Skip to main content

Interface: SkillRuntime

Defined in: packages/core/src/skill.ts:277

The runtime a bot weaves in for a set of skills: the hoisted tools (every skill's tools plus the shared load_skill/unload_skill loaders) and a per-turn buildContext that emits the always-on skill index plus the full instructions of any active (keyword-matched or explicitly loaded) skill.

Shape-compatible with the bot's systemPromptBuilder pipeline — the same seam MemoryBankLike uses.

Properties​

tools​

tools: Record<string, ZodToolDefinition>

Defined in: packages/core/src/skill.ts:279

Hoisted tools: each skill's tools (tagged) + the load_skill/unload_skill loaders.

Methods​

buildActive()​

buildActive(input): Promise<string>

Defined in: packages/core/src/skill.ts:293

The instructions of every skill active for this input (keyword, host request, or loaded earlier), as one block, or '' when none is. Records what was applied (see getAppliedActivations). Changes as the active set does, so it belongs after the cache breakpoint (Baleybot sends it as a message just before the person's latest message).

Parameters​

input​

string

Returns​

Promise<string>


buildContext()​

buildContext(input): Promise<string>

Defined in: packages/core/src/skill.ts:295

buildIndex followed by buildActive — for hosts that build one prompt fragment.

Parameters​

input​

string

Returns​

Promise<string>


buildIndex()​

buildIndex(): string

Defined in: packages/core/src/skill.ts:285

The skill index: one line per skill (name and description), and how to load one. Byte-stable — it never changes with which skills are active, so it belongs in the cached system prompt.

Returns​

string


getActivations()​

getActivations(input): SkillActivation[]

Defined in: packages/core/src/skill.ts:304

The skills whose instructions would be injected for this input, with the reason. Keyword match takes precedence as the reason (it's the immediate trigger); otherwise the skill is active because of a prior load_skill call. Pure — safe to call for observability without side effects. For what was actually injected on the latest turn (excluding a skill whose instructions threw), use getAppliedActivations.

Parameters​

input​

string

Returns​

SkillActivation[]


getActiveSkills()​

getActiveSkills(): string[]

Defined in: packages/core/src/skill.ts:329

The skills active right now: loaded, requested this call, or keyword-matched on the latest input.

Returns​

string[]


getAppliedActivations()​

getAppliedActivations(): SkillActivation[]

Defined in: packages/core/src/skill.ts:312

The activations the most recent buildContext actually injected into the prompt — the truthful basis for skill_activated events. Differs from getActivations only when a skill's instructions threw: it is active but its body was omitted, so it is not reported here. Returns [] before the first buildContext call.

Returns​

SkillActivation[]


getState()​

getState(): SkillState

Defined in: packages/core/src/skill.ts:337

The skills loaded via load_skill and not unloaded, in load order.

Returns​

SkillState


reset()​

reset(): void

Defined in: packages/core/src/skill.ts:335

Forget every skill activated via load_skill. Keyword triggers are stateless and unaffected. The bot calls this from clearHistory() so a fresh conversation does not inherit skills loaded in a previous one.

Returns​

void


restore()​

restore(state): void

Defined in: packages/core/src/skill.ts:343

Replace the loaded set with state.loaded (a saved getState). Unknown names are dropped with one warning — a skill renamed or removed since the state was saved must not break the conversation.

Parameters​

state​

SkillState | undefined

Returns​

void


setActivationListener()​

setActivationListener(listener): void

Defined in: packages/core/src/skill.ts:320

Register a listener fired the instant a skill is newly activated by a load_skill call mid-turn — so the bot can emit a same-turn skill_activated event for the model's deliberate choice (the pre-turn sweep can't see it, since load_skill runs during the tool loop). Replaces any previous listener; pass undefined to clear.

Parameters​

listener​

((activation) => void) | undefined

Returns​

void


setDeactivationListener()​

setDeactivationListener(listener): void

Defined in: packages/core/src/skill.ts:327

Register a listener fired the instant a skill stops being active mid-turn: unload_skill (or a restore) dropped it and no keyword or host request keeps it active. The bot uses it to return the skill's tools to on-demand loading. Replaces any previous listener; undefined clears.

Parameters​

listener​

((skill) => void) | undefined

Returns​

void


setRequested()​

setRequested(names): void

Defined in: packages/core/src/skill.ts:350

Activate these skills for the next turn(s) until called again (the host's per-call pre-activation; reason requested). Not persisted in getState. Pass undefined or [] to clear. Unknown names are ignored with one warning.

Parameters​

names​

readonly string[] | undefined

Returns​

void


skillToolNames()​

skillToolNames(): string[]

Defined in: packages/core/src/skill.ts:354

Every tool name contributed by any skill (not the loaders).

Returns​

string[]


toolsFor()​

toolsFor(skill): string[]

Defined in: packages/core/src/skill.ts:352

The tool names a skill brings (its own declaration), or [].

Parameters​

skill​

string

Returns​

string[]