@rules/routing.md

Codex Skill

<output_language>

Default all user-facing deliverables, saved artifacts, reports, plans, generated docs, summaries, handoff notes, commit/message drafts, and validation notes to Korean, even when this canonical skill file is written in English.

Preserve source code identifiers, CLI commands, file paths, schema keys, JSON/YAML field names, API names, package names, proper nouns, and quoted source excerpts in their required or original language.

Use a different language only when the user explicitly requests it, an existing target artifact must stay in another language for consistency, or a machine-readable contract requires exact English tokens. If a localized template or reference exists (for example *.ko.md or *.ko.json), prefer it for user-facing artifacts.

</output_language>

Defaults

Parameter	Default
Model selection	Use Codex CLI default unless the user explicitly asks for -m / --model
Profile	Use CLI default unless the user explicitly asks for -p / --profile
Sandbox	--sandbox read-only for analysis, --sandbox workspace-write only when the user asks Codex to edit
Approval policy	Interactive mode only — use CLI default unless the user explicitly asks for -a / --ask-for-approval
Headless mode	codex exec
Resume target	codex exec resume --last for the latest non-interactive run, codex resume --last for the latest interactive session

Do NOT ask the user for model, profile, or approval policy unless explicitly requested.

Routing

Use this skill when the request actually needs Codex CLI or a separate Codex session.

• Read rules/routing.md before building a command when the request might be out of scope.
• Route away to direct editing or another skill when the user wants generic writing, documentation cleanup, or local edits without needing the codex CLI itself.

Examples

Positive triggers:

• "Use Codex to review this repository and summarize the risks."
• "Run codex exec and analyze this architecture."
• "Continue the last Codex session and ask it to finish the patch."
• "Run codex review against the base branch and list the blocking issues."

Negative triggers:

• "Rewrite this runbook for readability."
• "Create a new skill for our repo."

Boundary trigger:

• "Research Codex sandbox modes and tell me what they do." Use this skill only if the user wants the codex CLI involved; otherwise route to research or direct documentation work.

Critical: Non-Interactive Mode

Always use the codex exec subcommand for non-interactive Codex runs. Calling codex "prompt" without a subcommand starts the interactive TUI.

# Non-interactive
codex exec --sandbox read-only "your prompt here"

# Interactive TUI
codex "your prompt here"

codex exec reads the prompt from the argument or stdin and prints the result to stdout when finished.

Running a Task

Read references/recipes.md for concrete command recipes before changing sandbox mode, resuming a session, or adding extra directories.

Sandbox Mode Selection

Flag	When to use
--sandbox read-only	General analysis, review, planning, or structured output with no file writes
--sandbox workspace-write	Only when the user explicitly wants Codex to modify files in the workspace
--sandbox danger-full-access	Only after explicit approval, and only in isolated environments
--dangerously-bypass-approvals-and-sandbox	Only after explicit approval, and only in environments that are externally sandboxed

Approval Policy Selection (Interactive Mode Only)

-a / --ask-for-approval is only available on the top-level interactive codex command, not on codex exec. For non-interactive runs, control safety through --sandbox instead.

Flag	When to use
-a untrusted	Interactive only — Codex must escalate to the user for any non-trusted command
-a on-request	Interactive only — Codex asks for approval when it decides it needs to
-a never	Interactive only — Codex never asks; pair only with a restrictive --sandbox

When the user does not specify, omit -a and let the CLI default apply.

Command Discipline

• Start from codex exec --sandbox read-only "your prompt here".
• Add -m <model> or -p <profile> only when the user explicitly asks.
• Use --json only when the user wants machine-readable JSONL output.
• Use --output-schema <FILE> when the user wants the final response to follow a JSON schema.
• Use --add-dir <path> when the task needs files outside the launch directory; pair with a writable sandbox only when edits are explicitly requested.
• Use -C <dir> / --cd <dir> to set the working root when the user names a different directory.
• Ask before using --dangerously-bypass-approvals-and-sandbox.
• Prefer --sandbox workspace-write over bypassing approvals for normal file edits.

Resuming a Session

codex exec resume --last "continue the previous task"
codex exec resume <session-id> "continue with this follow-up"
codex resume --last       # interactive picker for the latest TUI session

Use codex exec resume --last for the latest non-interactive run. Use codex exec resume <session-id> when the user wants a specific session or the latest run is not the right one. When resuming, keep the existing session behavior unless the user explicitly asks to change the model, profile, sandbox, or approval policy. Use codex fork --last only when the user wants to branch from the existing session instead of reusing it.

Code Review

codex review --uncommitted "review the local changes and list blocking issues"
codex review --base main "review this branch against main and summarize risks"
codex review --commit <sha> "review only the changes in this commit"

codex review is read-only by design; do not pair it with --sandbox workspace-write or --dangerously-bypass-approvals-and-sandbox.

After Completion

• Summarize the result, including any warnings or partial output.
• Tell the user they can resume with codex exec resume --last or codex exec resume <session-id>.
• Ask whether to continue, adjust the prompt, or switch back to direct work.

Critical Evaluation

Treat Codex as a colleague, not an authority.

• Trust your own grounded knowledge when you are confident.
• Verify disagreements with current docs or primary sources before accepting a claim.
• Remember that a separate Codex session can still be wrong or stale.
• Let the user decide when there is genuine ambiguity.

Error Handling

• command not found: codex: tell the user OpenAI Codex CLI is not installed.
• Auth errors: use codex login or check the configured OpenAI credential path.
• Sandbox blocks: retry with the appropriate --sandbox only when the user wants that behavior.
• Approval blocks (interactive codex only): retry with a less restrictive -a only when the user wants that behavior.
• Session not found: run codex exec resume (without --last) to use the picker, or switch to codex resume --last for the latest interactive session.
• Invalid flag or model errors: check codex --help or codex exec --help and retry with supported options.
• Outside a Git repo: add --skip-git-repo-check only when the user has accepted that the run will operate without repo guardrails.