v1.0.3 to v1.1.0

---

name: xnurta-create-ai-group

description: >-

Create a NEW AI managed group (托管组) for Amazon Sponsored ads and place campaigns

under it when required. Routes by ad type - SD via create_sd_ai_managed_group (campaigns

required); SP/SB via

save_sp_sb_ai_managed_group (create mode). Use when the user wants to create / set up /

新建 / 建一个 a managed group, or put campaigns under AI management for the first time,

even without the exact words "managed group". This creates the group itself - it is NOT

for enabling or editing an existing group (use xnurta-edit-ai-group) or deleting one (use

xnurta-delete-ai-group).

metadata:

---

# Create AI Managed Group

Create a new AI managed group. SD creation requires campaigns; SP/SB creation may

create an empty group or include campaigns. This is a

**write, non-idempotent** operation: calling it twice creates two groups. Read

[`references/platform-notes.md`](references/platform-notes.md) once before your

first write - it covers auth, the shared response envelope, the 30s-timeout rule,

and how errors come back (`data.error` + an optional `recoveryHint` that isn't always

populated, so some stay generic).

## The one thing that decides everything: ad type

Creation logic differs by ad type, and they use **different tools**. Your first job

is always to determine the campaign type, then route:

| Ad type | Tool | Read before building the call |

|---|---|---|

| **SD** (Sponsored Display) | `create_sd_ai_managed_group` | [`references/create-sd.md`](references/create-sd.md) |

| **SP / SB** (Sponsored Products / Brands) | `save_sp_sb_ai_managed_group` (leave `aiGroupId` empty -> create mode) | [`references/create-sp-sb.md`](references/create-sp-sb.md) |

How to determine the ad type, in order of preference:

1. The user says it ("SP 托管组", "for my Sponsored Display campaigns").

2. Infer it from the campaigns to be added - look them up and read `campaignType`.

3. If still unknown, **ask the user** - don't guess. A group is created for one ad

type; picking the wrong tool creates the wrong kind of group.

> SP and SB both go through `save_sp_sb_ai_managed_group`, but a few capabilities

> are SP-only (see `create-sp-sb.md`). SB is not SP - don't copy SP-only fields

> into an SB group.

> **Reads use the full signature.** Every `get_entity_metadata` call requires

> `profileIds`, `entity`, and `userContext` -

> `get_entity_metadata(profileIds=[<id>], entity='campaign', userContext='<why>')`.

> The shorthand `get_entity_metadata(entity='campaign')` fails; always pass all three.

## Disambiguate before writing (when in doubt, ASK - don't guess)

Everyday words - in English OR Chinese - map to different fields. Before setting a value,

be sure which field the user means; if it could be more than one, list the options and

ask - don't guess.

This section lists **business meanings only** - the exact field depends on ad type, so

confirm the meaning first, then use the ad-type reference to pick the field.

- **"budget" / "预算"** - decide the sense AND whether it's a **target value** or an

**increase**:

- **托管组总预算 / group total budget** (= sum of the group's enabled campaigns' daily

budgets) and **a single campaign's daily budget** are **target values**. Note: **a

single campaign's daily budget cannot be set through the managed-group tools** (it's

campaign-level - use the platform).

- **按表现调预算 / performance (dynamic) budget** is an **increase cap on top of the

current budget, NOT a target** (fixed `+$num` or `+num%`); its scope depends on **预算

重新分配** (OFF = per enabled campaign; ON = whole group). See

[`references/coupling-rules.md`](references/coupling-rules.md).

- **预算重新分配 / budget reallocation** - a switch that also changes the scope of 按表现调预算.

Which field, at create time:

- **SD** uses `budget` + `budgetChange` (a boolean toggle) - **there is no `budgetType`

at create** (`budgetType`/`budgetRatio` are edit/batch only).

- **SP/SB create has no group-budget input at all** - only budget-related action-space

switches. If the user wants an SP/SB group's spend budget set at creation, tell them

that's not a create-time field here.

Never map "set budget to X" (a target) onto the dynamic-budget increase value; if the

meaning isn't clear, ask. When enabling 按表现调预算 at create, **set 预算重新分配

explicitly** - its scope (per-campaign vs whole-group) depends on it; don't rely on an

unknown default. To preview a cap, base it on the **enabled** campaigns among those

you're adding (use each one's `dailyBudget`).

- **"target / goal" / "目标"**: 推广目标 `targetType` (1 growth / 2 stability / 3 volume /

4 legacy) vs 目标 ACOS (`acos`). **Create has no `roas` field** - if the user says

"目标 ROAS", do NOT build `roas`; tell them create can't set a target ROAS directly, and

ask whether they want to set the promotion target or convert it to a target ACOS. Also:

"ACOS 优先模式 / ACOS priority" is the strict-ACOS switch (see the ACOS-priority item

below), **not** setting a target ACOS value - don't confuse them.

- **AI on/off at creation**: see "AI on vs off at creation" below - only start AI if the

user explicitly asked to create-and-start.

- **"ACOS 优先模式 / strict ACOS / ACOS priority"** (`bidPerformanceStrictAcosStatus`) - a

sub-switch of 按表现调价, **SP only**, and only triggers with `targetType=2` +

`bidPerformanceStatus=1` + `aiPersonality>=3`. It's a tradeoff (may underspend / reduce

orders) - confirm with the user before enabling. See

[`references/coupling-rules.md`](references/coupling-rules.md).

## Workflow (same for every ad type)

1. **Resolve inputs to IDs and codes.**

- **Campaigns -> internal `campaignId`.** Look them up with

`get_entity_metadata(profileIds=[...], entity='campaign', userContext='...')`.

Each row has both `campaignId` (internal auto-increment int) and

`amazonCampaignId` (Amazon's long string ID). **Always pass `campaignId` into

`campaignIds`; never `amazonCampaignId`** - the Amazon ID is a ~20-digit value

that overflows the int32 the tool expects and the call fails. The lookup also

gives each campaign's `campaignType` (routing) and `aiGroupId` (conflict check).

- **A `campaignId` from `get_ads_perf` may be the Amazon long ID, not the internal

one.** Never feed a `campaignId` taken from `get_ads_perf` straight into

`campaignIds` - first map it through `get_entity_metadata` (match on

`amazonCampaignId`) to the internal `campaignId`. Only the internal int works.

- **One profile, one ad type.** A create targets a single `profileId`, and every

campaign must belong to that profile and share the same `campaignType`. If the

user mixes profiles or mixes SP + SB (or gives a name that matches campaigns in

several stores), stop and clarify - don't silently pick a store or mis-route.

- **Chinese terms -> codes.** Map Chinese goal wording ("推动增长", "活动冲量",

"保持订单稳定", "激进人格") to the right field + code via

[`references/enum-i18n.md`](references/enum-i18n.md) before building the call,

and translate back when you confirm/report.

2. **Map the request to a single business meaning** (see "Disambiguate before writing").

Especially for budget/target wording, decide exactly what the user wants set and which

field it is (after ad type is known). **If more than one meaning is plausible, STOP and

ask - don't build a create on a guessed interpretation.**

3. **Pre-flight checks:**

- **Name uniqueness.** The `aiGroupName` filter is a `like` (substring) match, not

exact - query `get_entity_metadata(profileIds=[...], entity='aiGroup',

filters={"aiGroupName": {"like": "%<name>%"}}, userContext='...')`, then compare names

**exactly** in the returned rows. If an exact match exists, ask for a different

name rather than letting the create fail.

- **Already-managed campaigns (checkable).** A campaign already in another managed

group shows a non-empty `aiGroupId` in its metadata - flag those, let the user

decide.

- **Budget-management conflict (NOT pre-checkable).** Campaign metadata does not

expose budget-management membership, so you cannot verify this up front - only

the backend create can reject it. Don't claim conflicts are fully cleared; say

the budget-management check happens server-side.

4. **Confirm before creating - show everything that will take effect.** Echo the

**complete** config, not just the basics: ad type, group name, the campaigns (by

name), `targetType`/`optimizeType`, target ACOS, budget settings, `aiPersonality`,

`campaignNameSign`, and **every supported action-space switch you're enabling**

(bid / budget / target / struct optimization). For anything you're not

setting, say it will use the platform default - but **don't invent specific default

values** (the tool schema doesn't define them); only state a concrete default if

you've read it back or it's documented. Get an explicit go-ahead - and note that **the

user answering an earlier clarification question is not itself authorization to create**;

you still need an explicit yes on this full preview before calling the create tool. This

matters most when AI will start on (`aiStatus=1`) - those switches immediately affect

live delivery and spend.

- **Turn AI on only if the user explicitly asked to start it.** If they said "create

and start" / "启动", set AI on. If they only asked to set up / create the group or

place campaigns into it (or didn't say), default AI **off** (`aiStatus` / `status`

= 0) and state that in your confirmation. Don't start automation on settings the

user didn't confirm.

5. **Build and call the routed tool** using the exact **write** field names + enum

values (ad-type reference + `field-reference.md`; write names != read names).

> **Create is non-idempotent - never blind-retry.** On any failure, timeout, or

> missing response, the group **may already have been created**. Before retrying,

> re-query by name (`get_entity_metadata entity='aiGroup'`, exact-match the

> `smartCreationName`/`aiGroupName`); only create again if it truly does not exist.

> A blind retry creates a duplicate group.

**Before sending, self-validate the front-end-only rules MCP bypasses** (see

platform-notes "MCP bypasses the platform UI's validation"):

- **Action-space support** - only enable a capability that supports AI for this ad

type ([`references/action-space-matrix.md`](references/action-space-matrix.md));

if the user asked for one that isn't supported for their ad type, tell them and

skip it - don't send a silently-ignored field.

- **Budget** - sanity-check what you reliably can (positive; JP integer-only;

multi-campaign minimum scales with campaign count). Exact per-site/account ranges

are only partly known and the backend may enforce them, so relay backend range

errors rather than hard-blocking on an incomplete table

([`references/budget-limits.md`](references/budget-limits.md)).

- **aiPersonality** `1`-`5`, and **>=3 when `targetType=3` (volume / 冲量)**.

- **Coupled fields** - enabling a switch requires its companion fields, and any

range must have min <= max (see [`references/coupling-rules.md`](references/coupling-rules.md)).

- **Invalid values are rejected by the backend now - but pre-validate anyway** so the

user gets a clear message instead of a downstream error (prod-confirmed 2026-08-13):

`acos` must be > 0 and in range (`0`, negatives, and over-limit are all rejected);

`aiPersonality` outside `1`-`5` is rejected; `campaignIds` is capped at **1000** per

group; and a coupled field sent without its companion is rejected. At **create** the

real budget couplings are: **SD budget** = `budgetChange` + `budget`; **SD dynamic

budget** = `budgetDynamicStatus` + `numType` + `num`; **SP/SB dynamic budget** = the

action-space switch + `budgetNumType` + `budgetNum`. (`acosType`/`budgetType`/

`budgetRatio` are **edit/batch** fields - do **not** put them in a create call.)

Catch these up front rather than leaning on the backend error.

- **Word-list settings are not supported.** Do not send branded, non-branded,

competitor, harvest-blacklist, or negative-target-blacklist fields, even if the

routed schema exposes them. Tell the user to configure word lists in the platform.

6. **Verify it landed - group AND campaigns.** Don't trust the envelope alone:

- Re-read the group (`entity='aiGroup'`) -> confirm it exists with the intended

top-level settings.

- Re-read `entity='campaign'` and check each intended campaign's `aiGroupId` now

points at the new group - the group can exist with fewer campaigns than you sent.

- **SP/SB: a `success` response can still leave an EMPTY group.** `save_sp_sb_ai_managed_group`

may create the group without immediately binding `campaignIds`. So after create,

count how many of your campaigns actually have `aiGroupId` = the new group. **If the

bound count is 0 (or short), re-attach via edit mode**: call `save_sp_sb_ai_managed_group`

again with the returned `aiGroupId` (> 0) and the full `campaignIds`, then re-verify.

- **Reading back binding: query per campaign, not one big `in` list.** In some

environments an `campaignId: {"in":[...]}` filter throws a backend type error.

If a batched filter errors, fall back to per-id `campaignId = <id>` reads (or another

confirmed-usable field). This applies to both the pre-create mapping and this check.

- If operation-log read access is available, query `get_operation_log` and confirm

the create is recorded against an identifiable token user. The server supplies

`changedBy`; never send or fabricate it. If logs cannot be read, state that audit

verification was not performed.

## AI on vs off at creation

`aiStatus` (SP/SB) / `status` (SD) decides whether AI starts optimizing immediately.

Turn it on **only when the user explicitly asked to create-and-start**. If they just

want the group set up (or didn't say), default to AI **off** and say so in your

confirmation - turning AI on means it starts adjusting bids/budgets and spending against

the target right away.

> Note: an "off" group reads back as `aiStatus=2` ("AI Turned Off"), not `0`. Don't

> treat a non-zero `aiStatus` on read as "it's on" - `1` = running, `2` = off.

## Enum discipline

`optimizeType` / `targetType` / `status` / `aiStatus` / `targetHarvestStatus` /

`numType` / `aiPersonality` and the `aiActionSettings` switches are closed enums.

Use only the values listed in

[`references/field-reference.md`](references/field-reference.md) - do not invent or

infer values. Passing an unlisted value makes the call fail, and the error comes back

generic (no field-level hint), so validate against the dictionary **before** sending

rather than relying on the error to tell you what was wrong. For mapping the user's

Chinese wording to these codes (and back), use

[`references/enum-i18n.md`](references/enum-i18n.md).

## Response & errors

Success looks like `{ "isError": false, "data": { "status": "success", "result": {

"aiGroupId": <new id> } } }` - capture `aiGroupId` for the verify step.

Errors come back as `{ "isError": true, "data": { "error": "...", "recoveryHint":

"..." } }`. Relay `recoveryHint` when present, but it isn't always populated - some

come back generic - so still map the common ones yourself:

| Symptom | Likely cause | What to do |

|---|---|---|

| duplicate-name / "Duplicate group name" | name already used in this profile | ask for a different name |

| "Campaign conflict" | a campaign is already in another AI group / budget mgmt | surface which campaign; drop it or free it first |

| generic `business_error` | often an out-of-range/invalid field value | re-check every value against `field-reference.md` |

| timeout (~30s, no response) | the create may have already applied downstream | **verify with a read before retrying** - a blind retry can create a duplicate group |

## Reference files

- [`references/create-sd.md`](references/create-sd.md) - SD create (`create_sd_ai_managed_group`) fields + example

- [`references/create-sp-sb.md`](references/create-sp-sb.md) - SP/SB create (`save_sp_sb_ai_managed_group`): core fields, SP-vs-SB differences, action-space **coupling rules** + example

- [`references/field-reference.md`](references/field-reference.md) - exact **write** field names + enum values (write names != read names)

- [`references/action-space-matrix.md`](references/action-space-matrix.md) - which action-space capabilities are supported (AI / Rule / none) per SP / SB / SD

- [`references/coupling-rules.md`](references/coupling-rules.md) - companion-field couplings, the 按表现调预算 / 预算重新分配 scope relationship, and group-total-budget rescale behavior

- [`references/budget-limits.md`](references/budget-limits.md) - site/account-type budget ranges (front-end rules MCP bypasses)

- [`references/enum-i18n.md`](references/enum-i18n.md) - 中文 <-> English <-> code mapping (parse Chinese requests -> codes; render codes -> Chinese)

- [`references/platform-notes.md`](references/platform-notes.md) - shared write-tool behavior (auth, response envelope, timeout, error shape)