Model tasks and helper models

Optional Agent.create({ modelExecution: { helpers: { extractor: { connector: 'helper-account', model: 'gpt-4o-mini' } } } }) keeps the main model fixed and enables an explicit invokeHelper() call. Helpers receive only a text packet, have no tools or provider continuation state, and require a locally validated JSON schema. Default helper output is 512 tokens with at most one repair; unsupported settings and oversize packets fail before dispatch. Schema IDs and compiler caches are isolated per invocation. Supported Anthropic extended thinking validates its mode and token budget before building the packet. Google GenerateContent/Vertex thinking is included in ledger output/cost totals and helper output checks; legacy provider response/event fields remain unchanged.

Pass one host-owned ModelTask through { modelTask } to managed/direct run and stream, helper calls, or explicit rollover to share reservations. Sharing is explicit: a helper without modelTask owns an independent task, even during another active execution. Missing or partial external usage stays unknown in the ledger; failed/cancelled external outcomes remain visible. Terminal stream usage survives later failure or cancellation without being counted twice. Tracked calls snapshot request records before admission. Google/Vertex candidate count multiplies output reservations and the configured aggregate per-call cap. Helper repairs retain selected evidence and fail before dispatch if the expanded packet cannot fit. Mixed external pricing remains unknown through final snapshots. Optional limits are calls, tokens, absolute deadline and estimated model token cost. Reservations use full provider input capacity plus bounded output, so large-window models need large temporary allowances. Unknown usage/prices remain explicit; external fees and distributed/external runtime admission are outside the guarantee. With no configuration, existing defaults, signatures, hooks and sessions remain unchanged.

Read the complete recipe for named connector setup, all options, cancellation/late settlement, registry versus upper-bound costs, helper fallback, external usage, ownership and save/load boundaries. Run npm run example:purpose-models and npm run example:shared-budget after a build for offline demonstrations; --live on the purpose example explicitly opts into billable OpenAI calls. AMOS adds /helpers configuration and usage inspection. Opt-in experimental automatic strategies are described in the context strategy recipe. Configuring a compactor alone keeps the algorithmic default; selecting StructuredSummaryStrategy uses that role during bounded pre-call preparation. The bounded comparison is a functional pilot, not a general quality or cost-reduction claim.