ObsidianVault 4-Agent Architecture

Claude Code / Antigravity / Codex / Cursor | 2026-05-20 | v2.0 | Detail view

Active PJ

AI Agents

Claude Skills

Codex Skills

AG Skills

Shared

Sched Tasks

Sub-agents

MCP Servers

CLI Tools

39+

Claude Code (Primary)

Codex (OpenAI)

Antigravity (Gemini)

Cursor (Planned)

Vault (Shared)

Human Layer

Founder (tadashi.kudo) — PM / Tech / Policy | Obsidian GUI | Terminal | Browser

↓

AI-AGENT.md (Neutral Canon — 248 lines)

Map
26 PJ / 4 categories

Rules
15 [I] + 12 [G] rules

Layer 2
7 event triggers

Triggers
7 code-change triggers

Keywords
13 query keywords

↓

Adapter Layer (Runtime-specific diffs only)

Claude Code

CLAUDE.md (199 lines)

Home: ~/.claude/
Agents: 22 (role-based)
Skills: 50/50 (at cap)
Sched: 40 cron tasks
Model: Sonnet 4.6 default
MCP: 12+ servers
Hooks: 5 lifecycle hooks

Codex

AGENTS.md (63 lines)

Home: ~/.codex/
Agents: 12 (function-based)
Skills: 27
Browser: Built-in Use
Model: GPT-5.5 / 5.4
MCP: 10+ servers
Hooks: 4 lifecycle hooks
Plugins: 10 (gmail, drive, etc.)

Antigravity

ANTIGRAVITY.md (55 lines)

Home: ~/.gemini/
Skills: 27 (Codex bulk copy)
Artifacts: brain/<conv>/
Sub: research / self
Model: Gemini 2.5 Pro
MCP: Sidecars only
Sidecar: figma-dev-mode

Cursor

.cursor/rules/ (planned)

Home: .cursor/
Status: Planned
Role: IDE-native editing
Sync: Claude → Cursor
Script: sync-skills-to-cursor.sh
Model: TBD

↓

Shared Resources

Vault Core

01_Projects/ 26 PJ
02_Knowledge/ ADR/playbooks
memory/ auto-memory
Daily Notes/

Code Repos

~/workspace/pj/ 22+ repos
~/workspace/scripts/ shared
~/workspace/cli-registry/

Content Store

$GDRIVE_WORKSPACE/contents/
$GDRIVE_WORKSPACE/datasets/

External Svc

SupabaseOracleNeon CloudflareVercelStripe GitHubX APISlack GmailNotionGA4

[I] Commit Sovereignty Rule (2026-05-20): Each agent commits ONLY its own changes. No proxy commits. Shared resources (memory/, 01_Projects/) are committed by the agent that made the change. Exception: explicit Founder instruction only.

Obsidian Vault = Shared Data Layer

Obsidian はエージェントではない。4つのAI Agentが読み書きする永続データレイヤー（第二の脳）であり、人間がObsidianアプリで閲覧・編集するUIでもある。

Positioning Diagram

        Human (Founder)

        Obsidian GUI · Terminal · Browser · IDE
      
↓ reads / edits via GUI
OBSIDIAN VAULT (Git-managed)
Rules Layer
AI-AGENT.md (neutral canon)
CLAUDE.md / AGENTS.md /
ANTIGRAVITY.md (adapters)
Knowledge Layer
01_Projects/ (26 PJ)
02_Knowledge/ (ADR/playbooks)
Daily Notes / Weekly Notes
Memory Layer
memory/MEMORY.md (index)
memory/feedback_*.md
memory/pj*.md / working-memory

        ↓    ↓    ↓    ↓
        
git clone / pull / push — each agent operates on the same repo
      
Claude Code
~/.claude/
Codex
~/.codex/
Antigravity
~/.gemini/
Cursor
.cursor/

Obsidian IS

永続データレイヤー — 4Agent全員が読み書きする唯一の共有ストレージ
ルールの正本置き場 — AI-AGENT.md（中立正本）+ 各adapter
PJ状態の共有層 — 01_Projects/ に全PJのindex/status
意思決定の記録層 — ADR・feedback・lessons learned
記憶の永続化層 — memory/ をgit経由で全Agent共有
人間の閲覧UI — Obsidianアプリでグラフビュー・検索・リンク
Git同期の媒体 — GitHub経由で全Agentの変更を統合
パッシブな存在 — 自ら何も実行しない。Agent/人間が動かす

Obsidian IS NOT

エージェントではない — 自律的に動作しない。処理能力ゼロ
実行環境ではない — コードを実行しない（それはAgent/ターミナルの仕事）
MCP/APIの提供元ではない — vault-rag MCPが検索を提供するが、Obsidian自体ではない
リアルタイム同期基盤ではない — git push/pullベースの結果整合性
コードリポジトリではない — コードは ~/workspace/pj/ に分離
コンテンツストアではない — HTML/画像/動画は $GDRIVE_WORKSPACE に分離

Vault Directory Roles

Directory	Role	Primary Writer	Reader	Governance
`AI-AGENT.md`	全Agent共通ルール（中立正本）	Claude (primary)	All 4 agents	[I] rules require Founder approval
`CLAUDE.md`	Claude adapter（差分のみ）	Claude	Claude only	Commit Sovereignty
`AGENTS.md`	Codex adapter（差分のみ）	Codex	Codex only	Commit Sovereignty
`ANTIGRAVITY.md`	Antigravity adapter（差分のみ）	AG	AG only	Commit Sovereignty
`01_Projects/`	26 PJ の index / status / docs	All (per PJ)	All	変更者がcommit
`02_Knowledge/`	ADR / playbooks / references / tech	All	All	変更者がcommit
`memory/MEMORY.md`	メモリインデックス	Claude (primary)	All	150行超で /compact-memory
`memory/feedback_*.md`	運用フィードバック（恒久ルール）	Claude + Human	All (RO)	Founder corrections only
`memory/pj*.md`	PJ詳細メモリ	Claude (primary)	All	—
`memory/working-memory.md`	未完了タスク・一時ログ	All	All	OCC via memory_write_guard.py
`Daily Notes/`	日次ノート	Claude + Human	All	/session-end が自動転記
`Weekly Notes/`	週次ノート	Human (primary)	All	—
`Qiita/`	Qiita記事草稿	Claude (sched-task)	Claude	qiita-daily-writer が管理
`00_Inbox/`	未整理メモ一時置き場	Human	All	定期整理
`03_Archive/`	完了PJ・古いノート	—	検索対象外	—
`04_Lab/`	短期試行メモ（コード不要）	All	All	育てるならRnD/へ移動

Storage Separation: Obsidian vs Others

Obsidian Vault

What goes here:

PJ状態・ステータス・index
意思決定記録（ADR）
Agentルール・adapter
auto-memory / feedback
playbooks / references
Daily / Weekly Notes

同期: Git (GitHub) — 全Agent push/pull

~/workspace/pj/

What goes here:

アプリケーションコード
PJ固有の CLAUDE.md / .codex/
テスト・CI設定
package.json / pyproject.toml
PJ固有のspec / docs

同期: Git (GitHub) — PJごとにrepo分離

$GDRIVE_WORKSPACE

What goes here:

HTML視覚化・設計図
画像・動画・PDF
KWプランナー (xlsx)
CSV / ZIP データセット
LP / バナー等コンテンツ

同期: Google Drive — ローカルマウント

Agent ↔ Obsidian Interaction

Agent	Read Method	Write Method	Search Method	Bootstrap
Claude	`Read` tool (direct file I/O)	`Write` / `Edit` + auto git commit+push	`vault-rag search_vault` + `find_similar_notes`	Fully automatic (SessionStart hook loads AI-AGENT.md + MEMORY.md)
Codex	`Read` tool + `mcp-obsidian`	`Write` / `Edit` + manual git	`vault-rag search_vault` (same MCP binary)	Hook (session_start.py) + fallback: vault-context-bootstrap skill
AG	`Read` tool (file I/O)	`Write` + manual git	File search (no vault-rag MCP)	Manual: user runs vault-context-bootstrap skill
Cursor	IDE file explorer	IDE editor + manual git	IDE search	TBD: .cursor/rules/ auto-loaded

Key Insight: Obsidian は「場所」であり「アクター」ではない。Agent が自律的に思考・実行する一方、Obsidian は静的にデータを保持するだけ。この分離により、Agent を追加・削除しても Vault のデータ構造は不変。4Agent → 5Agent になっても、新しいadapter.md を1ファイル追加するだけで済む。

Agent Comparison

	Claude Code	Codex	Antigravity	Cursor
Provider	Anthropic	OpenAI	Google	Anysphere
Primary Role	Master runtime. Full PJ lifecycle, scheduled tasks, Agent Teams, MCP orchestration	Review (GPT-5.5), Dev (GPT-5.4). Browser Use built-in. Independent second opinion	Research, exploration, creative artifacts. Lightweight brainstorming	IDE-native coding. Fast inline edits (planned)
Adapter	`CLAUDE.md` 199L	`AGENTS.md` 63L	`ANTIGRAVITY.md` 55L	`.cursor/rules/` TBD
Config Home	`~/.claude/`	`~/.codex/`	`~/.gemini/`	`.cursor/`
Sub-agents	22 (role-based)	12 (function-based)	2 (research, self)	TBD
Skills	80 (50 skills + 30 cmds)	27	27 (Codex copy)	TBD
Scheduled Tasks	40 cron jobs	None (receives from Claude)	None	None
MCP Servers	12+ (vault-rag, code-rag, gitnexus, Chrome, Slack, Gmail, Calendar, Notion, Drive, Playwright, Preview, computer-use)	10+ (vault-rag, obsidian-vault, gitnexus, sqlcl, playwright, notion, xactions, fudosandb, stitch, mcp-youtube)	Sidecars only (figma-dev-mode)	TBD
Plugins	N/A (MCP-based)	10 (github, computer-use, documents, spreadsheets, presentations, gmail, drive, calendar, slack, browser/chrome)	N/A	N/A
Browser	Chrome MCP + Playwright + Preview + computer-use	Browser Use (built-in, first choice) + Playwright MCP	Limited	Built-in preview
Memory Write	RW primary writer + auto-commit+push	RW + own `~/.codex/memories/`	RW vault + own `brain/` artifacts	TBD
Cost Model	Max Plan ($200/mo) or API usage	ChatGPT Plus (rate-limited, no per-call)	Gemini Pro / API credits	Cursor subscription
Strength	Deep context, agentic orchestration, scheduled automation, full MCP ecosystem	Strong reasoning (5.5), built-in browser, independent review, image generation	Fast research, creative exploration, Figma integration	IDE integration, fast inline edits

Sub-agent Roster

Claude (22) — Role-based

ceopmdevdesigneranalystbloggermarketingnewssalessecuritycomplianceproposalform-salesgc-articlegc-note-articlekarte-articlegovinsight-matchercodex-workerlp-builderai-consultantweb-provisionerbrand-voice-base

Codex (12) — Function-based

content-researchercreative-production-orchestratordesign-reviewerexplorerimage-studioimplementation-orchestratorlanding-page-builderreviewerrisk-auditorstrategy-plannervault-note-editorvisual-director

Antigravity (2)

researchself

research: read-only investigations
self: parallel work clones

MCP Server Distribution

How MCP servers are shared (or not) across agents. Each agent has its own MCP config; servers with the same name share the same binary but are configured independently.

Vault & Code Intelligence

vault-rag

✓ .mcp.json

✓ config.toml

—

Vault semantic search (bge-m3 + sqlite-vec)

obsidian-vault

—

✓ mcp-obsidian

—

Vault full-text read via mcp-obsidian

code-rag

✓ Claude native

—

Code search across ~/workspace/pj/ (22 repos)

gitnexus

✓ .mcp.json

✓ config.toml

—

Code dependency / impact analysis (PJ03, PJ19, PJ29, PJ34)

sqlcl

—

✓ Oracle SQL MCP

—

Oracle DB queries via SQLcl

Browser & UI

Claude-in-Chrome

✓ Extension

—

Chrome DOM control (navigate, find, form, JS exec)

Claude Preview

✓ Built-in

—

Dev server preview + screenshots

computer-use

✓ Native app

✓ Plugin

—

Desktop control (screenshot, click, type)

playwright

✓ .mcp.json

✓ config.toml

—

Headless browser automation (E2E, complex flows)

Browser Use

—

✓ Built-in plugin

—

Codex native browser (first choice for web)

figma-dev-mode

—

✓ Sidecar

Figma design inspection

Communication & Productivity

slack

✓ MCP

✓ Plugin

—

Slack send/read/search

gmail

✓ MCP

✓ Plugin

—

Gmail search/draft/labels

google-calendar

✓ MCP

✓ Plugin

—

Calendar list/create events

google-drive

✓ MCP

✓ Plugin

—

Drive search/read/list files

notion

✓ .mcp.json + global

✓ config.toml

—

Notion page/DB CRUD

Domain-specific

xactions

✓ CLI

✓ config.toml

—

X (Twitter) post/delete/search

mcp-youtube

—

✓ config.toml

—

YouTube transcript extraction

fudosandb

—

✓ mcp-remote

—

Real estate DB API (PJ29)

stitch

—

✓ Google Stitch

—

Google Stitch API

openaiDeveloperDocs

—

✓ Remote

—

OpenAI API docs reference

scheduled-tasks

✓ Built-in

—

Claude scheduled task management

MCP Sharing Architecture

Key insight: MCP servers are NOT shared processes. Each agent has its own config pointing to the same binary/endpoint independently.

Pattern	How it works	Examples
Same binary, separate config	Both agents point to the same executable (e.g. `vault-rag-mcp-single`) but configure it in their own settings file	vault-rag, gitnexus, playwright, notion
Plugin equivalence	Claude uses MCP, Codex uses curated plugins that wrap the same APIs	slack, gmail, calendar, drive
Agent-exclusive	Only one agent has access (by design or not yet migrated)	code-rag (Claude only), Browser Use (Codex only), figma (AG only)
Tooling-adapter abstraction	`02_Knowledge/playbooks/ai/tooling-adapters.md` maps abstract roles to runtime-specific tool names	"Vault semantic search" → Claude: `mcp__vault-rag__search_vault` / Codex: same MCP

CLI & Script Sharing

All scripts live in ~/workspace/scripts/. The CLI Registry (~/workspace/cli-registry/registry.toml) is the source of truth for which runtime can use which tool.

Guard Levels

HARDBLOCK Blocks execution without confirmation (billing=paid)

WARN Shows warning, allows override

LOG Records usage silently

HARDBLOCK Scripts (billing=paid)

Script	Purpose	Claude	Codex	Guard	Requires
`dalle_gen.py`	DALL-E / GPT-image-1 generation ($0.04-$0.12/img)	✓	✓	HARDBLOCK	OPENAI_API_KEY
`oci-a1-launch.sh`	Oracle Cloud A1 instance launch	✓	—	HARDBLOCK	OCI_TENANCY, etc.
`gemini`	Gemini CLI (pay-per-call)	✓	✓	HARDBLOCK	GEMINI_API_KEY
`stripe`	Stripe CLI (production billing)	✓	—	HARDBLOCK	STRIPE_SECRET_KEY

WARN Scripts (high-risk writes)

Script	Purpose	Claude	Codex	Requires
`supabase`	Supabase CLI (db push / migration)	✓	✓	SUPABASE_ACCESS_TOKEN
`kw_volume.py`	KW volume check (SEO gate)	✓	✓	DATAFORSEO_*
`vault-claude-update.sh`	Vault git commit & push	✓	—	—
`vercel`	Vercel deploy CLI	✓	—	VERCEL_TOKEN

Shared Tools (both runtimes)

Script	Purpose	Guard
`ga4_cli.py`	Google Analytics 4 reporting	LOG
`gsc_cli.py`	Google Search Console data	LOG
`gitnexus`	Code dependency analysis MCP	LOG
`vault-rag-mcp-single`	Vault semantic search server	LOG
`markitdown`	Convert files to Markdown	LOG
`scrapling`	Web scraping toolkit	LOG
`firecrawl-cli`	Firecrawl web crawler	LOG
`grok_cli.py`	Grok API caller	HARDBLOCK
`x_api_search.sh`	X API search	LOG

Claude-only Tools (20+)

x_api_post.shx_api_delete.shx_post_scheduled.pyx_update_cookie.sh qiita_post.pyqiita-daily-writer.shcodex_orchestrator.py nl_subscribers_cli.pyxserver_domain.py morning_briefing_orchestrator.shsync-skills-to-cursor.sh backup-vault-to-icloud.shbackup-workspace-to-gdrive.sh claude-session-analyzer.pygenerate-claude-agents.pygenerate-codex-agents.py safe-delete.shcleanup-claude-cache.sh

CLI Registry Architecture

      ~/workspace/cli-registry/registry.toml (source of truth)

        ↓ guard.sh reads billing / guard_action fields

        ↓ auto-generates Vault/02_Knowledge/references/ai-tools/cli-inventory.md

        ↓ runtime = ["claude", "codex"] controls which agent can use each tool

        ↓ codex_safe = true/false gates Codex access to dangerous tools

Configuration Hierarchy

        ~/.claude/CLAUDE.md (Global Rules 199L) — Claude-only top priority
      
↓ overrides

        AI-AGENT.md (Neutral Canon 248L) — ALL agents read this
      
↓ adapter diff only
CLAUDE.md
Vault adapter
AGENTS.md
Codex adapter
ANTIGRAVITY.md
AG adapter
.cursor/rules/
Cursor (planned)
↓ PJ-local overrides
PJ-local CLAUDE.md / .codex/ / .cursor/

Shared Resource Access Matrix

Resource	Claude	Codex	Antigravity	Cursor	Notes
`memory/MEMORY.md`	RW	RW	RO	TBD	Claude is primary writer
`memory/feedback_*.md`	RO	RO	RO	RO	Human-only write
`memory/working-memory.md`	RW	RW	RW	TBD	OCC via memory_write_guard.py
`01_Projects/`	RW	RW	RW	TBD	Commit sovereignty applies
`02_Knowledge/`	RW	RW	RW	TBD
`~/workspace/pj/`	RW	RW	RO	RW
`~/workspace/scripts/`	RW	RW	RO	TBD	Gated by cli-registry
`~/.claude/`	OWNER	DNT	DNT	N/A	Do Not Touch
`~/.codex/`	DNT	OWNER	DNT	N/A	Do Not Touch
`~/.gemini/`	DNT	DNT	OWNER	N/A	Do Not Touch
`$GDRIVE_WORKSPACE`	RW	RW	RW	TBD	Content store
`cli-registry/`	RW	RO	RO	TBD	Claude manages registry

Sync Strategy

Direction	What Syncs	Mechanism	Frequency	Notes
Claude → Cursor	Skills (read-only replica)	`sync-skills-to-cursor.sh [--dry-run]`	Manual	One-way push
All ↔ All	AI-AGENT.md rules	Each reads at session start	Every session	rules-only, not agent bodies
All ↔ All	`memory/` via git	Git auto-commit+push	Continuous	Commit sovereignty rule
Claude ↔ Codex	Agent bodies	NOT synced	Never	Different philosophies (role vs function)
All	CLI tools	`cli-registry/registry.toml` runtime field	On script create	Guard actions enforced per-agent
All	Tooling adapters	`tooling-adapters.md` lookup table	Reference	Abstract role → runtime-specific MCP name

Hooks & Session Bootstrap

Claude Code Hooks

Event	Action
SessionStart	Auto-load AI-AGENT.md summary + auto-memory (MEMORY.md)
UserPromptSubmit	Prompt validation hooks
PreToolUse:Bash	Guard checks (git add -A deny, etc.)
PostToolUse:Write	Preview panel trigger
Stop	Layer 2 check + session cleanup

Bootstrap: Fully automatic. Hook + auto-memory + CLAUDE.md chain.

Codex Hooks

Event	Script
SessionStart	`session_start.py` — Vault context injection
UserPromptSubmit	`user_prompt_submit.py`
PreToolUse:Bash	`pre_tool_use_bash.py` — Guard checks
Stop	`stop_layer2_check.py`

Python path: Must use /opt/homebrew/bin/python3 (Apple 3.9.6 breaks type hints).

Fallback: vault-context-bootstrap skill if hook fails.

Antigravity Bootstrap

Step	Action
1. Manual	User runs `vault-context-bootstrap` skill
2. Read chain	AI-AGENT.md → MEMORY.md → user_profile → feedback → PJ
3. Artifacts	Complex outputs → `~/.gemini/antigravity/brain/<conv-id>/`

No auto-hooks. Bootstrap is always explicit.

Cursor (Planned)

Step	Action
1. Rules	`.cursor/rules/` auto-loaded by IDE
2. Sync	`sync-skills-to-cursor.sh` (one-way from Claude)

TBD: adapter file, hook system, MCP config.

Safety Guardrails Comparison

Guardrail	Claude	Codex	Antigravity
git add -A / --all / . deny	✓ Hook + settings.json deny	✓ pre_tool_use hook	N/A (manual git)
API billing gate	✓ CLAUDE.md [I] + cli-registry hardblock	✓ guard.sh	✓ ANTIGRAVITY.md safety
X post confirmation	✓ Draft + confirm mandatory	✓ Same rule via AI-AGENT.md	✓ Same
Destructive git commands	✓ settings.json deny list (force push, reset --hard, etc.)	✓ pre_tool_use hook	N/A
Secret leak prevention	✓ AI-AGENT.md [I] + .gitignore	✓ Same	✓ Same
Commit sovereignty	✓ AI-AGENT.md [I] (2026-05-20)	✓ Same	✓ Same

RAG & Code Intelligence Stack

3つの独立システムが Vault ノート・ソースコード・コード依存関係を検索可能にしている。全てローカル実行（外部API課金なし）。

vault-rag Notes

1,320

vault-rag Chunks

28,083

code-rag Repos

code-rag Files

1,051

code-rag Chunks

4,096

gitnexus Repos

gitnexus Nodes

14,560

gitnexus Edges

24,547

vault-rag

対象	Obsidian Vault ノート (Markdown)
Engine	bge-m3 + sqlite-vec
Binary	`vault-rag-mcp-single`
Index	1,320 notes → 28,083 chunks
Tools	`search_vault`, `find_similar_notes`, `entity_mentions`, `query_graph`
Agent	Claude Codex
用途	概念・文脈検索。「PJ33のAUC改善経緯」等

code-rag

対象	~/workspace/pj/ 全コードリポ
Engine	bge-m3 + sqlite-vec
Index	26 repos / 1,051 files → 4,096 chunks
Tools	`search_code_tool`, `find_similar_code_tool`
Agent	Claude only
用途	コード意味検索。「Stripe webhook handler」等

gitnexus

対象	主要5リポのコード依存関係
Engine	Graph (AST解析 → nodes + edges)
Binary	`gitnexus mcp` (npm)
Index	825 files / 14,560 nodes / 24,547 edges / 627 communities
Tools	`impact`, `context`, `query`, `route_map`, `api_impact`
Agent	Claude Codex
用途	影響範囲分析。「この関数を変えたらどこが壊れる？」

How They Relate

          Obsidian Vault

          1,320 Markdown notes
        
          ~/workspace/pj/

          26 code repos (1,051 files)
        
↓ chunk & embed
↓ chunk & embed   ↓ AST parse

          vault-rag

          sqlite-vec
28,083 vectors
        
          code-rag

          sqlite-vec
4,096 vectors
        
          gitnexus

          graph DB
24,547 edges
        
        ↑ 全てローカル実行（bge-m3 embedding model はマシン上で推論）↑

gitnexus Indexed Repos

Repo	Files	Nodes	Edges	Communities	Indexed	Status
PJ03 politech-os	248	6,337	13,537	392	2026-05-02	OK
PJ19 gcportal	267	4,156	5,447	128	2026-05-18	1 behind
PJ25 pj25-reserve	125	1,306	1,792	42	2026-05-18	1 behind
PJ29 residential-hazard	107	1,919	2,509	50	2026-05-18	6 behind
PJ34 ReraFlowAIAutopilot	78	842	1,262	15	2026-05-18	1 behind

vault-rag PJ Breakdown (top 12)

PJ19

185

PJ07

120

PJ08

PJ01

PJ03

PJ11

PJ34

PJ25

PJ22

PJ09

PJ05

PJ29

Gap: code-rag は Claude専用。Codex はコード意味検索できず grep / gitnexus に頼っている。Antigravity は vault-rag も code-rag も未接続で、ファイル直接読みのみ。

Skills & Commands Overview

commands/ vs skills/ — Claude Code 固有の区別

commands/ — スラッシュコマンド

ユーザーが /session-end と明示入力して起動。~/.claude/commands/*.md に格納。30件

skills/ — 自動マッチスキル

自然言語の意図から Skill tool が自動判定して起動。~/.claude/skills/*/SKILL.md に格納。50件（上限）

scheduled-tasks/ — 自動実行タスク（v2.0 追加）

~/.claude/scheduled-tasks/<id>/SKILL.md に格納。cron で自動起動。36件。人間レビュー薄い = A分類相当リスク。/skill-rotate Part C で棚卸し。

Codex / Antigravity はこの区別がなく、全て skills/ 一本。Claude Code だけが3系統（skills + commands + scheduled-tasks）に分かれている。以下の集計では Claude の commands + skills を統合して比較する（scheduled-tasks は別管理）。

Claude Total

50 skills + 30 cmds

Codex Skills

AG Skills

3-Agent Shared

Codex∩AG Only

Unique Total

3-Agent Shared (21) — 全Agent共通

同名が Claude（skills/ or commands/）、Codex、Antigravity の全てに存在。実装は各Agent固有。

Code Review / Dev (6)

browser-to-api migrate-fan-out plan-review two-stage-review postmortem session-review

Design / UI (4)

claude-design-prompt frontend-design responsive-check ux-upgrade

SEO (2)

seo-rank-tracker design-review

X / SNS (4)

buzz-copy form-sales x-post x-longform-post

Research (1)

jp-research

Ops / Session (4)

session-end stop-report skill-health skill-optimizer

Claude: commands/ 配置 / Codex+AG: skills/ 配置

Codex ∩ AG Only (4)

Codex と AG には存在するが Claude（skills/ にも commands/ にも）にはない。

ai-video-factcheck claude-context-sync seo-content-pipeline vault-context-bootstrap

Claude は hook で Vault 文脈を自動読み込みするため vault-context-bootstrap 不要。seo-content-pipeline は Claude の SEO 9段階スキル群で代替。

Claude Only (59) — skills: 34 + commands: 25

Claude にのみ存在。S = skills/ C = commands/

Code Review / Dev (8)

^S adversarial-debate ^S claude-api ^S codex-review ^S diagnose ^S grill-me ^S zoom-out ^C quality-gate ^C dev-preview

Design / UI (7)

^S ban-pattern ^S clone-website ^S diagram-viz ^S html-arch-diagram ^S tailwind-patterns ^S vibe-design ^C slide-design

SEO / Content (8)

^S seo-kw-research ^S seo-competitor-analysis ^S seo-outline ^S seo-article-writer ^S seo-fact-check ^S seo-quality-check ^S seo-rewriter ^S seo-thumbnail

Media / YouTube (2)

^S youtube-summarizer ^S yt-competitive-analysis

Research / Data (5)

^S deep-research ^S defuddle ^S notebooklm ^C autoresearch ^C research

GovCloud / GEO (4)

^S gc-rag-refresh ^S gcinsight-newsletter ^S geo-citation-monitor ^S geo-llmstxt

X / SNS (1)

^C x-research

Ops / Vault (14)

^S commit ^S cross-pj-kpi ^S handoff ^S repo-snapshot ^S restore-scheduled-tasks ^C analyze ^C archive-done ^C brainstorm ^C cleanup-inbox ^C compact-memory ^C morning-briefing ^C sort-meetings ^C gws ^C model-route

Harness / Meta (10)

^C harness-audit ^C harness-init ^C harness-kb-sample ^C hook-profile ^C loop-start ^C loop-end ^C loop-status ^C skill-create ^C skill-rotate ^C update-agent-guide

Claude Code のハーネス管理・スキルライフサイクル系。他Agentには不要

Codex Only (2)

codex-dev-pipeline codex-primary-runtime

Codex 固有のランタイム起動・開発パイプライン

AG Only (4)

antigravity-dev-pipeline antigravity-primary-runtime

AG 固有のランタイム起動・開発パイプライン（Codex版のリネームコピー + swiftdata-pro, swiftui-pro は Apple Dev 特化）

Apple Dev (2)

swiftdata-pro swiftui-pro

Gemini 2.5 のネイティブ Swift 知識を活用

Skill Sharing Architecture

Aspect	How it works
Claude 2系統	`commands/` (30) = `/name` 明示呼び出し / `skills/` (50) = 自然言語自動マッチ。重複ゼロ・完全排他。Codex/AG はこの区別なし
Body sync	NOT synced. 同名スキルでも中身は各Agent独自実装。MCPツール名・モデル・パイプライン構成が異なる
AG = Codex コピー	`copy_codex_skills.py` で Codex skills/ を一括コピー。25/27件が Codex と同名。Agent名のみリネーム（codex-dev-pipeline → antigravity-dev-pipeline）
Claude → Cursor	`sync-skills-to-cursor.sh` で一方向コピー（read-only replica）
命名規約	同じ目的のスキルは同名にする（将来の棚卸しのため）。ただし実装は自由
設計思想の違い	Claude: 細分化 (SEO 9段階 + ハーネス管理cmds) / Codex: 統合パイプライン (seo-content-pipeline 1本) / AG: Codex clone + Swift 特化
Dedup audit	未実装。Risks タブ参照 — 手動で同名スキルの乖離を確認中

Key Insight: AG は Codex のほぼ完全コピー（25/27件一致）。独自性は swiftdata-pro / swiftui-pro の Apple Dev 特化のみ。Claude だけが commands/ という別系統を持ち、ハーネス管理・メモリ管理・スキルライフサイクル管理といった「メタ運用」機能が commands/ に集中している。

Product (Revenue) 01_Projects/Product/

PJ03 politech-os (SaaS)

PJ25 Cabstay (rental)

PJ26 AI site builder

PJ29 bukken-score (SaaS)

PJ33 Teinou (boat AI)

PJ34 ReraFlow AI Autopilot

PJ36 expired patent DB

PJ38 Umanou (horse AI)

Media (Acquisition) 01_Projects/Media/

PJ19 GCInsight (SEO)

PJ22 reform-navi (SEO)

PJ24 DX navi (matching)

PJ31 denshi-karte (SEO)

Infrastructure 01_Projects/Infrastructure/

PJ01 Vault auto-improve

PJ05 X auto-post

PJ06 X Articles

PJ07 AI news summary

PJ08 Morning briefing

PJ09 AI management

PJ11 Content marketing

PJ14 Maps lead gen

PJ15 VaultIndex (RAG)

PJ17 AI Video X

PJ37 Corporate site

Scheduled Tasks (40)

Category	Tasks	#
SEO/Content	gcinsight-, pj19-, pj22-, seo-, qiita-*	14
X/Social	x-tweet-, x-reply-, x-algo-, x-weekly-	5
Business	pj34-, pj25-, pj29-, job-hunting-	7
Vault/Meta	vault-, session-audit, wtf-triage, harness-, weekly-*	8
News/Briefing	morning-briefing, firehose-collect, model-usage-daily	3
KPI	pj09-kpi-daily, pj09-ceo-reflection	2
Research	github-weekly-trending	1

Cognitive Load Risks

1. Context Fragmentation

4 agents each maintain separate context windows. Decisions in Antigravity won't auto-propagate to Claude's scheduled tasks. AI-AGENT.md mitigates but real-time sync is absent.

2. Memory Write Conflicts

memory/ shared via git. Simultaneous writes to working-memory.md cause merge conflicts. Mitigation: memory_write_guard.py --merge (optimistic concurrency).

3. Skill/Agent Sprawl

Claude 116 entries (50 skills + 30 commands + 36 scheduled-tasks) + Codex 27 + AG 27 + 36 sub-agents across 3 runtimes. skills-usage.json v2 で3層トラッキング統合済み（2026-05-20）。

4. Adapter Drift

AI-AGENT.md updated but adapters lag. context-drift-review exists but manual. No automated 4-adapter drift detection.

5. MCP Config Drift

Same MCP server configured independently in Claude (.mcp.json) and Codex (config.toml). Version/args can diverge silently.

6. CLI Registry Coverage Gap

New scripts may be created without registry.toml entry. Guard enforcement only works for registered scripts.

7. Bootstrap Asymmetry

Claude: fully automatic hooks. Codex: hooks with fallback. Antigravity: fully manual. Risk of Antigravity operating without Vault context.

Mitigations

Already in Place

AI-AGENT.md as neutral canon
Adapter diff-only pattern
Runtime home isolation (~/.claude/ etc.)
feedback_*.md read-only for all agents
Commit sovereignty rule [I]
CLI registry with guard_action + runtime fields
tooling-adapters.md abstraction layer
sync-skills-to-cursor.sh
context-drift-review keyword
vault-context-bootstrap skill (Codex/AG)

Consider Adding

Automated adapter drift CI (daily context-drift-review)
MCP config parity checker (Claude .mcp.json vs Codex config.toml)
Cross-agent skill dedup audit
Antigravity auto-bootstrap hook
Cursor adapter bootstrap
CLI registry pre-commit hook (new scripts must be registered)
Agent task routing decision log
Shared MCP version pinning

更新履歴 — Version History

v2.0 2026-05-20 — 大規模構造リファクタリング

区分	変更内容	影響
新規	Scheduled Tasks を第3管理層として正式 INDEX 化（36件）	skill-index.md / CLAUDE.md / skill-management.md
新規	Commands 全30件を skills-usage.json に seed 登録（以前は3件のみ）	skills-usage.json v2
改善	skills-usage.json を v1（3スキーマ混在）→ v2（skills / commands / scheduled_tasks 3セクション）に正規化	トラッキング精度向上
改善	A分類スキル3件（form-sales / gc-rag-refresh / gcinsight-newsletter）に placeholder エントリ追加	Eval 監視の実効化
改善	/skill-rotate に Part C（Scheduled Tasks 棚卸し）追加予定	skill-rotate.md
改善	A分類基準に「外部サービス依存（Codex exec 等）」を追記	skill-management.md
修正	Sched Tasks カウント 40→36（registry-snapshot.json 実数に合わせ修正）	ヘッダー stat カード
改善	Skills タブに scheduled-tasks/ 3層目の説明を追加	Skills タブ

v1.1 2026-05-20 — 数値修正（Opus + GPT-5.5 レビュー検出）

区分	変更内容	修正値
修正	AG Skills カウント	4 → 27（実 ls 計測）
修正	Shared スキル数	16+2 → 21
修正	Unique Total	88 → 90（数式再計算: 59+21+4+2+4=90）
修正	AG Only 件数	(2) → (4)
新規	Commands 管理を /skill-rotate Part B に統合（以前は棚卸し対象外）	skill-rotate.md / skill-index.md
修正	skills-usage.json: commands が skills キーに誤分類（session-end 等3件）を修正	skills-usage.json
修正	skill-management.md: dev-monitor を A分類リストから削除（2026-05-18 アーカイブ済み）	skill-management.md

v1.0 2026-05-18 以前 — 初版

4-Agent Architecture の初版ダッシュボード。Claude Code / Codex / Antigravity / Cursor（予定）の構造を可視化。Skills / MCP / CLI / RAG / PJ Map / Risks の11タブ構成。