adk-evals（adk 评估）

ADK Evals Skill

What are Evals?

Evals are automated conversation tests for ADK agents. Each eval defines a scenario — a sequence of user messages or events — and asserts on what the bot should do: what it says, which tools it calls, how state changes, what gets written to tables, and more.

Evals run against a live dev bot (adk dev), so they test the full stack — not mocks.

When to Use This Skill

Use this skill when the developer asks about:

• Writing evals — file format, assertions, turn types, setup
• Running evals — CLI commands, filtering, output interpretation
• Testing specific primitives — how to test actions, tools, workflows, conversations, tables, state
• The testing loop — write → run → inspect traces → iterate
• CI integration — exit codes, --format json flag, tagging strategies
• Eval configuration — idleTimeout, judgePassThreshold, judgeModel

Or when you are developing an ADK bot and need to write the equivalent of unit/end-to-end tests.

Trigger questions:

• "How do I write an eval?"
• "How do I test my workflow?"
• "How do I assert that a tool was called with specific params?"
• "My eval is failing, how do I debug it?"
• "How do I test that the bot stays silent?"
• "How do I run evals in CI?"
• "How do I seed state before an eval?"
• "How do I trigger a workflow in an eval?"

Available Documentation

How to Answer

1. Writing an eval → Read eval-format.md for structure and assertions
2. Running evals → Read testing-workflow.md for CLI commands and output
3. Testing a specific primitive → Read test-patterns.md for the relevant section
4. Debugging a failure → Combine testing-workflow.md (inspect traces) + eval-format.md (check assertion syntax)

Quick Reference

Eval file structure

import { Eval } from '@botpress/adk'

export default new Eval({
  name: 'greeting',
  type: 'regression',
  tags: ['basic'],

  setup: {
    state: { bot: { welcomeSent: false } },
    workflow: { trigger: 'onboarding', input: { userId: 'test-1' } },
  },

  conversation: [
    {
      user: 'Hi!',
      assert: {
        response: [
          { not_contains: 'error' },
          { llm_judge: 'Response is friendly and offers to help' },
        ],
        tools: [{ not_called: 'createTicket' }],
        state: [{ path: 'conversation.greeted', equals: true }],
      },
    },
  ],

  outcome: {
    state: [{ path: 'conversation.greeted', equals: true }],
  },

  options: {
    idleTimeout: 20000,
    judgePassThreshold: 4,
  },
})

Turn types

Assertion categories

CLI commands

adk evals                        # run all evals
adk evals <name>                 # run one eval
adk evals --tag <tag>            # filter by tag
adk evals --type regression      # filter by type
adk evals --verbose              # show all assertions
adk evals --format json          # JSON output for CI

adk evals runs                   # list recent runs
adk evals runs --latest          # most recent run
adk evals runs --latest -v       # with full details

Critical Patterns

✅ Every turn needs user or event

// CORRECT
{ user: 'hello', expectSilence: true }
{ event: { type: 'payment.failed' }, expectSilence: true }

❌ expectSilence alone is not a valid turn

// WRONG — missing user or event
{ expectSilence: true }

✅ Assert tool params to verify correct extraction

// CORRECT — verifies the LLM extracted the right values
{ called: 'createTicket', params: { priority: { equals: 'high' } } }

❌ Only asserting the tool was called

// INCOMPLETE — doesn't verify params were correct
{ called: 'createTicket' }

✅ Use outcome for post-conversation state and table assertions

// CORRECT — final state checked once after all turns
outcome: {
  state: [{ path: 'conversation.resolved', equals: true }],
  tables: [{ table: 'ticketsTable', row_exists: { status: { equals: 'open' } } }],
}

❌ Checking tables in per-turn assertions when the write happens at the end

// WRONG — table may not be written until after all turns
conversation: [
  {
    user: 'Create a ticket',
    assert: { tables: [{ table: 'ticketsTable', row_exists: { status: { equals: 'open' } } }] },
  },
]

✅ Seed state to test conditional behavior without running setup turns

// CORRECT — start in a known state
setup: {
  state: {
    user: { plan: 'pro' },
    conversation: { phase: 'support' },
  },
}

❌ Using conversation turns to set up state (slow and fragile)

// WRONG — depends on the bot correctly processing setup turns
conversation: [
  { user: 'I am on the pro plan' },      // hoping bot sets user.plan
  { user: 'I need help with billing' },   // actual test turn
]

Example Questions

Writing evals:

• "Write an eval that tests my createTicket tool is called with the right priority"
• "How do I assert that the bot stays silent after an internal event?"
• "How do I test a multi-turn conversation where context is retained?"

Running evals:

• "How do I run only regression evals?"
• "How do I see which assertions failed and why?"
• "How do I integrate evals into GitHub Actions?"

Debugging:

• "My eval says the tool wasn't called but I think it was — how do I check?"
• "How do I inspect what the bot actually did during an eval?"

Per-primitive:

• "How do I test a workflow that uses step.sleep()?"
• "How do I verify a row was written to a table after a conversation?"
• "How do I test that state changed from the seeded value?"

Response Format

Match depth to the question.

Simple questions ("what assertions are available?", "how do I run evals?")

Answer directly — show the relevant table or CLI command. Don't generate a full eval file for an informational question.

Writing an eval

1. Show the complete new Eval({}) call with realistic field values
2. Include imports (import {Eval} from '@botpress/adk')
3. Briefly explain non-obvious assertions — skip if the assertion is self-explanatory
4. Suggest the CLI command to run it: adk evals <name>

Debugging a failing eval

1. Ask for or show the failing assertion (expected / actual diff)
2. Suggest opening traces in the Control Panel to see what the bot did
3. Identify whether the issue is in the eval assertion or the bot's behavior