[SaulDoesCode]

This is the redesign agreed on 14 Oct (docs/next.md, "Decisions"). Status and what's left are in docs/next.md under "Resuming".

Data

• Entry graph store (core/src/graph/, format in docs/format.md):
  • one store for every entry, with timestamps;
  • a 5-bit packed codec with references and function calls (with tail calls);
  • edges, borrowing and dedup; posting-list indexes; compaction.
  • Size: 17.1 MB against the old store's 54.0 MB on the same data.
• Gematria is checked against weum. Atbash is removed; the modes are ordinal, reversed and frequency.
• Byte sums are a separate measure.
• Imports: weum's all-of-it.json. Backup writes texts.json + other.json. Old saves are not read.

One app API

• core/src/ops/ holds the op code; the desktop app and the hub both use it. The ops are listed in docs/app-api.md.

The app (Dioxus 0.7, app/)

• Desktop and web: desktop runs over the graph in process; the web build runs over the hub's WebSocket and is served by the hub at /app/.
• Views: memory (at parity with weum), space (three.js), lists, reader, notes (markdown), tags, stats, models (the spec's outline), filters, settings.
• Settings: zen and top-bar autohide, palettes, backup, weum import.
• Bevy app and old JS web app: removed. Release builds now ship anend-app.
• Screenshots: in app/docs/shots/.

Hub

• Server options: --mode hub|app|both, --join invite|apply|open, private/unlisted/public visibility, /health.
• Docs and deployment: docs/deploy.md, Docker.
• CSP change: the content security policy now allows wasm, eval and blob scripts. Dioxus's JS interop needs eval, and the space loads three.js from a blob. docs/security.md covers it.

Models (core/src/classify/)

• Declarative model spec (docs/models.md): generative heads with a cached decoder, critics, guidance and reinforce, and codecs for numbers and pictures. Gematria inputs are opt-in.
• Pretrained Llama-family import: reads safetensors and BPE tokenizer.json, for llama, mistral, qwen2 and qwen3. Tested on synthetic checkpoints only, because model downloads are blocked in the build environment. The example to run on a real model is core/examples/pretrained.rs.
• Optional GPU backend (--features gpu, wgpu): inference and training. It is bit-exact against the CPU on llvmpipe and opt-in per run. core/examples/gpu_bench.rs is for real hardware.

Checks

GitHub Actions doesn't run on this account, so CI shows red without running. These were run locally on ce4351b, and all pass:

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace: core 135, hub 36, app 7
• the wasm32 web build

Left

• Saved patches (docs/wip/): two unfinished rounds:
  • gematria presets off by default;
  • further GPU work.
• Not yet tried: an input method (IME) on a real system; real pretrained checkpoints; a real GPU.
• Docs and UI: docs/guide.md still describes the old apps; the models view waits for classifiers to move onto the spec.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Tq7kHwoCFpfspEmb6rmcAF

[SaulDoesCode]

CI (ci / "app and hub") fails before any step runs.

Every run since 14 Oct has ended in about 3 seconds with no runner assigned (runner_id: 0, no steps, no log). That includes the pushes to master after #5, #6 and #7. GitHub never started the job, so this isn't a code failure. The usual causes are Actions minutes used up, a spending limit, or a failed payment, so check Settings → Billing → Actions. Actions may also be disabled for the repo.

I ran the same steps locally on this branch at 56df925/cc1519f, and all passed:

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace: core 120, hub 35, app 8
• the wasm32 web build

I can't re-run the job from here. Once Actions can start jobs again, the next push will run CI.

---

Generated by Claude Code

[SaulDoesCode]

Update: the room runs natively on the Linux desktop, with no JavaScript (up to faf92ce)

The space on the Linux desktop is now Rust, drawn with wgpu under the page. space.js and three.js are no longer in that build.

• The room in Rust (render/, anend-render): it takes every command space.js took and sends the same events, so the windows and menus didn't change. A parity guard checks that every command is handled. It covers:
  • the spread with zones; texts with their underlines;
  • blocks with their tiles (texts, recipes, pictures), bends, spheres, baked lights, the sky and fog;
  • the gizmo, prefab copies as one group, Shift+click and box selection, groups;
  • wires made by dragging, rays, timers and zone events;
  • sounds, mixed natively (symphonia and cpal; building on Linux needs libasound2-dev);
  • animations, walking, the grapple, superflight, the prefab ghost;
  • the prefab menu's pictures and the textures window's previews.
• Culling: back faces, meshes outside the view, and meshes hidden behind others (occlusion queries) are left out. Blocks past 120 units are drawn as billboards; past 400 they are neither drawn nor hit by rays. Both distances are exact.
• Speed: the spread runs on threads. A room of 2,000 texts went from about 100 ms to 30 ms a frame on four cores here. Frame times on a real GPU haven't been measured yet.
• Polish: each view was looked at with a 50,000-text save. Two faults were found and fixed: the room showed through behind the other views, and the empty-room hint covered a room of prefab copies.
• Still JavaScript: macOS, Windows and the web still use space.js for now (questions 118). What remains on the Linux desktop is listed in parity.md.

Shots: app/docs/shots/native-*.png, polish-*.png. The guide's space section describes the native room.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Tq7kHwoCFpfspEmb6rmcAF

---

Generated by Claude Code