ln-004-agent-config-sync（ln 004 Agent 配置同步）

Paths: File paths (shared/, references/, ../ln-*) are relative to skills repo root. If not found at CWD, locate this SKILL.md directory and go up one level for repo root.

Agent Sync (Standalone Utility)

Type: Standalone Utility Category: 0XX Shared

Synchronizes skills and MCP server configurations from Claude Code (source of truth) to Gemini CLI and Codex CLI. Creates symlinks for skills, copies/converts MCP settings.

When to Use This Skill

• After adding/removing MCP servers in Claude Code settings
• After installing new plugins in Claude Code
• First-time setup of Gemini CLI or Codex CLI alongside Claude Code
• Periodic sync to keep all agents aligned

Input Parameters

Workflow

Detect OS → Discover Configs → Sync Skills → Sync MCP → Report

Phase 0: OS Detection

Paths by OS:

Phase 1: Discover Current State

1. Read Claude settings (check both sources, merge):

- Parse ~/.claude.json → extract mcpServers block (primary source, app state file) - Parse ~/.claude/settings.json → extract mcpServers block (fallback, user settings file) - Merge: primary overrides fallback by server name - If no mcpServers in either → WARN "No MCP servers configured in Claude", skip MCP sync

1. Read target configs (if exist):

- Gemini: Parse ~/.gemini/settings.json → extract existing mcpServers - Codex: Parse ~/.codex/config.toml → extract existing [mcp_servers.*]

1. Detect installed plugins:

- Glob ~/.claude/plugins/*/plugin.json → list plugin directories - Also check if skills repo itself is a plugin source

1. Check existing symlinks:

- ~/.gemini/skills → exists? points where? - ~/.codex/skills → exists? points where?

1. Show current state: Current State: | Agent | Skills | MCP Servers | Config Exists | |-------|--------|-------------|---------------| | Claude (source) | 5 plugins | 4 servers | yes | | Gemini | no link | 2 servers | yes | | Codex | → ~/.claude/plugins | 4 servers | yes |

Phase 2: Sync Skills (symlinks/junctions)

FOR EACH target IN (gemini, codex) WHERE target in targets:

1. Determine link path:

- Gemini: ~/.gemini/skills - Codex: ~/.codex/skills

1. Check if already linked correctly:

- IF link exists AND points to correct source → SKIP (already synced) - IF link exists AND points to wrong source → WARN, ask user before replacing - IF regular directory (not link) exists → WARN "Target is a real directory, not a link. Skip to avoid data loss."

1. Determine source:

- IF single plugin: link directly to plugin dir - IF multiple plugins: ask user which plugin to share

1. Create link: OS Command Windows cmd /c mklink /J "{target_path}" "{source_path}" macOS / Linux ln -s "{source_path}" "{target_path}"
2. Verify: Check link exists and resolves correctly

Phase 3: Sync MCP Settings

Source: ~/.claude.json (primary) + ~/.claude/settings.json (fallback) → merged mcpServers

3a: Claude → Gemini (JSON → JSON)

1. Read Claude mcpServers as JSON object
2. Read Gemini settings.json (or create {} if missing)
3. Merge strategy: Claude servers override Gemini servers by key name. Gemini-only servers preserved.
4. Write updated settings.json

Conversion: Transport type mapping required:

Gemini-only fields (not mapped from Claude):

• timeout — request timeout in ms (default: 10min)
• trust — bypass tool confirmations
• includeTools / excludeTools — tool whitelist/blacklist

3b: Claude → Codex (JSON → TOML)

1. Read Claude mcpServers as JSON object
2. Read Codex config.toml (or create empty if missing)
3. Convert each server: Claude JSON field Codex TOML field Notes command command Same args args JSON array → TOML array env [mcp_servers.{name}.env] Nested table type: "http" + url url Streamable HTTP (Codex auto-detects by url presence) type: "sse" + url url SSE → Streamable HTTP (same url field) headers http_headers Different key name! Static header values Example conversion: "context7": {"command": "npx", "args": ["-y", "@upstash/context7-mcp"]} [mcp_servers.context7] command = "npx" args = ["-y", "@upstash/context7-mcp"]
4. Merge strategy: Claude servers override. Codex-only servers preserved.
5. Write updated config.toml

Codex-only fields (preserve as-is, no Claude equivalent):

• bearer_token_env_var — Codex bearer token auth
• enabled_tools / disabled_tools — Codex tool filtering
• startup_timeout_sec / tool_timeout_sec — Codex timeouts
• enabled / required — Codex server lifecycle
• env_vars / env_http_headers — Codex env-based headers

Phase 4: Report

Sync Complete:
| Action | Target | Status |
|--------|--------|--------|
| Skills symlink | Gemini | Created → ~/.claude/plugins/... |
| Skills symlink | Codex | Already linked |
| MCP sync | Gemini | 4 servers synced (2 new, 2 updated) |
| MCP sync | Codex | 4 servers synced (1 new, 3 updated) |

Critical Rules

1. Claude = source of truth. Never write TO Claude settings. Read-only source.
2. Non-destructive merge. Target-only servers/settings preserved. Only Claude servers added/updated.
3. No data loss. If target is a real directory (not symlink) — warn and skip, never delete.
4. Backup before write. Before modifying any config file, create .bak copy.
5. Dry run first. If dry_run=true, show all planned actions without executing.
6. Ask on conflict. If symlink points to different source — ask user, don't auto-replace.

Anti-Patterns

• Writing TO Claude settings from Gemini/Codex (reverse sync)
• Deleting target-only MCP servers during sync
• Creating symlinks inside symlinks (circular)
• Modifying config files without backup

Definition of Done

• Claude settings read successfully
• Skills symlinks created/verified for each target
• MCP settings synced with format conversion (JSON→TOML for Codex)
• Backup files created before any config modification
• Report shown with all actions and warnings

Version: 1.1.0 Last Updated: 2026-03-09