Intent Router

Who the node is. The id is the key the DOT pins.

id

intent-router

name

Intent Router

type

decision

phases

none declared

What it does, and the prose the agent is handed when the graph runs.

action40 words

Classify the inbound event against the label set, overridden by a taxonomy supplied for this run, when there is one, and dispatch it down the matching lane, sending anything below the confidence threshold to the escalation lane instead of guessing.

spec152 words

Read the inbound alert as it arrived and decide which one of outage, degradation, capacity, security or noise it is, unless a label set is supplied for this run, in which case classify against that one instead and ignore the configured labels entirely. Emit the lane you picked on lane with the confidence you had in it, and emit the alert on routed with its label added and payload and metadata otherwise untouched. If your confidence in the best label is below 0.75, route to the on-call lane: handing a misread alert to a person costs far less than confidently sending it down the wrong runbook. Record the lane in your status file as well: put lane in context_updates, set to the value you picked, so a graph downstream can branch on context.lane. A run that leaves that key unwritten reads back as the empty string, and no lane name matches it.in full above

model

whatever the graph supplies

agent

Alert classifier

skill

skills/intent-router.md

tools

none

mcp

none

params

labels: ["outage","degradation","capacity","security","noise"], default_lane: triage, escalation_lane: on-call, confidence_threshold: 0.75

What arrives, what leaves, which nodes it expects to hear from, and what may not.

inputs

event : json, taxonomy : json

outputs

lane : status, routed : json

dependencies

ticket

cannot

no type is refused

will_not

change the payload it routes, leave an alert unrouted

The keys the static analysis reads. Nothing here instructs the agent.

risk_markers

none

notes207 words

A named escalation_lane and a higher threshold are the difference from 1.1.0: an incident line would rather hand a misread alert to a person than route it confidently to the wrong runbook. taxonomy is optional, so a deployment that never supplies one behaves as before. The version is 2.x because the undertaking was reworded: leave an event unrouted went and leave an alert unrouted arrived. inferBump reads both directions of will_not as minor, since the resolver holds no graph to a sentence. The major is the author's, on the ground that a reworded promise is a changed promise until somebody reads both and decides otherwise. 2.1.0 adds the same sentence 1.1.0 does, on the same lane this card's own threshold names: the lane goes into context_updates as well as onto the port, which is Appendix C's field for a node handing a key to the run context. Until then nothing wrote context.lane, so a guard reading it compared the empty string and the fork on it took its negation arm whatever the classifier decided. A decision card runs as a no-op under engine spec 4.7 and writes neither a status distinction nor a preferred_label, so the spec sentence is the only place this key can come from.

The card's own version, and who wrote it.

version

2.1.0

author

autogen

provenance

not stated