A zero-config project workspace for Claude Code (github.com)

Meridian is a zero‑config workspace that makes Claude Code deterministic and persistent by injecting project context, rules, and tasks from your repo — without changing how you talk to Claude. Installation is literally copy two folders (.claude and .meridian) into your project, make the Python hook scripts executable, and open the repo in Claude Code. No API keys, no service wiring, no new commands or special phrasing: Meridian’s hooks (claude-init.py, session-reload.py, post-compact-guard.py, plan-approval-reminder.py, pre-stop-update.py) automatically inject the system prompt, CODE_GUIDE.md (plus project‑type add‑ons and an optional TDD add‑on), relevant docs, backlog, active task state, and an append‑only memory.jsonl so Claude always resumes with full context after compaction or reload. Technically, Meridian turns best‑effort guidance into enforced workflow: tasks are created as .meridian/tasks/TASK-### folders (brief, approved plan, timestamped context), the task‑manager and memory‑curator scripts reliably persist state, and guards block tool use or session exit until Claude syncs notes, tests, and builds. Configurable modes (hackathon|standard|production) and tdd_mode govern rule precedence (baseline → add‑on → TDD). For the AI/ML community this means more reproducible, auditable agent-driven development: fewer context losses, consistent adherence to project standards, traceable decisions in memory.jsonl, and optional TDD enforcement — all while preserving the natural chat UX developers already use.