04 — Intake Triggers
The events and sources that feed the triage system, from connector syncs to AI session outputs.
Triage is triggered by events from multiple sources. All paths converge at the same triage boundary.
Trigger Sources
1. MCP Connector Session Save
When an MCP tool call completes, the connector saves session memory.
Path: MCP Connector → POST /v1/triage/session → Knowledge Service → triage-bot extraction
Payload:
{
"session_id": "uuid",
"summary": "Session summary text",
"memories": [{"text": "extracted memory"}]
}
Processing:
- Workers AI classifies priority, category, sentiment
- All items written as
pending(no auto-approve) - If spam or low confidence →
discarded
2. Conversation Archive
When a user archives a conversation in the AI Conversations system.
Path: IQ Hub POST /conversations/:id/archive → triage-bot extraction
Processing:
- Reads all messages from the conversation
- Workers AI extracts knowledge items (insights, facts, decisions, etc.)
- Each extraction written to
triage_resultsaspending - Optional Slack notification
3. Runtime TriageAgent
The Durable Object TriageAgent in iw-agent-runtime receives events from Twin sessions, worker handoffs, and manual triggers.
Path: Twin/Agent event → TriageAgent.triage() → classify() → emitTriageInput() → Continuity service
Processing:
- Simple pattern-based classification (not LLM)
- Routes to memory, proposal, signal, or governance
- Memory/signal routes emit to Continuity service
- Items stored in Durable Object SQLite
4. Manual Trigger
Admin or automated system triggers triage via API.
Path: POST /v1/triage/run → Knowledge Service → consolidation pipeline
Payload:
{
"type": "hourly|daily|weekly"
}
Processing:
- Fetches approved triage items in time window
- Groups by category
- Generates consolidated memories using Workers AI
- Writes to
consolidated_memoriestable
5. Triage Workflow (Durable)
The Cloudflare Workflow TriageWorkflow processes high-stakes items that require cognitive evaluation.
Path: Workflow queue → TriageWorkflow.run() → TriageAgent.execute() → route
Processing:
- Cognitive evaluation across 4 dimensions
- Routes to HITL DLQ or terminal rejection
- No direct canonical writes
Intake Convergence
All triggers converge at the same boundary:
All Sources → Triage Intake → classify → stage → governance review
The source determines the extraction method (LLM vs pattern), but all outputs are candidates — never approved organizational memory.