Conventions

• Atomic style applies to Claude's TUI replies, not to source files, comments, or documentation. Source files follow the codebase's own conventions.
• CLAUDE.md in any project should hold only meaningful context for that codebase — not general reminders, not duplicated tool lists. Keep it lean.
• No AI bylines in commit messages or PR descriptions.
• The scratchpad (.claude/.scratchpad/<slug>/) is LLM working memory — ephemeral, gitignored, not for human consumption. Durable decisions go in docs/. It is a slug-keyed bundle, not a dated one: atomic scratchpad new <slug> --purpose <plan|implement|fix|diagnose|review> creates or extends it, seeding only what that purpose still needs, so a task worked across several phases accumulates in one directory instead of a new one per phase per date. atomic scratchpad path <slug> prints its location, list [--archived] enumerates bundles, and archive <slug> retires one.
• Session reports, reminders, and archived bundles live outside the repository, under ~/.atomic/<project-key>/ — one home per clone, shared by every worktree of it. atomic where --json resolves the current branch's report path, the reports root, the reminders directory, and the archive root for this project; nothing should construct those paths by hand. Run atomic migrate --show-log for the history of how this and other state layouts changed.
• .claude/ paths throughout these docs describe the default (Claude Code) harness. Under another agent the same layout resolves beneath that harness's directory (.pi/.scratchpad/, .pi/project/, …) via harness detection — experimental; see concepts.
• Tests verify intent, not behavior. A test that still passes when the business logic changes is wrong.
• tmp/ is for throwaway experiments and ad-hoc verification scripts. Not a scratch directory for checked-in work.
• When /subagent-implementation is about to start significant work, it prompts whether to use an isolated worktree. Already inside .claude/worktrees/* (or any linked worktree)? It skips the prompt.