Summary

• bug(a2a): message/send returns completed task with no artifacts when LLM response is empty #2302 — message/send returned completed with no artifacts on second and subsequent requests. Root cause: after AgentTaskProcessor consumed LoopbackEvent::FullMessage and broke out of the recv loop, a trailing LoopbackEvent::Flush (emitted by flush_chunks()) remained buffered in output_rx. The next request consumed this stale event first, produced an empty ArtifactChunk, and broke — never seeing the real LLM response. Fix: drain output_rx with try_recv() after every recv-loop exit so no stale events survive into the next request.

• bug(a2a): daemon PID file not cleaned on abnormal exit — restart requires manual cleanup #2295 — A PID file left by a crashed or SIGKILL'd daemon blocked restarts with WARN: failed to write PID file: File exists. Fix: before writing the PID file, read any existing file and check whether the stored PID refers to a live process (kill -0). If stale → remove and proceed. If alive → bail with an actionable error message.

Changes

• crates/zeph-core/src/daemon.rs — add is_process_alive(pid: u32) -> bool (Unix: kill -0, non-Unix: always false)
• src/daemon.rs — drain loop after recv loop (while let Ok(_) = handle.output_rx.try_recv() {}); stale-PID check in run_daemon before write_pid_file
• src/tests.rs — 5 new tests covering both fixes
• CHANGELOG.md — [Unreleased] section updated

Test plan

• cargo +nightly fmt --check — passes
• cargo clippy --features full --workspace -- -D warnings — passes
• cargo nextest run --config-file .github/nextest.toml --workspace --features full --lib --bins — 6915/6915 passed
• New tests: is_process_alive_current_process, is_process_alive_nonexistent_pid, loopback_stale_flush_drained_after_full_message, stale_pid_detection_dead_process, stale_pid_detection_live_process