mirror of
https://github.com/agentscope-ai/ReMe.git
synced 2026-10-23 05:35:30 +00:00
a0f66c1779
The refresh chain decides what is worth proposing but cannot know what the host actually surfaced. Let a host agent report an agenda item back, so the next round withholds it instead of proposing the same thing twice. - proactive_feedback job: zero-LLM intake, idempotent upsert keyed on (topic_id, external_ref) and falling back to (topic_id, agenda_date, channel); status only moves forward out of a terminal value, and a failed report never blocks a retry - truth source: optional ``deliveries`` section in daily/_proactive.yaml, trimmed by retention_days then max_deliveries, and salvaged entry by entry with the rest of the file - reader side: proactive_read withholds agenda items inside delivery_cooldown_days and returns delivered_ids so a caller can see the suppression instead of guessing; include_delivered opts out - agenda side: a reported candidate leaves the set before reminders are pinned and before the editor call, so the cooldown cannot be argued away by narrative coherence, and it is suppressed with an auditable "already delivered ..." reason - tests: receipt intake, reader-side withholding and agenda-side suppression, each paired with an empty-deliveries no-change case, plus a round-trip of the report shape a consumer actually sends Reporting stays optional: a host that never calls proactive_feedback keeps an empty deliveries section and sees no behaviour change. |
||
|---|---|---|
| .. | ||
| fixtures/config | ||
| integration | ||
| unit | ||