Summary

• boxlite run <img> <fast-cmd> against the cloud (Go runner) intermittently returned exit 0 with empty stdout for commands that exit almost instantly (e.g. uptime). The command ran inside the box, but its output was silently dropped before reaching the client (~80% over the cloud path; local boxlite serve masked it via low attach latency).
• Root cause is in the REST client WS attach pump (src/boxlite/src/rest/litebox.rs): when the attach WebSocket drops before an exit frame (a proxy / high-latency cut), attach_ws_pump probes GET /executions/{id}; on a Terminal ("completed") result it emitted the exit code and returned without re-attaching — dropping the stdout still buffered in the runner's replay backlog.
• Fix: on a disconnect-then-Terminal probe, re-attach once, immediately (no backoff) so the runner replays its backlog and we exit via its authoritative exit frame. Bounded by a terminal_drain_attempted flag, with fallback to the probed exit code so a runner that never sends a closing exit frame can't hang the pump.

Independent of boxlite-ai#563 (an sdks/c FFI drain fix): this is the REST client path, so the bug reproduced on both main and the boxlite-ai#563 branch. They are orthogonal fixes for the same user-visible symptom.

Tradeoff

If the first connection delivered partial stdout before being cut, the re-attach replays the runner's full backlog, so the already-delivered prefix can be duplicated. This matches the existing StillRunning/Unavailable reconnect semantics (the runner's streamBus replays without an offset) and is strictly preferable to silent loss. Eliminating it entirely needs offset-based replay in the attach protocol (out of scope here).

Test plan

• New regression test ws_terminal_probe_after_cut_must_not_drop_buffered_stdout (reuses the in-process mock-WS harness): cuts the first attach, reports completed/exit 0, serves the backlog only on re-attach.
• Two-side verified on a main base: fails without the production change (left:"" right:"hello\n", clean exit 0), passes with it.
• Existing WS attach tests pass: ws_clean_exit_emits_result, ws_close_without_exit_falls_back_to_status, ws_text_error_frame_logs_but_continues.
• cargo fmt --check + cargo clippy -p boxlite --features rest --lib -D warnings clean on the change.

Note: ws_watchdog_fires_when_idle fails in my local sandbox (the mock server's accept loop is blocked by the test's own 2s sleep, blowing the test's 3s budget). It reproduces on pristine main/boxlite-ai#563 and is unrelated to this change.

🤖 Generated with Claude Code