Enum AssistantEvent 

pub enum AssistantEvent {
    InferenceStarted {
        model: String,
        attempt: u32,
        turn: u32,
    },
    InferenceRetry {
        model: String,
        attempt: u32,
        reason: String,
        backoff_ms: u64,
    },
    ModelServed {
        model_id: String,
        local_last_resort: bool,
    },
    Text(String),
    ToolCall {
        call_id: String,
        sequence: u32,
        name: String,
        params: Value,
    },
    ToolResult {
        call_id: String,
        sequence: u32,
        name: String,
        ok: bool,
        content: String,
    },
    Done {
        text: String,
    },
    Error(String),
    AuthRequired {
        reason: AuthRequiredReason,
        message: String,
    },
    GoalEvaluated {
        iteration: u32,
        met: bool,
        grounded: bool,
        reason: String,
    },
}

Streamed events from one loop run. emit is called synchronously as the loop progresses; a chat caller forwards these to agent.chat.event, a CLI caller prints them.

Variants

InferenceStarted

A model call is about to begin. Emitted before awaiting the generator so a host can distinguish live inference from a wedged assistant loop.

Fields

model: String

attempt: u32

turn: u32

InferenceRetry

A transient provider failure will be retried after a bounded backoff.

Fields

model: String

attempt: u32

reason: String

backoff_ms: u64

ModelServed

Attribution for one completed model turn. Emitted before that turn’s text/tool events so every rendered answer can be tied to its actual serving model, including transparent on-device degradation.

Fields

model_id: String

local_last_resort: bool

Text(String)

The model’s free-text for a turn (may be empty when it only calls tools).

ToolCall

A tool is about to run. call_id is generated by the loop rather than trusted from the provider, and sequence is its one-based position in this turn. The matching result carries both values.

Fields

call_id: String

sequence: u32

name: String

params: Value

ToolResult

A tool finished. ok is false for a failed/denied/cancelled call.

Fields

call_id: String

sequence: u32

name: String

ok: bool

content: String

The uncapped observation. Every external projection MUST redact and bound this before emitting it; the model transcript gets a separate capped copy.

Done

Terminal: the model answered with no further tool calls.

Fields

text: String

Error(String)

Terminal: the run failed (inference/transport error).

AuthRequired

Terminal: the turn cannot reach Parslee inference because of the ACCOUNT, and says which remedy applies.

Typed apart from AssistantEvent::Error because the two ask different things of a host. An error is something that went wrong; an auth_required is a step the person can take, and a host can render it as a sign-in card rather than a red row. It is terminal in the same sense Done/Error are — the chat service sends it INSTEAD of the done/error frame, never before one.

Fields

reason: AuthRequiredReason

message: String

Host-facing. Names the remedy in plain language.

GoalEvaluated

Goal-loop verifier result after one iteration. This surfaces CAR’s grounded completion evidence to CLI/chat hosts instead of hiding it in tracing logs.

Fields

iteration: u32

met: bool

grounded: bool

reason: String