NoteSprite

NoteSprite is a vault-native AI agent for Obsidian that helps solo power users manage program, project, and meeting notes using a wizard-generated ontology (folders + templates + frontmatter conventions). It compounds knowledge over time with heading-level provenance and a safe propose/approve editing workflow.

Who this is for

Solo power users who want:

• A consistent structure for projects, meetings, decisions, risks, assumptions (and optional concepts like objectives/themes)
• A "deep work" system for sensemaking and executive-ready outputs (not a task manager)
• Safe AI-assisted note creation and refactoring with reviewable diffs and rollback
• Compounding over time: weekly/monthly synthesis driven by deltas, not repetition

Not currently aimed at:

• Team collaboration
• Task management/reminders
• User-managed model tuning or provider-specific advanced controls

What it does

1) Wizard-generated vault ontology (adaptive)

• Setup wizard asks about your role, domain, and ways of working
• Generates a vault structure (folders), templates, and conventions that match your preferences
• Three entry modes:
  • Start fresh — generate a new ontology from scratch
  • Evolve existing — rerun the wizard against your current ontology to adjust policies, evolve structure, or rebuild fresh
  • Adopt existing vault — analyse an established vault and generate an ontology that matches what you already have (see Adopting an existing vault below)

2) Task-driven agent UI (one main view)

A single primary view (typically pinned to the right panel) that starts with Task Cards. The core Phase 1 workflows are:

• Assistant — general vault-aware chat and ad-hoc queries
• Capture & Reconcile (formerly Ingest) — meeting notes or documents → structured notes + entity updates
• Review / Brief — structured intelligence brief from vault content, with citations
• Populate Ontology — interview-driven bootstrap to create initial ontology instances

Additional tasks (e.g. Identify Risks, Prep for Meeting, Weekly Synthesis, Monthly Review) are contract-derived and adapt to your vault ontology. Tasks can be contract-defined prompt-based skills or richer UI workflows.

3) Local-first indexing + retrieval with citations

• Indexes your vault into a local SQLite database (sql.js)
• Uses hybrid retrieval (lexical + lightweight embeddings)
• Responses can include heading-level citations back to your notes (configurable)

4) Safe writes: propose → approve → rollback

• The agent never silently edits your vault
• All writes go to a Proposal Buffer and accumulate in a collapsible queue
• You review diffs, approve selected/all, and can rollback by run ID
• Deletes are supported but explicitly gated by policy

5) Change over time (deltas)

• Runs can produce delta artefacts (stored under the plugin system directory)
• Entity notes maintain a Timeline section (policy-driven)
• Weekly/monthly review tasks can aggregate deltas to surface what changed

---

How it works (high-level)

• Home/Agent View: task cards, home sections (past chats, recent changes, suggested next actions), streaming output, proposal queue, and activity trace
• Vault Contract (generated by wizard): a versioned contract stored in the plugin directory that defines:
  • enabled entity types, paths, templates, frontmatter conventions
  • timeline/delta policies
  • task catalogue and defaults
  • write safety boundaries
• Indexing: heading-level chunking + lightweight embeddings into local SQLite
• Tasks: outcome-oriented definitions that parameterise prompt + retrieval + output destinations
• Tools: ontology-aware operations that enforce the contract and safety policy, with subagent delegation for context-heavy work (research, entity scanning, patch QA)
• Patch engine: apply with snapshot + rollback, with run artefacts stored on disk