graphrag-vault-init

name	graphrag-vault-init
version	1.2.0
description	vault (ナレッジグラフ) の初期構築。system vault (コード/プロダクト知識)・project vault (時限イニシアチブ)・principal vault (当事者=法人/常設ユニットの恒久知識) の3種に対応。「vault を作りたい」「初期構築したい」「新しいプロジェクトを管理したい」「リポジトリを索引したい」「全社ルールの置き場がほしい」で発火。

Vault Initial Setup

Creates a new vault and populates initial nodes. Routes to the appropriate flow based on vault type.

Prerequisites

• Embedding endpoint must be reachable (Ollama / LM Studio with nomic-embed-text).
• CLI launcher: node --experimental-strip-types ${CLAUDE_PLUGIN_ROOT}/graphrag/cli.ts <verb> [args]

Hereafter $CLI = the launcher above, $REF = ${CLAUDE_PLUGIN_ROOT}/references.

Vault Types — Choose First

The axis is lifecycle × 拠り所 (what grounds the knowledge). Three presets, enumeration closed:

Type	拠り所	Purpose	Schema	Example
system	人工物	Code/product knowledge (passive)	schema: system (13 node types)	A product platform, an API service, an internal tool
project	企て	Time-bounded initiative (active)	schema: project (16 node types)	認可取得プロジェクト, API renewal
principal	当事者	The deciding/signing entity — perpetual	schema: principal (14 node types)	法人の手続き規範 (調達しきい値・稟議), standing contracts, 全社方針

project vault scope: Goal reaches achieved / abandoned → vault lifecycle closes (read-only archive). Do NOT create vaults for systems/products or teams/orgs.

principal vault scope: never closes (business withdrawal is a Decision, not a vault closure). Admission test: is the author / signer / allocator of this knowledge the operating entity itself? "Useful company-wide" alone is NOT admission. No Task/Milestone — time-bounded work routes to a child project vault. 与件 (law / regulation / standard) has NO preset of its own: write Source(source_kind: regulation) pointers + entity-authored Constraint/OK/Decision inside principal (or project) — see $REF/schema-quickref-principal.md. Setup follows the Project Vault Path below (same family, minus Task/Milestone steps); when bootstrapping, deliberately wire ≥1 cross-vault ref from a live project vault and carve ≥1 Source(regulation)→Constraint chain — these consumption events are what validate the vault.

---

System Vault Path

For system vaults (code/product knowledge), use the carve command:

$CLI carve --root <repo-path> --system <system-name>

This runs: index → concept extraction → quality gate. See $REF/indexing-and-carving.md for the full procedure.

VAULT.md for system vault:

---
name: <system name>
schema: system
vault_slug: <slug>
---

Close the initial build by presenting the plugin's operating conditions (${CLAUDE_PLUGIN_ROOT}/docs/operating-conditions.md) — the runbook of what the project must keep doing for the tool to stay effective (wire knowledge to files, write enforcers, register "later", commission periodic audits…), each tied to which feature dies without it. A vault that is built once and never wired decays into a diary.

The rest of this document covers project vault setup only. For system vault schema, read $REF/schema-quickref-system.md.

---

Project Vault Path

Step 0: Create Thin System Vaults for Dependencies

Before creating a project vault, ensure each system vault it will reference already exists with Deliverable stubs. This resolves the chicken-and-egg problem (project vault can't reference non-existent Deliverables).

# For each system vault the project depends on:
mkdir -p <system-root>/vault/Deliverable
# Write VAULT.md + Deliverable node files (title + summary only, no edges needed yet)

Bootstrap exception: hand-writing VAULT.md and stub node files is sanctioned ONLY at this step — an empty vault has no writer to go through yet. From the vault's first real write onward, the parent skill's "DO NOT edit vault/ directly" rule applies with no exceptions (all writes via add-* / commit-mutation).

Step 1: Create Directory Structure

<project-root>/
  VAULT.md           ← vault profile (sibling of vault/)
  vault/             ← node files go here

VAULT.md placement rule: vaultProfilePath() resolves to path.dirname(vaultDir)/VAULT.md. If vault is at <root>/vault, then VAULT.md must be at <root>/VAULT.md (NOT inside .graphrag/). Misplacement causes fallback to system schema.

Step 2: Write VAULT.md

---
name: <project name>
schema: project
vault_slug: <slug>
---
<1-2 line project description>

Required fields:

• name: Human-readable project name
• schema: project — selects the project schema. Omitting this causes system schema to be used, which will reject project node types. A vault's schema is its kind: system or project.
• vault_slug: Cross-vault ref namespace. Short kebab-case. Immutable once set.

Optional fields:

• parent: vault_slug of the single parent program/project this sub-project is contained by. Same-schema, single-parent, genuine structural containment only (not a business/marketing grouping under a product name) — see Decision Criteria → "the parent field".

Step 3: Gather Information & Populate — Model Division Strategy

Initial construction has three distinct cognitive tasks. Use the right model for each:

Phase	Task	Model	Why
3a. Node extraction	Extract nodes from information sources (Confluence, Jira, Slack, etc.)	extraction model — a fast, cheaper tier (e.g. Sonnet), as subagent	Pattern recognition / information extraction. High volume, lower abstraction.
3b. Edge modeling	Wire has_premise, risks_in, depends_on, cross-vault ref edges between nodes	modeling model — the strongest available tier (self)	Conceptual modeling. Requires reverse reasoning ("what depends on this assumption?").
3c. Theme extraction	Identify cross-vault themes and wire encompasses edges	modeling model (self)	Highest abstraction. Requires seeing patterns across multiple vaults simultaneously.
3d. Gap reification	Materialize implicit concepts mentioned in Themes/descriptions but not yet nodes	modeling model (self)	Themes often name shared bottlenecks that the extraction pass didn't extract as Resource/Constraint nodes. Ask: "What compute, personnel, facilities, or budget does this Theme's activities consume?"

Recommended flow (from a single session on the strongest available model):

1. Create directory structure + VAULT.md (Steps 1-2 above)
2. Dispatch extraction subagent(s) for node extraction:
   - Provide information sources (URLs, page IDs, documents)
   - The subagent extracts Goal/Stakeholder/Milestone/Risk/Assumption/Agreement/Constraint/Task/Source nodes
   - The subagent writes via commit-mutation
3. Review the extraction output, then perform yourself (modeling model):
   - has_premise edges (which Goals/Decisions depend on which Assumptions?)
   - cross-vault ref edges (which Tasks require which system vault Deliverables?)
   - risks_in edges (which Risks threaten which Tasks/Goals/Milestones?)
4. If multiple project vaults exist, extract Themes:
   - Read nodes across all vaults
   - Identify recurring cross-project patterns
   - Create Theme nodes with encompasses edges in each relevant vault
5. Gap reification — for each Theme, check:
   - Does the Theme description mention a shared resource/bottleneck? → create Resource node (category: asset/people/budget)
   - Does it mention an implicit constraint? → create Constraint node
   - Wire requires/constrains edges from existing Tasks to the new nodes
   - This catches implicit shared resources (compute clusters, specialist personnel pools,
     partner capacity) that source documents mention as activities but never name as resources

Why not the extraction model for edges? Extracting "what this document says" (nodes) is different from reasoning "what conceptual dependency exists between these two nodes" (edges). A fast extraction model excels at the former but tends to miss reverse-direction implications (e.g., "this Goal has_premise that Assumption" requires understanding that if the Assumption breaks, the Goal is at risk).

Extraction-model delegation: what it CAN and CANNOT do:

Handles well	Misses (supplement in later steps)
Node extraction from source docs	Resource nodes (source docs describe activities, not underlying resources)
Extraction-type edges: achieves, targets, responsible_for, documented_by	Inference-type edges: has_premise, risks_in (requires reverse reasoning)
Cross-vault ref matching (if given the full ID list)	Cross-vault ref discovery (cannot infer which Deliverables are needed)
certainty assignment on Assumptions (if given the 4-level definition)	Theme extraction (requires cross-vault abstraction)

Extraction-subagent prompt tips:

• Include a complete mutation plan example (project vault version) — extraction accuracy jumps with concrete examples
• Provide the cross-vault Deliverable ID list from Step 0 system vaults
• Include the certainty 4-level definition: Established (confirmed fact) / Expected (high confidence from evidence) / Assumed (unverified premise) / Speculative (hope or guess)
• For Agreements with source backing: use raw_content + raw_content_status: copied_from_summary when derived_from type pairs don't allow direct linking

Information to gather

Collect from Confluence, Jira, Slack, Google Slides, meetings, etc.:

1. Goal: What is the project trying to achieve? By when?
2. Stakeholder: Who is involved? Owners, approvers, partners
3. Milestone: Key checkpoints and dates
4. Risk: Known risks
5. Assumption: Premises being relied upon (certainty field is required: Established/Expected/Assumed/Speculative)
6. Agreement: External commitments (contracts, grants, partnerships)
7. Constraint: Immutable conditions (regulations, deadlines, budget caps)
8. Cross-vault Deliverable: Which system vault Deliverables does this project depend on?

Step 4: Populate Initial Nodes (extraction-subagent phase)

Use commit-mutation for batch creation. All distilled nodes (Decision/RejectedOption/Risk/OK/Agreement) require a derived_from edge to a Source or Investigation with raw_content.

$CLI commit-mutation <plan.json>

For the full initial-population plan template (Goal / Stakeholder / Milestone / Assumption with required certainty / Agreement / Source nodes + achieves / cross-vault requires edges), use $REF/mutation-templates.md §Initial population.

Step 5: Verify

Smoke-test retrieval:

$CLI ask "What is this project's goal?"
$CLI ask "What are the biggest risks?"
$CLI ask "Who is responsible?"

Resource gap audit — Resources are the most under-extracted node type because source documents describe activities ("data collection", "model training") without naming the underlying resources ("compute cluster", "field engineers"). Run this check:

1. List all Task nodes that have zero requires edges to any Resource
2. For each, ask: "What physical, computational, human, or financial resource does this task consume?"
3. Create missing Resource nodes (category: people/budget/asset/time) and wire requires edges
4. Pay special attention to shared resources that multiple Tasks (possibly across vaults) compete for — these are the highest-value Resource nodes for conflict detection

Schema Reference

For the full project vault schema (16 node types, 23 edge types, state vocabulary, cross-vault ref format), read $REF/schema-quickref-project.md. For principal vaults (14 node types, 20 edge types, 与件 pattern), read $REF/schema-quickref-principal.md.

Decision Criteria

"Should this be a vault?"

• ✅ Time-bounded initiative (has start and end) → project vault
• ✅ Independently deployed system → system vault
• ❌ Team / org → NOT a vault (use Stakeholder)
• ❌ Sub-component that doesn't release independently → Component in parent system vault

"Both nested levels are valid vaults — which one owns a given node?" (the parent field)

When a containment relationship holds and both the container and the contained legitimately exist as vaults (a subsystem that releases on its own; a sub-project with its own lifecycle), a node can plausibly be filed in either — neither is wrong. A node-to-node edge cannot resolve this, because the ambiguity is about vault scope, not about a link. Declare the containment in the child's VAULT.md:

parent: <parent vault_slug>

This lets a knowledge-gathering crawler choose the narrowest correctly-scoped vault for each node instead of guessing. Rules — keep them strict or the tree rots into a junk-drawer pointer:

• Single parent only. If you can't name exactly one containing vault, it's not a parent — it's a dependency (cross-vault ref) or a crosscut (Theme/Concern).
• Same schema. project→project, system→system. A project under a system is a cross-vault Deliverable ref, not parentage.
• Genuine containment, not a collective label. The child must be literally part of the parent, not just grouped with peers under a name the business uses now. "Systems A/B/C are currently sold as product X" is a business framing that drifts when the product is renamed or repackaged — wiring it into parent makes the tree churn with org/marketing changes. Test: would the part-of relation survive dropping the product name? If it holds only because they're "called X together," it's a crosscut (Theme/Concern), not parent.
• No lifecycle cascade. parent is organizational only; archiving stays independent (a child may outlive its parent).
• Not goal-alignment. "Which goal does this serve" stays refines (Goal→Goal) + cross-vault ref; "which concern cuts across many vaults" stays Theme. parent is purely which vault contains which.

Validate with xref-check — it reports parent status (resolved / orphan / self / schema-mismatch / cycle / unresolvable).

"Decision or OperationalKnowledge?"

• Consciously compared alternatives → Decision
• Learned through operation → OperationalKnowledge
• Immutable external condition → Constraint
• Unverified premise → Assumption

"Goal two-layer pattern"

Vision Goal (stays active) + gate Goal (achieved/abandoned), connected by refines:

Goal: Strengthen hiring pipeline (active, never closes)
  refines → Goal: Run event successfully (achieved @ 8/23)
  refines → Goal: Send 10 interview offers (achieved @ 8/25)

"How to express blocked?"

Task has no blocked state. Create a Risk node for the blocker and connect via risks_in → Task.

"Agreement retreat (negotiation failed, restart)"

Do NOT reverse state. Expire old Agreement → create new at exploring. Connect with supersedes if appropriate.

"When does the vault close?"

Top-level Goal reaches achieved or abandoned. All Tasks are completed or cancelled. Vault becomes read-only archive. System vault Deliverables live on.