Skip to content

BACKLOG.md

Latest commit

 

History

History
93 lines (78 loc) · 4.85 KB

BACKLOG.md

File metadata and controls

93 lines (78 loc) · 4.85 KB

Backlog

Ordered, not-yet-started work. Git history is the audit trail; this file is only what is still open. Last reviewed 2026-06-01.

Acto runs on Cloudflare Workers + D1. Persistence-heavy work is built once on that substrate; the platform-independent AI improvements below can land anytime.

1. AI layer (platform-independent — landable now)

These change how Gemini is called and how a beat is delivered; none depend on the hosting platform, so they carry over to Cloudflare unchanged.

  • Structured model output. Call Gemini with responseMimeType: 'application/json' and a responseSchema derived from StorySceneSchema (lib/domain/schemas.ts). Removes the markdown-strip + JSON.parse step in app/actions/generateStoryScene.ts and the MALFORMED_OUTPUT error class it produces.
  • System instruction. Pass the storyteller preamble in lib/promptUtils.ts as a systemInstruction rather than concatenating it into the user turn.
  • Progressive media delivery. Return passage + choices as soon as the text beat is ready; load image and audio lazily instead of blocking on Promise.all in app/actions/generateStoryScene.ts. The store already tolerates a missing imageUrl / audioBase64.
  • Eval harness. Golden scenarios -> generate -> assert structural validity, with optional LLM-as-judge for coherence and variety, so prompt changes are measurable rather than vibes-only. This is the main missing technique for an AI-first product.
  • (Stretch) Stream the prose with generateContentStream for word-by-word display, once text generation is split from media.

2. Finish the Cloudflare cutover

Acto is deployed on Cloudflare Workers + D1 at https://acto.tre.systems (Worker acto, database acto-db). What remains is mostly account-side configuration.

  • Register OAuth callback URLs for the new domain. In each provider console add https://acto.tre.systems/api/auth/callback/{github,google,discord} (GitHub Developer Settings, Google Cloud Console, Discord Developer Portal). Until then, sign-in fails with a redirect-URI mismatch.
  • Enable CI auto-deploy. Create a Cloudflare API token (Workers-deploy scope) and add it as the CLOUDFLARE_API_TOKEN repo secret; the deploy job in .github/workflows/cloudflare.yml skips until it is set.
  • Retire the Fly app. The Fly machine still runs the old build; fly apps destroy acto once Cloudflare is confirmed. The Fly-capable code is preserved on the fly-fallback branch.
  • Re-add the PWA via Serwist. @ducanh2912/next-pwa is not OpenNext-compatible and is disabled in next.config.js; adopt @serwist/next (the pattern Comprehendo uses).
  • Drop dead deps. Remove @google-cloud/text-to-speech and @ducanh2912/next-pwa from package.json once the Serwist PWA lands.

3. Persistence & observability (built once, on D1)

  • Persist generated stories server-side: share/replay for users, a corpus for the eval harness, and creator visibility into what is being generated. Generated beats currently live only in one browser's localStorage.
  • AI-call telemetry. Record tokens, latency, model, and outcome per call; surface in the existing admin panel. Three paid APIs run per beat with no structured record today.
  • Error tracking. Add a service such as Sentry; production failures are currently invisible.

4. Livestream MVP (on the Cloudflare substrate)

Spec: docs/LIVESTREAM_MVP.md. The MVP is streamer-mediated — chat-command voting and a local story-driving agent, with the streamer approving each beat — so the first version reads votes from platform chat and needs a streamer control surface, not a high-fan-out public web UI.

  • Server-authoritative session + beat state (D1). A Durable Object per session is the right primitive if live in-app web voting is added later — they are built for one coordinator per room.
  • Streamer control surface: approve / edit / reroll / veto a beat before it publishes.
  • Keep hosted-model usage optional and capped; keep the local-agent path first-class.

Other engineering

  • Re-enable admin-panel e2e tests. Regenerate the auth fixtures and remove the describe.skip (see docs/TESTING.md).
  • Externalize rate limits. Move the per-API daily limits out of lib/rateLimitSqlite.ts into environment-driven config.
  • Drop the dead maxTokens field. MODELS[].maxTokens (lib/modelConfig.ts) is unused; the real cap is maxOutputTokens in lib/ai/googleAiService.ts. Wire it through or remove it.
  • NextAuth v4 -> Auth.js v5 whenever auth is next touched; v4 is on the legacy line.