Roadmap

Designed and partly prototyped in the deployment this scaffold came from; not in the scaffold yet.

Done: engines as subagents (agents/, architecture § Engines) with the action gate enforced by a hook. Next for them: replay a day's stream through each engine and score the blocks in the nightly fold, so prompt changes are measured rather than guessed; then the local-model shadow below, engine by engine (interpretation and epistemic first: no tool calls, fixed labels).

• Local-model shadow. A shadow Anna on a local GPU (Ollama speaks the Messages API) fed the day's stream overnight, with no actions and no replies, scored per engine against the frontier run: right tool, right facts, honest "don't know", voice, latency. The outcome: which parts a one-GPU home can run locally. Where models disagree is where Anna was unsure, which becomes a morning salience list ("dreaming").
• Event wakes. Home Assistant, the UPS, or a ⛔ check pushes an event over the relay and wakes the brain at any hour to gather context. The system's own alert always reaches the person first.
• The canon off-site. Client-side encrypted backups of the record and state (restic, append-only) to a second location.
• Proof of memory. OpenTimestamps over the record's commits, so the history provably existed when it says it did.
• More senses. Cameras as eyes, proxy errors as skin, WAN health as breath.
• The chat app as a setting. The bot's channel is a Claude Code plugin, but Telegram is wired into anna-tg, anna-notify, the bot's settings and install.sh. A bot.channel setting (plugin, token file, and a notify method per channel) would make another chat app a config change instead of an edit.
• merlin during setup, via podman. /anna-setup asks whether to install merlin now: rootless podman quadlets (Postgres + pgvector, a CPU embedder, the indexer on a timer) shipped in merlin's own repo, registered with Claude Code as the merlin MCP server. Until then, the merlin guide covers it.