askcodex

You have the askcodex CLI. It sends one request to an OpenAI model through the user's ChatGPT/Codex subscription and returns a text answer, an image, or a transcript. Turn the user's request into a complete brief, run one focused call, check the result against the brief, and hand the user the useful output with its saved path.

Canonical copy: skill/ in https://github.com/JairoTorregrosa/askcodex (this is a mirror; change it there, then re-sync). Local delta: the description's not-for clause points to second-opinion.

Context you must respect

• This skill describes askcodex 0.3.0 or later. Check askcodex --version once per session. 0.2.x has no --schema, --verbosity or --background, prints the ~700 KB raw catalog inside models --json, and gives errors without error.backend; 0.1.x also hides the GPT-6 models. If it is older, tell the user to update instead of working around it: in an askcodex checkout, git pull && ./install.sh; without one, download the latest release binary from https://github.com/JairoTorregrosa/askcodex/releases (verify SHA256SUMS), or clone it and run ./install.sh (needs Rust 1.88+).
• Every call is one-shot and stateless. ask sends only the prompt and optional --instructions: the model gets no tools, files, browsing, repository, or earlier conversation. A file path in a prompt is just text; paste the contents instead.
• Everything in a prompt, a reference image or an audio file leaves the machine for OpenAI's backend. Before you paste logs, exports, code or documents, always remove credentials, tokens, keys and passwords, even when the user supplied them for the task: replace them with placeholders. Also remove personal or customer data the task does not need; if the task needs it, or you cannot tell, ask the user first.
• Every call spends the user's subscription quota. GPT-6 Astra is the most expensive model.
• A second model's answer is input to your work. Verify its facts, review its code, and look at its images before you report success.
• askcodex uses the credentials codex login saved to ~/.codex/auth.json; it reads no API key. Never print credentials.

Facts already verified (2026-10-02)

• askcodex models --no-refresh lists what the account can use, with each model's description, Codex's default effort and any retirement date. On 2026-10-02: gpt-6.1-sol, gpt-6-astra, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-daybreak-blue-latest, gpt-5.5 (retires 2026-10-14), and the hidden gpt-reserve and codex-auto-review. All take text; ask sends text only. OpenAI ships models every few days: trust the live catalog over this list.
• ask defaults to gpt-6.1-sol at medium. The catalog lists it for every plan; if a plan refuses it with a model-availability error (OpenAI's Codex docs have offered Free and Go only Luna), pass --model gpt-6-luna. askcodex never switches models on its own.
• --effort takes low, medium, high, xhigh, max, or none. The catalog also lists ultra, which is Codex's multi-agent mode; the backend rejects it. A level a model does not take fails with HTTP 400 (none on gpt-6-astra and gpt-6.1-sol; gpt-6-sol and gpt-6-luna accept it although the catalog omits it).
• --verbosity low|medium|high sets answer length and detail. Unset, the backend answers at medium; Codex itself asks for low. On the same prompt low used 240 output tokens and high 513.
• --schema FILE makes the backend enforce a JSON Schema in strict mode, even against a prompt that asks for prose. Strict rules: every property in required, "additionalProperties": false on every object, and an optional value adds "null" to its own type (["string","null"], ["integer","null"]; an enum must list null too). A schema that breaks them fails before generation with error.backend.code: "invalid_json_schema" and a message naming the key. With --json, .result.json is the parsed answer. askcodex re-checks its structure against the schema, so an answer that does not parse or does not match is a response_invalid failure, never a result.
• image create and image edit return one PNG per call of about 1.57 megapixels; the prompt sets the aspect ratio. --background transparent or opaque forces the background (both honored); unset, the prompt decides. The result reports the file's real width, height and alpha_channel. There is no model, size, quality, format, or count flag, because the backend ignores those fields.
• transcribe takes a WAV file up to 25 MiB and has no model or language selector.

Do this

1. Pick the command and read its guide before you write the prompt:

   Result	Command	Read first
   Answer, analysis, writing, code, or review	askcodex ask "prompt" --model <slug> --effort <level> (add --verbosity low for a concise answer)	prompting-text.md and the model's guide below
   Classification or extraction you will parse	askcodex ask - --schema /tmp/askcodex/<name>.schema.json --json < brief.txt, then read .result.json	Same as above
   New image	askcodex image create "prompt" -o /tmp/askcodex/<name>-v1.png (add --background transparent for a cutout)	prompting-images.md and model-gpt-image-2.md
   Image edit from one to five PNG references	askcodex image edit "prompt" -i ref.png -o /tmp/askcodex/<name>-v2.png	Same as a new image
   Audio transcript	askcodex transcribe recording.wav	transcription.md
   Which models exist, their default effort, what retires	askcodex models --no-refresh (a short list with descriptions)	Nothing; no prompt is involved
   Quota, account, auth status	askcodex usage / whoami / auth status, each with --json --no-refresh	Nothing

2. For ask, pick the model and read its guide:

   The task	Model	Guide
   Most work: coding, review, debugging, analysis, drafting	gpt-6.1-sol (default)	model-gpt-6.1-sol.md
   The hardest or most consequential single deliverable	gpt-6-astra	model-gpt-6-astra.md
   Clear, repeatable extraction, classification, or transformation	gpt-6-luna	model-gpt-6-luna.md
   Authorized defensive security work	gpt-daybreak-blue-latest	model-gpt-daybreak-blue-latest.md
   Reproducing a GPT-6 Sol result	gpt-6-sol	model-gpt-6-sol.md
   Reproducing a GPT-5.6 result	gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna	sol, terra, luna
   Anything on gpt-5.5	move to gpt-6.1-sol (gpt-6-luna if the plan refuses it)	model-gpt-5.5.md

   prompting.md indexes every guide. Pin --model and --effort whenever you compare outputs or need to reproduce one.

3. Write a complete brief: the deliverable, its audience, the constraints, what "done" means, the source material itself, what must stay unchanged, and the output format.
4. Run one focused request. Create /tmp/askcodex/ first and give every artifact a descriptive name with a version suffix, so an edit never overwrites a useful result. For a long-running job, run it in the background or in tmux and save its output to a file.
5. Inspect the result against the brief. Make the next request about one specific observed problem, and keep the last useful version.

Handle the output

• Images: always pass -o /tmp/askcodex/<name>-vN.png. With --json, .result.width, .result.height and .result.alpha_channel come from the saved file; alpha_channel: true only says transparency is possible, so look at the image before you call it a cutout.
• Long text answers: --json > /tmp/askcodex/<name>.json, then jq -r .result.text /tmp/askcodex/<name>.json. Keep large JSON and image payloads out of the conversation.
• --json, --events, and --no-refresh are global flags and work after the subcommand.
• Semantic --json success output is {schema_version:1,command,result}. Read .result. Do not pass --backend (it adds the raw backend response, ~700 KB for models) unless the user is debugging the protocol.
• ask --json gives one final envelope with .result.model, .result.effort, .result.verbosity, .result.text, .result.usage (possibly null), and .result.json with --schema. The catalog is at .result.models (slug, description, default_reasoning_level, upgrade); quota windows are at .result.rate_limit.
• --events prints newline-delimited JSON: ask sends text_delta events and one final result event on success; other commands send their result. Check the exit status: partial text is not success. Do not combine --events with --json.
• transcribe prints the text, or .result.text with --json.
• Stdin prompts (-) and raw --body - accept up to 16 MiB of UTF-8. File inputs must be regular files (symlinks to them work; devices and FIFOs do not). image edit accepts up to five PNG references totaling 25 MiB; renaming a JPEG to .png does not convert it, and the CLI checks the file signature. These caps are client memory limits, not server limits.
• raw <METHOD> <PATH> --body '<JSON>' is for a backend operation the user explicitly needs. It accepts only the backend origin, leaves JSON and stream bytes unwrapped, and does not support --events.
• reference prints the CLI reference locally, without credentials or network.

Rules

• On a nonzero exit, report the error and fix its cause before another attempt. Do not loop, fabricate a result, or silently switch models.

• With --json or --events, a failure prints {schema_version:1,error:{code,message, http_status?,backend?}} on stderr and no success result. Never turn a partial stream into a completed answer. error.backend carries the backend's own reason, filtered: only code, type, param, message, detail and error (or incomplete_details.reason), each cut to 400 bytes. When it is absent or reads as cut off, tell the user the full cause could not be read instead of guessing. Act on it:

  The diagnostic says	Cause	Next step
  backend.code: "unsupported_value", backend.param: "reasoning.effort"	The model does not take that effort	Pick a level from the model's guide and retry once
  backend.code: "invalid_json_schema"	The --schema file breaks a strict rule	Fix the key the message names; no answer was generated
  backend.detail: "... not supported when using Codex with a ChatGPT account."	Unknown slug, or a model this plan cannot use	Check askcodex models; ask the user before using another model
  backend.code: "cyber_policy" or another policy code	A safety classifier blocked the request	Report it; do not reword the request to get around it
  backend.incomplete_details	The answer was cut short	Report it; narrow the task or ask for less output, then retry once
  code: "rate_limited" (HTTP 429)	Quota or rate limit	Check askcodex usage --json --no-refresh and tell the user

• Use --no-refresh for read-only account and auth checks. askcodex auth refresh rotates the credentials and rewrites the auth file; run it only when the user asks for a refresh.
• Write into a project directory only when the user asks for that destination.

Report

Tell the user the command you ran, the outcome, the artifact path if one was saved, and any limitation that affects the result. For ask, add the model and effort. For an image, give the file's real dimensions and never name a serving model: the backend does not report one. Show an image when the user needs to judge it visually.

Install

The binary is ~/.local/bin/askcodex. From a checkout, ./install.sh installs it without touching credentials; --skills agents, --skills claude, or --skills all also installs this skill, and --check-auth checks credentials without refreshing. See the repository and the generated command reference.