Why

Channels Phase 2 (cross-device attach, per docs/channels-design.md) needs two foundations before any remote client can exist: a protocol way for that client to declare which device a prompt was typed on (otherwise every message in a shared session gets attributed to the hosting gateway's device), and a dialable address in the presence records so a second device that discovers a live session knows where to connect. Both are inert for today's local-only clients and activate the moment a remote client ships.

What changed

• tui_gateway/server.py: prompt.submit accepts an optional sender_device param, sanitized by a new _sanitize_sender_device (non-strings → empty, whitespace collapsed, 80-char cap — it renders as a chat label, so it is not trusted verbatim). Stashed on the session under history_lock, popped by _run_prompt_submit, handed to the agent per turn.
• agent/conversation_loop.py: the turn's user-message dict carries the declared sender; the API-message builder strips sender_device exactly like reasoning/finish_reason/_thinking_prefill so strict providers (Mistral et al) never see it.
• run_agent.py: the session-DB flush passes sender_device for user rows — explicit remote sender wins; absent stays None so append_message's local-device auto-stamp (schema v17) applies unchanged.
• hermes_cli/web_server.py: new _presence_advertise_endpoint(host, port) + _primary_lan_ip(). Non-loopback dashboard binds advertise ws://<addr>:<port>/api/ws into HERMES_SESSION_PRESENCE_ENDPOINT via setdefault (operator-set env always wins). Wildcard binds resolve the primary outbound IPv4 via the UDP-connect trick (no packet sent, no DNS, no mesh tooling); loopback binds advertise nothing because the address is meaningless off-machine.
• Tests: tests/cli/test_presence_advertise_endpoint.py (5 cases) + sanitizer cases in test_concurrent_attach.py.

How to review

1. _sanitize_sender_device + the prompt.submit stash and _run_prompt_submit pop (one lock-scoped handoff).
2. conversation_loop.py — the user-dict tag + the API strip; confirm the strip sits with its four siblings.
3. run_agent.py flush — confirm None for non-user rows preserves assistant/tool NULL semantics.
4. web_server.py — the two helpers and the setdefault at startup (placed right after app.state.bound_host/port).

Evidence

• test_wildcard_bind_without_detectable_ip_advertises_nothing pins the fail-closed behavior; test_explicit_sender_wins_over_auto_stamp (from feat(state): messages.sender_device — device-level sender attribution (F-003 slice 1) #129) already pins the DB-level override this feeds.

Verification

• tests/cli/test_presence_advertise_endpoint.py + tests/tui_gateway/test_concurrent_attach.py + tests/tui_gateway/test_protocol.py + tests/test_sender_attribution.py — 79 passed.
• py_compile clean on all four touched modules.

Risks / gaps

• The full param path (remote ws client → prompt.submit → stored row) has no end-to-end test until a remote client exists — accepted scope, the three layers are unit-pinned and the integration work is the Phase 2 desktop client tracked in F-003-multi-participant-channels.
• prompt.background does not yet accept sender_device — accepted scope, background prompts are gateway-originated (no remote typer) today.
• Advertised endpoints assume the ws path /api/ws — low risk, it is the only mount point and the constant lives beside the helper.
• Token/auth hardening for non-loopback binds stays as designed in docs/channels-design.md (the --insecure opt-in contract is unchanged here) — covered by the Phase 2 listen-auth item on F-003-multi-participant-channels.

Collaborators

• @OmarB97 (operator)
• Claude Fable 5 (Claude Code)