Struct HandoffMode 

pub struct HandoffMode { /* private fields */ }

Implementations

impl HandoffMode

pub fn new(evaluator: Arc<dyn PolicyEvaluator>) -> HandoffMode

Construct the mode with an injected governance policy evaluator.

Trait Implementations

impl Mode for HandoffMode

fn validate_client_envelope( &self, session: &Session, env: &Envelope, ) -> Result<(), MacpError>

RFC-MACP-0010 §5.1(3): a client-submitted HandoffAccept carrying implicit = true MUST be rejected, and the synthetic accept’s message_id namespace is reserved.

Gated to semantics_rev >= 2 so rev <= 1 wire behavior is byte-identical: a legacy session’s client could send an implicit-accept:-prefixed id and be accepted, and its history must stay replayable under the semantics it was accepted with.

Two rules, in this order:

1. any message_id in the reserved namespace -> InvalidEnvelope (checked first, and for every message type: the squat works through SessionStart, Commitment and HandoffContext too);
2. HandoffAccept whose payload decodes with implicit = true -> InvalidPayload — the same code handle_message returns for the same envelope today, so the rev-2 error surface does not shift.

Rule 2 is what the mode itself will stop being able to enforce once the runtime synthesizes implicit accepts: at rev >= 2 a well-formed implicit accept with the deterministic message_id is exactly what dispatch must accept on replay, and the mode cannot tell client provenance from runtime provenance. This boundary can.

fn due_synthetic_envelope( &self, session: &Session, now_ms: i64, ) -> Option<Envelope>

RFC-MACP-0010 §5.1(2)-(3): the synthetic implicit HandoffAccept, due once the outstanding offer’s implicit_accept_timeout_ms has elapsed in unsuspended session time.

Gated to semantics_rev >= 2. Rev <= 1 sessions get None forever and keep resolving through the interim in-Commitment path, so their histories replay to the outcome they were accepted with.

Decision vs. timestamp — two different functions, on purpose

The decision (“has the timeout elapsed?”) is the scalar implicit_accept_elapsed_ms, which is exact for this purpose: elapsed(T) >= timeout iff T >= D, because every pause banked since the offer lies inside [offered_at, T].

Only the timestamp needs the interval walk Session::unsuspended_deadline. §5.1(3) fixes it at “offer acceptance time + timeout + suspended time within the window”, which is not the naive offered_at + timeout + banked_since_offer: that form counts pauses beginning after the deadline — fully reachable, since SuspendSession/ResumeSession are RPCs and need no session-scoped message — and it would make a permanently-recorded timestamp depend on when it happened to be observed.

The two are consistent in the safe direction: an under-recorded suspension_intervals vec (see unsuspended_deadline’s under-count invariant) can only move D earlier, never later, so the debug_assert!(D <= now_ms) below holds even on a legacy rev-2 snapshot.

Cost

One mode_state decode per call for handoff sessions with a bound timeout, and an allocation only when an envelope is actually due. A cached “next deadline” flag on the session was considered and rejected as premature: it would need a writer on the resume path, which is not mode-dispatched at all.

fn authorize_sender( &self, session: &Session, env: &Envelope, ) -> Result<(), MacpError>

Authorize the sender for this message. Modes can override to customize authorization (e.g., allowing orchestrator bypass for Commitment messages).

fn on_session_start( &self, session: &Session, _env: &Envelope, ) -> Result<ModeResponse, MacpError>

fn on_message( &self, session: &Session, env: &Envelope, ) -> Result<ModeResponse, MacpError>

fn on_message_at( &self, session: &Session, env: &Envelope, ctx: &MessageContext, ) -> Result<ModeResponse, MacpError>

Kernel entry point: on_message plus the runtime’s macp_core::mode::MessageContext (acceptance clock). Defaulted to plain on_message so most modes ignore it; modes that need a trustworthy time source (Handoff) override this instead of reading the forgeable Envelope.timestamp_unix_ms. The runtime and replay always call this, with the same clock value that the log entry records.