fix(config): release only Claude's MCP servers on a root move; read the recorded root in import and skill tracking

[Rererr]

Summary

Two follow-ups to #728, from the review's third pass. A re-init that moves the Claude Code root released the teamai-managed MCP servers of every MCP-capable tool, not only Claude's: reconcileMcpForConfig(…, { removeAll: true }) walks every tool of the config it is handed, so Codex, Cursor and the rest lost their servers until the next pull put them back. The release now hands it a team config narrowed to Claude. And two readers still hard-coded ~/.claude: import --from-claude scanned ~/.claude/rules, and skill-use tracking (skillExistsOnDisk) only knew the static ~/.claude/skills, so a relocated Claude Code's rules were not importable and its skills were rejected as not installed. Both now read the recorded root: import through one helper, resolveMemberToolRoots (the resolution the local agent already used — project config governing the directory, else user scope — moved into config.ts so the readers agree), and tracking through the same helper, given the hook's directory, so a skill under the root that governs the project the hook fired in counts.

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature causing existing behavior to change)
• Documentation only
• Refactor / internal cleanup

Test Plan

• npx tsc --noEmit passes
• npx vitest run passes (4500 tests)
• Added/updated tests for the change

New/updated tests: init.test.ts (the MCP release is handed a team config whose toolPaths hold Claude only), usage-tracking.test.ts (+2: a skill installed only under the given toolRoots.claude counts as present and not otherwise; a slash command from a hook whose directory's project config records a relocated root is tracked when the skill lives only there; a project config without a record follows the user-scope one; a hook that reports a directory that no longer exists — a deleted worktree — still records, under user scope), import-local-relocated-root.test.ts (new, 1: import --from-claude scans the recorded root's rules and not ~/.claude/rules).

mcp-reconcile.test.ts (+1: removeAll on a config narrowed to Claude removes Claude's server only — Cursor's file and manifest rows stay — the runtime counterpart of the init assertion).

Mutation checks: handing the full team config to the MCP release fails 1 test; dropping the recorded root from skillExistsOnDisk fails 1; not passing the resolved roots from trackSlashCommand fails 1; dropping the root from the import scan fails 1; probing a missing hook directory without the existence guard fails 1 (Cannot use simple-git on a directory that does not exist).

Real-CLI end-to-end (built dist/index.js, isolated HOME, local bare repo bridged from a synthetic HTTPS URL, Claude and Codex both installed, a team MCP server in mcp/mcp.yaml, seed toolPaths declaring mcp for both tools):

• main (55b71ef): init → pull lands the server in ~/.claude.json and ~/.codex/config.toml (1 each). Re-init with CLAUDE_CONFIG_DIR=~/.claude-a prints Removed 2 teamai-managed MCP server(s) and, before the next pull, Codex's count is 0; the pull restores it.
• This branch: the same re-init prints Removed 1, Codex stays at 1 throughout, ~/.claude.json goes 1 → 0 and ~/.claude-a/.claude.json gets the server on the next pull. Same on the second move (~/.claude-a → ~/.claude-b) and on the blank-variable clear.

Related Issues

Follow-up to #728 (third review pass, findings 1 and 5).

Notes for Reviewers

• Why narrow the team config rather than the reconciler. The reconciler's removeAll contract ("every tool of this config") is what uninstall relies on; the release is the one caller that means a single tool, so the narrowing belongs at the call site.
• resolveMemberToolRoots keeps the local agent's resolution order (project config governing the directory, else user scope) and only moves it into config.ts; the local agent's behavior is unchanged. A project config without a record follows user scope on purpose: it is the rule init applies when it fills a project config in, and a user-scope init may record the root after the project was set up. Tracking calls it with the hook's directory (resolveHookCwd), so the hook's directory, not the process cwd, decides.
• Not in this PR, left from the same review pass: the release still keys the root change on the settings path alone (unset → explicit ~/.claude skips the MCP move), only releases delivered model config in user scope, and resolves the old root from dirname(settings). Each has a narrow trigger (an explicit default, HTTP mode in project scope, a customized nested settings path); they are separate changes if wanted.
• Behavior change in the local agent, small and intended: memberToolRoots now goes through the same helper, so a workspace path that no longer exists resolves to the user-scope record instead of throwing from simpleGit. Existing directories behave as before. A project config that is present but unreadable still falls back to user scope here (as the inline code did); resolveConfigForDir treats that case as "no config" — left as is, noted for a later pass.
• Missing hook directory. resolveMemberToolRoots checks that the directory exists before probing for a project config, the way resolveConfigForDir already does: simpleGit throws on a directory that is gone (a deleted worktree), and a slash command from such a hook must still be recorded.
• Docs: one sentence in both usage guides; the --from-claude option text in src/index.ts names the recorded root and skill-data/core/references/commands.md is regenerated from it (npx vitest run commands-reference -u). The setup skill already describes the recorded root.

🤖 Generated with Claude Code

[github-actions]

• [P2 non-blocking] src/config.ts:366 skips the existence check when dir is undefined. If a hook omits cwd while the process is inside a deleted worktree, resolveConfigForDir() correctly falls back to user scope, but resolveMemberToolRoots() then calls detectProjectConfig(undefined) and simple-git throws. Resolve dir ?? process.cwd() first, then apply pathExists() to that target.

The PR description includes a sufficient test plan and real-CLI end-to-end verification record.