Cultivate — purposeful work and thinking sessions¶

Start with your own material¶

Open Home → Start with my material (or Resume my inquiry). Choose an existing note with the literal title/path picker, or capture one real idea. No Canvas setup, system installation or bulk tagging is required. Only one inquiry is retained; ordinary Cultivate remains available separately.

1. Choose a starting note and state your question if useful. Add other notes explicitly.
2. Keep selected-only scope or deliberately include one recorded-link hop. Inspect directed relations and recorded evidence. Opening a note does not mark it consulted; I used this material is your explicit attestation and determines the references in an outcome.
3. Write your provisional response and uncertainty. Confidence is optional. Continuing, insufficient evidence, or stopping are valid dispositions. Sufficient for now is a separate reversible decision; adding support never supplies it.
4. Save progress or Save and pause acknowledges a local checkpoint. Save response snapshot creates Markdown, not an overwrite. Retry pending work after an error; inspect the pending snapshot before abandoning an ambiguous operation.
5. Cultivate without a purpose keeps the original moves, recipe and friction settings. Switching modes retains the in-memory inquiry; save it successfully for restart continuity.

Editors remain mounted during metadata/context refresh. Loading and errors are distinct from an empty scoped result. Actual device walkthroughs and the consented adoption pilot remain pending; automated tests are not evidence of user adoption or improved understanding.

Recovery guarantees¶

Observed unambiguous renames follow the exact file. Missing/offline renames are not guessed from a basename; explicitly choose replacement context. Observed deletion remains unavailable even if a new file occupies that path. Newly excluded references cannot be read on retry. Abandoning an ambiguous operation requires confirmation and removes only its bookkeeping, never a potentially created note. This is local recovery, not distributed locking across devices or an atomic transaction with sync.

Create-only outcomes¶

Outcome snapshots use a create-only Vault boundary. An exact complete-content retry is recognized; different or externally modified content is a conflict and is never overwritten. Outcome folders must already exist, and unsafe/traversal/config paths are rejected. A snapshot is ordinary editable Markdown, not a live back-sync into the inquiry. New edits require a deliberately new snapshot, not an automatic append, replacement, filename suffix loop or extra file on retry.

Local checkpoint contract¶

The runtime keeps the editable draft outside settings until Save or Pause requests a checkpoint. All plugin-data writes use one serial queue, including journal/settings writes. Only the acknowledged revision is durable; newer input remains dirty, and a failure leaves it editable. One inquiry persists across restart independently of the bounded judgement log or whether logging is on. Unsupported/corrupt storage requires explicit reset. Clearing bookkeeping never deletes outcome notes. Unloading is not an acknowledgement: save successfully before closing to guarantee restart continuity.

Context bounds¶

The internal inquiry projection starts with selected notes only. Optional one-hop scope adds direct recorded neighbors of those seeds, never recursive expansion. Purpose prose is not a semantic search. Each candidate names its actual directed relation. Opening material is not an attestation that it was consulted, understood, or accepted.

One gather inspects at most 1,000 adjacency/relation/claim/source records, 200 material notes, 50 candidates and 100 combined relation/evidence rows. Truncation is explicitly reported; the bounded sample follows index encounter order, not an exhaustive global ranking. Missing/excluded endpoints and linked sources are not presented as evidence. Indexed context may omit inline enrichment, especially on mobile. No result means none found in this scope by this method, not universal absence.

State and provenance¶

One inquiry is independent of the activity log. Its optional purpose, selected notes, consulted notes, authored response and unresolved gaps have their own versioned local state. Only an explicit human decision marks a response sufficient for now; changing the question reopens it without deleting the previous response or decision context. Support edges and scores never settle it.

Markdown outcomes retain authored or explicitly accepted/modified provenance, consulted references only, and honest uncertainty. No confidence is invented. Draft limits are 4,000 purpose characters, 64,000 response characters, 32,000 gap characters, 20 selected and 200 consulted notes. Invalid or unsupported stored data is reported, not silently discarded; over-limit text is not truncated.

Ordinary thinking sessions¶

ZettelFlow is an engine that makes knowledge evolve. Cultivate is that engine made a daily practice: a short, guided thinking session that takes one idea and makes it measurably more connected and mature (#309).

Where the canvas wizard and quick-capture serve creation, and the dashboards serve diagnosis, Cultivate serves the middle of the lifecycle — DEVELOP → REVIEW → CONSOLIDATE — that used to be passive. It doesn't just tell you what to do; it walks you through doing it.

Cultivate is laid out as a dashboard (#620): the idea under cultivation is the one accent card, beside a Notes by stage card that is also the stage filter, and the five moves reflow as a grid of cards rather than a tall column. Like the rest of the Home surface it fills the pane — collapse Obsidian's side panels and the cards spread into two or three columns instead of a narrow centred strip. See the surfaces page.

Starting a session¶

• Home surface → Cultivate mode, the Cultivate — start a thinking session command, or the ribbon menu (🌱). Home offers Cultivate without a purpose alongside the own-material entry.
• ZettelFlow reviews your most embryonic ideas first (#589): the target is ordered by lifecycle stage — fleeting → literature → permanent → developing → evergreen → archived — and within a stage by how connected the note is, deterministically. You develop the rawest ideas first, so the review feels intentional rather than arbitrary. Another idea moves on to the next one in order.

Choosing a stage, and seeing the shape of your vault (#589)¶

Above the target sits a per-stage distribution — one bar for every lifecycle stage, counting every note in your vault (evergreen and archived included), so you see the whole shape of your thinking at a glance. The bars are the selector: click one and Cultivate narrows the review to that stage; click Any stage to clear it. Each bar's magnitude comes from a level class, so the chart respects your theme rather than painting a fixed pixel width.

The choice is remembered between sittings — it is a setting, cultivateStage, that ships with the control (you never hand-edit YAML for it), defaulting to any stage. When a chosen stage has no notes left, the selector and distribution stay on screen with a quiet "No notes at this stage yet." — never an empty surface you cannot get out of.

Peek at a note without leaving (#594)¶

Every note name in Cultivate — the target and the connect/challenge candidates — supports Obsidian's native Page preview: hold Ctrl/Cmd and hover to see the note in a popover without opening it. Clicking still opens the note as before — the preview is an addition, not a replacement.

This started in Cultivate and is now the rule everywhere a note name is clickable — Home, Ask-your-graph, Reasoning paths, Agency review, Health, Resurface, the Evidence map and the Evolution timeline. See Mobile & accessibility → what the code guarantees for the shared helper and the guardrail that keeps a new surface from forgetting it.

The moves¶

Each move is a real, one-click operation on the target note — nothing is invented:

The session refines as you act: after you link a note the connect list shrinks; after you advance the state the next one is proposed. The header shows the idea's degree and maturity — the before/after is a consequence of the moves, never an invented score. Advancing a state (or adding a source/connection) also records a development event for the thinking heatmap.

Two redraws, and never one guard for both (#580)¶

It said refines live for a year and did not. The card only listened to the vault-wide metadataCache resolved event, which a frontmatter-only write may never fire — so you could advance a note to literature and the button would still offer to advance it to literature. All five moves were affected; advance is simply the one whose result changes the whole card.

There are now two named redraws, the distinction the thinking space already paid for once:

And the state says what it became: the chip carries the state, pulses once when it changes, and a line beneath it reads was fleeting, now literature. An emoji can only say where a note is; a promotion is a fact about two states.

What it deliberately is not: there is no celebration, no level, no progress bar toward permanent and no count of notes per state — a locale scan over every cultivate_ and lifecycle_state_ string holds that line in both languages. The lifecycle is a cycle, not a ladder: evergreen goes back to developing for rework and archived revives to fleeting, so there is no final state and the advance control never runs out of something to offer.

Ask before revealing¶

#338, epic #335. On by default; one toggle in Settings → Cultivate turns it off.

Three of the five moves used to hand you their answer first — connect listed related notes, challenge listed contradictions, source announced the gap. A session could organise an idea without ever transforming it. So those three now ask a question before they reveal anything:

Question and advance deliberately get none: a question already is your own thought, and advancing a lifecycle state is a decision you are already making. Friction goes only where the system would otherwise answer for you — that is what makes it deliberate rather than a confirmation dialog.

Reveal needs something written — that is the commitment — and records the answer in the judgement record. Skip reveals the move and records nothing: a skip is not a judgement, and you can always skip.

Nothing here writes to your note. On challenge, what you wrote pre-fills the existing counterpoint field so your thinking is not thrown away and you can still save it with one click; on the other two it stays ephemeral rather than adding note noise. A different idea is a different session, so the prompts come back.

Momentum¶

The session header shows a streak and the size of the cultivation queue. Since #339 the streak counts days you exercised judgement — a verdict on an AI proposal, or an answered friction prompt — not days something happened in the vault. This describes recorded verdict days, not a measure of understanding. See cognitive agency.

Principles¶

• Offline-first. Every move is graph-derived. The optional AI actions (challenge-idea, synthesize) are never required — Cultivate works fully without them.
• Layering. The session is a pure projection (buildCultivationSession, re-exported from the Knowledge State surface); the writes live in a Workflow-Engine CultivationService, so the Experience view only reads state (the #266 guard).

Architecture¶

architecture/knowledge/cultivate/cultivationSession.ts   (pure: session + target + readyToCultivate)
  → re-exported via architecture/knowledge/state
architecture/plugin/services/CultivationService.ts        (the writes: link / question / counterpoint / source / advance)
architecture/components/core/cultivate/CultivateModeRenderer.ts  (the Cultivate mode on the Home surface)

README vocabulary for this page: Cultivate (thinking sessions).