🧰 jay-plugin

Jay's Claude Code toolbox — a personal collection of hooks, commands, skills, and agents for everyday Claude Code workflows, packaged as a single plugin so every machine stays in sync.

Instead of copying install scripts between machines, register the marketplace once and sync everything with git push → plugin update.

English | 한국어

Installation

/plugin marketplace add jyh4358/jay-plugin
/plugin install jay-plugin@jay-plugin

⚠️ If you previously installed these hooks manually, remove the old entries (Stop hooks in ~/.claude/settings.json and ~/.claude/commands/tokens.md) to avoid double execution.

Requirements: Python 3.9+ · macOS (for notifications only — the token reporter works everywhere)

Components

Type	Name	Description	Status
Hook (Stop)	plan-usage	Per-turn token usage report as a system message after every response	✅
Hook (Stop · Notification)	stop-notify	macOS notification on response completion (Glass) and when Claude is waiting for your input — permission requests, questions (Ping)	✅
Command	/tokens	Ranks the current session's turns by weighted usage and explains why the heavy ones were heavy	✅
Statusline	statusline	Model · git branch/worktree · 5h/weekly plan usage bars with reset timers · session time · context usage	✅
Skill	/commit	Analyzes staged changes, generates a Korean Conventional-Commits message (optional Jira ID), confirms, then commits	✅
Skill	/record	Records session work into the My Track Record Obsidian vault as resume-ready task notes — why/how/result structure, short interview, confidentiality filter	✅
Skill	/record-review	Monthly vault maintenance — re-evaluates resume/impact on work notes, checks frontmatter & wikilink integrity across work/learning notes (report → confirm → apply)	✅
Skill	/learn	Captures things learned mid-session (or user-specified concepts) into the vault as knowledge notes — fixed frontmatter, free-form body, "why" required, fact-checked against primary sources, always-on generalization filter	✅
Agents	—	(planned)	🚧

New components go into their directory (hooks/ commands/ skills/ agents/) — push, then update the plugin on each machine.

plan-usage — token usage reporter

Shows what every turn actually consumed, right after it finishes:

━━━━━━ 이 턴 플랜 토큰 ━━━━━━
Q: 사용하는 모델 기준으로 출력되게 할 수 있어?
모델: claude-opus-5
소진(신규) ~62,948   재사용(캐시) 10,071,993   가중≈1.21M (Opus입력 환산)
  ├ 출력 생성 24,287
  └ 신규 입력 38,661
       질문+시스템 3,059 / 도구결과 5,449 / 루프재입력 30,153
  도구별 유입(추정):
    Read        ~   3,644  (1회)
세션 누적: 소진 ~138,536  가중≈2.25M  (9턴)

Key concepts

• Weighted usage — a single index in Opus-input-token equivalents. Reflects model tier (haiku 0.2x / sonnet 0.4–0.6x / opus 1x / fable 2x), output 5x, cache reads 0.1x — a more honest proxy for plan consumption than raw token counts.
• Subagent-aware — reads agent-*.jsonl transcripts and attributes their usage to the turn that spawned them.
• Ledger — ~/.claude/token-usage/<session>.json, ~0.5 KB per turn. Stores numbers only (plus the first 200 chars of each question) — never conversation content.
• Future-proof pricing — built-in table → family inference for brand-new models (opus/sonnet/haiku/fable) → user override via ~/.claude/token-usage/prices.json. No code edits when models or prices change.

CLI reference

SCRIPT=~/.claude/plugins/…/jay-plugin/scripts/plan_usage.py   # installed path

python3 $SCRIPT --rank <session-id>   # recomputed ranking for one session
python3 $SCRIPT --rank-all            # all sessions (--since YYYY-MM-DD / --json)
python3 $SCRIPT --log  <session-id>   # read the stored ledger
python3 $SCRIPT --repair              # reconcile ledger against transcripts (--yes to apply, keeps .bak)

Notes

• Figures are local approximations. The authoritative plan-usage number is /usage.
• Once Claude Code's transcript retention (default 30 days) cleans a session, recomputation (--rank/--repair) is no longer possible — the ledger keeps the last known values.

statusline — plan usage HUD

Model: Fable 5 | branch:main
  5h:[█░░░░░░░]7%(4h32m) wk:[█░░░░░░░]17%(5d15h) | session:6m | ctx:[█░░░░░░░░░]9%

Two rows: the first holds the identity segments that vary in length (model, branch/worktree), the second the near-fixed-width numbers. A single row overflows a split pane and gets cut with ..., and wrapping it would move the segment boundaries on every refresh — a fixed split keeps each value in the same place.

Renders entirely from the JSON Claude Code pipes to the statusline command — no network calls, no dependencies (rate limits require a subscription/OAuth login; the segment is skipped for API-key users). Colors: green → yellow (≥70%) → red (≥90%). session is colored by elapsed time alone (≥60m yellow, ≥120m red) and formatted 7m / 2h5m / 5d3h; context pressure is already reported by the ctx segment next to it. Inside a linked git worktree the branch segment becomes branch:feat-x (wt:feat-x).

Plugins can't register a statusline automatically, so add this once per machine to ~/.claude/settings.json (after installing the plugin):

"statusLine": {
  "type": "command",
  "command": "python3 \"${CLAUDE_CONFIG_DIR:-$HOME/.claude}/plugins/marketplaces/jay-plugin/scripts/statusline.py\""
}

stop-notify — macOS notifications

Two distinct sounds so you can tell them apart without looking:

Event	Sound	Meaning
Response finished (Stop)	Glass	done — come check the result
Waiting for you (Notification)	Ping	action needed — permission request, question, or idle prompt

The notification title is the session title (custom title first, then AI title).

---

한국어 안내

Jay의 Claude Code 툴박스 — 일상 워크플로에서 반복해 쓰는 훅·커맨드·스킬·에이전트를 하나의 플러그인으로 묶어 여러 맥에서 동일하게 쓰기 위한 개인 컬렉션. 수동 설치 스크립트를 기기마다 옮기는 대신, 마켓플레이스 등록 한 번으로 설치하고 git push → 플러그인 업데이트로 모든 기기를 동기화한다.

설치

/plugin marketplace add jyh4358/jay-plugin
/plugin install jay-plugin@jay-plugin

⚠️ 이전에 수동으로 설치했던 맥이라면 다음을 제거해야 이중 실행되지 않는다: ~/.claude/settings.json 의 Stop 훅 2개(plan_usage, stop-notify), ~/.claude/commands/tokens.md

구성 요소

종류	이름	설명
훅 (Stop)	plan-usage	매 턴 종료 시 토큰 사용량 리포트를 시스템 메시지로 표시
훅 (Stop·Notification)	stop-notify	응답 완료(Glass음), 입력 대기 — 권한 요청·질문 — 시(Ping음) macOS 알림
커맨드	/tokens	현재 세션의 가중 사용량 랭킹 + 왜 무거웠는지 분석
상태줄	statusline	모델 · git 브랜치/워크트리 · 5h/주간 플랜 사용량 막대(리셋 타이머) · 세션 시간 · 컨텍스트 사용률 표시
스킬	/commit	변경사항 분석 → Conventional Commits 한국어 커밋 메시지 생성(Jira ID 지원) → 확인 후 커밋
스킬	/record	세션에서 한 작업을 Obsidian vault(My Track Record)에 무엇을/왜/어떻게/성과 구조의 작업 노트로 기록 — 부족한 맥락 인터뷰 + 회사 기밀 보안 필터
스킬	/record-review	vault 월간 점검 — 작업 노트 resume·impact 재평가, 작업·학습 노트 frontmatter·wikilink 무결성 점검 (리포트 → 확인 → 반영)
스킬	/learn	세션 중 새로 알게 된 것·개념 정리를 학습 노트로 기록 — frontmatter 고정 + 본문 자유, '왜/원리' 필수, 1차 출처 검증(틀린 지식은 수정하여 기록), 일반화 필터 상시 적용

주요 개념

• 가중 사용량: Opus 입력 토큰으로 환산한 단일 지수. 모델 티어(haiku 0.2x / sonnet 0.4~0.6x / opus 1x / fable 2x), 출력 5x, 캐시읽기 0.1x 반영.
• 서브에이전트 포함: agent-*.jsonl 을 읽어 각 턴에 시각 기준으로 귀속.
• 장부: ~/.claude/token-usage/<세션ID>.json — 턴당 ~0.5KB 요약만 저장 (대화 본문 없음, 질문 앞 200자만).
• 단가(가중치 원천): 내장 표 → 계열 추정 → ~/.claude/token-usage/prices.json 오버라이드 순. 새 모델이 나와도 계열로 자동 추정하고, 가격이 다르면 prices.json 에 한 줄 추가.

주의

• 수치는 로컬 근사치다. 플랜 실제 소진율의 권위값은 /usage.
• transcript(~/.claude/projects/**)가 보존 기간(기본 30일)으로 정리되면 해당 세션의 재계산은 불가능해지고, 장부의 기존 값만 남는다.

License

MIT