"Autonomously Improving Claude Code Without Humans" 前作 continued. This time, we're discussing mechanisms for recognizing environmental bloat.
`~/.claude/` As skills accumulated underneath, launchd jobs multiplied, and projects exceeded double digits, I could no longer instantly answer "how many layers does the whole system have and what's running now?" Rather than running `ls` every morning, I aimed for a state where opening Obsidian gives me an overview, so I wrote `~/.claude/scripts/env-map.sh`.
Here's a snapshot of the current state:
- Plugin bundle skills: 1,004
- Auto-generated skills (auto/): 77
- launchd jobs (`com.shun.*.plist`): 24
- Projects I'm tracking: 11
Last-update times and git states vary across skills, jobs, and projects, and the cost of checking them accumulates. I wanted a single page that could answer "what did today's autopilot do?" and "what's the branch status of that project?"
The diagram has 3 layers.
Output destination is Obsidian Vault (`~/Documents/claude-obsidian/wiki/meta/environment-map.md`). vault-auto-ingest (4:55) picks it up and auto-commits, so version control comes for free.
The script design philosophy is written in comments.
# 何が起きても生成を完走させる（個別の収集失敗は "?" で degrade）。set -e は使わない。
set -uo pipefail
The key point is not using `set -e`. Even if something partially fails—like MCP connection checks—we fill in `?` and output to completion.
When launched from launchd, PATH is only `/usr/bin:/bin:/usr/sbin:/sbin` or so. `node`, `claude`, and `jq` can't be found. We explicitly construct PATH at the script's beginning.
NVM_BIN="$(ls -d "$HOME"/.nvm/versions/node/*/bin 2>/dev/null | sort -V | tail -1)"
export PATH="$HOME/.local/bin:/opt/homebrew/bin:/opt/homebrew/sbin:${NVM_BIN:+$NVM_BIN:}/usr/local/bin:/usr/bin:/bin:/usr/sbin:/sbin"
Since nvm changes paths when you upgrade the Node version, we dynamically resolve the latest binary with `sort -V | tail -1`. Hard-coding paths breaks on the next version upgrade.
The minimal launchd PATH trips up "all CLIs installed by users." Stacking fallbacks in order—`~/.local/bin` (uv, claude, etc.) → Homebrew → nvm—ensures you hit something in any environment setup.
Collecting Claude Environment
AUTO_SKILLS="$(ls "$HOME/.claude/skills/auto/" 2>/dev/null | grep -vc README)"
AGENTS="$(find "$HOME/.claude/plugins" "$HOME/.claude/agents" -path '*/agents/*.md' \
-o -path "$HOME/.claude/agents/*.md" 2>/dev/null | wc -l | tr -d ' ')"
PLUGINS="$(jq -r '.enabledPlugins // {} | length' "$SETTINGS" 2>/dev/null || echo '?')"
HOOK_EVENTS="$(jq -r '.hooks // {} | keys | length' "$SETTINGS" 2>/dev/null || echo '?')"
LAUNCHD="$(ls "$HOME/Library/LaunchAgents/"com.shun.*.plist 2>/dev/null | wc -l | tr -d ' ')"
MCP requires special attention. `claude mcp list` attempts live connections, so startup is slow and failures happen. We cap it with a timeout.
MCP_OK="?"
if have claude; then
_mcp="$(timeout 12 claude mcp list 2>/dev/null)"
[ -n "$_mcp" ] && MCP_OK="$(printf '%s' "$_mcp" | grep -c 'Connected')"
fi
After 12 seconds via `timeout 12`, we give up and leave it as `?`. Generation doesn't stop.
Project State
We fetch branch, last commit date, and uncommitted change count per repository, reflecting them in node colors.
proj_meta() {
local path="$1"
PROJ_EXISTS=0; PROJ_BRANCH="-"; PROJ_LAST="-"; PROJ_DIRTY=0
[ -d "$path" ] || return
PROJ_EXISTS=1
if git -C "$path" rev-parse --git-dir >/dev/null 2>&1; then
PROJ_BRANCH="$(git -C "$path" rev-parse --abbrev-ref HEAD 2>/dev/null || echo '-')"
PROJ_LAST="$(git -C "$path" log -1 --format=%cd --date=format:%Y-%m-%d 2>/dev/null || echo '-')"
PROJ_DIRTY="$(git -C "$path" status --porcelain 2>/dev/null | wc -l | tr -d ' ')"
fi
}
Projects with uncommitted changes show orange; projects that don't exist on disk show red.
echo ' classDef dirty fill:#3a2a00,stroke:#e8a33d,color:#fff;'
echo ' classDef gone fill:#3a1a1a,stroke:#e06666,color:#fff;'
Just opening Obsidian in the morning, you can see at a glance which projects have uncommitted changes.
Sanitizing Mermaid Labels
If hostnames or chip names contain `()` or `[]`, Mermaid's parser breaks. We pass all labels through a sanitization function.
san() { printf '%s' "$1" | tr '"[]()#|<>' ' ' | tr '\n' ' ' | sed 's/ */ /g; s/ *$//'; }
Strings like `Apple M4 Pro (14-core)` are output safely.
Image of the Generated Diagram
graph LR
PC["💻 PC 環境<br/>MacBook · macOS 15.x"]
CC["🤖 Claude 環境<br/>plugins 14 · launchd 24"]
PJ["📦 プロジェクト環境<br/>11 repos"]
PC --> CC
CC -->|builds / runs| PJ
PC -.->|ローカル開発| PJ
The Claude environment breakdown looks like this (measured values):
graph TD
CC["🤖 Claude Code"]
SK["🧩 Skills"]
SK --> SKa["auto: 77"]
SK --> SKp["plugin: 1004"]
CC --> SK
CC --> AG["🎭 Agents"]
CC --> PL["🔌 Plugins enabled"]
CC --> HK["🪝 Hooks"]
CC --> MCP["🔗 MCP connected"]
CC --> AUTO["⚡ launchd 24 jobs"]
<key>StartCalendarInterval</key>
<array>
<dict><key>Hour</key><integer>4</integer><key>Minute</key><integer>50</integer></dict>
<dict><key>Hour</key><integer>8</integer><key>Minute</key><integer>10</integer></dict>
</array>
<key>RunAtLoad</key><false/>
It fires at 4:50 to place output before vault-auto-ingest at 4:55. ingest commits it directly to Obsidian, so no git operations needed. 8:10 is a catch-up in case 4:50 is skipped due to overnight sleep. It's idempotent, so multiple fires are harmless. `RunAtLoad: false` suppresses immediate firing on plist load.
Logs are consolidated into a single file.
<key>StandardOutPath</key>
<string>~/.claude/logs/env-map.launchd.log</string>
<key>StandardErrorPath</key>
<string>~/.claude/logs/env-map.launchd.log</string>
If `generated (N lines)` appears on the last line, you're good.
[2026-06-20 04:50:03] environment-map.md generated (218 lines)
- Adding `set -e` exits on MCP timeout → Use only `set -uo pipefail`. MCP and git failures continue via `?` or `-`
- Hard-coded NVM path broke on Node update → Dynamically resolve latest binary with `sort -V | tail -1`
- Mermaid labels with `()` cause parse errors → Pass all labels through a `san()` function. We hit this with chip names.
- Falling behind vault-ingest gives yesterday's output → Dual setup: fire early at 4:50, catch-up at 8:10
- Missing project path changes → Visualize with `(未検出)` red nodes. Update the list when PRI changes.
- Output directory doesn't exist, script fails → Place `mkdir -p "$(dirname "$OUT")"` right before `mv`
- Comma-joined node IDs with empty spaces break Mermaid → Format with `sed 's/ /,/g'` before passing to `class`
- When skills, launchd jobs, and projects exceed double digits, prepare one Mermaid diagram page for overview
- Minimal PATH on launchd startup: explicitly construct it at the script's beginning. For nvm, assume version upgrades and resolve dynamically.
- Slow operations like MCP connection checks: wrap with `timeout`, degrade failures with `?` to prioritize completion
- Placing output before vault-ingest means version control and commits happen automatically
- Visualizing uncommitted project changes with node colors means opening Obsidian in the morning tells you what actions to take
Combining this article with 死活監視（automation-health-check） gives you a single-page morning snapshot of "what's running, what's broken, and what's left undone."
Lily (@bokuwalily) — Individual developer. Building automation infrastructure with Claude Code while mass-producing iOS apps and web services.
◼︎ Apps I've made are summarized at **ポートフォリオ** 📱
◼︎ New releases and development behind-the-scenes shared on X **@bokuwalily** 🐦
◼︎ OSS: **github.com/bokuwalily** 🐙
Your ❤️ and shares are my motivation!
