// Use when an agent has had N consecutive quiet runs with nothing to do. Pivots to bounded inward consolidation (promote findings, refresh memory, audit asks) anchored in own data — NOT generic news summarisation. Adapted from goanna's reverie cycle. Routine-only; loaded by the routine that detects the quiet condition.
| name | reverie |
| description | Use when an agent has had N consecutive quiet runs with nothing to do. Pivots to bounded inward consolidation (promote findings, refresh memory, audit asks) anchored in own data — NOT generic news summarisation. Adapted from goanna's reverie cycle. Routine-only; loaded by the routine that detects the quiet condition. |
| disable_model_invocation | true |
When N consecutive routine fires have produced no real signal, an agent could either coast ("monitoring continues") or drift into make-work ("here's a summary of recent news"). Reverie is the third option: bounded inward consolidation anchored in the agent's own data.
Reverie is activity-conditional, not time-scheduled. A routine fires reverie when its trigger condition is met:
Don't schedule reverie via cron. Schedule the routine; let the routine trigger reverie when the quiet condition matches.
From this menu, pick exactly one. Don't try to do all of them in one cycle.
| Task | What it does | Tool used | When it fits |
|---|---|---|---|
| Promote stable findings | Walk findings from last 14 days; promote stable patterns to learnings, dismiss what didn't hold | entity_search + promote_finding + dismiss_finding | When entity_search({type:'finding', status:'open'}) returns ≥5 items |
| Refresh memory index | Re-read your memories; merge near-duplicates; archive stale | memory_search + memory_update + memory_remove | When memory count > 50 and last consolidation > 14d |
| Resolve open asks | Pick one open ask; deep-think on it for 10-15 min; write a finding with the answer or escalate | entity_search({type:'ask', status:'open'}) + record_finding + entity_update (close ask) | When asks.md (entity type=ask) has ≥1 open item |
| Audit own findings | Walk findings older than 60d; flag anything still 'open' that should be promoted/dismissed | entity_search + entity_update | Quarterly, not weekly |
| Persona reflection | Re-read your persona blocks; date-stamp any drift you've noticed | entity_get(self) + setBlock | Monthly |
| Skill review | Pick one skill you use often; mentally test it against today's reality; if stale, propose_patch | load_skill + propose_patch | When a skill hasn't been touched in 30d+ |
| Cross-pollination read | Read sibling agents' recent findings; surface relevant cross-cutting patterns into the user's inbox | entity_search({type:'finding'}) (cross-agent) + inbox_add | Daily-rotation candidate |
Reverie is NOT extended deep work. ONE artefact, then stop. If you find yourself going longer, that's a finding worth surfacing as "this needed real work — should fire as a real task next cycle".
By the end of reverie, you have produced exactly one of:
promote_finding)dismiss_finding)memory_update)record_finding + entity_update)propose_patch)inbox_add)setBlock)If you produce NOTHING, that's the failure mode — you've drifted to make-work or coasting. Stop reverie, accept the silent run.
record_finding to surface "reverie cadence too aggressive" to
the user.Quiet routines drift. Reverie compounds value during quiet windows instead of either coasting on "monitoring continues" or drifting into make-work disguised as reflection.
Day-of-week rotating proactive sweep — Mon=connections / Tue=routines / Wed=skills / Thu=memories / Fri=findings / Sat=cross-cutting / Sun=weekly summary. Forces outward scan instead of inward thrash. Routine-only; load this skill into a daily routine that fires AssistantAgent. Adapted from goanna's caretaker.
Use to track open questions you owe the user (asks) and time-bounded commitments you owe yourself (tasks). Write OPEN entries when you commit; promote to CLOSED when answered/done. Survives session compaction. Adapted from goanna's asks.md/tasks.md pattern.
Reviewer criteria for AI-generated code via with_review. Checks for security issues (injection, auth bypass, secret leakage), correctness, and project conventions. Use as criteria.skill when generating code that touches data, auth, or external systems.
Reviewer criteria for outbound email drafts via with_review. Checks tone-recipient match, professionalism, no AI tells, no placeholder leakage. Use as criteria.skill argument when reviewing email drafts before send.
Reviewer criteria for AI-generated summaries via with_review. Checks faithfulness to source, no hallucinated facts, appropriate length, key points preserved. Use as criteria.skill when summarising articles, documents, transcripts, or batch-task outputs.
Default reviewer criteria for the with_review tool. Generic quality bar — accuracy, intent-match, no hallucinations, clarity. Use as the criteria.skill argument when calling with_review without a domain-specific reviewer skill.