5dive

Fixed

Three grok provisioning gaps surfaced by a live smoke test:

• 5dive-agent-start now seeds /home/agent-<name>/.grok/auth.json from /home/claude/.grok/auth.json (or $PROFILE_STATE_DIR/.grok/auth.json under a bound profile) at every boot. Previously the auth-gate in cmd_create passed because the type-level shared credential satisfied it, but the agent's own ~/.grok/auth.json was never populated — so grok couldn't actually talk to xAI on first launch. Mirrors the codex seed block; mtime-gated so host-side 5dive auth login grok re-rotations propagate on the next agent restart.
• install_channel_for_grok_agent now copies notify-user/SKILL.md into ~/.grok/skills/notify-user/ for --channels=telegram agents, so the comms loop self-starts on the first DM (no manual nudge needed). Mirrors the claude-side seed in preseed_claude_agent.
• Default skills find-skills + 5dive-cli now install for grok agents too. Upstream npx skills add rejects --agent grok with "Invalid agents: grok", so a new manual-install fallback (git clone --depth=1 + cp -r) in install_default_skill_for_agent and cmd_skill_add handles types upstream doesn't recognize. _skill_needs_manual_install is the single switch — add new types there when upstream rejects them.