Miadi Factory — The Author's Desk (v4 draft: revised with terminal-agent discoveries)#

Read this first — intent and authority#

Create a functional prototype, not a fictitious integration demo. It is the next screenwalk experiment after Episode 120's interface prototype. The attached cinematic splash art is the visual design reference. BACKSTAGE (ceremony/research) → AUTHOR'S DESK (human writing) → ONSTAGE (fiction; optional documented provenance doorway). Preserve warmth, imaginative atmosphere, accessible hotspots and a usable mobile view. The primary end-to-end interaction MUST work without external services.

Human's authorship law, quoted from source: “I don't want the robot to write for me. What I want is the robot repeating me, which is different.” Agents can accurately quote, arrange, retrieve, ask, visualize, suggest separately and perform supporting engineering; they must NEVER silently replace human-authored text, self-approve a storybeat, or publish. Preserve the difference between exact quote, human writing, AI proposal, agent report, independently inspected source, and story fiction.

Handoff provenance: The human deliberately asked an assistant to draft a broad versioned prototype prompt before sending it to a terminal agent who knows the repo, to supply library/package/API facts. Langfuse v1 is the initial intent; terminal agent produced detailed v2 and trimmed v3. This v4 resolves the tension, retaining verified discoveries without turning the prompt into an unmanageable internal inventory. Existing versions remain historical; no automatic new project implementation or deployment is authorized by this prompt's creation.

Stack and readiness rules#

Use Next.js App Router, React, TypeScript, Tailwind/shadcn, accessible semantic controls and typed adapters. Prefer existing repo stack when running within it. Every tile shows a truthful state: LIVE (runtime read or import succeeds), AVAILABLE (source/package identified but not connected), DEMO (local sample), PROPOSED, or UNAVAILABLE. In-code dependency or npm release existence is NOT runtime integration. Attach source and verification time to capability facts. Keep secrets server-only. Use MockAdapter to ensure pages work even if network, packages, or credentials are absent. Prefer imports only after inspecting actual exports and running typecheck; do not invent signatures.

Confirmed or reported integration leads (not unconditional imports)#

Terminal agent asserted on 2026-10-10 (Langfuse prompt v2/v3) these npm packages and versions; public npm search independently corroborates the core NCP/Wampum package names, while remaining exports and live APIs require repo inspection:

• Research/episode: @miadi/episodic-memory-schema@0.14.1 (typed episode, screenwalk, review refs), @miadi/episode-vessel@0.4.2 (server filesystem/capabilities), @miadi/inquiry-weave@0.16.6 (chronicle inquiry links), @miadi/reference-ui@0.1.2 (reference links/chips), @miadi/transcription@0.3.4 (transcriptions), @miadi/hooks-interpreter@0.3.2 (session interpretations), @miadi/annotate-core@0.2.0, @miadi/annotate-ui@0.2.0 (signed word/block marks), @miadi/attention-core@0.1.1, @miadi/attention-ui@0.3.1 (unresolved owner decisions).
• Story: @miadi/ncp-story-studio@0.8.1 (NCP schema/graph/store/agents; verify subpath exports), @miadi/wampum-narrative-engine@0.1.1 (peer narrative model with co-valid readings), @miadi/ncp-wampum-bridge@0.1.1 (crossings and loss/report), @miadi/voice-client@0.2.4 (optional readback of author text), passages@0.3.11 (optional Twine book derivation).
• These are leads from the terminal agent, not evidence that every export or compatible version is usable in the prototype. Inspect and pin only what you install. The Wampum Narrative Bridge is the user's prototype and should not be presented as representing Indigenous communities or implementing an Indigenous cultural protocol without contextual authority.
• Review service: public review /review/<uuid>; agent reports public JSON /review/<uuid>/json and ?v=N. Test the actual response/shape before coding parser; do not assume a LIST endpoint. Public reviews should need no secret token; use a separate server-only adapter for any authenticated private review operations after documentation.
• Earlier Episode 120 prototype: https://miadi-factory-interface-analysis.vercel.app/ and issue https://github.com/jgwill/Miadi/issues/757. These links were supplied by the agent; GitHub issue independently confirms the original v0.app project. Test deployment reachability before calling it LIVE; provide a fallback link to the issue/reviews.
• Miadi Factory episode read routes reported: /api/chronicle/episodes/{ref}/reviews, /foundations, /files, /screenwalks; verify base URL, response contracts, availability and permissions. Medicine Wheel internals remain private; never call internal services from the browser.
• Two Langfuse domains: CEREMONY for Backstage read-only observations and prompts; STORY for opt-in author-writing session traces. Server env CEREMONY_LANGFUSE_* and STORY_LANGFUSE_*; do not leak keys, prompt bodies, private text, or cross-project data. Provider-neutral TracingAdapter can expose summaries after authorization. A story's optional source doorway uses selected PUBLIC-safe references and never implies fiction is factual.
• miadi-tracing: / miadi-notebook: / miadi-prototype: schemes are proposals until implemented. Canonical existing schemes should be rendered by verified reference utilities; never invent persistence identities.

User experience — all hotspots operate#

HOME: accessible image-inspired hero, working Backstage / Desk / Onstage regions, exact-word labels, breadcrumb navigation, clean mobile cards and an Integration Map showing each adapter's factual state plus evidence. BACKSTAGE six pages:

1. /backstage/screenwalks: browse/filter seeded real review references; select a source and 'Bring to my desk'. Screenwalk source card and relevant linked episode; a safe link to earlier Episode 120 prototype.
2. /backstage/plot-threads: free experimental canvas with keeper, concern, plot events, narrative scenes, execution-branch links, evidence state. Optional upper/lower timeline squares with intermediary connectors as design experiment. Only when the actual Wampum, NCP and bridge packages load should the mode use their real structure/crossing report. Show LOST/AMBIGUOUS mappings rather than silently inferring equivalence. Keep plot concerns distinct from agent execution branches. Do not copy culturally significant patterns without owner-approved context.
3. /backstage/agents: read-only agent roster and parent trunk/branch board showing source, state, results and owner decisions; demo Asterion, Marks, Jev examples are LABELED demo. No dispatch.
4. /backstage/reviews: curated public list + working detail/version comparison when fetch succeeds; robust JSON schema guard and error states. Actual general listing needs separately documented endpoint. No client-side API tokens.
5. /backstage/evidence: provenance cards and status Verified/Reported/Interpretation/Proposed; linked Episode 120 and Author's Desk traces. Optional server-side Ceremony Langfuse metadata adapter; never claim a submitted observation is retrievable until read succeeds.
6. /backstage/chronicle: episode and cycle cards, known links; read-only adapter for actual episode routes only after verified. Medicine Wheel concepts with context; no generic cultural authority claim.

AUTHOR'S DESK six human-controlled writing surfaces: /desk My Take, A Question, A Spark, A Character, A Scene, A Story Hook. Each has editable human-authored text, scratch/draft/selected state, attached source references, tags, manual save/revision/export/delete and local-only storage with clear warning and export backup. Non-AI 'Echo my words' reproduces exact selected text; suggested questions/alternate phrasings go in a separate clearly marked suggestion panel and NEVER touch original words without explicit human editing. A source quote retains literal text and provenance. Optionally preview candidate NCP storybeat/storypoint via real validated schema only once imports work; request human acceptance before moving it onstage.

ONSTAGE: /onstage author-selected story projects and Act I Setup / Act II Development / Act III Climax + Resolution scene/beat board; manual reordering, preview and saving. Audience Preview hides documentary sources. Optional 'Behind the story' doorway shows only author-selected public-safe citations and labels source/evidence versus invented events. Interactive story engine and Twine rendering are optional future integrations, not prerequisites or unearned deployment claims.

Canonical navigation examples#

Author's Desk review: https://miadi-review-service.vercel.app/review/d7dcde0c-8ab2-4cd7-98d6-a2a937ff28fd Episode 120 reviews: https://miadi-review-service.vercel.app/review/aa3bb198-59b2-4108-a452-36d25603d5e3 and https://miadi-review-service.vercel.app/review/e29a4fc3-302a-4309-8ad9-39085ee26283 Episode 120 issue: https://github.com/jgwill/Miadi/issues/754 ; earlier prototype: https://github.com/jgwill/Miadi/issues/757 Author's Desk Ceremony trace: https://cloud.langfuse.com/project/cmkgk10zy009tad077uc42ixf/traces/f06fa8abaf8e0c68cf71ffcbd320fa1d

Minimum usable end-to-end slice / tests#

1. Open the home splash; navigate to a real seeded Screenwalk card, optionally retrieve public review or use clearly marked DEMO fallback.
2. 'Bring to my desk' creates a reference, not AI-authored story text. User writes a sentence in My Take; autosuggestions don't overwrite it; explicitly save and export.
3. User manually selects a fragment as a scene/storybeat; show a correct NCP preview only if its package validates it, otherwise a neutral, NON-NCP preview. Add it to the Act board by human action.
4. Audience Preview shows a coherent demonstration scene without evidence exposed; optional Behind the story shows only the selected safe source link.
5. Typecheck and exercise every major route/hotspot, keyboard navigation, local data persistence/backup, no client secrets, no fake Connected badges, failures and provenance disclosure.
6. End with an Implementation & Gap Report: actual installed/imported/tested packages, what remains mock, verified API response shapes, packages/contracts missing or needing NEW package creation (proposal only: purpose, owner, narrow interface, dependencies), and exactly three human decisions needed. Do not create or publish packages, modify remote issues, run agents, or deploy without authorization.

Keep the V1 experience approachable. Discoveries from terminal agent enrich adapters; they must not bury the Author's Desk's essential fact: the HUMAN writes.