intention-kernel package
Build provider-neutral agents that interpret intentions, validate plans, execute capability effects and publish grounded, durable responses.
Import these runtime APIs from intention-kernel. For scenario evaluation, assertions, reports, and test helpers, see intention-kernel/testing.
Classes
Class | Description |
|---|---|
Raised when a public definition cannot be compiled safely. | |
Raised when durable state cannot be read or committed safely. | |
Raised by a host adapter only when it knows an external effect was not applied. | |
Raised when an external effect may have happened and cannot be retried safely. | |
Raised when a bounded turn exceeds its configured execution budget. | |
Base class for exceptional failures exposed by Intention Kernel. | |
Raised when required kernel adapters or execution limits are invalid. | |
Raised when a model provider cannot produce a usable result. | |
Raised when data crossing a trust boundary violates its schema. |
Functions
Function | Description |
|---|---|
Validate and brand an agent ID. | |
Validate and brand a capability ID. | |
Create a provider-neutral kernel whose compiled agents run on private LangGraph state. | |
Define and shallow-freeze a single-agent manifest. | |
Define and freeze a capability contract without executing it. | |
Define immutable contextual model guidance without coupling it to a provider. | |
Define and freeze a host-owned policy. | |
Define a vendor-neutral schema and normalize all validation results. | |
Brand a durability-adapter-generated effect ID after checking it is non-empty. | |
Brand a host- or capability-generated evidence ID after checking it is non-empty. | |
Validate and brand a fact type. | |
Return every distinct kernel instruction variant with immutable default content. | |
Brand a host- or capability-generated interaction ID after checking it is non-empty. | |
Validate and brand a policy ID. | |
Render bounded literal instructions with exact, one-pass | |
Brand a host-generated thread ID after checking it is non-empty. | |
Brand a host-generated turn ID after checking it is non-empty. |
Interfaces
Interface | Description |
|---|---|
Unresolved work persisted between turns. | |
Public definition compiled into an immutable runtime agent. | |
Host-defined objective whose completion is represented by one confirmed fact. | |
Complete declarative progression contract for one compiled agent. | |
Host-authored safe copy used only when a model cannot ground a control response. | |
Structured artifact returned to a channel or UI. | |
Capability-authored response copy with explicit evidence annotations. | |
Versioned input factory required for server-authorized capability activation. | |
Input available when a capability is invoked by trusted progression authority. | |
Executable operation independent of its infrastructure adapters. | |
Metadata accepted when a capability exposes observable domain progress. | |
Turn-scoped publisher for sanitized capability and nested-workflow events. | |
Runtime services available to a capability implementation. | |
Declarative set of related capabilities presented as one bounded interaction. | |
One executable member exposed inside a capability group interaction. | |
Model-facing semantic guidance attached to one capability contract. | |
Expected domain failure that is safe to expose through the response composer. | |
Model-reference requirements for capabilities that operate on active server-owned choices. | |
Runtime services available to a capability implementation. | |
Immutable agent executable produced by IntentionKernel.compile(). | |
Authorized completed operation exposed to the response composer. | |
Context reference used to resolve a conversational expression. | |
Explicit contradiction found inside the turn or against confirmed facts. | |
Persisted conversation message. | |
Host-owned persistence, locking, deduplication and effect ledger contract. | |
Stable identity supplied when coordinating an external effect. | |
Result stored for duplicate turn replay. | |
Scope held while one thread is serialized by a durability adapter. | |
Stable authority passed only to the external operation being coordinated. | |
Completed effect value together with its durable receipt. | |
Durable evidence for an idempotent external effect. | |
Sink for sanitized kernel events. | |
Evidence that supports an interpreted request or business claim. | |
Candidate fact returned by a capability before reducer validation. | |
Fact a capability promises to publish. | |
Confirmed fact persisted in the conversation checkpoint. | |
Reference to a versioned fact. | |
Fact required before a capability can execute. | |
Complete semantic interpretation of one user turn. | |
Evidence extracted by the model for an interpreted intention. | |
Isolated runtime facade that compiles declarative agent manifests. | |
Constructor properties for IntentionKernelError. | |
One user objective interpreted from the current turn. | |
A single user-facing information or confirmation request. | |
Structured answer to the currently active interaction. | |
Server-owned selectable option. | |
Minimal durable state owned by Intention Kernel. | |
Host-supplied clock used for durable timestamps. | |
Observable event emitted by the kernel. | |
Host-supplied source of unique, observable runtime identifiers. | |
Bounded execution settings. Zero is supported only for recentMessageLimit. | |
Provider and host adapters used by one isolated kernel instance. | |
Expanded capability contract supplied only after the operation was selected. | |
Compact capability entry used only to select relevant operations. | |
Adapter contract for any structured-output model provider. | |
One auditable example attached to contextual model guidance. | |
Canonical, read-only state available to a host-owned guidance selector. | |
Semantic guidance selected before capability routing; it never grants execution authority. | |
Result of evaluating one host-owned contextual guidance selector. | |
Stable versioned instruction contract, independent of provider and storage. | |
Exact instruction definition and execution values for one model request. | |
Synchronous instance-scoped resolver; storage loading belongs to the host. | |
Provider-neutral structured model request. | |
Validated provider response before kernel schema validation. | |
Provider metadata returned for observability and cost accounting. | |
Pure extension point for host-owned business constraints. | |
Durable status of the configured primary objective. | |
Durable fact-bound occurrence of one progression rule. | |
Fact-triggered target made available by the host application. | |
Kernel-owned progression state persisted with the conversation checkpoint. | |
Pinned host revision resolved from an already loaded snapshot. | |
Port-free, evidence-bearing input for response composition. | |
Evidence annotation for one exact factual span in the assistant message. | |
Validated semantic meaning of the current turn supplied to response composition. | |
Vendor-neutral runtime schema used at every trust boundary. | |
Public input for one compiled-agent turn. | |
Input accepted by defineSchema(). | |
Model-visible projection of one contextual guidance policy selected by the host. | |
Validated response returned by a compiled agent. | |
Public result committed exactly once for one turn. | |
Host-verified answer produced by a clicked interaction control. | |
Normalized schema issue independent of the validation vendor. |
Variables
Variable | Description |
|---|---|
Package version for runtime diagnostics; verified against the installed artifact during packaging. |
Type Aliases
Type Alias | Description |
|---|---|
Stable agenda item identifier. | |
Stable agent identifier. | |
Nominal string used to keep unrelated kernel identifiers distinct at compile time. | |
Stable capability identifier such as | |
Result returned by a capability. | |
Stable effect identifier. | |
Durable lifecycle of an external mutation. | |
JSON-safe context attached to a public kernel error. | |
Stable evidence identifier. | |
Controls how a persisted fact is exposed to model tasks. | |
Stable fact type such as | |
Stable intention identifier. | |
Stable interaction identifier. | |
JSON Schema representation accepted by model adapters. | |
Kernel-owned lifecycle request that never invokes a business capability. | |
Task-specific capability projection supplied to a model gateway. | |
Decision returned by a pure policy. | |
Stable policy identifier. | |
Stable plan step identifier. | |
Stable thread identifier supplied by the host. | |
Stable turn identifier supplied by the host. | |
Result of validating untrusted data. |