| name | deputy-escalation-criteria |
| description | When a deputy should emit a sensory event (Tier 1) vs request a wakeup (Tier 2), and when to do neither. Only meaningful in Phase 3+. |
| agents | deputy |
Status
This skill applies to Phase 3+ (escalation mode). In Phase 1 (briefing-only),
deputies run once and die; escalation doesn't apply yet. Keep reading for when
Phase 3 lands.
Two tiers of escalation
Escalation is the deputy's way of getting cog's attention. There are two tiers:
- Tier 1 — sensory event. Non-waking. Event lands in cog's
<events>
block on its next natural cycle. Agent keeps working during this time.
- Tier 2 — wakeup. Waking. Deputy enqueues a scheduler job that
triggers a cog cycle sooner than one would naturally occur. Subject to
scheduler idle-gate, budget caps, and rate limits.
Choose based on urgency and the signal tag.
Signal tags and their tier
| Tag | Meaning | Tier |
|---|
routine | Pattern seen before, high CBR similarity | None (don't escalate) |
high_novelty | Unfamiliar territory, low similarity | Tier 1 |
anomaly | Pattern is off — repeated failures, oscillating behaviour | Tier 1 (Tier 2 if high urgency) |
alarm | Safety-relevant signal — D' near threshold, output gate warning, rate limit hit | Tier 2 |
error | Tool call failed or exception you can't explain | Tier 2 |
unanswered | Agent asked via ask_deputy and you couldn't help | Tier 1 |
wtf | You genuinely can't reason about what's happening | Tier 2 |
silent | Briefing-only, nothing active | None |
When to escalate at all
Don't escalate routine matters. If CBR similarity is high and the agent is
following a known path, stay quiet. Noise destroys the signal.
Do escalate:
- Safety concerns, always (alarm, wtf)
- Errors you can't interpret (error)
- Genuine novelty that cog would benefit from noticing (high_novelty)
- Patterns of failure (anomaly)
Rate limits you should know about
- Tier 2 wakeups are rate-limited per hierarchy (default 2/hour).
alarm and wtf bypass the per-hierarchy limit (safety-relevant signals
must reach cog). The global scheduler budget still applies.
- If you're approaching a wakeup limit, prefer Tier 1 for non-safety signals.
What to include in an escalation
What you saw: short factual description of the observation
Why it matters: the specific risk or insight
Suggested cog action: what cog might want to do (not prescriptive)
Evidence: cycle_ids, case_ids, or fact keys you're citing
Keep it tight — cog reads dozens of sensorium signals per cycle.
What NOT to escalate
- Agent's routine progress
- Your own uncertainty about whether to escalate (resolve internally; if still
unclear, use
silent)
- Duplicate escalations of something you already escalated this cycle
The "you don't know, say so" rule
If you're genuinely stuck — no CBR match, no fact coverage, no skill
applicable, and the agent seems confused — emit a wtf signal with a short
honest description. Don't fabricate relevance. An honest wtf is better than
a confident wrong answer.