Aether.

The problem on file

Obsidian is a local-first markdown notebook with a graph view bolted on. It is great as an editor and weak as a system. The graph does not feed anything else. The notebook does not know what happened yesterday. Cross-vault queries are manual.

The reasoning layer that would summarise a week of notes, surface emerging themes, or answer a question that spans projects is missing entirely, and adding it would cost a paid LLM subscription per device.

Aether is the rebuild. Same local-first markdown-on-disk substrate (the vault is a Syncthing folder that any editor can open).

A NestJS API watches the directory, indexes every save into Postgres with pgvector embeddings and a bidirectional link table, and a Next.js PWA renders the graph, the day-by-day Pulse reports, and a presentation mode called Show. The council from Shellwire is embedded in the api so the LLM bill stays at zero.

What it is, in numbers

Exhibit A·Captured 2026-05-21 07:39 UTC·Source aether.alkenacode.devlive

The marquee piece, the Pulse loop

Every evening at 20:00 Africa/Nairobi a cron fires the Pulse worker. It walks the notes the watcher has indexed since the last Pulse, runs a council pass (synthesize-best, councilSize 3 from the shellwire-derived provider matrix).

It then writes a new markdown file with aether: pulse in the frontmatter, which the watcher re-indexes but the aggregator excludes from tomorrow's input, so the loop is bounded.

@Cron('0 20 * * *', { timeZone: 'Africa/Nairobi' })
async runDailyPulse() {
const notes = await this.notes.indexedSince(this.lastPulseAt());
if (notes.length === 0) return;

const result = await this.council.orchestrate({
  task: 'pulse-synthesis',
  tier: 'complex',
  strategy: 'synthesize-best',
  councilSize: 3,
  prompt: this.buildPulsePrompt(notes),
});

await this.vault.writePulseNote({
  body: result.response,
  frontmatter: { aether: 'pulse', day: today() },
});
this.bus.emit('pulse:complete', { noteId, latencyMs: result.totalLatencyMs });
}

The aesthetic of those Pulse files is the point. They are not chat transcripts. They are short editorial-pass essays in the same Fraunces-italic voice the rest of the interface uses, written in the user's own vault, version-controlled with the rest of the markdown.

The hidden ninety percent

In production