ask-coding-cli

<!-- Example prompts: /ask-coding-cli Review my implementation plan /ask-coding-cli Ask codex whether this error handling is right /ask-coding-cli Get claude's take on this migration --> # Ask a Coding CLI Use the agent the user names. Otherwise pick one that is not the agent you are running in — Codex from Claude Code, Claude Code from Codex — since a second opinion from the same model adds little. Never start the CLI of the agent you are running in; that nests a whole session. For your own agent, use its native subagent instead (the Agent tool in Claude Code, a native subagent in Codex). Every path takes a self-contained prompt: the other agent sees none of this conversation, so whatever it cannot infer has to be in the prompt. Keep it focused — the relevant files, the decision at stake, and the constraints — rather than dumping the whole codebase. ## Codex Always run the latest luna model on the fast service tier — a small model gains most from the lower latency. `scripts/codex-model.nu` in this skill's directory resolves the newest slug of a family (default `luna`) from Codex's live catalog, so no model name is pinned here: ```sh codex exec --skip-git-repo-check --ephemeral \ -m "$(<skill-dir>/scripts/codex-model.nu)" -c service_tier=fast "<task>" ``` Inside Codex, a native subagent takes `agents.default_subagent_model` from `nix/modules/home/programs/codex/` instead. `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. ## Claude Code ```sh claude -p --model fable --permission-mode plan "<question>" ``` Claude Code has no model-list command; its aliases are the runtime resolution. `fable`, `opus`, `sonnet`, and `haiku` each map to the newest model of that tier, so pass an alias, never a full model ID. `--output-format json` reports the resolved ID under `modelUsage` when you need to confirm it. `claude -p` is billed per use, so reserve it for decisions worth a second opinion rather than grunt work. ## Grok ```sh grok -p "<question>" ``` Grok's default is already its latest model (`grok models` lists them), so pass no `-m`. For searching X, use the `web-fetch` skill instead; that is where Grok's native X index earns its place. ## Cursor ```sh cursor-agent -p --mode plan "<question>" ``` Leave the model on `auto`, Cursor's default, which routes to its current models; `cursor-agent models` lists the alternatives when the user names one. `--mode plan` (and `--permission-mode plan` for Claude Code) keeps a second opinion read-only; drop it only when the user hands the agent a task to carry out. ## Briefing and using the answer - 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. - For a second opinion, ask for an independent assessment rather than agreement with your plan. 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. Let the user make the final call.