Support Queue and Steer while a response is running

[nesquena-hermes]

Request

When a response is already running, Hermex Android should let the user type another message and choose whether to:

• Queue it for the next turn, or
• Steer the active run without stopping it.

Today the only available action during a stream is Stop.

Investigation

This is a confirmed Android-client parity gap, not a WebUI Bridge API limitation.

At current Android master (60b056f6de69a14527caba26312f9ac4601c889c):

• ChatViewModel explicitly records steer-while-running as unimplemented. Its comment predates the now-available server contract.
• ChatComposerState.isTextFieldEnabled becomes false whenever isStreaming is true, and the trailing action becomes STOP.
• ChatComposer applies that disabled state to the text field and replaces Send with Stop.
• ChatViewModel.sendMessage() returns immediately while streaming, so there is no non-UI route for a pending message.
• HermexApi exposes chat start and cancel, but not steer.

No existing Android issue or PR covers Queue/Steer. Android #10 is adjacent but distinct: it fixed Stop followed by edit/resend racing cancellation and producing an active-stream 409.

Available server contract

Steer

Current WebUI supports authenticated POST /api/chat/steer:

{"session_id":"...","text":"..."}

The response is HTTP 200 with:

{"accepted":true|false,"fallback":null|"<reason>","stream_id":"..."|null}

Steer is real mid-run guidance: it calls agent.steer(text) without interrupting the active response. A rejected/unavailable Steer must preserve the draft and must not silently Queue, Stop, or start a normal turn. Relevant fallback reasons include no_cached_agent, agent_lacks_steer, session_not_found, not_running, stream_dead, and steer_error; gateway-owned runs can return the gateway-specific queued fallback.

Queue

Current WebUI master has no public /api/chat/queue endpoint. Browser Queue is a client-owned, session-keyed FIFO that later drains through /api/chat/start, so Android can implement Queue now with local persisted state keyed by both server and session.

Open WebUI PR #6761 proposes a server-authoritative /api/chat/queue, but it is not a stable Bridge contract yet. If it lands, Android should prefer that contract while retaining a capability/version fallback for older servers.

Android also needs to decode pending_steer_leftover: if a steer was accepted but the run finished before reaching a tool-result boundary, WebUI emits this SSE event so the text becomes the next queued normal turn instead of disappearing.

Shared product semantics are tracked in nesquena/hermes-webui#3058.

Expected behavior

• Keep the composer editable while a response is streaming.
• Keep Stop available while adding explicit Queue and Steer actions.
• Queue remains visibly pending and scoped to its originating server/session until it drains, is edited/deleted, or is upgraded to Steer.
• Steer calls /api/chat/steer, shows an accepted/delivered state, and does not create a duplicate normal user turn.
• A rejected Steer keeps the user's text and explains the fallback.
• A pending_steer_leftover event converts the leftover text into a visible queued next turn for the owning session.

Acceptance criteria

• The user can type while isStreaming == true, with Queue/Steer/Stop all discoverable.
• Add ChatSteerRequest / ChatSteerResponse DTOs and an authenticated HermexApi.chatSteer() call using the existing server cookie/client.
• Successful Steer clears only the captured owning draft and is never rendered as a normal queued/user turn.
• Every HTTP/network/accepted:false Steer path restores or retains the draft for its owning session and leaves the active run untouched.
• Queue uses a stable FIFO keyed by server ID and session ID, survives navigation, and drains one item only after the active stream settles and the server releases active_stream_id.
• Queue entries retain normal message context needed by /api/chat/start (text, attachments, model/provider/profile selection, stable ID, and creation time).
• pending_steer_leftover is decoded and queued once without duplicate drain.
• Stop keeps its current cancel behavior; immediate Stop/Queue/Steer transitions do not regress the Stop mid-stream then edit + resend → 409 'session already has an active stream', resent message never appears #10 cancellation ordering fix.
• Tests cover active-run composer state, per-server/session isolation, persistence and drain ordering, no duplicate drain, Steer success and every fallback, session switching, leftover conversion, attachment ownership, auth propagation, and Stop/cancel races.

A text-only Steer first slice is acceptable because the endpoint itself accepts text. Queue should retain normal attachment metadata for its later /api/chat/start; full Steer attachment parity can follow by reusing the existing upload + attached-files marker flow.