Review: fix(ui): stop New Task from mirroring the previous session's live stream

Summary

Solid, well-scoped fix for #1580 — the root cause (global isStreaming/streamingContent/agentSteps in the chat store, written by a backgrounded ChatView until its delayed unmount) is correctly diagnosed and addressed on both halves: reset the globals at swap time and guard every stream-write path by sessionId. The guard coverage is complete: all five paths the description names plus three nested handlers (ChatView.tsx:1009,1141,1165) check currentSessionId before writing, and the guard keys off the store's currentSessionId — which flips synchronously on switch — rather than the 220ms-lagged displayedSessionId, so stale writes are blocked from the instant the user navigates away, with resetStreaming() as the final cleanup. No blocking issues.

Issues

🟢 Repeated guard expression could be a named helper (ChatView.tsx:286,682,726,967,1041,1009,1141,1165)
useChatStore.getState().currentSessionId !== sessionId appears eight times. A tiny local reads better and documents intent at each call site — purely optional:

const isStale = () => useChatStore.getState().currentSessionId !== sessionId;

Then each guard becomes if (isStale()) return;. Skip if you prefer the explicit form.

🟢 Switch-away-then-back mid-stream is worth a manual pass
resetStreaming() wipes streamingContent on every swap, so returning to a still-streaming session before it completes relies on the per-instance streamBufferRef / server reload to restore partial content. Not a regression and likely fine, but it's an edge case the two unchecked manual-test items don't cover — worth a quick look when you run them.

Strengths

• Guard keyed on the store's currentSessionId (not the lagged displayedSessionId) is the right call — the comment at App.tsx:500 makes the 220ms transition relationship clear.
• resetStreaming lands as one atomic set() in the store (chatStore.ts:191) rather than three separate calls scattered across the view, and chatStore.test.ts covers both the populated→clean transition and the no-op-safe path.
• CI pause is reversible-by-design and self-documenting: the false && guards plus the trigger comment spell out exactly what to restore (Agent Hub: Publish Python agent wheels to PyPI #1179), and the paired test_agent_pypi_publish.py update keeps the assertion in sync with the OIDC/id-token: write reality instead of the removed token secret.

Verdict

Approve. Clean root-cause fix, full guard coverage, tests updated alongside both the UI and CI changes. The two 🟢 notes are optional — the guard-helper is cosmetic and the switch-back case just deserves a glance during the manual test you've already queued.