The vault under `docs/ontology/` (this repo's dogfood) — or the user's own vault when one is selected — is the **shared mental model** between the developer and the AI agent. When code grows or shifts, the vault has to follow, otherwise the graph drifts and stops being a useful map.

Pull concepts out of prose (meeting notes, specs, PR descriptions), checking for duplicates first. Writes nothing until a person picks. Use for "extract from this" · "fold this document in".

The next step. Propose, with evidence, what would add the most to this vault. Writes nothing until a person picks. Use for "what should I fill in next" · "find the gaps" · "what's missing".

Review this vault. Find what is broken or disconnected, report it in plain language, and give the exact call that fixes each one. Writes nothing. Use for "what's wrong here" · "check the vault" · "is this healthy".

A **ramp** is the allowed value ladder; an **off-ramp value** is hand-selected outside it; a **gate** is automated enforcement. A 2026-08-03 inventory found 300+ off-ramp values, but their cause was four gate holes:

Build UI from this repository's primitives and ramps in a deterministic order, then prove the rendered result with the matching instruments and gates.

Review only the Atlas structural commitments selected by design routing, using relevant seats, built evidence, optional conflict rebuttal, and one accountable guardian.

When Atlas design routing finds a structural commitment, sketch three different directions in text before code, including the status quo, and let the owner choose one.

Measure topology-map drag, pan, and zoom with a deterministic harness that proves it grabbed a real node instead of silently panning the background.

Verify canvas and UI motion with a real macOS recording, uniform 30fps frames, a visual phase strip, and frame-to-frame pixel-diff statistics.

Read a wiki page through a user-registered third-party MCP, classify it with absorb_document dry-run, obtain approval, and land only approved candidates with the source URL cited.

Build a trustworthy first ontology from an empty or near-empty ontology-atlas vault using only Atlas MCP evidence. Use when the user asks to analyze a codebase, bootstrap/fill its ontology, extract product meaning from a repository, or when a requested ontology task finds only starter nodes. Separate observed implementation facts from proposed meanings, define and cite every domain/capability, answer competency questions, obtain independent source-hidden qualification and user approval, then write only the exact released plan with batch tools. Route mature vaults with 20+ curated nodes to ontology-sync instead; this is a workflow threshold, never a vault or project node limit.

Extract a small, evidence-bound set of ontology candidates from prose, check the existing vault for duplicates, obtain user approval, and land only the approved nodes and relations.

Measure Atlas ontology quality on an unfamiliar repository with a source-hidden handoff and four baseline measurements. Use for changes to meaning-construction rules or MCP read/write behavior that can change vault contents, or an explicit ontology-quality field trial. Skip UI work and wording-only clarifications that preserve the evidence, approval, and write contracts.

Write a subagent brief that isolates ports, files, worktrees, scratch output, baselines, and primary evidence before parallel work begins.

Route Atlas product work from observable change facts and one human-recovery outcome; use a compact solo pass for reversible work.

Live-verify layout across tablet, laptop, and wide breakpoints by measuring rects, elementFromPoint occlusion, bottom-tab reserve, and screenshots.