Codex Subagent

Hand off a task to Codex CLI for autonomous execution. Codex is a capable coding agent that can implement features, fix bugs, refactor code, investigate issues, and review changes.

Session Info

Session ID: ${CLAUDE_SESSION_ID} Output directory:

~/.claude/codex/${CLAUDE_SESSION_ID}/

Output files:

• progress-{timestamp}.jsonl

  - Streaming JSONL events (for monitoring)

• summary-{timestamp}.txt

  - Final agent message only (for results)

Parse Arguments

Arguments: $ARGUMENTS

Parse the arguments:

• Everything before flags is the task description

• --model <model>

  or

  -m <model>

  : Override model (only if explicitly requested)

• --sandbox <mode>

  : Override sandbox mode (only if explicitly requested)

Available models (only pass if user explicitly requests):

• gpt-5.2-codex

  - Default for Codex CLI, optimized for agentic coding (user's current config)

• gpt-5.2

  - Flagship model, best for complex professional tasks

• gpt-5-mini

  - Fast, cost-efficient (replaces o4-mini)

• o3

  - Deep reasoning model for complex multi-step problems

Sandbox modes (only pass if user explicitly requests):

• read-only

  - No file modifications allowed

• workspace-write

  - Can modify files in workspace

• danger-full-access

  - Full system access (use with caution)

Default behavior: Omit

--model

and

--sandbox

flags to use the user's

~/.codex/config.toml

preferences.

If no task description provided, ask the user what they want Codex to do.

Assess Task Complexity

Based on the task description, determine context depth:

Minimal context (single file fix, simple query):

• Task mentions specific file or function
• Clear, scoped objective
• Example: "fix the type error in auth.ts", "add logging to handleRequest"

Medium context (feature work, multi-file change):

• Task involves a feature or component
• May touch multiple files
• Example: "add input validation to the API", "refactor error handling"

Full context (architectural, investigation, unclear scope):

• Task is exploratory or investigative
• Scope unclear or potentially large
• Example: "figure out why tests are flaky", "improve performance"

Gather Context

Git State (always include)

pwd
git rev-parse --show-toplevel 2>/dev/null || echo "Not a git repo"
git branch --show-current 2>/dev/null || echo "N/A"
git status --short 2>/dev/null | head -20

For Medium/Full Context

git diff --stat 2>/dev/null | tail -20
git log --oneline -5 --since="4 hours ago" 2>/dev/null || echo "No recent commits"

For Full Context Only

Summarize relevant session context:

• What was being worked on
• Key decisions or approaches discussed
• Files that were read or modified
• Any blockers or open questions

Generate Codex Prompt

Create a prompt using CTCO structure (Context → Task → Constraints → Output format) optimized for GPT-5.2.

Structure:

<context>
Working directory: {cwd}
Repository: {repo_name}
Branch: {branch}

{git_status if relevant}
{recent_changes if medium/full context}
{session_summary if full context}
</context>

<task>
{task description from arguments}
</task>

<constraints>
- Implement EXACTLY and ONLY what is requested
- No extra features, refactoring, or "improvements" beyond the task
- Read relevant files before making changes
- Run tests/linters if available to validate changes
- If task is ambiguous, state your interpretation before proceeding
</constraints>

<output>
After completing the task, provide a structured summary (≤5 bullets):
- **What changed**: Files modified and nature of changes
- **Where**: Specific locations (file:line when relevant)
- **Validation**: Tests run, linters passed, manual verification
- **Risks**: Any potential issues or edge cases to watch
- **Next steps**: Follow-up work if any (or "None")
</output>

Execute Codex

First, ensure output directory exists:

mkdir -p ~/.claude/codex/${CLAUDE_SESSION_ID}

Check if we're in a git repo:

git rev-parse --show-toplevel 2>/dev/null && echo "IN_GIT_REPO" || echo "NOT_GIT_REPO"

Build and run the command:

codex exec --json \
  -o ~/.claude/codex/${CLAUDE_SESSION_ID}/summary-{timestamp}.txt \
  {required_flags} \
  {optional_flags} \
  - <<'CODEX_PROMPT'
{generated_prompt}
CODEX_PROMPT > ~/.claude/codex/${CLAUDE_SESSION_ID}/progress-{timestamp}.jsonl

Output handling:

• --json

  streams progress events to stdout → redirected to

  progress-{timestamp}.jsonl

• -o

  writes only the final message →

  summary-{timestamp}.txt

Flag rules:

• If NOT in a git repo: add

  --skip-git-repo-check

  (required)
• Default: add

  --full-auto

  (enables workspace-write sandbox + auto-approval)
• If user requests read-only: use

  --sandbox read-only

  instead of

  --full-auto

• If user requests specific model: add

  -m <model>

• If user requests danger-full-access: use

  --sandbox danger-full-access

  instead of

  --full-auto

Run via Bash tool.

Background vs Foreground

Always background tasks that might take >30 seconds:

• Any task touching multiple files
• Investigation/debugging tasks
• Tasks requiring test runs
• Feature implementations

Use

run_in_background: true

in the Bash tool call.

After backgrounding:

• Inform the user the task is running
• Do NOT immediately check output
• Wait for user to ask about status, OR continue with other work
• When checking, use the token-efficient methods below

Foreground only for trivial tasks (<30 seconds expected):

• Single-line fixes
• Simple file reads
• Quick queries

Monitoring Execution

Two files are created:

• progress-*.jsonl

  - Streaming JSONL (verbose, for progress checking)

• summary-*.txt

  - Final message only (clean, for results)

Token-efficient monitoring (CRITICAL):

# Check if still running (line count growing = active)
wc -l < ~/.claude/codex/${CLAUDE_SESSION_ID}/progress-*.jsonl

# Quick progress check - last 3 events only
tail -n 3 ~/.claude/codex/${CLAUDE_SESSION_ID}/progress-*.jsonl

# Check if summary exists (means Codex finished)
ls ~/.claude/codex/${CLAUDE_SESSION_ID}/summary-*.txt 2>/dev/null

Do NOT:

• Read the entire progress file
• Use

  tail -f

  (streams indefinitely, wastes context)
• Check more than once per 30 seconds for long tasks

When checking on background tasks:

1. First: check if summary file exists (finished?)
2. If not finished:

   wc -l

   on progress file to confirm activity
3. If needed:

   tail -n 3

   on progress for current status

Return Result

When Codex completes:

1. Read the summary file (already contains only the final message):

   cat ~/.claude/codex/${CLAUDE_SESSION_ID}/summary-*.txt
   

2. Parse Codex's structured summary
3. Report concisely to the user (3-6 sentences or ≤5 bullets)

Format:

## Codex Result

**Status:** {success/error/partial}

**What changed:**
- {file1}: {change summary}
- {file2}: {change summary}

**Validation:** {tests/linters run, results}

**Risks/Notes:** {if any, otherwise omit}

**Next steps:** {if any, otherwise "None"}

Keep the summary concise. If Codex produced verbose output, distill to essentials.

Examples

Simple fix

/codex fix the null pointer in utils/parser.ts line 42

Minimal context, quick execution.

Feature implementation

/codex add rate limiting to the /api/submit endpoint

Medium context, may take a few minutes.

Investigation

/codex investigate why the CI build fails on arm64

Full context, potentially long-running, consider backgrounding.

With model override (only when explicitly requested)

/codex --model o3 design a caching strategy for the database queries

Read-only mode (for review/analysis tasks)

/codex --sandbox read-only review the authentication implementation

Skips

--full-auto

, uses read-only sandbox for safe exploration.