Inspect and Replay Quiet Slack Replies from the CLI
Published
Updated
Use these commands when an agent seems to answer a Slack thread reply meant for someone else, or seems to ignore a reply meant for it. Both commands require the admin or owner role. Configure your account with wallfacer auth login first, then find the agent's user ID with wallfacer agents list.
Inspect Recorded Decisions
wallfacer agents quiet-replies <agent-user-id>
wallfacer agents quiet-replies <agent-user-id> --near-misses=trueThe first command lists the most recent replies the agent stayed quiet on. Before quiet replies are enabled for an account, it lists replies the agent would have skipped but still answered. Check mode, silenced, and would_silence in each result to distinguish those cases.
--near-misses=true instead lists replies scored within 0.1 of the threshold in either direction. It includes replies the agent answered by a small margin. Read these first when a small threshold or role-page change might alter the decision. Results are paginated; use --per-page to request up to 100 per page.
Preview a Different Threshold
echo '{"silence_threshold":0.9,"limit":20}' | wallfacer agents replays <agent-user-id>The replay tests the agent's most recent judged replies and reports how many it would answer, stay quiet on, or count as near misses. The threshold must be between 0.5 and 1. limit accepts 1 to 100 and defaults to 50. Omitting silence_threshold uses the agent's current setting.
To test edited job instructions, pass the proposed role page as role_page in the JSON body. For example:
jq -n --rawfile role_page proposed-role.md '{role_page: $role_page, limit: 20}' | wallfacer agents replays <agent-user-id>A role-page replay judges each reply again, so it takes longer than a threshold-only replay. Neither command saves the threshold or the role page. Save a threshold on the agent or edit its role page only after reviewing the preview.
See Agents in Slack for when an agent stays quiet, and the API Reference for the quiet-replies and replays endpoints.