The installed estack-read-agent-history 4.1.4 returns zero sessions/messages for current Codex Desktop rollouts that contain response_item messages (user input_text and assistant output_text) but no event_msg user_message/agent_message mirrors. normalize_rollout drops response_item message entirely, so a September 29 session-report returns an empty morning despite genuine prompts.
Expected: normalize response_item message when the event mirror is absent, convert input_text/output_text to the shared text shape, and deduplicate by message ID or verified mirror identity when both streams exist. Ignore injected environment/AGENTS messages, approval review sessions, heartbeat jobs, and replayed child-agent context. Verified the underlying records using lib.parser.iter_lines, then used a temporary compatibility adapter to reconstruct the bounded morning.
The engagement engine also credits >5-minute gaps when the assistant was working; Elliot's requested accounting explicitly excludes those and adds one minute for isolated prompts. This is a separate configurable accounting requirement, not grounds to silently change the default for other users.
The installed estack-read-agent-history 4.1.4 returns zero sessions/messages for current Codex Desktop rollouts that contain response_item messages (user input_text and assistant output_text) but no event_msg user_message/agent_message mirrors. normalize_rollout drops response_item message entirely, so a September 29 session-report returns an empty morning despite genuine prompts.
Expected: normalize response_item message when the event mirror is absent, convert input_text/output_text to the shared text shape, and deduplicate by message ID or verified mirror identity when both streams exist. Ignore injected environment/AGENTS messages, approval review sessions, heartbeat jobs, and replayed child-agent context. Verified the underlying records using lib.parser.iter_lines, then used a temporary compatibility adapter to reconstruct the bounded morning.
The engagement engine also credits >5-minute gaps when the assistant was working; Elliot's requested accounting explicitly excludes those and adds one minute for isolated prompts. This is a separate configurable accounting requirement, not grounds to silently change the default for other users.