[tiensonqin]

Summary

Add explicit loopback attach for native test hosts and a separate public-XCTest iOS runner. Existing semantic defaults and native protocol v1 remain unchanged. LUI's Android host and DEBUG iOS read-only probe are integrated separately in logseq/lui#329; MyGo remains unsupported on its current SDK pin.

Live.connect_tcp ~host ~port / drive --tcp HOST:PORT accepts only IPv4 loopback, localhost mapped without DNS, or IPv6 loopback, and is mutually exclusive with other CLI transports. Forwarding and host startup remain explicit; transport never implies native capabilities or semantic fallback.

Native actions now share one deadline across writing, socket backpressure, framing and completion; timeout/failure poisons and closes the session. Per-call nonblocking sends preserve concurrent buffered readers. Fatal protocol errors now fail assertion/final polling too. Existing Live.create calls remain compatible; an optional deadline-aware send_line_until supports blocking custom transports, whose callback prerequisites are documented. Close callbacks are idempotent.

apple/ provides a standalone UI-testing-only Xcode generator and runner. A DEBUG app exposes read-only snapshots on loopback; all input stays in a separate XCTest process, using current unique native identifiers, native text replacement and correlated snapshot/result completion. An already-running atomic XCTest call cannot be forcibly interrupted. Guides distinguish sampled settling from network/application idle and list unsupported keys/editors/scenes. Apple CI compiles the runner with build-for-testing without executing UI tests or installing an app.

Validation: dune build @all, dune runtest (22 tests), Node syntax, Xcode project generation and git diff --check pass. The two new boundary regressions fail before the fixes. Real-socket Unix/TCP/WS backpressure/deadline and fragmented-read probes pass; they are protocol evidence, not native UI evidence. The standalone XCTest runner passes Apple-SDK compilation. Latest browser/CDP regression recording on 2aeaf98 passes and replaces earlier evidence in the existing PR comment. Native Android/iOS device input remains unverified.

Link to Devin session: https://app.devin.ai/sessions/9bf68438070747fb8b5c0c6718836a56
Open in Devin Desktop: https://app.devin.ai/desktop/session/9bf68438070747fb8b5c0c6718836a56?variant=devin
Requested by: @tiensonqin

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

Browser regression passed on 2aeaf98 after the bounded-write/fatal-poll fixes, through the real CLI → WebSocket → CDP → visible Chromium path with an isolated fixture.

Native click/tap, Unicode/empty replacement, focus/keys, wheel, held drag, auto-scroll and PNG capture passed immediate scenario assertions. Primary CLI exited 0 with 45 trusted / 0 synthetic observed events.

Held native drag	Primary completion

Failure protections and scope

Ambiguous, disabled, occluded, hidden and stale targets failed without executing the sentinel. Semantic-only native rejection, finite-animation completion, settle bounds and completion deadline checks passed.

A temporary harness listener-start race was corrected outside the repository and rerun. Exact Unicode content was asserted, not full glyph appearance (the VM lacks Chinese/emoji fonts). This is browser regression evidence, not send-backpressure, sticky malformed-protocol, TCP/Unix or mobile-device verification; those protocol seams have separate shell evidence.

Written by Devin