/obsidian-vault

obsidian-vault

Voice-inline capture into a Sutando-owned Obsidian vault. The voice agent calls `add_to_vault(kind, body, title?)` directly — no core round-trip, no Obsidian plugin required. Filesystem-direct: Obsidian's watcher picks up the change instantly when the vault is open.

Vault

Lives at `<workspace>/obsidian-vault/` (`<workspace>` resolved via `bash scripts/sutando-config.sh workspace` — per `sutando.config.local.json`, defaulting to `<repo>/workspace/`; `$SUTANDO_WORKSPACE` no longer honored as of v0.8 / #1440). Auto-created on first capture, with a `.obsidian/` marker dir so Obsidian recognizes the folder as a vault.

Layout

Everything Sutando writes lives under the `Sutando/` subfolder, by kind:

<workspace>/obsidian-vault/
  .obsidian/                                 ← marker; Obsidian populates on first open
  Sutando/
    Notes/<slug>-<YYYY-MM-DDTHHMMSS>.md      kind="note"   → standalone file w/ frontmatter
    Tasks.md                                 kind="task"   → appended checkbox
    Thoughts/<YYYY-MM-DD>.md                 kind="thought" → appended timestamped block

This subfolder convention keeps Sutando's writes out of the way of anything else you put in the vault later.

Triggers (what the voice agent listens for)

• "save this as a note" / "note that X" → `kind="note"`
• "add to my tasks" / "todo: X" / "remind me to X" → `kind="task"`
• "remember this thought" / "log this idea" → `kind="thought"`
• Ambiguous capture intents → the tool description picks `thought` for stream-of-consciousness, `task` for action-shaped, `note` otherwise.

One-time setup in Obsidian

Open Obsidian → **File → Open vault → Open folder as vault** → pick `<workspace>/obsidian-vault`. Obsidian will remember it. The vault appears empty until you trigger your first capture.

Opt-in: agent-state mirror + nightly dream

`add_to_vault` (the voice-inline capture tool) is always available — it only writes when you explicitly say "save this as a note" / "todo: ..." / "thought: ...".

Two automatic features are **opt-in via env var** and OFF by default:

• `src/obsidian-mirror.py` — one-shot CLI that sweeps `tasks/` + `results/` + `notes/` + the pending questions (via the declared store adapter) into `Sutando/Agent/`. No background process; runs once and exits. Schedule it yourself via `crons.json` at whatever cadence you want.
• Nightly `dream.py` cron — Opus-4.7-judged cross-linking (inline `(cf. [[X]])` citations + tiered `## Strongly Related` / `## Related` / `## See also` footer block). Sweeps the mirror first, then judges.

Both are gated by `SUTANDO_OBSIDIAN_MIRROR`. To enable, add to `.env`:

SUTANDO_OBSIDIAN_MIRROR=1

Then either:

• One-shot now: `python3 src/obsidian-mirror.py` (single sweep + exit).
• Or add an `obsidian-dream` entry to your per-host `crons.json` to run on a cadence (it is not included in the default template). The nightly job sweeps + judges in one go.

CLI flags worth knowing:

• `python3 src/obsidian-mirror.py --force` — bypass the env gate (used by `run_dream` voice tool).
• `python3 src/obsidian-mirror.py --since 1h` — only sync sources modified in the last hour (faster sweep for frequent crons).

The on-demand voice tool `run_dream` *bypasses* the gate — explicit user invocation always wins.

What's not in this skill (yet)

• **Search / weekly roundup / multi-file edits** — those are the "core" half of this integration. Plan: add `scripts/search.py`, `scripts/daily-roundup.py` driven via `work()` so heavier ops don't block the voice turn.
• **Obsidian Local REST API** — community plugin that exposes HTTP endpoints. We don't use it. Filesystem-direct is simpler and has no plugin dependency. Could be a future opt-in for read-side flows.

Loader

Loaded by `src/inline-tools.ts:loadSkillManifestTools()` at voice-agent startup. To pick up changes: restart voice-agent and reconnect the web client (Gemini caches the tool list at session start).