بنقرة واحدة
use-codex
Run Codex CLI as a subagent with GPT-5.6 at low or medium effort.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
القائمة
Run Codex CLI as a subagent with GPT-5.6 at low or medium effort.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
Codex CLI delegation model floors, sandboxes, and acceptance gates.
Cursor delegation model floors, modes, and acceptance gates.
Parent-only orchestration: plan, route slices by capability floor, keep truth-judgment; never read or write application code.
Use the generic delegator workflow with the Cursor model mapping. Use when the user wants Cursor-based delegation, delegator-cursor, or the delegator workflow backed by Cursor subagents.
Use the generic delegator workflow with the Codex model mapping. Use when the user wants Codex-based delegation, delegator-codex, or the delegator workflow backed by Codex CLI/subagents.
Certainty-scored pull request reviews with reasoning and final ratings. Use when reviewing a PR, branch diff, or when the user asks for a PR review.
استنادا إلى تصنيف SOC المهني
| name | use-codex |
| description | Run Codex CLI as a subagent with GPT-5.6 at low or medium effort. |
| disable-model-invocation | true |
You have access to the Codex CLI, as a cheaper source of GPT-5.6 calls.
Available models:
gpt-5.6-sol: biggest model, expensive but powerfulgpt-5.6-terra: default; balanced for everyday workgpt-5.6-luna: fast and affordablegpt-5.4-mini: smallest and cheapest for simple tasksTreat Codex as a subagent: hand it a prompt, wait for the result, report back.
Pick model, effort, and sandbox. Use the default model unless the task calls for another listed model. Effort is low or medium (default medium). Sandbox is read-only (default, review), workspace-write (edits), or danger-full-access (network/broad).
Done when: model, effort, and sandbox are chosen.
Run codex exec synchronously using the command shape below. Repo-exploration tasks at medium routinely take 3–7 minutes; set the shell's block/wait time above the expected runtime (e.g. 540s for a task that reads code and writes a doc). Background only if required; then timeout is 150s (low) or 600s (medium).
Done when: the process exits and stdout is captured. Killing early yields empty output — never treat that as success.
Report. Summarize stdout. On non-zero exit, stop and report; do not retry without direction. Mention the session can be resumed. Done when: the user has the outcome (summary on success, failure report on error).
codex exec \
--skip-git-repo-check \
-m <model> \
--config model_reasoning_effort="<low|medium>" \
--sandbox <read-only|workspace-write|danger-full-access> \
[-C <DIR>] \
"prompt" \
</dev/null 2>/tmp/codex-stderr.log
Always: --skip-git-repo-check, effort low|medium, </dev/null (stdin must be closed or Codex hangs), 2><file> (hides thinking tokens from output but keeps them diagnosable — on failure, tail the log instead of flying blind; 2>/dev/null is acceptable only for throwaway calls).
failed to initialize in-process app-server client: Operation not permitted (exit 1, empty stdout). Requesting full_network is NOT enough. Always run the Shell tool with required_permissions: ["all"] on the first try.--sandbox workspace-write, network is still blocked inside Codex, so Gradle/npm downloads and build daemons fail (SocketException: Operation not permitted). Don't ask Codex to verify with networked builds or tests unless using danger-full-access; run the verification yourself afterward.test -s <output-file>) rather than trusting exit code alone.Inherits model, effort, and sandbox — do not re-pass config unless the user asks:
echo "follow-up" | codex exec --skip-git-repo-check resume --last 2>/dev/null
Any flags go between exec and resume.