ask-codex

<!-- Example prompts: /ask-codex Review my implementation plan /ask-codex Is this the right approach for error handling? --> # Ask Codex This skill is shared by Codex Desktop, Codex CLI, and other agents. Choose the execution path that matches the current agent: - **Inside Codex (Desktop or CLI):** delegate to a native subagent. Never run `codex exec` from here; that starts a nested Codex process. - **Outside Codex:** use the Codex CLI path below for an independent agent. Both paths take a self-contained prompt: the subagent sees none of this conversation, so whatever it cannot infer has to be in the prompt. ## Native subagent path Use one native subagent for either job. Ask it to return the requested facts with citations (`URL` or `file:line`), and ask for verbatim quotes where exact spelling matters. For a second opinion, ask for an independent assessment, then compare it with your own reading and preserve disagreements in the response. For delegated grunt work, ask for the small conclusion rather than forwarding the whole transcript. The repository's Codex configuration sets native subagents to `gpt-5.6-luna` with `max` reasoning and the `fast` service tier. Keep those defaults unless the user requests another model or effort. ## Codex CLI path Use this path only outside Codex: ```sh codex exec "<question>" ``` Leave the model alone — the reasoning effort in Codex's own config is what makes the answer worth having. Treat the reply as one data point: compare it against your own reading of the codebase, report both views to the user with the disagreements intact, and prefer established project patterns where they conflict. ## Delegated grunt work ```sh codex exec --skip-git-repo-check --ephemeral -m gpt-5.6-luna "<task>" ``` For work whose output is bulky but whose conclusion is small: web search, page reading, codebase investigation, summarising long output, drafting a commit message from a diff. A cheap model is the point — the subscription pays for it and only the answer costs tokens here. `codex exec --help` carries the rest; reach for `-o` to capture the final message alone, and `--sandbox`/`--add-dir` to widen what it may touch. Measured, and not visible in the help output: - Keep `model_reasoning_effort` at `medium` or above. At `low` it answered a "latest release version" search from stale knowledge. - Never `--disable multi_agent`. Native web search runs through it, so disabling it makes a search task hang indefinitely with no tool call. - Shell commands only reach the network under `--sandbox workspace-write` with `-c sandbox_workspace_write.network_access=true`. The native web search needs neither. - `exa` cannot run there at all: no sandbox mode reaches its 1Password key, so run exa in this session. `ax` and `tgrab` live in `~/.agents/skills/web-fetch/` — prepend that to PATH when the task needs them. Briefing it: - Say which facts to return, and demand a citation for each — a URL, or a `file:line`. - Ask for verbatim quotes wherever exact spelling matters: API options, flags, version numbers. A digest paraphrases them away. - The `web-fetch` skill holds the search tool selection; name the choice rather than leaving it open.