sync-configs（同步配置）

description: Sync Claude Code configuration to Codex CLI and Gemini CLI

SYNC-CONFIGS

Synchronize your AI-assisted development workflow configuration from Claude Code (canonical source) to Codex CLI and Gemini CLI.

Your Mission

Read Claude Code configuration and generate adapted versions for Codex and Gemini, ensuring consistent philosophy and workflow while respecting each tool's unique strengths.

Process

1. Audit Current State

First, examine what exists:

# Claude Code (source)
ls -la ~/.claude/CLAUDE.md
ls ~/.claude/commands/*.md | wc -l
ls ~/.claude/skills/*/SKILL.md | wc -l

# Codex CLI (target)
cat ~/.codex/AGENTS.md | wc -l
ls ~/.codex/prompts/*.md 2>/dev/null | wc -l

# Gemini CLI (target)
cat ~/.gemini/GEMINI.md | wc -l
ls ~/.gemini/commands/*.toml 2>/dev/null | wc -l

2. Generate AGENTS.md for Codex

Adapt the core philosophy from CLAUDE.md into AGENTS.md format. Key adaptations:

Structure:

• AGENTS.md should be concise but complete
• Include Ousterhout framework
• Include persona definitions
• Include red flags checklist
• Reference Codex-specific tools (rg, ast-grep)

Content to port:

• Software Design Philosophy section
• Complexity Management principles
• Module Design guidance
• Red flags (Manager/Util/Helper, pass-throughs, etc.)
• Persona vocabulary (Carmack, Jobs, Torvalds, Hara, Ousterhout)

Write to: ~/.codex/AGENTS.md

3. Generate GEMINI.md for Gemini

Create a comprehensive GEMINI.md (currently empty). Key adaptations:

Structure:

• Similar depth to CLAUDE.md
• Leverage Gemini strengths: shell interpolation !{cmd}, multimodal, Google grounding
• Include full Ousterhout framework
• Include persona definitions

Content to port:

• Full Software Design Philosophy
• Complexity Management
• Module Design
• Red flags checklist
• Persona vocabulary
• Tool usage guidance (adapted for Gemini capabilities)

Write to: ~/.gemini/GEMINI.md

4. Sync Agents (New: 15 agents)

Sync all 15 agents (8 specialists + 7 personas) to Codex and Gemini:

For Codex (~/.codex/agents/):

• Create .md files for each agent
• Keep full agent personality, philosophy, and checklists
• Adapt tool references (use rg, ast-grep instead of Claude-specific tools)
• Full sync - preserve all content, examples, red flags

For Gemini (~/.gemini/system-instructions/):

• Create .txt files for each agent
• Convert to natural language system instructions
• Full personality transfer - all quotes, philosophy, checklists
• Adapt format but preserve depth (multi-paragraph system instructions work great)

Agents to sync:

*Domain Specialists:*

1. complexity-archaeologist.md → agents/complexity-archaeologist.md + system-instructions/complexity-archaeologist.txt
2. data-integrity-guardian.md → agents/data-integrity-guardian.md + system-instructions/data-integrity-guardian.txt
3. api-design-specialist.md → agents/api-design-specialist.md + system-instructions/api-design-specialist.txt
4. test-strategy-architect.md → agents/test-strategy-architect.md + system-instructions/test-strategy-architect.txt
5. error-handling-specialist.md → agents/error-handling-specialist.md + system-instructions/error-handling-specialist.txt
6. state-management-analyst.md → agents/state-management-analyst.md + system-instructions/state-management-analyst.txt
7. dependency-health-monitor.md → agents/dependency-health-monitor.md + system-instructions/dependency-health-monitor.txt
8. documentation-quality-reviewer.md → agents/documentation-quality-reviewer.md + system-instructions/documentation-quality-reviewer.txt
9. infrastructure-guardian.md → agents/infrastructure-guardian.md + system-instructions/infrastructure-guardian.txt

*Master Personas:* 10. grug.md → agents/grug.md + system-instructions/grug.txt 11. carmack.md → agents/carmack.md + system-instructions/carmack.txt 12. jobs.md → agents/jobs.md + system-instructions/jobs.txt 13. torvalds.md → agents/torvalds.md + system-instructions/torvalds.txt 14. ousterhout.md → agents/ousterhout.md + system-instructions/ousterhout.txt 15. fowler.md → agents/fowler.md + system-instructions/fowler.txt 16. beck.md → agents/beck.md + system-instructions/beck.txt

Codex format (.md - keep full content):

# Agent Name

Philosophy quote...

## Core Concepts

[Full agent content preserved]

## Review Checklist

- [ ] Check 1
- [ ] Check 2
...

Gemini format (.txt - natural language system instruction):

You are [Agent Name]. [Philosophy quote]

Your role: [description]

When reviewing code, you:
- [Checklist item 1]
- [Checklist item 2]
...

[Full philosophy section in narrative form]

Red flags to watch for:
- [Flag 1]
- [Flag 2]
...

[All wisdom quotes and examples in natural language]

CRITICAL: Both formats get 100% of content. Format changes, depth stays identical.

5. Port Core Commands

Identify commands that should exist in all three tools:

Core workflow (always sync):

• prime, spec, plan, execute, ship
• ultrathink, carmack, aesthetic
• quality-check, triage, observe
• flesh, architect, debug

Claude-specific (don't sync):

• Commands that heavily use subagents
• Commands that depend on Claude-specific skills

Adaptation rules:

For Codex (prompts/*.md):

• Keep markdown format
• Add YAML frontmatter with name, description, aliases, enabled
• Reference Codex tools (rg, ast-grep, gemini CLI for research)
• Remove Claude-specific skill references

For Gemini (commands/*.toml):

• Convert to TOML format
• Use shell interpolation: !{cat TODO.md}, !{ls -F}
• Use {{args}} for arguments
• Keep prompts concise but complete

6. Apply Sync Policy

SYNC (must be consistent across all 3 CLIs):

• Ousterhout principles as default lens
• Red flags checklist
• All 15 agent personalities (8 specialists + 7 personas)
• Persona definitions and philosophies
• Core workflow commands
• Commit/PR standards
• Testing philosophy
• Code review checklists

DIVERGE (preserve tool strengths):

• Claude: Full subagent ecosystem via Task tool, skill library, MCP integrations, parallel agent composition
• Codex: Reasoning effort settings, simpler execution model, direct agent .md files
• Gemini: Shell interpolation !{cmd}, multimodal analysis, Google grounding, system instruction .txt files

7. Generate Sync Report

Output a clear report:

## Sync Report

### Base Configuration
- AGENTS.md: [UPDATED/CREATED/UNCHANGED] - [line count] lines
- GEMINI.md: [UPDATED/CREATED/UNCHANGED] - [line count] lines

### Agents Synced (15 total)

| Agent | Codex (.md) | Gemini (.txt) | Notes |
|-------|-------------|---------------|-------|
| complexity-archaeologist | ✅ | ✅ | |
| data-integrity-guardian | ✅ | ✅ | |
| api-design-specialist | ✅ | ✅ | |
| test-strategy-architect | ✅ | ✅ | |
| error-handling-specialist | ✅ | ✅ | |
| state-management-analyst | ✅ | ✅ | |
| dependency-health-monitor | ✅ | ✅ | |
| documentation-quality-reviewer | ✅ | ✅ | |
| infrastructure-guardian | ✅ | ✅ | |
| grug | ✅ | ✅ | Full personality preserved |
| carmack | ✅ | ✅ | Full personality preserved |
| jobs | ✅ | ✅ | Full personality preserved |
| torvalds | ✅ | ✅ | Full personality preserved |
| ousterhout | ✅ | ✅ | Full personality preserved |
| fowler | ✅ | ✅ | Full personality preserved |
| beck | ✅ | ✅ | Full personality preserved |

### Commands Synced

| Command | Codex | Gemini | Notes |
|---------|-------|--------|-------|
| ultrathink | ✅ | ✅ | |
| execute | ✅ | ✅ | Carmack + Ousterhout composition preserved |
| plan | ✅ | ✅ | Grug complexity review preserved |
| simplify | ✅ | ✅ | 4-agent composition preserved |
| debug | ✅ | ✅ | Specialist routing preserved |
| spec | ✅ | ✅ | Jobs + domain experts preserved |
| groom | ✅ | ✅ | 15-agent parallel audit preserved |
| ... | | | |

### Commands Skipped (Claude-specific)
- [command]: [reason - e.g., depends on Task tool subagent spawning]

### Manual Attention Needed
- [any issues or warnings]

### Next Steps
1. Review generated files in ~/.codex/agents/ and ~/.gemini/system-instructions/
2. Test agent invocation in Codex and Gemini
3. Test commands in each tool
4. Run `/sync-configs` again after adding new agents or modifying commands

Adaptation Guidelines

Philosophy Adaptation

When porting philosophy content:

1. Preserve the core - Ousterhout principles, complexity focus, red flags
2. Adapt the examples - Use tool-appropriate references
3. Maintain the tone - Concise, direct, zero fluff
4. Include personas - Define Carmack/Jobs/Torvalds/Hara/Ousterhout vocabulary

Command Adaptation

When porting commands:

1. Preserve ALL richness - Same personas, examples, red flags, philosophy, output formats
2. Adapt the mechanics - Shell interpolation for Gemini, tool refs for Codex
3. NEVER strip content - Gemini TOML should have identical depth to Claude md (only format changes)
4. Test the output - Verify commands work in target tool

CRITICAL: The Gemini TOML format can hold just as much content as Claude markdown. Converting format does NOT mean reducing content. A 400-line Claude command should become a 400-line Gemini command. Every persona quote, every red flag, every example must transfer.

Anti-pattern to AVOID: Stripping rich Claude prompts down to bare-bones Gemini commands. This destroys the prompt engineering that makes commands effective.

TOML Format for Gemini

CORRECT (Rich, complete prompt engineering):

description = "Deep critical evaluation of plans for simplicity and system health"
prompt = """
# ULTRATHINK

> **THE MASTERS OF SIMPLICITY**
>
> **Steve Jobs**: "Simple can be harder than complex..."
> **John Ousterhout**: "The most fundamental problem in computer science is problem decomposition..."
> **John Carmack**: "It's done when it's right..."

You're an IQ 155 principal architect who's seen 50+ systems collapse under their own complexity...

## Your Mission
[Full mission with context question]

## The Ousterhout Framework

### 1. Complexity Analysis
[Full framework with bullet points, red flags]

### 2. Module Depth Evaluation
[Full evaluation criteria with formula, examples]

[... ALL sections from Claude version ...]

## Red Flags Checklist
- [ ] Shallow modules
- [ ] Information leakage
[... complete list ...]

## Output Format
[Full structured output template]

## Philosophy
[Rich closing philosophy section]
"""

WRONG (Stripped down, lost all value):

description = "Evaluate plans"
prompt = """
# ULTRATHINK

> Simple is better

Review the plan for issues.

## Process
1. Check complexity
2. Find problems
3. Report findings
"""

The first example preserves all prompt engineering. The second destroys it. Always use the first pattern.

When to Run

• After adding/modifying agents in ~/.claude/agents/
• After significant changes to Claude Code commands
• After adding new skills you want everywhere
• After updating CLAUDE.md philosophy sections
• After modifying agent personalities or review checklists
• Periodically to catch drift (monthly)

Example Sync

Running /sync-configs should:

1. Read ~/.claude/CLAUDE.md → generate adapted AGENTS.md and GEMINI.md
2. Read all 15 agents from ~/.claude/agents/*.md → generate:

- ~/.codex/agents/*.md (full content, adapted tool refs) - ~/.gemini/system-instructions/*.txt (natural language format)

1. Read ~/.claude/commands/ultrathink.md → generate:

- ~/.codex/prompts/ultrathink.md (adapted) - ~/.gemini/commands/ultrathink.toml (converted)

1. Read ~/.claude/commands/groom.md → generate:

- ~/.codex/prompts/groom.md (15-agent composition adapted) - ~/.gemini/commands/groom.toml (15-agent invocation patterns converted)

1. Report what changed (agents, commands, configs)

The goal: run codex or gemini and get the same quality bar, same philosophy, same 15-agent perspectives, same workflow - just adapted for each tool's strengths.