Claude host livemsg delivery (Mode 2)

Turn-boundary delivery for directed live messages uses the Stop hook in hooks/hooks.json and .claude-plugin/hooks.json (P29 dual-sync).

Turn delivery (default ON)

At each turn boundary, the Stop hook runs:

bin/harness inbox check --team "${HARNESS_LIVEMSG_TEAM:-default}"

The recipient agent id is resolved from (in order):

1. --agent flag (explicit CLI / generated hooks)
2. HARNESS_LIVEMSG_AGENT environment variable
3. Stop hook stdin JSON session_id (Claude host)

When there are zero unread messages, the command produces no stdout (silent hook). Trust contract (sanitize, non-instruction disclaimer, 4096B cap) is applied in go/cmd/harness/inbox_check.go before any inject payload is emitted.

Host identity fallback chains

Host	Invocation	Team resolution (first match)	Agent resolution (first match)	Unresolved
Claude	inbox check (Stop hook; --team from env default)	--team flag → HARNESS_LIVEMSG_TEAM → default	--agent flag → HARNESS_LIVEMSG_AGENT → Stop stdin session_id	silent exit 0 (no stdout)
Codex / Cursor	inbox check --from-env (generated turn hook)	deliveryidentity.Resolve() (HARNESS_LIVEMSG_* both set, else breezing BREEZING_SESSION_ID / role) → HARNESS_LIVEMSG_TEAM → default	same Resolve agent → HARNESS_LIVEMSG_AGENT → hook stdin session_id	stderr livemsg: identity unresolved (...) then exit 0

Codex and Cursor hooks use --from-env so one generated hooks.json works across sessions; stdin session_id covers standalone sessions when livemsg/breezing env is absent (Phase 122.2).

Live monitor (~5s poll) — opt-in, default OFF

bin/harness inbox monitor is not wired in the tracked Claude hooks. Enable it manually (for example in .claude/settings.local.json) only when you want a blocking SessionStart monitor stream. It remains on the self-audit allowlist (CCHKnownHooks) when injected locally.