@agent-native/creative-context

Shared creative-library ingestion, governed reusable contexts, immutable evidence, hybrid retrieval, brand context, and generation provenance for Agent Native creative apps.

Creative Context is a Toolkit capability module installed on demand. It lives in its own npm package so apps only take the governed corpus, retrieval, UI, and lifecycle metadata they need.

The package turns Google Slides, Figma, Notion, websites, and uploaded PPTX/DOCX/PDF/PNG/JPEG/WebP/GIF/safe-SVG files into a versioned SQL corpus. Binary originals, localized visual fallbacks, and bounded UI thumbnails live in private blob storage; full-slide and full-frame source renders are temporary QA inputs and are discarded. Text remains directly greppable in SQL; PostgreSQL deployments add full-text candidates and pgvector visual similarity in the same DATABASE_URL database.

Install

pnpm add @agent-native/creative-context

Add the server plugin once per app:

// server/plugins/creative-context.ts
import { setupCreativeContext } from "@agent-native/creative-context/server";

export default setupCreativeContext({ appId: "slides" });

Mount the shared Library tab on the app's existing Agent page:

import { AgentTabsPage } from "@agent-native/core/client";
import { createCreativeContextAgentTab } from "@agent-native/creative-context/client";

export default function AgentPage() {
  return <AgentTabsPage extraTabFactories={[createCreativeContextAgentTab]} />;
}

Importing ./server registers the package actions through the framework's package-action registry, applies additive corpus migrations, registers shareable sources/profiles/contexts/packs, starts resumable import processing, and conditionally contributes published brand context to agent prompts. Each owning app also registers a server-side native capture adapter so its Share popover can submit a deck, design, document, asset, or dashboard by resource id without sending native payloads through a generic action. The prompt contribution is omitted structurally whenever application state contains { contextMode: "off" }.

Governed contexts

A brand profile contains approved identity and DNA. A Creative Context is the durable, shareable collection people maintain (for example Default, Marketing, Product, or Sales). A context pack is the immutable exact-version receipt for one generation.

• Every personal or organization scope has one non-archivable Default context.
• Requests use Default plus at most one specialty context. Explicit selection wins; an app binding or semantic match supplies the automatic specialty.
• open contexts publish editor submissions immediately. review and admins-only contexts keep the proposed version private to its submitter and context administrators until approval.
• The currently approved snapshot remains active while a newer version is pending. Removal affects new retrieval immediately without rewriting old packs.
• Publishing into a broader context requires source-management permission and explicit confirmation. Acceptance writes an immutable context-owned copy; it never changes the original resource's sharing.

The standard UI is the Context tab in the framework Share popover plus the context-first Library. The Library rail selects Default or a specialty and its Items, Sources, Approvals, and Settings views show only safe, bounded previews: slide filmstrips, design frames, rendered document structure, access-scoped media, and dashboards populated with synthetic preview data.

Exact app-native data is a separate private blob capability. Generic list, item, pack, status, media, and search responses never expose that capability. Only the owning app's typed clone action can resolve it, verify the stored hash and resource identity, and create a new editable resource through that app's normal persistence path. Imported Google Slides and Figma artifacts continue to use the stricter trusted-compiler and native-reassembly path.

Retrieval behavior

• Portable SQL uses normalized grep modes (allTerms, anyTerm, phrase, regex) and weighted title/summary/body scoring.
• PostgreSQL adds a tsvector/GIN candidate lane and one configured multimodal embedding family backed by pgvector in the same database.
• SQLite keeps corpus and lexical/caption search only. Visual queries fail with a clear PostgreSQL+pgvector requirement; the package never creates another SQLite or vector database.
• Every context-backed generation resolves Default plus zero or one specialty, applies the specialty ranking boost, and snapshots exact (itemId, itemVersionId) evidence, lane scores, selection reason, and context ids into an immutable context pack.

Context packs are intentionally ownable and independently shareable instead of inheriting broad source visibility. A pack records one generation's request, selected evidence, and influence scores, which can reveal private work intent even when every underlying source is organization-visible. Only the creator can see a new pack; collaborators receive it through an explicit pack share. Ordinary source evidence remains subject to current source access and revocation. Context-owned published snapshots remain replayable at their pinned version for authorized pack viewers after membership removal.

Isolated deployments

Apps in one workspace use the shared database in-process by default. If an app is deployed against a separate database, set CREATIVE_CONTEXT_A2A_URL to the app that owns the corpus and configure either the shared A2A_SECRET on both deployments or a matching org-scoped CREATIVE_CONTEXT_A2A_KEY. The fallback forwards only bounded resolve, validate, generation-record read, and generation-record write operations through the authenticated A2A protocol. Caller identity is signed and re-resolved by the receiver, all request and response shapes are strictly validated, and malformed or timed-out responses fail clearly instead of falling back to untracked context. Optional CREATIVE_CONTEXT_A2A_TIMEOUT_MS is clamped to 1–60 seconds and defaults to 30 seconds. Do not place these values in client code or application state.

CREATIVE_CONTEXT_A2A_URL is opt-in: without it, no A2A request occurs. A saved or one-generation contextMode: "off" also structurally suppresses the remote path.

Trust and lifecycle

Imported text is evidence, never instructions. Only human-published, validated structured brand fields can enter the typed <brand-context> prompt block. Restricted upstream items default to review and exclusion. Organization copies require an organization admin or verified container owner plus a confirmation of the exact container and item count. Disconnecting or deleting a source tombstones it synchronously so retrieval stops immediately; private blobs are purged asynchronously. Native context submissions use the same fail-closed private-blob rule and never fall back to SQL payload storage.

Exports

| Subpath | Purpose | | -------------- | ------------------------------------------------------------ | | . | Shared types and generation-context helpers | | ./actions | Registered action map | | ./client | Library panel, Agent tab, context chip, state hooks | | ./connectors | Connector registry plus public Slides/Figma native compilers | | ./embeddings | Multimodal family adapters and bake-off inputs | | ./eval | Retrieval, quality, and hard-gate evaluation helpers | | ./jobs | Durable import, enrichment, purge, and rebuild workers | | ./messages | Shared localized Library UI messages | | ./schema | Additive Drizzle corpus schema and migrations | | ./server | Setup plugin, action registration, prompt/resource wiring | | ./store | Access-scoped corpus stores | | ./vector | PostgreSQL/pgvector adapter and fail-clear guard |