Glossary | Awaken Agents documentation

Term	中文	Description
`Thread`	会话线程	Persisted conversation + state history.
`Run`	运行	One execution attempt over a thread.
`Phase`	阶段	The committed run status: `Phase { Running, Waiting, Ended(EndCause) }`. It is not a sequence of named lifecycle steps — the points a hook observes within a step are the separate `PhaseHookPoint` enum.
`PhaseHookPoint`	阶段钩子点	The five points a phase hook may observe within one step: `StepStart`, `BeforeInference`, `AfterInference`, `AfterTool`, `StepEnd`.
`EndCause`	结束原因	Why a run ended: `NaturalEnd`, `MaxSteps`, `Cancelled`, `Stopped(reason)`, or `Error(Failure)`.
`Runtime`	智能体运行时	The embeddable execution engine (`awaken-runtime`). `Runtime::run` consumes a `ExecutableAgentSnapshot`; it owns the model/tool loop, commit boundary, permission gate, and resolver.
`AgentConfig`	智能体规约	Serializable agent definition authored through config: `instructions`, `model_binding`, `tool_ids`, `plugin_ids`, per-plugin config, and context policy.
`ExecutableAgentSnapshot`	快照	The resolved, serializable agent snapshot the runtime installs and executes; carries a `CatalogFingerprint`.
`AgentEvent`	智能体事件	The neutral, post-commit projection of a run (seven variants); each protocol `Transcoder` maps it to a wire format.
`Key`	状态地址	The plain string newtype used by the persisted `Command`/`Store` core; it requires no registration.
`StateKey`	类型化状态视图	An optional typed whole-value view that fixes a key’s address, scope, merge policy, and value shape while producing the same `Command`. `FoldStateKey` adds deterministic typed updates.
`Command`	状态命令	One staged state transition `{ key, scope, merge, action }`, where `action` is `Set(value)` or `Remove`. Producers stage commands; the commit boundary validates and applies them.
`Scope`	键作用域	The declaration scope of a key: `Run`, `Thread`, `Shared`, or `Profile`. Binding to a concrete run/thread happens at commit.
`MergePolicy`	合并策略	How concurrent writes to one `(scope, key)` reconcile in a commit batch: `Disjoint` (later write replaces), `Commutative` (object values shallow-merge, non-object replace), or `Exclusive` (a second write in one batch is a conflict → `Failure::StateConflict`).
`Store`	状态映射	The materialized read side — a `(scope, key) → value` map rebuilt from committed commands for replay. Hooks and gates receive a read-only `&Store`.
`Plugin`	插件	A unit of runtime extension: declares a `PluginManifest` + `CapabilityBound` and contributes tools, state keys, phase hooks, tool gates, and run-end guards through `Plugin::resolve`.
`Contributions`	插件贡献	The bundle a `Plugin::resolve` returns — tools, state keys, phase hooks, tool gates, tool observers, run-end guards, and dynamic tools — merged into `ResolvedExecutionEnv`.
`CapabilityBound`	能力边界	The declared upper bound of what a plugin may contribute (tool ids, state keys, phase hooks, gates, guards, …). Actual contributions must be a subset, or the run fails closed with `Failure::CapabilityBound`.
`PhaseHook`	阶段钩子	An async hook bound to one `PhaseHookPoint`; returns a `HookReaction` that stages `Command`s. It runs at five points including `AfterTool` (which reacts to a tool result). At `BeforeInference` it injects request-only context by writing the `ContextMessages` state key, which is never committed to the transcript.
`ToolGateHook`	工具闸门钩子	A pre-execution hook that decides whether a tool call is allowed, blocked, short-circuited, suspended, or scheduled. It can only restrict, never grant.
`RunEndGuard`	运行结束守卫	A guard consulted at the natural-end boundary; returns `RunEndDecision::{Complete, Steer}`. The first guard that steers wins, otherwise the run ends.
`Tool`	工具	A model-callable capability with an id, JSON-Schema parameters, and a `call`/`invoke` implementation returning a `ToolOutput`.
`ToolDescriptor`	工具描述符	Model-visible tool metadata: id, description, parameters schema, content hash. Carries no executable handle.
`ToolOutput`	工具结果	A tool call’s result: `{ call_id, content, is_error, state: Vec<Command> }`.
`HookReaction`	钩子反应	What a phase hook returns: staged `Command`s plus committed messages. An `AfterTool` hook’s messages reach the transcript; request-only context is not a message here — it is the `ContextMessages` state key written at `BeforeInference`.
`ResumeTicket`	挂起票据	The committed record of a parked run (correlation/run/thread/snapshot ids, catalog fingerprint, `AwaitReason`, pending tool, deadline). A resume is accepted only on an exact match.
`AwaitReason`	Await reason	Why a Run is `Awaiting`: `ToolPermission`, `UserInput`, `ExternalEvent`, `RateLimit`, `ManualPause`, `ScheduledAction`, or `Delegation`.
`RunDelegationService`	智能体解析器	The port the runtime routes the delegation tool (`agent_run`) to; it runs native (in-process) or remote (A2A) sub-agents chosen by agent id, returning an `DelegationStep`.
`RunActivation`	运行激活	The immutable, serializable input to start a run: run id, thread id, snapshot, input, and trace.
`RunIngress`	运行入口	The delivery port for run submission and control. `DirectRunIngress` is the in-process path; `DurableRunIngress` (in `awaken-run-ingress`) is single-node durable with queue, recovery, and scheduled wake.
`ContextPolicy`	上下文策略	How the model-visible context window is bounded: `KeepAll` or `KeepLast { keep_last }`.
`ModelSpec`	模型规约	A catalog entry for a model: stable `id`, optional context window / max output / modalities / knowledge cutoff, and pricing. Referenced by a stable id from an agent’s model binding.
`Transcoder`	流转码器	A stateful per-protocol mapper from the `AgentEvent` stream to a protocol’s wire format; there is no single fixed wire envelope.
`TokenUsage`	令牌用量	Token consumption report from LLM inference.
`CancellationToken`	取消令牌	Cooperative token threaded through a run to request cancellation (`EndCause::Cancelled`).