Enum ServerMsg 

pub enum ServerMsg {
    Ask {
        req_id: String,
        parent_req_id: Option<String>,
        task_id: String,
        question: Value,
    },
    HookBefore {
        req_id: String,
        parent_req_id: Option<String>,
        task_id: String,
        agent: String,
        attempt: u32,
    },
    HookAfter {
        req_id: String,
        parent_req_id: Option<String>,
        task_id: String,
        agent: String,
        attempt: u32,
        result: Value,
    },
    Spawn {
        req_id: String,
        parent_req_id: Option<String>,
        task_id: String,
        agent: String,
        attempt: u32,
        capability_token: String,
        worker_handle: Option<String>,
        worker: Option<WorkerBinding>,
        directive: String,
    },
}

Server → client push messages on WS /v1/operators/:sid/ws. Each variant pairs with a ClientMsg reply carrying the same req_id (except HookAfter, which is fire-and-forget).

Variants

Ask

SeniorBridge.ask request.

Fields

req_id: String

Correlation key the client must echo back in ClientMsg::Answer.

parent_req_id: Option<String>

Reserved for nested-ask correlation; currently always None (see [current_parent_req_id]).

task_id: String

Task the question originates from.

question: Value

Free-form question payload produced by the engine middleware.

HookBefore

SpawnHook.before request (= the client returns OK / NG via ack).

Fields

req_id: String

Correlation key the client must echo back in ClientMsg::HookAck.

parent_req_id: Option<String>

Reserved for nested-ask correlation; currently always None.

task_id: String

Task whose spawn is being gated.

agent: String

Agent ref about to be spawned.

attempt: u32

1-based dispatch attempt counter for this agent step.

HookAfter

SpawnHook.after notification (= no client ack, fire-and-forget).

Fields

req_id: String

Correlation key (informational only — no reply is expected).

parent_req_id: Option<String>

Reserved for nested-ask correlation; currently always None.

task_id: String

Task the spawn belonged to.

agent: String

Agent ref that was spawned.

attempt: u32

1-based dispatch attempt counter for this agent step.

result: Value

Worker result payload observed after the spawn completed.

Spawn

Operator.execute request (= delegates the whole spawn to an external Operator, via OperatorDelegateMiddleware). The client replies with the WorkerResult-equivalent (= value + ok) in spawn_ack.

Thin control channel (the Spawn thin-control axis): the server sends only the capability_token. system_prompt / prompt are NOT carried in the WS payload. The MainAI (WS Client) forwards the token to the SubAgent, and the SubAgent hits /v1/worker/prompt + /v1/worker/result itself with Authorization: Bearer <capability_token> — fetching prompt / system and posting the result (= heavy payloads go over HTTP; WS stays purely thin control).

capability_token is CapToken::encode() form (= URL-safe base64 of serde_json): a session token with Role::Worker + ["*"] scopes + 600s TTL. The HMAC sig is verified server-side by verify_token_for_task — a self-contained capability token (= no server lookup required).

directive (= immediate instruction for the MainAI; fix for observation #7): Under thin-push discipline, if the payload were only routing fields, the MainAI (a large LLM) would fire the drift “I have a token → I should fetch it myself” / “I got the prompt → I should embed it literally into the SubAgent” 100% of the time (= bias accumulation across 50–100 parallel agents dulls decisions). To structurally remove this drift, a literal instruction text — “launch a SubAgent, hand it the token + endpoint, and let the SubAgent do the fetch / execution / post” — is explicitly embedded into the payload (= implicit convention → literal statement).

This field carries natural-language text intended for the MainAI to read (= not a JSON schema target for parsing). See operator_ws::session::default_spawn_directive() for the server-side default text.

Fields

req_id: String

Correlation key the client must echo back in ClientMsg::SpawnAck.

parent_req_id: Option<String>

Reserved for nested-ask correlation; currently always None.

task_id: String

Task the delegated spawn belongs to.

agent: String

Agent ref the Operator is asked to execute.

attempt: u32

1-based dispatch attempt counter for this agent step.

capability_token: String

CapToken::encode() form Bearer credential for the worker HTTP endpoints (see the variant doc above for the thin-control contract).

worker_handle: Option<String>

Short handle (= wh-XXXXXXXX, 12 chars). An alternate Bearer path paired with capability_token. When /v1/worker/submit receives a handle in Bearer, the server resolves nonce → task_id via the worker_handles map (the short-handle switchover — removes base64 copy-paste accidents). SubAgents (mse-worker) should use this field instead of capability_token as the recommended path.

worker: Option<WorkerBinding>

Worker binding resolved from the Blueprint at compile time. None never reaches the wire on the WS thin path (compile-time gate, see Operator::requires_worker_binding), but the field stays optional for forward compatibility.

directive: String

Literal natural-language instruction for the MainAI (see the variant doc above for why this is embedded in the payload).

Trait Implementations

impl Debug for ServerMsg

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

Formats the value using the given formatter.

impl Serialize for ServerMsg

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.