Skip to content

BACKLOG.md

Latest commit

 

History

History
66 lines (53 loc) · 3.36 KB

BACKLOG.md

File metadata and controls

66 lines (53 loc) · 3.36 KB

Backlog

Outstanding work and intentional gaps. Check here before "fixing" something that is already a tracked, deliberate quirk. The patterns the code follows are in ARCHITECTURE.md.

Deliberate gaps (intentional — context, not bugs)

  • One shared world, by design. All real-time state (eyes, boxes, subscribers) lives in a single EventHub Durable Object resolved by idFromName("global") — in-memory and ephemeral, no external store needed. For multiple worlds, shard by DO name rather than adding a backing store.
  • Agent POV thumbnails are host-only. Only the host renders each agent's offscreen view, so the AIAgentViews HUD stays blank on viewer clients — that redundant per-frame work is exactly what the host model removes.
  • next dev serves the UI only. The real-time hub (/api/events + the DO) runs only under the Workers runtime; use npm run preview to exercise it.
  • Auto-deploy paused. CI deploys only when the DEPLOY_ENABLED repo variable is true (currently unset). The planeo.tre.systems custom domain is already configured in wrangler.jsonc and takes effect once deploy is re-enabled.

Outstanding work

Prioritised, P1 (highest value) → P3. None are known bugs in shipped behavior.

P2 — maintainability & robustness

  • Single top-level animation-tick loop. Three useFrame updaters (boxStore, eyesStore, useAIAgentController) run independently; one tick would unify frame-rate handling.
  • Seamless host handoff. On host change the sim freezes briefly and cube velocities aren't migrated (boxes resume at rest); snapshot velocity into the host event for a clean handoff.
  • Broader useShallow. The frame-driven container selectors (Eyes, Box, useAIAgentController) still read whole records — low impact, since the over-render is masked by the frame loop.

P3 — minor

  • Wire-format version field on EventSchema. No version discriminant, so a long-lived DO can't detect client contract drift during a rollout — add one when the wire format next changes.

Product / features

  • Multi-agent text chat. useAiChat routes human chat only to agents[0]; let other agents answer typed messages too (they already act and speak via the vision loop).
  • Richer agent behavior. Memory across decisions, agent-to-agent dialogue, and actions beyond move / turn.
  • User-to-user chat. Humans currently only see AI messages.
  • Persistent profiles. Identities are ephemeral nanoid ids today.
  • Speech UI polish. The Chirp3 TTS path works; there are no per-message playback/status indicators in the chat UI.
  • SSE → WebSocket + DO hibernation. A bidirectional transport with hibernation would cut idle cost and simplify egress; a bigger change — evaluate when needed.
  • Re-add PWA. next-pwa was removed (no service worker today); a Workers-compatible SW such as Serwist could restore offline/installable support.

Ops / hygiene

  • Dependency bump. npm audit flags 19 issues, but all are dev/build-tooling transitive (eslint, playwright, wrangler, postcss toolchains) or non-exploitable in our usage (the prod uuid finding needs a buf argument we never pass) — none reach the deployed worker. Clear them on a deliberate npm run deps:update pass rather than a risky 28-package audit fix.