold, when there is no public URL to hand import_media (prefer import_media whenever the bytes already sit at an https URL). TWO steps: (1) this tool reserves a media item and returns { id, uploadUrl, expiresAt } — a short-lived PUT URL. PUT the raw file bytes to uploadUrl over HTTPS with your declared Content-Type (≤ ~5 GB per PUT). (2) call finalize_upload({ mediaItemId: id }) to start processing. Body: filename, contentType (allowed enum), size (bytes). Thereafter it is async exactly like import_media: poll list_media for processingStatus, or register a webhook (create_webhook on media.completed). Use the returned id with get_media_context, suggest_briefs, or as a composition source.

Register a webhook so long-running jobs notify you instead of blocking on wait_job — the right pattern for fan-out/async agents. Pass url + events — ONE terminal event per job kind with status INSIDE the payload (render.completed, compose.completed, media.completed, suggest.completed) plus checkpoint.ready (a held compose paused for review). DELIVERY: a POST { event, timestamp, projectId?, data } with Standard-Webhooks headers (webhook-id, webhook-timestamp, webhook-signature: v1,<base64>); the secret (whsec_<base64>) is returned ONCE — verify the RAW body with new Webhook(secret).verify(rawBody, headers). ACTIVATION is synchronous: CueFrame POSTs a signed webhook.verify challenge and your endpoint must reply 2xx echoing data.token — otherwise it returns status:"pending" and receives NO events (recover with verify_webhook, no delete/recreate). Read the cueframe://webhook-events resource (or GET /webhooks/events) for the full contract + every event's data shape BEFORE building your receiver.

Delete a brief. Compose jobs already run from it are unaffected (they snapshot).

Delete a composition (and its version history) from a project. REFUSES the ACTIVE composition (409) — make another one active first, or delete the project. Use list_compositions to see which is active.

Delete a registered critic by its criticId. It stops running in the Director gates and score_composition immediately.

Permanently delete a media item from your library (cleanup). Irreversible. Compositions referencing it will fail to render — only delete media no composition still needs.

Permanently delete a project you created (cleanup). Removes its composition and render history. Irreversible — only call on a project you own and no longer need.

Delete a webhook subscription (stop receiving its deliveries). Irreversible — the signing secret is invalidated.

Derive a reframed SIBLING composition for a different aspect ratio (the agent-native Auto-Reframe / Magic-Resize). Pass projectId + body { fromCompositionId, targetFormat: { aspectRatio, fps?, resolution? } }. CueFrame copies the source composition's authored intent and content-aware-reframes every clip for the new aspect (point/face/active-speaker focus), then saves a new independent composition in the same project. Returns { compositionId } — render it with create_render, or refine it with apply_composition. Use this to make 9:16 + 1:1 + 4:5 cuts from one 16:9 edit. (To only change resolution within the SAME aspect, pass create_render.format instead — that is a scale-only export override, not a new composition.)

Discover the exact CompositionOp vocabulary accepted by apply_composition. Omit type for a compact list of operation names and descriptions; pass type for only that operation's JSON Schema. Read this before authoring edits or captions.

Queue brand-token extraction from a website URL — ASYNC: returns a jobId immediately; then wait_job(kind="brand_kit_extract", id=jobId) for the tokens (colors/fonts/title/description, fetched + regex-parsed from the site's HTML/CSS — NOT a screenshot/vision pass). It does NOT persist — review the tokens, then call create_brand_kit to save a kit. A blocked (private/SSRF), non-2xx, non-HTML, or slow (>20s) URL degrades to EMPTY token arrays (a succeeded job) rather than erroring — JS-rendered SPAs typically extract empty; author the kit manually from the compiled CSS in that case.

Finalize a direct upload once you have PUT the bytes to the uploadUrl from create_upload. Confirms the bytes landed and starts ingest: image/asset become ready immediately (status "complete"); audio/video enter processing (status "processing"). Returns { id, status, processingPhase }. 400 if no bytes were uploaded (PUT first) or the item is not yours. Then call wait_job(kind="media", id=mediaItemId), or use a media.completed webhook.

Generate one first-class Library media item from a prompt. Choose generator: text-to-video | text-to-image | image-to-video | text-to-music | text-to-speech | text-to-sfx | auto. Put supported presentation controls in body.style and first/last/image references in body.refs. Returns { id, status:"generating", estimateUsd? }; returns 402 if the estimate exceeds the per-call ceiling. Async — call wait_job(kind="media", id=mediaItemId) or use a webhook (media.completed — status rides the payload). Use the returned id like any other media item.

ORIENT FIRST — show THIS key's account before you spend a metered call: the org it is bound to, the current plan, a per-feature entitlements map, and per-call cost ceilings. READ `balance`, NOT `included`. CueFrame is credit-funded: rendering (entitlements.exports), generation (generate), preview stills (compose_capture), the judge (compose_verify) and the Director (compose) have NO separate per-plan allowance — they all spend one shared wallet, named in `fundedBy` (`credits`; generation is metered in provider-cost units against that wallet). Each `balance` restates that ONE wallet in that feature's own unit at `creditCost` credits per unit, so they are alternative ways to spend the same money, not four budgets. Each entitlement also carries `listUsdPerUnit`: the dollar quote for ONE unit before you call (0 = a plan grant that draws no credits, e.g. resource_import — stock search + import are free; null = not quotable up front). Check the requested cost against the balance. For generation, also check generationLimit: free accounts have a separate provider-cost cap even with a positive wallet. A billing_required error with reason free_generation_cap requires a credit-pack purchase through normal checkout. `included:false` means the plan genuinely lacks the feature. Every new org starts with a free credit grant, so an account with an empty usage history can still render, generate, preview and score. costCeilings.generateMediaUsdPerCall is the generate_media per-call USD cap. No args — identity is resolved from your key (never echoed from input).

Read a brief by id — the persisted intent plus its current quote (compose + estimated render).

Read a checkpoint packet — the durable approval unit a held compose pauses at (wait_job reports awaiting_input with {checkpointId, stage, version}). Carries the stage artifact, presigned storyboard stills (REVIEW THE EVIDENCE, not raw JSON), critic verdicts, the money triple {thisStepUsd, jobSpendSoFarUsd, estimatedJourneyTotalUsd} to relay to your human, and the recorded decision once resolved. When the Brief's gates set reviewMode:'hosted' the packet also carries reviewUrl — a browser review page where your human decides directly (relay THAT link instead of raw stills; you learn the decision via checkpoint webhooks / wait_job). Answer an awaiting packet with resume(checkpointId, ...).

Read an exact component source as complete JSON. Supply sourceVersionId from the selected catalog card to read that immutable revision, retaining sourceSnapshot with its complete versioned definition, manifest and template beside its exact pin and schema metadata. For an explicitly selected builtin, also supply builtinCollection:'native'|'curated-definitions' and its exact sourceVersionId; without builtinCollection resolution is scoped to the authenticated org. An unavailable exact pin refuses; never retry a head or another scope. To fork, keep the version-4 snapshot whole, change its componentId (the template root's data-composition-id too) and template, and send it to create_component. Builtins always require an exact collection and pin; ID-only builtin reads are refused. Declared slots are bound per placement in execution.assets, never caller params.

READ BACK a composition as JSON — the edit itself: tracks, clips, card HTML, reframe segments, captions, markers, format. This is what create_render encodes, so it is also what you hand to validate_composition (which needs a full inline composition body), what you diff after an edit, and what you archive/share so someone else can re-render the SAME video. Read-only and free. Authoring stays op-based (apply_composition) — do not round-trip this object back as a write. A composition too large to fit one MCP result degrades to a STRUCTURAL SUMMARY (every track and clip still listed; card html and caption words elided as lengths) with a header saying so — read it over REST if you need the verbatim body. Defaults to the project's ACTIVE composition; pass `compositionId` to read a specific one (the same id apply_composition takes as `composition_id`, and preview_frame / create_render take as compositionId) so an A/B variant can be read back, not just written. Pass includeDependencies="true" on a v1 guarded endpoint to read composition_id, revision and dependency tokens. With componentExecutionVersions=["3","4"], this envelope also returns expectedBrandIdentity from the persisted project authority; carry it unchanged into the matching negotiated write. Preserve those observed tokens when authoring; never refresh the basis to hide a conflict.

Fetch bounded authoring context for a media item: dimensions/thumbnail, compact face roster, transcript counts/sample, temporal grids, and fact readiness. The full transcript body is deliberately elided so this read cannot overflow MCP; follow with get_transcript for exact paginated word timings.

Read prepared subject tracks and behind-subject mattes for one media item. For an exact readiness check, pass kind + startSec + endSec together using the same SOURCE window sent to prepare_media; exact.state is pending | ready | failed | available | unavailable. If subjectTrack was prepared without intent, omit all three filters and check the whole-source subjectTrack in facts instead. Wait for the required fact to be ready before preview_frame / score_composition.

The org's legible taste in ONE read: brand kits, registered critics (the machine-checkable standard both the Director's gates and score_composition enforce), their reference media (presigned thumbnails), and the recent checkpoint-decision log. Read this BEFORE composing for an org you haven't worked with — it is the house style.

Read exact source-time transcript words in bounded utterance pages. Word start/end values are seconds from the source media start. Pass limit (20 recommended), then pagination.nextCursor as after; the formatter omits the unbounded fullText duplicate.

The A2 usage ledger: every metered billing event newest-first with attribution — featureId, amount, the project/job that produced it, and the event's LIST price in USD (a plan allowance may zero it on the invoice). Answer "what did project X cost this period?" with a query instead of summing your own quotes. Filter by projectId, jobId, and/or a [from, to) epoch-ms window. Requires the billing:read scope.

Import a media item into your library from a public URL — the agent-friendly ingest path (single call, no upload protocol). Pass url (https) + filename (+ optional contentType). Returns { id, status:"importing" }. Processing is async: call wait_job(kind="media", id=mediaItemId), or register a webhook (create_webhook on media.completed — status: complete|failed rides the payload). Use the returned id with get_media_context, suggest_briefs, or as a composition source.

Import a stock candidate (candidateId from search_resources) into a project as an ORG-OWNED media item — provider/externalId/license provenance preserved, bytes deduped across orgs. Pass projectId + candidateId. Returns { id, status:"importing"|"ready" }; processing is async — call wait_job(kind="media", id=mediaItemId), then use the returned id as a composition source. Gated on the resource_import plan feature (402 billing_required if your plan lacks it). COST: 0 credits — search and import are both free for a keyed caller (get_account → entitlements.resource_import.listUsdPerUnit is 0); imported bytes count toward the storage gauge. 422 candidate_mismatch if the candidate expired — re-run search_resources.

List briefs. DEFAULT shows human-authored briefs only; kind:"clip" shows the machine-minted clip briefs (suggest_briefs output), kind:"all" shows everything. Each carries its quote.

Use projection:'component-library' for the unified source-free library page, returned as complete JSON with exact scoped definition/preset/recipe cards and pagination. Choose origin:'builtin' for the maintained builtins or origin:'library' for the authenticated serving home; use libraryKind, category, q, limit and after to filter/page. Preserve each card identity.scope and identity.revision, its source/definition pin or exact recipe document identity; same names/IDs in different scopes remain distinct. Resolve the selected exact source with sourceVersionId (plus builtinCollection for an explicitly selected builtin), never a latest-head or another-home lookup. Presets identify their exact presetVersionId, definition pin and historical document witness; recipes identify their complete exact document and member source pins. This projection contains no source bodies or legacy thumbnail substitution. Omit projection for the historical catalog protocol below. DISCOVER WHAT TO BUILD WITH — list the component catalog: the built-in primitive catalog plus your org-installed components. Desktop also includes bundled kind:'component-package' entries: these are complete reusable authored packages; insert them with the desktop apply_template tool using placement.template and explicit insert.startTime/idPrefix, media bindings and parameter values. Do not pass their ids to clip.add, get_component_source, or cueframe://component resources. Their draft status does not certify visual review or licensing. For hosted builtin preset/recipe cards, call resolve_builtin_component with the exact collection, id and revision, then apply_template with template:<builtin id> and body:{projectId,documentSha256,insert:{startTime,idPrefix},...}. Hosted tool request fields live under body; desktop tool shapes may differ. The agent-safe default page is 20 entries; use pagination.nextCursor as `after`, or narrow with category, kind, tier, useCase, mood, q, or intent. Results are recommended-tier-first when filtered. Primitive/installed entries carry the exact placement hint for apply_composition and fixedCopy warnings; use get_component_source to fork a primitive when baked copy must change. Read cueframe://component/{id} for the full primitive/installed prop schema. SCENE SHOTS ARE NOT LISTED HERE: there is no built-in scene primitive, so do not look for one — author a shot with the scene.add op (see cueframe://scene-shot). The wrapper it registers then appears here as an ordinary INSTALLED component under these same rules.

List checkpoint packets newest-first (filter with composeJobId for one run's v1→vN gate history, or projectId). The version stack with per-version status + decisions IS the review history — nothing is deleted.

List every composition (video) in a project, with its format and which one is active. Use a returned compositionId with create_render(compositionId=...) to render a specific video, or apply_composition to edit it.

List registered critics: criticId, label, tier (rules|rubric|evaluator), phase, floor, referenceMediaIds. Rubric text / rules / evaluator URLs are write-only — never echoed. Manage with register_critic (upsert) / delete_critic.

List the media in your CueFrame library. Filter a real library with kind (video|audio|image|gif|asset|source|stock), source (upload|generated|import-url|stock), or tag — so you find YOUR footage, not e2e debris. Returns media item IDs and metadata; use an ID with get_media_context, suggest_briefs, or as a composition source.

List the plan catalog (tiers, intervals, product ids, prices) BEFORE a checkout, so create_checkout requests a real {plan, interval} instead of guessing an id. No args — reads the platform catalog.

List your projects — re-orient in an existing workspace instead of minting duplicates (steady-state entry, A9). Each row carries the id every other tool needs.

The fonts the RENDERER can actually draw — read this BEFORE authoring type, not after a refusal. Per family: the exact family name authoring accepts, the weight RANGES a shipped file declares (a static face has min===max; a variable face gives its axis range), the caption default weights, and whether the bytes are bundled or an aliased metric-compatible stand-in. Anything listed here passes validation by construction: it is derived from the same table the font gate refuses by. Two facts worth reading off it — every family's `styles` is ["normal"], because NO bundled family ships an italic FILE (a real italic comes only from a brand-kit font asset), and a weight outside every range is REFUSED, never synthesised. Contrast search_fonts, which searches the GOOGLE catalog and will happily return families the renderer cannot draw. `manifestVersion` identifies this exact face table; record it beside a composition to state what a font choice was validated against. Covers the bundled set only — a brand kit's sealed faces are per-org and per-render. No args, read-only, free.

Create a NEW composition (a new video) inside a project — a project is a workspace that holds many videos. Returns { compositionId, format, etag }. Pass an optional name and format (aspectRatio/fps/resolution); format defaults to the project's compositionFormat. The first composition becomes the project's active default. To author it, seed with apply_composition then compose({ fromComposition: true }) — your authored graphics ride as locked clips, carried verbatim; to render a specific one, pass its compositionId to create_render. (For a reframed sibling in another aspect, use derive_composition.)

Prepare the perception artifact an authored edit needs BEFORE sighted review. Pass body.kind="subjectTrack" for face/active-speaker reframing, or body.kind="matte" for a true zPlane:"behind-subject" graphic. Pass body.intent {startSec,endSec} using SOURCE time (the media trim, not timeline time); it is required for mattes and strongly recommended for subject tracks so the prepared fact exactly matches the clip and quote. This is the supported product path — never accept a preview that silently flattens depth or ignores speaker tracking. Returns a fact receipt; with intent, poll get_media_facts using the same kind/window until exact.state is ready; without intent, omit all three filters and find the whole-source subjectTrack in facts until status is ready (failed is terminal). Existing pending/ready facts are reused without another charge.