Chisle
코딩 에이전트 플러그인, 20개 실시간 작업에서 기본 모델 52% 활용, 간결한 페르소나와 출력 압축 제공
Trace는 Claude Code와 Codex 세션을 읽고, 수정할 때마다 표시해 CLAUDE.md에 반영해 반복을 방지합니다
Trace reads your Claude Code and Codex sessions, marks every time you corrected the agent, and writes the fix into CLAUDE.md so you stop repeating yourself.
개발 도구GitHub @brambach온라인운영자 미확인
써 보셨어요? 한 번만 눌러 주세요. 로그인은 필요 없어요.
저장소 정보는 아직 수집하지 않았습니다.
아직 감지된 업데이트가 없습니다. GitHub 릴리스가 나오면 여기에 보입니다.
Finds the corrections you keep repeating to your coding agent and writes them into CLAUDE.md
README 전체를 불러오는 중입니다.
README 전문 보기 ↗I tell my coding agent the same things over and over. Use pnpm. Run the tests with pnpm vitest run. Stop writing comments that restate the code. Each correction costs a turn, and the agent forgets it by the next session.
Trace reads the session files Claude Code and Codex write to disk, finds the corrections you keep repeating, and drafts the line for your CLAUDE.md that would have saved you the trouble. You apply it in one click, as a local edit or a pull request. Then Trace watches whether you ever have to say it again.
Try the demo → It runs on made-up sessions, and nothing you click is saved.
~/.claude/CLAUDE.md.gh, so your checkout and uncommitted work stay put.~/Trace/letters/ as markdown.pnpm install
pnpm dev
That starts the watcher and the dashboard at http://localhost:3000. Use Node 22. If pnpm says it skipped build scripts, run pnpm approve-builds and allow better-sqlite3.
Your sessions stay on your machine. The database lives at ~/Trace/trace.db, and Trace never calls a model.
pnpm demo # the dashboard on made-up sessions, nothing read from disk
pnpm letter # write this week's letter to ~/Trace/letters/ now
pnpm test # vitest across every package
Pre-v1. Every page works on real sessions and in the demo. The coach, replay, the letter and Codex support are new, and the correction rules will keep getting tuned against real habits.
Does Trace send my sessions anywhere?
No. It reads files on disk, writes to a local SQLite database, and touches the network only when you press Open PR (git push and gh).
Does it use an LLM to judge my prompts? No. Everything is counting and quoting. A fix only exists because you said the same thing in more than one session, and Trace shows you each time.
What if a drafted line is wrong? Dismiss it and Trace won't draft it again. Restore it from the same page if you change your mind.
Which agents does it read?
Claude Code from ~/.claude/projects/ and Codex from ~/.codex/sessions/.
Node 22 and strict TypeScript in a pnpm workspace. SQLite through better-sqlite3 and Drizzle, FTS5 for search, chokidar for the watcher, and a Next.js 15 dashboard with Tailwind and hand-drawn SVG.
apps/dashboard/ Next.js dashboard reading from SQLite
apps/watcher/ tails session files, writes rows and the Friday letter
packages/coach/ corrections, fixes, CLAUDE.md edits, PRs, the letter
packages/db/ schema, migrations and queries
packages/demo/ the made-up sessions behind the demo
packages/ingest/ turns parsed lines into rows
packages/parser/ Claude Code and Codex line parsing
packages/shared/ schemas, types and paths
The demo builds its database in memory from scripts in packages/demo, run through the same parser and ingest code as your real sessions.
MIT. See LICENSE.