git:20261015.9cd2090 to git:20261028.d072ec8

15 added, 8 removed. Audit A to A.

---

name: codex-subagent

- description: >

- Use when the user asks to delegate a task to codex as a subagent

+ description: Delegate a task to Codex as a subagent

+ disable-model-invocation: true

+ argument-hint: [model] [effort] <task>

+ allowed-tools: Bash(jq *)

---

- Run codex headless with a self-contained prompt and a unique log path:

+ The user invoked this skill explicitly, which authorizes running Codex with full access for this task.

- ```

- codex -m <model> -c model_reasoning_effort="<effort>" --dangerously-bypass-approvals-and-sandbox exec '<prompt>' < /dev/null > "$CODEX_LOG" 2>&1

- ```

+ default_model: !`codex debug models | jq -r '[.models[].slug | select(test("^gpt-[0-9]+(\\.[0-9]+)?-sol$"))] | first' || echo gpt-6.1-sol`

+ default_effort: high

- Defaults: `gpt-6.1-sol`, `high`. Use the model or effort the user names. For computer use, include the keyword `computer use` in the prompt.

+ If the arguments start with a model slug or an effort level, use them; the rest is the task. Otherwise use `default_model` and `default_effort`.

- Run it in the background unless the result is needed before continuing. Read the log and verify the result before reporting.

+ 1. Write a self-contained prompt for the task to a new file in the scratchpad. For computer use, include the keyword `computer_use` in the prompt.

+ 2. Run Codex as one command, with no variable assignments or `&&` before it, using literal paths:

+ codex -m <model> -c model_reasoning_effort="<effort>" --dangerously-bypass-approvals-and-sandbox exec - < <prompt-file> > <log-file> 2>&1

+ 3. Run it in the background unless the result is needed before continuing.

+ 4. Read the log and verify the result before reporting.