English | 한국어

Multi-repo background coding agent orchestrator — one root system, many trunks.

Pando is the largest known single organism on Earth: a quaking-aspen clone of ~47,000 stems sharing one root system. This project mirrors that — one orchestrator (the root) grows git worktrees (stems) across many repos, and coding agents implement tickets on each stem.

Feed it a Jira ticket (or a brief drafted from chat) and it runs:

SPEC → PLAN → TEST → IMPL ⇄ REVIEW → PR (draft)

Each stage runs inside an isolated git worktree, executed by a coding-agent CLI (Claude Code / Codex). Gates between stages are judged only by deterministic signals — exit codes, file artifacts, checksums. An agent claiming "done" is never trusted.

It processes many repos × many tickets in flight, distinguishing company repos (Jira/Confluence/Figma) from personal repos (brief-based) through per-repo profiles.

Early implementation. W1-W3 are complete; W4 n x n parallel scheduling is next. See the design docs under docs/ (written in Korean):

• research-v1.md — tooling & pattern research
• design-v2-multi-repo.md — n×n design built on reusable agent-skill assets
• repo-structure.md — repo layout & core interfaces
• engineering-standards.md — development methodology
• adr/ — architecture decision records

pnpm install
pnpm verify   # coverage + lint + types — required before every commit

Discipline: no implementation code without a failing test first (RED-GREEN-REFACTOR) · atomic commits (~100 lines) · architecture decisions go in an ADR. See CLAUDE.md.

This repository uses Git Flow:

• main: protected release branch
• develop: integration branch
• feature/*, release/*, hotfix/*: working branches
• release tags do not use a v prefix, for example 0.1

Commit messages are written in English. Run pnpm verify before every commit.