Codex

Codex Subagent

Session ID: ${CLAUDE_SESSION_ID} Output: ~/.claude/codex/${CLAUDE_SESSION_ID}/

Arguments

Task: $ARGUMENTS

Optional flags (only if user explicitly requests):

• --model <model>: gpt-5.2-codex (default), gpt-5.2, gpt-5-mini, o3
• --sandbox <mode>: read-only, workspace-write, danger-full-access

Omit flags to use user's config defaults.

Context Depth

Minimal: Specific file/function → git state only Medium: Feature/multi-file → add recent changes Full: Investigation/unclear → add session summary

Gather Context

Always:

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

Medium/Full:

git diff --stat 2>/dev/null | tail -20
git log --oneline -5 --since="4 hours ago" 2>/dev/null

Full only: Session summary (work done, decisions, blockers)

Prompt Structure

Use CTCO format (Context → Task → Constraints → Output):

<context>
Working directory: {cwd}
Repository: {repo_name}
Branch: {branch}
{git_status}
{recent_changes if medium/full}
{session_summary if full}
</context>

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

<constraints>
- Implement EXACTLY what is requested
- Read files before changing
- Run tests/linters to validate
- State interpretation if ambiguous
</constraints>

<output>
Summary (≤5 bullets):
- **What changed**: Files and changes
- **Where**: file:line references
- **Validation**: Tests/linters run
- **Risks**: Edge cases to watch
- **Next steps**: Follow-up or "None"
</output>

Execute

Setup:

mkdir -p ~/.claude/codex/${CLAUDE_SESSION_ID}
git rev-parse --show-toplevel 2>/dev/null && IN_GIT=true || IN_GIT=false

Run:

codex exec --json \
  -o ~/.claude/codex/${CLAUDE_SESSION_ID}/summary-{timestamp}.txt \
  {--skip-git-repo-check if not in git} \
  {--full-auto OR --sandbox <mode>} \
  {-m <model> if requested} \
  - <<'CODEX_PROMPT'
{prompt}
CODEX_PROMPT > ~/.claude/codex/${CLAUDE_SESSION_ID}/progress-{timestamp}.jsonl

Flags:

• Not in git: add --skip-git-repo-check
• Default: --full-auto (workspace-write + auto-approval)
• User-requested: --sandbox <mode> or -m <model>

Background vs Foreground

Background tasks (>30 seconds expected):

• Multi-file changes, investigations, tests, feature work
• Use run_in_background: true → returns task_id

Foreground tasks (<30 seconds):

• Single-line fixes, simple queries

Monitoring Background Tasks

Token-efficient approach:

1. Use TaskOutput(task_id, block=true) to wait for completion
2. Ignore TaskOutput's content (stdout redirected to progress file)
3. Read summary file directly: cat ~/.claude/codex/${CLAUDE_SESSION_ID}/summary-*.txt

The summary file contains only Codex's final message (token-efficient).

Progress checks (if needed before completion):

• TaskOutput(task_id, block=false) - check if still running
• tail -n 3 ~/.claude/codex/${CLAUDE_SESSION_ID}/progress-*.jsonl - last 3 events only

Do NOT read entire progress files or use tail -f.

Report Result

Read summary:

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

Report format (≤5 bullets):

**Status:** {success/error/partial}
**Changed:** {files and changes}
**Validation:** {tests/linters}
**Risks:** {if any}
**Next:** {follow-up or "None"}

Examples

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

# Feature work
/codex add rate limiting to the /api/submit endpoint

# Investigation (background)
/codex investigate why the CI build fails on arm64

# Model override
/codex --model o3 design a caching strategy

# Read-only
/codex --sandbox read-only review the auth implementation