Preserve shared-session viewer prompts that never reach the sharer

[warp-agent-staging]

Description

An Agent Mode prompt submitted by a shared-session viewer could disappear. Routing is decided from SharedSessionStatus, which stays ActiveViewer while the viewer websocket is reconnecting, so the prompt was handed to the viewer network — and Network::send_message_to_server returned silently for any stage other than JoinedSuccessfully. The prompt was neither delivered nor retained, and the input could stay frozen with a stale Warping... indicator indefinitely.

This makes the send observable and the prompt recoverable, using the existing queued-prompts panel as the single durable surface rather than introducing a second queue.

1. Viewer prompt acceptance is observable and correlated. send_message_to_server now returns ServerMessageSendOutcome::{LocallyQueued, Undeliverable} and warns on every local rejection with the network stage, the session id, and a content-free UpstreamMessage kind (an exhaustive static label — the message itself is never Debug-rendered, because prompt text and attachments are nested inside it). send_agent_prompt_request takes the caller's AgentPromptRequestId so the server's AgentPromptRequestInFlight echo can be matched to the prompt that was staged. try_send acceptance is explicitly not treated as delivery; the correlated acknowledgement is, with a five-second bound.

2. Undelivered prompts land in the existing queue. A rejected send, or one the sharer does not acknowledge within five seconds, becomes an unlocked QueuedQueryOrigin::DisconnectedViewer row carrying the prompt, its pending attachments, and the session + server conversation token it was addressed to. The input unfreezes immediately, and a turn that this submission itself started is no longer left reported as in progress — a turn that was already running when the prompt was submitted is left alone, so a genuinely active stream keeps its indicator.

3. Dispatch is transactional. Every prompt row carries a stable request id that commit_edit re-mints, and claim_prompt_head hands a row to exactly one dispatch attempt. Racing lifecycle events cannot both submit it, a failed attempt restores it at its original position, and a late or duplicate acknowledgement can only finalize the unchanged row it belongs to — never delete a newer edit.

4. Three explicit drain triggers. A same-session rejoin retries the FIFO head (after RejoinedSuccessfully has flushed buffered input updates); a fatal ambient end converts the head into the cloud follow-up that starts the replacement execution; and the replacement network's JoinedSuccessfully — not ExecutionSessionReady, which fires before it connects — continues the queue. Each validates the row's session and conversation token first and leaves a non-matching head queued rather than redirecting it.

5. Queued cloud follow-ups keep their attachments. The stale "cloud follow-up does not support attachments" comment and its drop are gone; the row's attachments now go through upload_pending_attachments_to_task before dispatch, and the row is restored on upload failure.

Linked Issue

APP-5223. The approved spec is committed on this branch at .agents/specs/APP-5223-buffer-disconnected-viewer-prompts.md and is unchanged by these commits.

Testing

Reproduction test, verified fail-before / pass-after. viewer_prompt_submitted_while_reconnecting_is_preserved_as_an_editable_queue_row drives a prompt through the real input → view → terminal-manager → network path with the viewer network in Stage::Reconnecting. Disabling the fallback-queueing branch makes it fail with exactly the pre-fix symptom (zero queue rows); it passes with the fix. It detects the defect rather than restating the fix.

Also added:

• viewer/network_tests.rs — Undeliverable for every non-joined stage and for a closed proxy channel; LocallyQueued only under the caller's request id; the drop diagnostic reports stage/session/kind and contains none of the sentinel prompt, attachment-name, or attachment-id content; the kind label does not vary with the payload.
• viewer/terminal_manager_tests.rs — the joined happy path leaves no fallback row, and an unknown or duplicate acknowledgement is a no-op.
• queued_query_tests.rs — the disconnected-viewer row stays unlocked and deletable; the head is claimable exactly once; a restored claim returns to its original index; retargeting preserves row identity, text, and attachments without duplicating; a locked, command, or being-edited head is not claimable; commit_edit retires the request id.

Repository gates (all from the repo root):

• ./script/format --check — clean
• ./script/check_no_inline_test_modules — clean
• cargo clippy --workspace --exclude warp_completer --all-targets --tests -- -D warnings — clean
• cargo clippy -p warp --all-targets --tests -- -D warnings — clean
• cargo clippy -p warp_completer --all-targets --tests -- -D warnings — clean
• cargo nextest run -p warp --no-fail-fast — 6225/6228 passed. The three failures (ambient_agent_headers_for_task_overrides_existing_cloud_agent_header, test_histignorespace_support_in_zsh, test_decorations_with_multibyte_chars) reproduce identically on master in the same sandbox and are unrelated to this change.
• cargo test --doc -p warp — passed
• The base was merged in and the touched-module tests re-run green afterwards.

Still outstanding before this is ready to merge

The spec's validation section lists 20 criteria. The ones below are not yet satisfied, and the PR is deliberately left as a draft because of them. They are listed rather than quietly dropped.

• Criterion 20 (running-UI verification). Not performed. It needs the real Warp GUI driven against a live shared ambient session — an authenticated Warp account, a real cloud agent VM, and the ability to induce both a transient websocket reconnect and reconnect exhaustion. None of that is reachable from the sandbox this was built in (no GPU, no Warp build, no credentials), so it needs a credentialed run before merge.
• Criteria 5, 7, 9, 11, 12, 13, 14, 15, 17. The behavior each describes is implemented, but the test matrices are not written yet: the async attachment race, the acknowledgement-timeout boundary cases, the rejoin mismatch matrix, the ineligible-fatal matrix, the replacement-join duplicate-send cases, ordering across all triggers under delayed futures, and the queued-cloud attachment success and failure paths.

Screenshots / Videos

None yet — see criterion 20 above.

Agent Mode

• Warp Agent Mode - This PR was created via Warp's AI Agent Mode

CHANGELOG-BUG-FIX: Agent prompts sent while a shared session is reconnecting are no longer lost; they are kept in the queued prompts panel, the input is released immediately, and they are retried when the session reconnects or continued as a cloud follow-up.

Originating thread: https://staging.warp.dev/conversation/613eb99f-d7d2-4fd8-9d29-f7e817cdee85