codex-oauth-plus-onboarding-skill/skill/codex-oauth-plus-onboarding/references/luban-one-by-one-runner-review-and-fixes-2026-05-11.md

Success-record review and Luban one-by-one runner fixes (2026-05-11)

Why this review was needed

BOSS asked to deeply revisit the known phone-verification success records and compare them with the current Luban flow.

The immediate trigger was a Luban France event where a code was received but OpenAI stayed on phone-verification, followed by a one-by-one test that ended with:

{"bought":4,"sms_received":0,"sms_timeout":2,"submit_no_sms_page":2}

The current OpenAI page then showed 糟糕，出错了！ on https://auth.openai.com/add-phone.

Success records reviewed

1. 5sim Vietnam success, 2026-05-10

Reference: openai-phone-free-country-vietnam-sms-success-codex2api-region-block-2026-05-10.md.

Key success properties:

• Strict country/area-code gate before buying and again before submit.
• Count only actual provider activations.
• If OpenAI reaches phone-verification, poll provider.
• Submit a strict 6-digit code.
• Consider phone success only when OpenAI leaves phone-verification and reaches about-you / profile.

2. 5sim Indonesia success at 335 activations, 2026-05-11

Reference: openai-phone-335-success-cpa-oauth-login-session-pitfalls-2026-05-11.md.

Key success properties:

• OpenAI phone codes should prefer strict 6-digit extraction.
• Do not mark provider activation as finished until browser leaves phone-verification.
• A wrong parser selected a 4-digit candidate from a provider blob; rechecking the finished activation and extracting a 6-digit code fixed it.
• After phone success, fill about-you fields explicitly (name, age, birthday).

Defects found in the current Luban one-by-one flow

Defect 1 — submit-code method not identical to successful 5sim method

The current Luban runner used a generic setVal with plain Event('input') in one path. The successful 5sim submit used:

const input=document.querySelector('input[autocomplete=one-time-code], input[name=code], input');
const setter=Object.getOwnPropertyDescriptor(window.HTMLInputElement.prototype,'value').set;
input.focus();
setter.call(input, code);
input.dispatchEvent(new InputEvent('input', {bubbles:true, inputType:'insertText', data:code}));
input.dispatchEvent(new Event('change', {bubbles:true}));
click exact '继续';

Fix: use this exact method for OpenAI phone-code submit.

Defect 2 — Bad Request / OpenAI error state was not fatal enough

The one-by-one run continued country gates after OpenAI became:

title: 糟糕，出错了！ - OpenAI
url: https://auth.openai.com/add-phone

The page had no usable country selector/tel input for later countries, producing repeated false country-gate failures.

Fix: add ensure_add_phone_ready():

• require add-phone URL, country select, and tel input;
• reject pages whose title/body match 糟糕|出错|Bad request|unknown_error|HTTP ERROR;
• try reload/direct navigate a few times;
• if still not ready, stop the runner before buying another number.

Defect 3 — Luban release retry was too shallow

France 34133707 initially returned:

停用失败,该供应商需要在:118 秒 后停止.
停用失败,该供应商需要在:43 秒 后停止.

The runner stopped retrying, requiring manual cleanup. Later manual check showed 该号码已释放.

Fix: reject_after_min_wait() now loops until:

• success, or
• 已释放, or
• wrong_status, or
• hard retry deadline.

It parses provider messages such as N 秒 后停止 and sleeps accordingly.

Defect 4 — provider/channel ambiguity was under-logged

The current OpenAI phone page can show WhatsApp delivery text:

重新发送 WhatsApp 消息

The runner must log page channel (whatsapp, sms, unknown). If provider reports a code while page says WhatsApp, classify failures as ambiguous until provider raw fields and OpenAI error text are inspected.

Defect 5 — submit_no_sms_page must not be treated as SMS delivery evidence

If OpenAI stays on add-phone after submit but does not show an explicit phone reject, classify as submit_no_sms_page, hold for the minimum Luban release window, release, then recover page. Do not count it as phone-verification SMS timeout.

Implemented optimized runner

New script:

scripts/luban_one_by_one_phone_verify.py

Key behavior:

1. Buys only one Luban request at a time.
2. Validates OpenAI add-phone readiness before each country and before buying.
3. Stops instead of buying when the page is in Bad Request/error state.
4. Holds unused Luban orders for at least 3 minutes before reject.
5. Retries reject until success/already-released/wrong_status or hard deadline.
6. Polls provider for strict 6-digit code.
7. On SMS received, writes the code only to a local file and logs sms_received without printing it.
8. Submits code with the proven 5sim InputEvent method.
9. If submit fails, keeps page and provider order for retry and exits with sms_submit_failed_hold_for_retry.
10. Does not set success unless OpenAI leaves /phone-verification.

Run example:

PYTHONUNBUFFERED=1 MAX_BUYS=10 PHONE_POLL_SECONDS=220 MIN_HOLD_SECONDS=180 \
  /Users/chick/homebrew/opt/python@3.11/bin/python3.11 \
  /Users/chick/.hermes/skills/software-development/codex-oauth-plus-onboarding/scripts/luban_one_by_one_phone_verify.py \
  | tee /tmp/luban_one_by_one_phone_verify.log

Optional priority override:

LUBAN_COUNTRY_PRIORITY='France,Vietnam,Portugal,Romania,Germany,Netherlands,Argentina'

Operational rule going forward

For Luban/OpenAI tests, use the optimized script above instead of ad-hoc /tmp/luban_one_by_one_stop_on_sms.py.

If the runner exits with sms_submit_failed_hold_for_retry, do not release the order and do not navigate away. Retry from the preserved page using the local code file and corrected submit helper.