Skip to main content

Interface: IAIProfile

Defined in: internal/ai-profile-store/types.ts:171

Tenant choices for one way of using AI. Passed as data to Assembler, ToolPolicy, LLMClient.

Properties​

allowedTools​

allowedTools: string[]

Defined in: internal/ai-profile-store/types.ts:183

Tool names the LLM may see. Subset of ToolRegistry.

approvals​

approvals: string[]

Defined in: internal/ai-profile-store/types.ts:185

Tool names that pause for a human before execution.

cheapModelId?​

optional cheapModelId?: string

Defined in: internal/ai-profile-store/types.ts:181

Model for helper calls (memory, rewrite). Empty = platform default.

id​

id: string

Defined in: internal/ai-profile-store/types.ts:173

The tenant's own id of the profile; what a call names in profileId (D103).

limits​

limits: IAIProfileLimits

Defined in: internal/ai-profile-store/types.ts:249

The loop and output caps of the profile.

modelId​

modelId: string

Defined in: internal/ai-profile-store/types.ts:179

Main model; must exist and be enabled in ModelRegistry.

name​

name: string

Defined in: internal/ai-profile-store/types.ts:177

Human label, e.g. "Careful refunds".

parallelToolCalls?​

optional parallelToolCalls?: boolean

Defined in: internal/ai-profile-store/types.ts:217

The tool calls of one model step run at the same time (D101).

An agent step that looks up three things waits for the slowest, not for the sum.

A model that asks for several tools in one step gets no result before the step is out, so the calls are independent by construction: the gateway decides every call first, runs every allowed one concurrently, stores the results in the order of the request and pauses once with every call that needs a human. The calls of a tool whose policy is a predicate are decided and run one by one after the others (D112): such a policy may read state the step changes, like "already refunded", and must see the earlier calls. Switch it off for a profile whose tools must not overlap (two writes to the same record, a rate-limited API): the calls then run one after the other, and a call that needs a human pauses the run before the calls after it are decided.

Example​

false - a chat profile whose tools write to one ledger and must not overlap

Default​

true

promptAddendum​

promptAddendum: string

Defined in: internal/ai-profile-store/types.ts:233

Tenant text appended to the CHAT_MAIN system prompt.

reviewAnswers?​

optional reviewAnswers?: boolean

Defined in: internal/ai-profile-store/types.ts:199

Every answer of a run() on this profile waits for a human before the end user sees it (D99).

Gates the words, as approvals gates the actions: a moderator reads the reply, a clerk checks the drafted letter.

The run pauses with a RESULT_REVIEW envelope after the answer passed the output checks; the reviewer approves, edits or rejects it through resume(). ask() on such a profile is refused: it has no run to pause. Leave it off for a chat whose answers the user reads directly.

Example​

true - a support bot whose replies an agent checks before the customer sees them

Default​

false

skills?​

optional skills?: ISkill[]

Defined in: internal/ai-profile-store/types.ts:247

Instructions the model loads on demand on every call of this profile (D124).

A profile's long rules without paying for them on every call.

Listed by name and description in the system prompt of a run(), loaded with dyna_aigw_load_skill when a task needs one; inlined whole in an ask(). Merged with the call's skills; a name present in both is a 400. Names must be unique here too: upsertProfile refuses a duplicate.

Example​

[{name: "refund_policy", description: "…", body: "…"}] - the shop's refund rules, read only when a refund comes up

Default​

undefined - no skills

templateOverrides?​

optional templateOverrides?: Record<string, string>

Defined in: internal/ai-profile-store/types.ts:251

Reserved v2 seam: per-key template body overrides.

tenantId​

tenantId: string

Defined in: internal/ai-profile-store/types.ts:175

The tenant the profile belongs to; a profile is never visible to another tenant.

workingMemory?​

optional workingMemory?: TWorkingMemoryStrategy

Defined in: internal/ai-profile-store/types.ts:231

How this profile's threads reach the model: raw sliding window or summary buffer.

One tenant, two bots: precise chat on one, cheap long threads on the other.

Overrides IWorkingMemoryConfig.strategy; IWorkingMemoryConfig.enabled stays gateway-wide. A thread keeps its stored summary when the strategy switches: SLIDING_WINDOW ignores it, a switch back resumes from it. Read by the built-in WorkingMemory; a host memory may read it or ignore it.

Example​

{type: EWorkingMemoryStrategy.SLIDING_WINDOW} - a support chat whose answers must quote the user's own words

Default​

undefined - the gateway's default strategy