llm-cli-gateway: Skill for Claude Code

Personal Agent Config Kit

Use the Kit for one developer who wants the same personal agent baseline on multiple workstations and repositories. It is a local, Git-synchronized personal configuration layer, not an organization, team, tenancy, or remote policy service.

Read the Kit guide before changing the baseline or recovering an interrupted attempt.

Enable and bootstrap

Configure each workstation's local gateway:

[personal_config]
enabled = true
baseline_path = "~/.agent-config"
max_stale_hours = 168

baseline_path must resolve to a non-home, non-symlinked descendant of the current user's home directory. Relative paths, /, ~, .. traversal, and existing symbolic-link components are rejected before permission hardening.

Use a private Git remote. config_init accepts an HTTPS URL without userinfo, authenticated through the normal Git credential helper, an SSH URL without a password authenticated through the normal SSH agent, or standard SSH shorthand. It rejects local paths, plaintext git://, helper transports, URL credentials, query or fragment components, and dash-leading hosts.

config_init({remote:"[email protected]:you/agent-config.git"})

For a new baseline, initialize it, use ordinary Git tooling to add a remote and first commit, then publish. Do not put secrets, tokens, .env files, or machine binding data in the baseline.

Synchronization workflow

1. Edit and commit the personal baseline with normal Git tooling.
2. Call config_publish() from the local gateway. It requires a clean named branch and never force-pushes.
3. On every workstation, call config_sync(). It fetches, fast-forwards only, verifies the committed tree, creates an immutable local release, then changes the active pointer atomically.
4. Call config_status() to inspect freshness, the active release ID, and the last sync error without exposing baseline paths or binding values.
5. Before consequential work, call explain_effective_config with the intended absolute workingDir or registered workspace. This is read-only scope inspection: it reports release, scope, provenance, and effective preferences, never instruction text or local paths.