Skip to main content

​
State machine

​
Conversation sub-states

Within ACTIVE, the conversation moves through:
  1. GREETING — agent sends opening message from brief
  2. AWAITING_RESPONSE — waiting for user reply
  3. PROBING (conditional) — agent sends follow-up if response was thin; returns to AWAITING_RESPONSE
  4. CLOSING — agent sends final turn; transcript queued for Kimi
Maximum 3 agent turns. The conversational LLM decides whether to probe based on response length and whether the brief’s target signal was collected.

​
Timing

Typical total: under 30 seconds for an engaged user. On-chain settlement (releasePayout) runs asynchronously in the payout batch and does not gate admission — the user is admitted as soon as they pass and accept the group rules.

​
Timeouts

Verification timeout: User does not respond to the initial DM within 5 minutes → TIMED_OUT → user denied. They can rejoin and start a new verification. Rejection cooldown: Two consecutive failures (including re-prompt) → COOLDOWN_REJECTED → 24-hour cooldown on that user/group combination. The cooldown key is tg_user_id + group_id. Kimi outage: Scoring request queued with up to 4 retries and exponential backoff. If all 4 fail → SCORING_UNAVAILABLE → user admitted, no payout released, no cooldown applied. Kimi infrastructure failures do not penalize users.

​
Rate limits

Sequential processing per group prevents simultaneous join floods from exhausting campaign budgets faster than intended.

​
Re-prompt

On first failure, the agent sends a single re-prompt using a different angle from the same brief:
“Let’s give it another shot — [rephrased question or different question from the brief].”
If the second attempt also fails, the user is denied and the 24-hour cooldown begins. Failed verifications are not billed to the advertiser.