description: Delegate code tasks and analysis to Codex CLI agents allowed-tools: Bash, Read, Glob, Grep argument-hint: [task description]

Codex Agents - Simple Code Tasks & Analysis

Triggered by: c, c2, c3, c5 at end of command, or /codex_agents

This skill runs parallel Codex CLI calls for code tasks and analysis. Codex handles simple code work so Claude can focus on complex decisions.

STEP 1: Announce with Orange Indicator (REQUIRED)

You MUST announce activation before doing anything else:

🟠 3 Codex agents activated

• Codex 1: [what this agent will do]
• Codex 2: [what this agent will do]
• Codex 3: [what this agent will do]

Parse the shorthand:

• c or c3 → 3 Codex agents (default)
• c2 → 2 Codex agents
• c5 → 5 Codex agents

STEP 2: Determine Sandbox Mode

Choose the appropriate sandbox based on the task:

Default to read-only unless the task explicitly requires code changes.

STEP 3: Run Codex CLI Commands (REQUIRED)

You MUST actually run Codex CLI commands. This is the whole point.

Single Codex Agent (c or c1)

codex exec -s read-only -m gpt-5.3-codex -c reasoning_effort=xhigh "Analyze src/ for potential bugs. List issues by file with line numbers." -o /tmp/codex_out.txt 2>/dev/null && cat /tmp/codex_out.txt

Multiple Codex Agents (c2, c3, c5)

Run in parallel using background jobs:

# Agent 1: Code analysis
(codex exec -s read-only -m gpt-5.3-codex -c reasoning_effort=xhigh "Review src/auth/ for security issues. Be concise." -o /tmp/c1.txt 2>/dev/null) &

# Agent 2: Test coverage analysis
(codex exec -s read-only -m gpt-5.3-codex -c reasoning_effort=xhigh "Check test coverage gaps in src/. List untested functions." -o /tmp/c2.txt 2>/dev/null) &

# Agent 3: Code quality
(codex exec -s read-only -m gpt-5.3-codex -c reasoning_effort=xhigh "Find code smells and dead code in src/. Be specific." -o /tmp/c3.txt 2>/dev/null) &

wait

echo "=== Codex 1: Security ===" && cat /tmp/c1.txt
echo "=== Codex 2: Test Coverage ===" && cat /tmp/c2.txt
echo "=== Codex 3: Code Quality ===" && cat /tmp/c3.txt

Code Review (Built-in)

# Review uncommitted changes
codex review --uncommitted 2>/dev/null

# Review against a branch
codex review --base main 2>/dev/null

Code Changes (workspace-write)

# Simple bug fix
codex exec -s workspace-write --full-auto -m gpt-5.3-codex -c reasoning_effort=xhigh "Fix the null pointer exception in src/utils/parser.js line 42" 2>/dev/null

# Generate tests
codex exec -s workspace-write --full-auto -m gpt-5.3-codex -c reasoning_effort=xhigh "Write unit tests for src/auth/login.js" -o /tmp/c_result.txt 2>/dev/null

Automatic Model Fallback

codex_with_fallback() {
    local sandbox="$1"
    local prompt="$2"
    local outfile="$3"

    # Primary: gpt-5.3-codex with xhigh reasoning
    result=$(codex exec -s "$sandbox" -m gpt-5.3-codex -c reasoning_effort=xhigh "$prompt" -o "$outfile" 2>&1)

    # Fallback 1: o4-mini
    if echo "$result" | grep -qi "error\|rate.limit\|quota"; then
        result=$(codex exec -s "$sandbox" -m o4-mini "$prompt" -o "$outfile" 2>&1)
    fi

    # Fallback 2: gpt-4.1-mini
    if echo "$result" | grep -qi "error\|rate.limit\|quota"; then
        result=$(codex exec -s "$sandbox" -m gpt-4.1-mini "$prompt" -o "$outfile" 2>&1)
    fi

    # All models failed → signal for Claude takeover
    if [ ! -s "$outfile" ] || echo "$result" | grep -qi "error\|rate.limit\|quota"; then
        echo "AGENT_FAILED"
        return 1
    fi

    cat "$outfile" 2>/dev/null
}

# Usage:
codex_with_fallback "read-only" "Analyze this code for bugs" "/tmp/c1.txt"

CLI Syntax Reference

# Analysis (read-only)
codex exec -s read-only -m gpt-5.3-codex -c reasoning_effort=xhigh "prompt" -o /tmp/out.txt

# Code changes (write access)
codex exec -s workspace-write --full-auto -m gpt-5.3-codex -c reasoning_effort=xhigh "prompt"

# Code review
codex review --uncommitted

# Specific working directory
codex exec -s read-only -m gpt-5.3-codex -c reasoning_effort=xhigh -C /path/to/project "prompt" -o /tmp/out.txt

# Flags:
# -s = sandbox mode (read-only | workspace-write)
# -m = model
# -o = output last message to file
# --full-auto = skip approval prompts (sandboxed)
# -C = working directory
# --skip-git-repo-check = allow running outside git repos

STEP 4: Report Results (REQUIRED FORMAT)

Your report MUST include the Codex commands you ran:

🟠 Codex Analysis Complete

Commands ran:

• Codex 1: codex exec -s read-only -m gpt-5.3-codex -c reasoning_effort=xhigh "..." -o /tmp/c1.txt
• Codex 2: codex exec -s read-only -m gpt-5.3-codex -c reasoning_effort=xhigh "..." -o /tmp/c2.txt
• Codex 3: codex exec -s read-only -m gpt-5.3-codex -c reasoning_effort=xhigh "..." -o /tmp/c3.txt

Findings:

• [Finding 1 from Codex]
• [Finding 2 from Codex]
• [Finding 3 from Codex]

Summary: [Brief synthesis of what Codex found]

What Codex Is Good For

What Codex Cannot Do (Use Claude Instead)

• Complex multi-step architectural changes
• Final decisions on design approaches
• Work requiring judgment across multiple systems
• Deployment and infrastructure changes
• Tasks needing conversation context

Codex vs Gemini vs Claude

Rule of thumb:

• Text/data analysis → Gemini
• Code-specific tasks → Codex
• Complex judgment → Claude

Combining with Other Agents

When the user says something like fix bugs x3 c3:

1. First announce all:

✅ 1 Claude subagent activated

• Agent 1: [Claude task requiring judgment]

🔵 2 Gemini agents activated

• Gemini 1-2: [bulk analysis tasks]

🟠 3 Codex agents activated

• Codex 1-3: [code tasks]

1. Run Claude subagents for judgment/complex work
2. Run Gemini agents via CLI for bulk analysis
3. Run Codex agents via CLI for code tasks
4. Synthesize all results

Graceful Degradation

If all Codex models fail (output file empty or contains AGENT_FAILED):

⚡ Claude takes over — perform the code analysis yourself using native tools:

1. Announce: ⚡ Claude takeover — Codex agents failed, Claude devralıyor
2. Read the code files using Read, Grep, Glob tools
3. Analyze the code directly
4. Report findings with the same format as Codex would

Degradation chain:

gpt-5.3-codex (xhigh) → o4-mini → gpt-4.1-mini → ⚡ Claude takeover

Do NOT skip failed agents silently. Always announce which agents failed and that Claude is taking over.

Error Handling

If Codex fails:

1. Check if codex is in PATH → which codex
2. Check authentication → codex login
3. Try with a different model → -m o4-mini
4. Check sandbox permissions → try -s read-only first
5. For git issues → add --skip-git-repo-check

If response is empty or cut off:

• Prompt too vague → be more specific
• Project too large → point to specific directory with -C
• Try splitting into smaller tasks across more agents