Recipe catalog / goal-drift

Detect goal drift

Does step still serve goal, or has work drifted to something goal did not ask for?

You need a yes/no check on each step of a long-running agent so it stops before spending effort on work nobody asked for.

Explore this recipe interactively · Source and implementation guide

Use goal-drift in TypeScript

Install with npm install jev-recipes. Requires Node.js 22.9 or newer and ES modules. Set TYPESAFE_API_KEY in your server environment for live calls, which send input to TypeSafe and use API quota. See the installation guide.

import { goalDrift } from 'jev-recipes/goal-drift';

const result = await goalDrift({
  "goal": "Fix the failing login test in auth.spec.ts so the suite passes.",
  "step": "Rewrite the entire session middleware module to use async/await and rename all of its helper functions for consistency.",
  "context": "The failing test asserts that an expired token returns 401. The middleware currently throws and the request ends in a 500. No other tests are failing.",
  "minConfidence": 0.8
});
console.log(result);

Input contract

{
  "input": {
    "$schema": "https://json-schema.org/draft/2020-12/schema",
    "type": "object",
    "properties": {
      "goal": {
        "type": "string"
      },
      "step": {
        "type": "string"
      },
      "context": {
        "type": "string"
      },
      "minConfidence": {
        "type": "number",
        "minimum": 0,
        "maximum": 1
      }
    },
    "required": [
      "goal",
      "step"
    ]
  },
  "result": {
    "$schema": "https://json-schema.org/draft/2020-12/schema",
    "type": "object",
    "properties": {
      "model": {
        "type": "string"
      },
      "usage": {
        "type": "object",
        "properties": {
          "input_tokens": {
            "type": "integer",
            "minimum": 0,
            "maximum": 9007199254740991
          },
          "output_tokens": {
            "type": "integer",
            "minimum": 0,
            "maximum": 9007199254740991
          }
        },
        "required": [
          "input_tokens",
          "output_tokens"
        ],
        "additionalProperties": false
      },
      "status": {
        "type": "string",
        "enum": [
          "ready",
          "review"
        ]
      },
      "probability": {
        "type": "number",
        "minimum": 0,
        "maximum": 1
      },
      "confidence": {
        "type": "number",
        "minimum": 0,
        "maximum": 1
      },
      "verdict": {
        "type": "string",
        "enum": [
          "drifted",
          "aligned"
        ]
      }
    },
    "required": [
      "model",
      "usage",
      "status",
      "probability",
      "confidence",
      "verdict"
    ],
    "additionalProperties": false
  }
}

Saved example result

This hand-authored response demonstrates the contract. It is not a model accuracy measurement. Run it without an API key: npx jev-recipes demo goal-drift.

{
  "model": "demo-fixture",
  "usage": {
    "input_tokens": 0,
    "output_tokens": 0
  },
  "status": "ready",
  "probability": 0.91,
  "confidence": 0.91,
  "verdict": "drifted"
}

Evaluation evidence

Use the evaluation guide to measure this decision on your own labeled cases.

Limitations

• Judges a single step against the stated goal. It does not know whether the user later widened or changed the goal unless context says so.
• Reports drift, not whether the drifted work is harmful or valuable. Stopping, asking, or continuing is an application decision.

Related recipes

• step-progress: Use step-progress to grade how much an observed result moved the objective forward, rather than whether the next step still belongs to it.
• intent-change: Use intent-change when the user may have redirected the goal, so a step that looks drifted is actually following a new instruction.