Deduplicate last assistant message after harness restart

[cbusillo]

Summary

After restarting/resuming the Every Code harness, the TUI can render the assistant's last response twice in the chat history.

Evidence

Observed on 2026-06-02 immediately after a harness restart on main:

• Header showed Every Code Lab, model GPT-5.5 (High), directory ~/Developer/code, branch main.
• The user's prompt Ok what are we working on now? appeared once.
• The assistant response beginning Right now the active workstream is dogfooding Auto Review after the diagnostics slice. appeared twice consecutively.
• The duplicate content looked identical and visible in the scrollback, not merely a transient repaint.

Why This Matters

Duplicate rendered assistant messages make restart/recovery feel unreliable and can confuse users about whether the model repeated itself, whether history replay duplicated a message, or whether hidden context/tool actions were also duplicated.

Desired Behavior

On resume/restart, each assistant response should appear once in visible history. If replay and live stream both reference the same response, the TUI should dedupe by stable turn/item/message identity rather than rendering both.

Questions To Answer

• Is this duplicate only a visible TUI replay/rendering artifact, or is the persisted conversation history itself duplicated?
• Does it affect only final assistant messages, or also reasoning/streaming inserts/tool cells?
• Is the duplication tied to restarting while the final answer is still being persisted or replayed?

Acceptance Criteria

• Add a focused resume/restart replay regression test for a final assistant answer.
• Verify persisted history contains one logical assistant response for the turn.
• Ensure replay plus live/last-message recovery does not double-render the same response.
• Confirm no duplicate hidden context/tool actions are emitted when the visible duplication occurs.