Enum AgentEvent 

#[non_exhaustive]
pub enum AgentEvent<S> {
    Started {
        run_id: String,
        tenant_id: TenantId,
        parent_run_id: Option<String>,
        agent: String,
    },
    ToolStart {
        run_id: String,
        tenant_id: TenantId,
        tool_use_id: String,
        tool: String,
        tool_version: Option<String>,
        input: Value,
    },
    ToolComplete {
        run_id: String,
        tenant_id: TenantId,
        tool_use_id: String,
        tool: String,
        tool_version: Option<String>,
        duration_ms: u64,
        output: Value,
    },
    ToolError {
        run_id: String,
        tenant_id: TenantId,
        tool_use_id: String,
        tool: String,
        tool_version: Option<String>,
        error: String,
        error_for_llm: RenderedForLlm<String>,
        envelope: ErrorEnvelope,
        duration_ms: u64,
    },
    Failed {
        run_id: String,
        tenant_id: TenantId,
        error: String,
        envelope: ErrorEnvelope,
    },
    Complete {
        run_id: String,
        tenant_id: TenantId,
        state: S,
        usage: Option<UsageSnapshot>,
    },
    ToolCallApproved {
        run_id: String,
        tenant_id: TenantId,
        tool_use_id: String,
        tool: String,
    },
    ToolCallDenied {
        run_id: String,
        tenant_id: TenantId,
        tool_use_id: String,
        tool: String,
        reason: String,
    },
}

Runtime events emitted by the agent during a single execute / execute_stream call.

Variants (Non-exhaustive)

Non-exhaustive enums could have additional variants added in future. Therefore, when matching against variants of non-exhaustive enums, an extra wildcard arm must be added to account for any future variants.

Started

Run opened. Sinks use this to mark span beginnings, allocate per-run state, and emit “session opened” telemetry.

Fields

run_id: String

Per-run correlation id (UUID v7). Stable for the duration of the run; matches the id on every subsequent event for this same call.

tenant_id: TenantId

Tenant scope this event belongs to (invariant 11 — every emit site stamps ctx.tenant_id().clone()). Audit / billing / replay consumers key off this field directly instead of correlating through a separate run_id → tenant_id lookup.

parent_run_id: Option<String>

Run id of the calling agent when this run was dispatched from a parent (sub-agent fan-out, supervisor handoff). None for top-level runs. LangSmith-style trace-tree consumers reconstruct the hierarchy from (run_id, parent_run_id) edges across these events.

agent: String

Agent identifier configured on AgentBuilder::name(...).

ToolStart

One tool dispatch began. Emitted by crate::agent::tool_event_layer::ToolEventLayer when wired into the tool registry. Absent when the layer is not wired (the agent runtime itself does not generate tool events).

Fields

run_id: String

Run correlation id.

tenant_id: TenantId

Tenant scope this event belongs to (see Started).

tool_use_id: String

Stable tool-use id matching the originating ContentPart::ToolUse.

tool: String

Tool name being dispatched.

tool_version: Option<String>

Tool version (Tool::version()) when the tool advertises one — useful for distinguishing behaviour changes between otherwise-identically-named tool revisions.

input: Value

Tool input (already JSON-validated by the tool’s schema).

ToolComplete

One tool dispatch finished successfully.

Fields

run_id: String

Run correlation id.

tenant_id: TenantId

Tenant scope this event belongs to (see Started).

tool_use_id: String

Stable tool-use id matching the corresponding ToolStart.

tool: String

Tool name (echoed for sink convenience).

tool_version: Option<String>

Tool version echoed from the matching ToolStart so sinks can correlate completion telemetry without retaining per-tool_use_id state.

duration_ms: u64

Wall-clock duration measured by the layer.

output: Value

JSON output the tool produced. Sinks that persist tool audit logs read this directly; PII redaction happens at the policy layer before this event is emitted, so the payload is safe for storage.

ToolError

One tool dispatch failed.

Fields

run_id: String

Run correlation id.

tenant_id: TenantId

Tenant scope this event belongs to (see Started).

tool_use_id: String

Stable tool-use id matching the corresponding ToolStart.

tool: String

Tool name (echoed for sink convenience).

tool_version: Option<String>

Tool version echoed from the matching ToolStart so sinks see the same provenance on the failure path as on success.

error: String

Operator-facing error message (Display form, includes vendor status, source chain). Sinks, OTel, and log destinations consume this.

error_for_llm: RenderedForLlm<String>

LLM-facing error message wrapped in a sealed RenderedForLlm carrier. The carrier’s constructor is pub(crate) to entelix-core, so the only path from a raw String to this field is entelix_core::LlmRenderable::for_llm — emit sites cannot fabricate model-facing content. The audit-log projection (Self::to_graph_event) extracts the inner rendering into GraphEvent::ToolResult so replay reconstructs the model’s view without re-leaking operator content (invariant #16).

envelope: ErrorEnvelope

Typed wire shape produced by entelix_core::Error::envelope. Bundles wire_code (i18n key / metric label), wire_class (responsibility split), retry_after_secs (vendor Retry-After hint), and provider_status (raw HTTP status) so sinks, audit replay, SSE adapters, and FE rate-limit timers all read one Copy value instead of pattern-matching the inner error variant. Patch-version-stable.

duration_ms: u64

Wall-clock duration measured by the layer.

Failed

Run terminated with the inner runnable’s error. The matching Started{run_id} is always present in the same stream. Caller-facing streams additionally surface the typed error via Result::Err; sinks see only this event.

Fields

run_id: String

Run correlation id.

tenant_id: TenantId

Tenant scope this event belongs to (see Started).

error: String

Lean error message (Display form).

envelope: ErrorEnvelope

Typed wire shape produced by entelix_core::Error::envelope — see ToolError for the field roster. Replay / audit / metric / SSE consumers route off this field instead of parsing error prose.

Complete

Run terminated successfully with the agent’s terminal state.

Fields

run_id: String

Run correlation id.

tenant_id: TenantId

Tenant scope this event belongs to (see Started).

state: S

Final state returned by the inner runnable.

usage: Option<UsageSnapshot>

Frozen UsageSnapshot of the entelix_core::RunBudget counters at the moment the inner runnable returned. None when no budget was attached to the entelix_core::ExecutionContext. Mirrors the usage field on crate::AgentRunResult so streaming and one-shot surfaces observe the same terminal artifact.

ToolCallApproved

HITL approver decided to permit one tool dispatch. Emitted by crate::agent::ApprovalLayer before the matching ToolStart fires. Only present when an Approver is wired (default agents skip approval and never emit this variant).

Fields

run_id: String

Run correlation id.

tenant_id: TenantId

Tenant scope this event belongs to (see Started).

tool_use_id: String

Stable tool-use id matching the originating ContentPart::ToolUse. Pairs with the subsequent ToolStart / ToolComplete / ToolError.

tool: String

Tool name being approved.

ToolCallDenied

HITL approver decided to reject one tool dispatch. The matching ToolStart does NOT fire — denial short-circuits the dispatch path. The agent observes the rejection as Error::InvalidRequest carrying the same reason.

Fields

run_id: String

Run correlation id.

tenant_id: TenantId

Tenant scope this event belongs to (see Started).

tool_use_id: String

Stable tool-use id of the rejected dispatch.

tool: String

Tool name being denied.

reason: String

Approver-supplied rationale.

Implementations

impl<S> AgentEvent<S>

pub fn to_graph_event(&self, timestamp: DateTime<Utc>) -> Option<GraphEvent>

Project this runtime event onto the durable audit-log shape GraphEvent. Returns None when the variant has no audit projection — Started, Complete, Failed are runtime-only lifecycle markers that do not belong in the per-thread audit trail.

The timestamp argument is supplied by the caller (typically Utc::now() at emit time) so this method stays pure: a single runtime event projected at two different points in time produces two distinct (but otherwise equal) GraphEvents.

Lossy projection notes — run_id, tool_version, and duration_ms are dropped because the audit log keys correlation by tool_use_id + timestamp and is not the home for runtime metrics. Operators who need run-level correlation in audit do it at the sink layer (e.g. by stamping a thread tag prior to append).

ToolError is mapped onto a GraphEvent::ToolResult with is_error: true and the error message carried as text content — preserving the same correlation key (tool_use_id) so a session replay can pair the failed dispatch back with the originating ToolCall.

pub fn erase_state(self) -> AgentEvent<()>

Erase the agent-state type parameter, replacing Self::Complete::state with the unit value. Every other variant rebuilds with identical field values — they carry no state. Enables a single audit / SSE / OTel sink (typed AgentEventSink<()>) to fan in from heterogeneous agents (Agent<ReActState>, Agent<SupervisorState>, …) through the StateErasureSink adapter.

Operators consuming the post-erasure event tree retain access to every header field (run_id, tenant_id, parent_run_id) and every per-variant payload (tool inputs / outputs, error envelope, usage snapshot) — only the agent’s terminal state is dropped, which is the field a state-agnostic sink could not type-erase anyway.

Examples

use entelix_agents::AgentEvent;
use entelix_core::TenantId;

let typed: AgentEvent<u32> = AgentEvent::Complete {
    run_id: "r1".into(),
    tenant_id: TenantId::new("t1"),
    state: 42_u32,
    usage: None,
};
let erased: AgentEvent<()> = typed.erase_state();
match erased {
    AgentEvent::Complete { state, .. } => assert_eq!(state, ()),
    _ => unreachable!(),
}

Trait Implementations

impl<S> Clone for AgentEvent<S>

where S: Clone,

fn clone(&self) -> AgentEvent<S>

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl<S> Debug for AgentEvent<S>

where S: Debug,

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl<S> PartialEq for AgentEvent<S>

where S: PartialEq,

fn eq(&self, other: &AgentEvent<S>) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl<S> Eq for AgentEvent<S>

where S: Eq,

impl<S> StructuralPartialEq for AgentEvent<S>