Classifiers

What it does

Classifiers show how a classification flow is behaving, so you can see auto, hold, escalate, and blocked before you trust a threshold.

Open Classifiers (/settings/classifiers). The page is read-only usage: runs, gate mix, a confidence histogram, shadow disagreement, and cost or tokens where they were recorded. It does not edit flows or question sets. That editor is not in the product. Nothing on this page mutates a flow. Actions still happen on Approvals or Decisions.

The screen is a thin client over one GET. There is no MCP tool for these stats.

MCP tools

No MCP tool reads classifier usage. Related agent tools are the extraction review tools on Meetings and extraction: extraction_metrics and extraction_pending. Those count extraction cards. They are not this histogram.

HTTP endpoints

Method	Path	Role
GET	/cockpit/api/classifiers/stats	Per-flow usage. GET only

The response is { "ok": true, "flows": [ ... ] }. Each flow includes gate counts (auto, hold, escalate, blocked_by_rule), confidence buckets, and cost or token fields when they were stored. An empty list means no scored runs yet, not a broken page.

Permissions

The screen uses the signed-in session. There is no classifier scope on the permission-set layer. Extraction tools, if you use them for the related queue, need approvals:read.

Row access is RLS by creator. Admins see all rows. Members see their own.

Agent example

There is no tool call. A session client reads the same route the page uses.

{
  "method": "GET",
  "path": "/cockpit/api/classifiers/stats"
}

Expected result: one object per flow that has runs, with a gate mix. Flows with no scored answers come back with an empty confidence histogram. The call does not change thresholds.

{
  "ok": true,
  "flows": [
    {
      "flow_key": "activity.quality",
      "gate_mix": {
        "auto": 0,
        "hold": 12,
        "escalate": 1,
        "blocked_by_rule": 0
      }
    }
  ]
}