| name | naming-studio-generate |
| description | Compute Chinese baby name candidates for a pending Codex Naming Studio request and write structured results back to the workbench. Use when the workbench widget posts a 处理起名请求 message, the request watcher reports a pending naming request, or the user asks to handle, retry, or finish a pending naming request. |
Naming Studio Generate
Workflow
A pending request means the user is waiting in the workbench loading state. Work fast and write results back; do not answer in prose only.
- Locate and immediately claim the request, retaining the returned
claimId for this turn:
- Widget follow-up message (primary mode): a chat message like 「处理起名请求 name-xxxx（项目目录 /path）…」 names both the
requestId and the projectDir. Call get_naming_product_state with that projectDir; use the named request if it is still pending, otherwise use latestPendingRequest. Then call claim_naming_product_request with its requestId and save the returned claimId.
- Watcher output (fallback mode):
scripts/watch-naming-request.mjs prints {projectDir, request}.
- Otherwise call
get_naming_product_state with the active projectDir and locate latestPendingRequest.
- Hosted mode recovery: if there is no
latestPendingRequest but latestActiveRequest.status is processing, call claim_naming_product_request for that request. This issues a fresh claimId and fences the dead or stale hosted turn.
Ignore requests whose status is superseded; a newer GUI click has replaced them.
- Read
profile (surname, gender, calendar, birth date/time, full-name length, preferred/blocked characters, tone sliders, filters) and plan — the ordered steps derived from the user's GUI selections. profile.fullNameLength is the product truth: 2 = 双字名 = full name 2 Chinese characters; 3 = 三字名 = full name 3 Chinese characters. For common one-character surnames, that means given is exactly 1 or 2 Chinese characters. Treat profile.tones as numeric truth: 0 means off, 1-39 较弱, 40-69 中等, 70-84 较强, 85-100 强. When the plan contains style-preferences, follow its human-readable strength summary.
- Execute the plan:
- Steps with a
skill field must run first. naming-studio-bazi derives the four-pillar chart and favorable elements from the birth time; naming-studio-research gathers external facts (popular-name blocklists, homophone audits).
- Steps without a
skill field are hard constraints; every generated candidate must satisfy all of them. name-length controls the exact length of given. style-preferences must shape candidate style, summaries, and the style metric. When the plan contains the skip-bazi step, do not run any five-elements reasoning: leave elements as an empty array, distribution/branches as null, and the first metric as null.
- Generate 8-12 Chinese given-name candidates yourself, using naming expertise: 姓名学、音律、字形、寓意、五行喜用与避讳. Do not call any external API and never ask for an API key — Codex's own reasoning is the model; research steps use built-in web search only.
- Call
save_naming_product_result with the same requestId, this turn's required claimId, and the candidates. If claimId is rejected, another claimant owns the request: reload state and abandon this turn's result. If the request was superseded, do not retry it; load state and handle the latest pending request instead.
- Tell the user briefly that the GUI has been updated, then end this turn. What happens next depends on the mode:
- Widget mode (the request came from a widget follow-up message): do NOT arm the watcher. The widget polls state every 1.6s and picks the result up itself; the next 生成好名 click posts a new follow-up message. Your turn simply ends.
- Hosted mode: do not start or re-arm a watcher, server, browser, or any other resident process. The host observes
.naming-product/trigger.json and starts a fresh turn when a new request arrives.
- Fallback (localhost) mode: do not re-arm a watcher. Tell the user that after their next 「生成好名」 click they can say 「处理起名请求」; that new user turn will run this workflow again.
Candidate Shape
{
"given": "景和",
"score": 94,
"grade": "极佳",
"elements": ["木", "火"],
"summary": "景星庆云，惠风和畅，光明温润。",
"risk": "风险低",
"poems": "景明春和，万物有光。",
"metrics": [24, 18, 14, 19, 9, 10],
"analysis": "日主与出生时空的简析。"
}
metrics order is: 八字五行、音律音调、字形结构、寓意内涵、避讳风险、风格匹配. Keep each value within its UI maximum: [25, 20, 15, 20, 10, 10]. risk must be one of 风险极低 / 风险低 / 风险中 / 风险高. Scores stay within 80-99 and must be consistent with metrics.
Failure Path
If generation fails for any reason, still call save_naming_product_result with the same claimId and result.error set to a short Chinese explanation so the workbench can leave loading state. If the claimId is rejected, reload and abandon rather than overwriting the newer claimant. Never leave the GUI spinning.
Taste Rules
Return structured data, not prose-only answers. The GUI is the product surface; chat is only the control plane.
