Onboarding is 4-5 commands, and the install instructions sit ~100 lines into the README

[MongLong0214]

The problem

Today, onboarding a repository is five separate commands, none of them discoverable from the top of the README:

git clone https://github.com/MongLong0214/commitlore ~/.commitlore
node ~/.commitlore/dist/commitlore.mjs --version
node ~/.commitlore/dist/commitlore.mjs doctor --fix
node ~/.commitlore/dist/commitlore.mjs hooks install
node ~/.commitlore/dist/commitlore.mjs index --rebuild

doctor --fix, hooks install, and index --rebuild are three separate commands a new user has to know exist, know the order of, and run every time they onboard a repository. The install instructions themselves — even after #39/#94 added a prebuilt-binary channel — sit roughly 100 lines into README.md, after the philosophy section, the trailer example, and the withdrawn-benchmark discussion.

Two projects the owner named as the standard to compare against:

• caveman — one command (curl -fsSL .../install.sh | bash), auto-detects every agent on the machine and installs for each, ~30 seconds, the install command is in the first screen of the README.
• ponytail — two commands, zero configuration afterwards.

The honest difference: those two are prompt rulesets with no per-repository state. CommitLore needs a git hook and an index per repository, so it cannot collapse to a single lifetime command the way they do. It can collapse to: one command to install, one command per repository.

Proposed fix

1. commitlore init — one command replacing doctor --fix + hooks install + index --rebuild. Runs all three, reports what it did and what it could not (never silently skips a step it failed — see doctor's own fix breaks git fetch and git pull, then reports ok #63 and inject --hook-input is byte-identical on malformed input and on no-records: silent fail-open #67, the same defect class), and is safe to re-run. The three commands stay available on their own.
2. Agent auto-detection in install.sh — after installing the binary, detect which coding agents are present (Claude Code, Codex, Gemini CLI, Cursor, Windsurf, opencode) and wire each one it can (plugin marketplace / .claude/settings.json for Claude Code, MCP registration for the rest), additively: never overwrite an existing config without saying so, never write one for an agent that is not installed.
3. README restructure, all four languages — the install command in the first screen, before the philosophy, the trailer example, and the benchmark discussion. Everything else stays, just moves below.

Ruled out

• Collapsing to a single lifetime command like caveman/ponytail — ruled out because the commit-msg hook and the SQLite index are per-repository state; a machine-wide install cannot set that up for a repository it has not seen yet.