Agentv Eval Writer

AgentV Eval Writer

Comprehensive docs: https://agentv.dev Promptfoo parity matrix: https://agentv.dev/docs/reference/promptfoo-parity/

Authoring Principle

Treat YAML as the canonical portable model. Prefer authoring .eval.yaml / EVAL.yaml first, then use TypeScript helpers, Python scripts, or executable graders only when they lower to the same fields or when the evaluation logic must actually run code.

Eval files define what is tested and how it runs: prompts, datasets, assertions, task fixtures, top-level providers, and suite run controls. Use field-local file refs such as tests: file://..., prompts: file://..., default_test: file://..., and environment: file://.... String-valued tests and string entries inside tests[] are raw-case refs for direct paths, directories, and globs. Run several full eval suites directly with CLI multi-file selection and tags. Use scoped run: on individual tests only for threshold, repeat, timeout_seconds, and legacy budget_usd; keep provider selection at top-level providers or CLI --provider, put suite budget caps under evaluate_options.budget_usd, authored concurrency under evaluate_options.max_concurrency, suite repeat policy under evaluate_options.repeat, coding-agent testbed setup under environment, provider environment overrides under env, and lifecycle hooks under extensions.

Use @agentv/sdk for TypeScript helper imports. Do not use @agentv/eval for new evals, examples, scaffolds, or skill guidance; it was a deprecated compatibility package and has been removed from this repository.

Authoring Checklist

• Put grading criteria in assert, not in test-level criteria. Plain assertion strings become an llm-rubric grader.
• Prefer plain assertion strings for semantic checks when the default rubric grader can judge them. Use type: llm-rubric for structured criteria, custom prompts, custom grader providers, or assertion-level transforms. Use type: agent-rubric when the grader itself must be an agent-capable provider that can inspect the workspace. Use type: script when grading must execute code.
• Put reference answers in tests[].vars.expected_output or default_test.vars.expected_output, and consume them with an explicit assertion such as type: llm-rubric with value: "Matches the reference answer: {{ expected_output }}". Do not write criteria, scoring instructions, or "the agent should..." rubric prose as the reference answer.
• For historical or repo-state evals, materialize the repo through a pinned environment setup recipe. Mentioning a SHA only in prompt prose is not enough because the agent needs an actual checkout to inspect.

Evaluation Types

AgentV evaluations measure execution quality — whether your agent or skill produces correct output when invoked.

For trigger quality (whether the right skill is triggered for the right prompts), see the Evaluation Types guide. Do not use execution eval configs (EVAL.yaml, evals.json) for trigger evaluation — these are distinct concerns requiring different tooling and methodologies.

Starting from evals.json?

If the project already has an Agent Skills evals.json file, use it as a starting point instead of writing YAML from scratch:

# Convert evals.json to AgentV EVAL YAML
agentv convert evals.json

# Run directly without converting (all commands accept evals.json)
agentv eval evals.json

The converter maps Agent Skills prompt text into AgentV prompt/vars data, promotes Agent Skills expected_output into explicit llm-rubric criteria or vars.expected_output when it is true reference data, and maps Agent Skills assertions to AgentV assert (llm-rubric checks). The generated YAML includes TODO comments for AgentV features to add (environment setup, script graders, rubrics, required gates).

After converting, enhance the YAML with AgentV-specific capabilities shown below.

From Chat Transcript

Convert a chat conversation into eval test cases without starting from scratch.

Input formats:

Markdown conversation:

User: How do I reset my password?
Assistant: Go to Settings > Security > Reset Password...

JSON messages:

[{"role": "user", "content": "How do I reset my password?"},
 {"role": "assistant", "content": "Go to Settings > Security > Reset Password..."}]

Select exchanges that make good test cases:

• Factual Q&A — verifiable answers
• Task completion — user requests an action, agent performs it
• Edge cases — unusual inputs, error handling, boundary conditions
• Multi-turn reasoning — exchanges where earlier context matters

Skip: greetings, one-word acknowledgments, repeated exchanges

Multi-turn format (when context from prior turns matters):

prompts:
  - - role: user
      content: "My name is Alice"
    - role: assistant
      content: "Nice to meet you, Alice!"
    - role: user
      content: "What's my name?"

tests:
  - id: multi-turn-context
    vars:
      expected_output: "Your name is Alice."
    assert:
      - type: llm-rubric
        value: "Matches the reference answer: {{ expected_output }}"
      - Correctly recalls the user's name from earlier in the conversation

Guidelines: preserve exact wording in vars.expected_output; aim for 5–15 tests per transcript; pick exchanges that test different capabilities.

Quick Start

description: Example eval
providers:
  - default

prompts:
  - "{{ prompt }}"

tests:
  - id: greeting
    vars:
      prompt: "Say hello"
      expected_output: "Hello! How can I help you?"
    assert:
      - type: llm-rubric
        value: "Matches the reference answer: {{ expected_output }}"
      - Greeting is friendly and warm
      - Offers to help

Eval File Structure

Required: tests (array or string raw-case path) or scenarios Optional: name, description, version, author, tags, license, requires, providers, prompts, default_test, timeout_seconds, evaluate_options, threshold, suite, environment, env, extensions, assert

Test fields:

Provider declarations: AgentV accepts Promptfoo-shaped provider entries where they map cleanly: strings such as openai:gpt-4.1-mini, object form with id, label, config, env, prompts, transform, delay, and inputs, and provider maps such as { "openai:gpt-4": { label, config } }. In object form, id is the backend/spec and label is the stable AgentV selection and result identity. AgentV-only runtime, provider-local environment, and provider hooks must be explicit extensions; direct Promptfoo runs do not execute them, and export must lower supported cases or reject unsupported ones clearly.

Prompt Templates and Vars

Use top-level prompts plus tests[].vars for the Promptfoo-compatible canonical input shape. Shared data defaults belong in default_test.vars; per-test vars override those defaults by key. AgentV renders every prompt with each test's merged vars, then expands the run across prompts, targets, tests, and repeat samples.

description: Prompt matrix example
providers:
  - default

prompts:
  - id: support-chat
    label: Support chat
    file: ./prompts/support-chat.json
  - id: terse
    label: Terse
    prompt: "Answer for {{ audience }} in one sentence: {{ question }}"

default_test:
  vars:
    audience: users
    category: support

tests:
  - id: password-reset
    vars:
      question: How do I reset my password?
      expected_output: Password reset guidance
    assert:
      - type: llm-rubric
        value: "Matches the reference answer: {{ expected_output }}"
      - Gives correct password reset guidance
  - id: admin-access
    vars:
      audience: admins
      question: How do I revoke a user's access?
      expected_output: Access revocation guidance
    assert:
      - type: llm-rubric
        value: "Matches the reference answer: {{ expected_output }}"
      - Gives safe access revocation guidance

Prompt templates can use {{ name }} or {{ vars.name }} placeholders. Use top-level names when matching Promptfoo-style prompt templates; use {{ vars.name }} when explicit namespacing is clearer.

Do not author direct input fields in normal eval YAML. tests[].input and top-level input are removed; write simple task text as a prompt template such as "{{ input }}" or "{{ vars.input }}" with tests[].vars.input: "Summarize X".

input_files is also direct-input convenience sugar. In prompt-template suites, model file-backed context as vars containing file paths or file:// references, then render those vars from the prompt template next to the input.

Shorthand forms:

• Prompt entries can be strings, message arrays, file references, or prompt objects.
• Put chat/system/user messages in prompts, not in tests[].input.
• vars.expected_output is a conventional reference-answer variable; explicit assertions decide how to grade it
• Use these canonical field names on disk; keep the wire format snake_case

Message format: {role, content} where role is system, user, assistant, or tool Content types: inline text, {type: "file", value: "./path.md"} File paths: relative from eval file dir, or absolute with / prefix from repo root File handling by provider type: LLM providers receive file content inlined in XML tags. Agent providers receive a preread block with file:// URIs and must read files themselves. See Coding Agents > Prompt format.

JSONL format: One test per line as JSON. Optional .yaml sidecar for shared defaults. See examples/features/basic-jsonl/.

Environment variables: Use {{ env.VAR }} templates in authored config. Missing vars resolve to empty string. Works in eval files, external case files, and environment configs. .env files are loaded automatically.

Output Transforms

Use Promptfoo-compatible transform when the target output needs shaping before grading. Common cases include converting a ContentFile such as an .xlsx spreadsheet into text before an llm-rubric grader runs.

prompts:
  - "{{ input }}"

default_test:
  options:
    transform: file://scripts/transforms/xlsx-to-csv.ts

tests:
  - id: spreadsheet-output
    vars:
      input: Generate the spreadsheet report
    assert:
      - Output contains the transformed spreadsheet text including the revenue rows

Transform placement:

• default_test.options.transform applies to every test unless overridden.
• tests[].options.transform overrides the inherited default transform for one test.
• Assertion-level transform applies only to that grader, after the test/default transform.

Do not author preprocessors or deprecated Promptfoo postprocess in current eval YAML. Use transform at the point that needs the shaped output.

Metadata

When name is present, the suite is parsed as a metadata-bearing eval:

name: export-screening        # required, lowercase/hyphens, max 64 chars
description: Evaluates export control screening accuracy
version: "1.0"
author: acme-compliance
tags: [compliance, agents]
license: Apache-2.0
requires:
  agentv: ">=0.30.0"

Shared Prompt Context

Put shared prompt instructions in top-level prompts and shared data in default_test.vars:

prompts:
  - - role: system
      content: |
        Read AGENTS.md before answering.
        Explain tradeoffs clearly.
    - role: user
      content: "{{ question }}"

tests: ./cases.yaml

# cases.yaml — each raw row supplies vars.question, or a compatibility input row
# - id: test-1
#   assert:
#     - ...
#   vars:
#     question: "User question here"

Tests as String Path

Point tests to an external file instead of inlining:

name: my-eval
description: My evaluation suite
tests: ./cases.yaml           # relative to eval file dir

The external file can be YAML (array of test objects) or JSONL.

Assert Field

assert defines graders at the suite level or per-test level. It is the canonical authored field for all graders:

# Mix exact checks with rubric shorthand when both matter.
assert:
  - type: is-json
    required: true
  - type: contains
    value: "status"
  - Correctly answers the user's question
  - Explains the reasoning clearly

tests:
  - id: test-1
    input: Get status
    assert:
      - type: equals
        value: '{"status": "ok"}'
      - Explains what the status means

Plain strings in assert are rubric criteria and are the preferred shape for qualitative agent behavior. Use deterministic assertions (contains, regex, is-json, equals) only for exact machine-verifiable outputs, and script graders when the check must inspect files, run commands, or validate structured state. Do not add a separate test-level criteria field. Legacy evals that still use criteria without explicit assert are loaded as a plain-string assertion for compatibility, but new evals should author the assertion directly.

For repo-state evals, combine a pinned checkout, a golden answer, and assertion shorthand:

environment:
  type: host
  workdir: ./agentv
  setup:
    command:
      - bash
      - ./scripts/materialize-repo.sh
      - ./agentv
      - https://github.com/EntityProcess/agentv.git
      - 5e3c8f46d80fe66b1a75659e4fd94e38a7e09215
    cwd: "."

prompts:
  - "{{ task }}"

tests:
  - id: verification-learning-capture
    vars:
      task: |
        The eval harness has prepared ./agentv at the commit before the
        verification guidance was added.

        Decide what durable repo change should be made and explain why.
      expected_output: |
        The durable repo change is to update .agents/verification.md with the
        reusable verification workflow lessons. AGENTS.md already routes this
        class of work to .agents/verification.md, so no extra AGENTS.md edit is
        needed unless that routing is missing.
    assert:
      - type: llm-rubric
        value: "Matches the reference answer: {{ expected_output }}"
      - The answer recommends updating .agents/verification.md rather than leaving the learning only in PR comments or private evidence.
      - The answer uses the pinned ./agentv checkout to verify the AGENTS.md routing.
      - The answer preserves the historical commit SHA as context.

Assertions and Reference Data

When assert is defined, only the declared graders run. For semantic checks, add plain rubric strings. If you need a custom LLM prompt or grader provider, declare llm-rubric explicitly:

prompts:
  - "{{ prompt }}"

tests:
  - id: mixed-eval
    vars:
      prompt: "Debug this function..."
    assert:
      - Explains why the bug happens
      - type: contains
        value: "fix"

vars.expected_output is passive reference data. It is available to graders through {{ expected_output }} and the script stdin payload, but it does not create an implicit LLM grading call by itself.

Common mistake: putting rubric prose in expected_output instead of an assertion:

prompts:
  - "{{ prompt }}"

tests:
  - id: bad-example
    vars:
      prompt: "What is 2+2?"
      expected_output: The assistant should explain why the answer is 4. # reference answer var, not a grader

Write this as:

prompts:
  - "{{ prompt }}"

tests:
  - id: good-example
    vars:
      prompt: "What is 2+2?"
      expected_output: "4"
    assert:
      - type: llm-rubric
        value: "Matches the reference answer: {{ expected_output }}"
      - The answer is 4 and explains the arithmetic briefly

Required Gates

Any grader can be marked required to enforce a minimum score:

assert:
  - type: contains
    value: "DENIED"
    required: true          # must score >= 0.8 (default)
  - type: llm-rubric
    required: true
    min_score: 0.6          # must score >= 0.6 (custom threshold)
    value:
      - id: accuracy
        outcome: Identifies the denied party
        weight: 5.0

If a required grader scores below its threshold, the overall case status is forced to fail.

Environment Setup/Teardown

Run scripts before/after each test. Define at suite level or override per case:

environment:
  type: host
  workdir: ./repo
  setup:
    command: ["bun", "run", "setup.ts", "sympy/sympy", "abc123"]
    cwd: "."
    timeout_ms: 120000
extensions:
  - file://scripts/teardown.mjs:afterAll

tests:
  - id: case-1
    input: Fix the bug
    metadata:
      source_repo: sympy/sympy
      source_commit: "abc123"

Lifecycle: environment setup → lifecycle extensions → target setup → agent → grading → teardown extensions → cleanup Merge: Case-level environment fields replace suite-level fields. Setup commands: setup.command is an argv command. command[0] is the executable and command[1...] are CLI arguments. Use ["bash", "-lc", "..."] when shell behavior is required. Setup stdin: AgentV sends environment metadata JSON on stdin for setup commands. Setup failure: aborts case. Teardown failure: non-fatal (warning). For SWE-bench-style evals, put operational checkout state under environment.setup.command CLI arguments; treat metadata.source_commit as informational only. A SHA in the prompt or metadata without a matching environment setup recipe is not an operational checkout.

Environment Lifecycle

Describe coding-agent testbeds with environment. Reusable recipes should live in field-local files and be loaded with environment: file://...:

environment: file://.agentv/environments/repo.yaml

# .agentv/environments/repo.yaml
type: host
workdir: ./repo
setup:
  command: ["bash", "./scripts/materialize-repo.sh", "https://github.com/org/repo.git", "main", "1"]
  cwd: "."
  timeout_ms: 120000

# .agentv/environments/docker-repo.yaml
type: docker
context: ./environment
dockerfile: Dockerfile
workdir: /app
setup:
  command: ["bash", "-lc", "bun install && bun run build"]
  cwd: "."
  timeout_ms: 120000

• type: host or docker
• workdir: path the target and graders should use
• setup: argv command, optional cwd, and optional timeout_ms for repository/testbed materialization
• Top-level env: provider/eval environment overrides
• environment.env: recipe-scoped process environment, distinct from top-level env
• extensions: lifecycle hooks such as beforeAll, beforeEach, afterEach, and afterAll

Grader Types

Configure via the assert array. Multiple graders produce a weighted average score.

script

- name: format_check
  type: script
  command: [uv, run, validate.py]
  cwd: ./scripts          # optional working directory
  provider: {}            # optional: enable LLM provider proxy (max_calls: 50)

Contract: stdin JSON -> stdout JSON {pass, score, reason, checks?: [{text, pass, score?, reason}]} Raw stdin uses snake_case and includes: input, expected_output, output (final answer string), messages, trace, trace_summary, token_usage, cost_usd, duration_ms, start_time, end_time, file_changes, workspace_path, config SDK handlers receive the same payload in camelCase: expectedOutput, traceSummary, tokenUsage, costUsd, durationMs, startTime, endTime, fileChanges, workspacePath. checks is an SDK/script convenience shape; public grading.json artifacts normalize checks into recursive component_results. When an environment prepares a workspace directory, workspace_path is the absolute path to that directory (also available as AGENTV_WORKSPACE_PATH env var). Use this for functional grading (e.g., running npm test in the prepared workdir). For deterministic workspace checks that fit normal Vitest expect(...) tests, prefer a plain verifier file and the built-in adapter:

- name: welcome_banner
  type: script
  command: [agentv, eval, graders/welcome-banner.test.ts]

AgentV infers the Vitest adapter for *.test.ts, *.spec.ts, and Vercel-style EVAL.ts files. Use the explicit agentv eval vitest subcommand only when you need adapter flags such as --cwd, --in-workspace, or --vitest-command. See the Script Graders docs for the full stdin/stdout contract.

llm-rubric

- name: quality
  type: llm-rubric
  prompt: ./prompts/eval.md     # markdown template or command config
  provider: grader_gpt_5_mini   # optional: override the grader provider for this grader
  model: gpt-5-chat            # optional model override
  config:                       # passed to prompt templates as context.config
    strictness: high

Variables: {{criteria}}, {{input}}, {{expected_output}}, {{output}}, {{metadata}}, {{metadata_json}}, {{rubrics}}, {{rubrics_json}}, {{file_changes}}, {{tool_calls}}

• Markdown templates: use {{variable}} syntax
• TypeScript templates: use definePromptTemplate(fn) from @agentv/sdk, receives context object with all variables + config
• Use provider: to run different llm-rubric graders against different named LLM providers in the same eval (useful for grader panels / ensembles)

assert-set

- metric: gate
  type: assert-set
  threshold: 0.7
  config:
    shared_setting: enabled
  assert:
    - metric: safety
      type: llm-rubric
      prompt: ./safety.md
      weight: 0.3
    - metric: quality
      type: llm-rubric
      weight: 0.7

Use assert-set for Promptfoo-aligned assertion grouping. Without threshold, the set passes only when every nonzero-weight child assertion passes. With threshold, the weighted aggregate score determines the set pass/fail result. Parent config is inherited by children, and child config keys override parent keys. Do not use type: composite; AgentV rejects it.

Skill And Trajectory Assertions

- type: skill-used
  value: csv-analyzer
- type: not-skill-used
  value:
    pattern: "web-*"
    max: 0
- type: trajectory:tool-used
  value:
    name: knowledgeSearch
    min: 2
- type: trajectory:tool-sequence
  value:
    mode: exact
    steps: [knowledgeSearch, documentRetrieve]
- type: trajectory:tool-args-match
  value:
    name: knowledgeSearch
    args: { query: "search term" }
    mode: partial

Use Promptfoo-compatible skill-used, not-skill-used, and trajectory:* assertions for new eval YAML. Do not author skill-trigger or tool-trajectory; AgentV rejects them with migration guidance. Per-tool latency checks from old tool-trajectory YAML do not have a Promptfoo trajectory equivalent yet; use a script assertion for that behavior.

field-accuracy

- name: fields
  type: field-accuracy
  match_type: exact          # exact | date | numeric_tolerance
  numeric_tolerance: 0.01    # for numeric_tolerance match_type
  aggregation: weighted_average  # weighted_average | all_or_nothing

Compares output fields against expected_output fields.

latency

- name: speed
  type: latency
  max_ms: 5000

cost

- name: budget
  type: cost
  max_usd: 0.10

token-usage

- name: tokens
  type: token-usage
  max_total_tokens: 4000

execution-metrics

- name: efficiency
  type: execution-metrics
  max_tool_calls: 10        # Maximum tool invocations
  max_llm_calls: 5          # Maximum LLM calls (assistant messages)
  max_tokens: 5000          # Maximum total tokens (input + output)
  max_cost_usd: 0.05        # Maximum cost in USD
  max_duration_ms: 30000    # Maximum execution duration
  target_exploration_ratio: 0.6   # Target ratio of read-only tool calls
  exploration_tolerance: 0.2      # Tolerance for ratio check (default: 0.2)

Declarative threshold-based checks on execution metrics. Only specified thresholds are checked. Score is proportional: passed / total assertions. Missing data counts as a failed assertion.

contains

- type: contains
  value: "DENIED"
  required: true

Binary check: does output contain the substring? Name auto-generated if omitted.

regex

- type: regex
  value: "\\d{3}-\\d{2}-\\d{4}"

Binary check: does output match the regex pattern?

equals

- type: equals
  value: "42"

Binary check: does output exactly equal the value (both trimmed)?

is-json

- type: is-json
  required: true

Binary check: is the output valid JSON?

llm-rubric

- Correctly identifies the denied party
- Provides clear reasoning

LLM-judged structured evaluation. Plain strings are the preferred shorthand. Use type: llm-rubric when you need weighted criteria, required: false, min_score, or score ranges. Rubric items support id, outcome, weight, and required fields. Use optional operator: correctness for positive support checks or operator: contradiction for guard criteria where omission is acceptable but incompatible claims fail.

See references/rubric-grader.md for score-range mode and scoring formula.

Suite-Level Quality Threshold

Set a minimum mean score for the eval suite. If the mean quality score falls below the threshold, the CLI exits with code 1 — useful for CI/CD quality gates. Use evaluate_options.repeat when each case should be attempted more than once.

evaluate_options:
  repeat:
    count: 3
    strategy: pass_any
    early_exit: false
threshold: 0.8

CLI flag --threshold 0.8 overrides the YAML value. Must be a number between 0 and 1. Mean score is computed from quality results only (execution errors excluded).

The threshold also controls JUnit XML pass/fail: tests with scores below the threshold are marked as <failure>. When no threshold is set, JUnit defaults to 0.5.

CLI Commands

# Run evaluation (requires API keys)
agentv eval <file.yaml> [--test-id <id>] [--target <name>] [--threshold <0-1>]

# Run with OTLP JSON file (importable by OTel backends)
agentv eval <file.yaml> --otel-file traces/eval.otlp.json

# Record live target output for later target substitution
agentv eval <file.yaml> --target live_agent --record-replay fixtures/target-output.jsonl
agentv eval <file.yaml> --target replay_agent

# Run a single assertion in isolation (no API keys needed)
agentv eval assert <grader-name> --agent-output "..." --agent-input "..."

# Import agent transcripts for offline grading
agentv import claude --session-id <uuid>

# Re-run only execution errors from a previous run
agentv eval <file.yaml> --retry-errors .agentv/results/<run_id>/.internal/index.jsonl

# Validate eval file
agentv validate <file.yaml>

# Compare completed runs
agentv results compare \
  .agentv/results/<baseline-run-id>/.internal/index.jsonl \
  .agentv/results/<candidate-run-id>/.internal/index.jsonl
agentv results combine \
  .agentv/results/<baseline-run-id> \
  .agentv/results/<candidate-run-id> \
  .agentv/results/<third-target-run-id> \
  --output .agentv/results/combined
agentv results compare .agentv/results/combined/.internal/index.jsonl
agentv results compare \
  .agentv/results/<baseline-run-id>/.internal/index.jsonl \
  .agentv/results/<candidate-run-id>/.internal/index.jsonl \
  --json

# Author assertions directly in the eval file
# Prefer simple assertions when they fit the criteria; use deterministic or LLM-based graders when needed
agentv validate <file.yaml>

Replay targets: Add provider: replay, fixtures: <jsonl>, and source_target: <live target label> in .agentv/targets.yaml. Optional suite, eval_path, and variant tighten lookup. The eval YAML and graders stay unchanged; replay only substitutes recorded target output, and graders run fresh.

TypeScript SDK Helpers

Use @agentv/sdk as the public lightweight SDK package for TypeScript/JavaScript helpers. SDK helpers must stay AgentV-native and lower to YAML/runtime contracts rather than introducing a second eval vocabulary.

YAML-aligned eval authoring

// evals/helper-suite.eval.ts
import { graders, type EvalConfig } from '@agentv/sdk';

const config: EvalConfig = {
  name: 'helper-suite',
  target: 'default',
  // TypeScript config loading lowers this to evaluate_options.repeat.
  repeat: {
    count: 3,
    strategy: 'pass_any',
    earlyExit: false,
  },
  threshold: 0.8,
  prompts: ['{{ task }}'],
  tests: [
    {
      id: 'json-answer',
      vars: { task: 'Return a JSON answer with a status field.' },
      assert: [
        graders.json({ name: 'valid-json', required: true }),
        graders.regex(/"status"\s*:/, { name: 'status-key' }),
      ],
    },
  ],
};

export default config;

The graders catalog returns ordinary assert entries such as type: is-json, type: regex, type: llm-rubric, and type: script. Explicit *.eval.ts and *.eval.mts files should default-export an EvalConfig; defineEval(config) is only an optional thin helper over that same shape. TypeScript config loading lowers camelCase fields such as expectedOutput, inputFiles, and maxSteps to canonical snake_case YAML/runtime keys.

If adapting Braintrust scores or DeepEval metrics, write small AgentV helper factories that return graders.* configs:

import { graders } from '@agentv/sdk';

export function ragFaithfulness() {
  return graders.llmRubric(undefined, {
    name: 'rag-faithfulness',
    provider: 'grader-provider',
    prompt: 'Grade whether the answer is supported by the retrieved context.',
  });
}

Use the helper in assert: [ragFaithfulness()]; do not create new YAML terms like scores.

defineAssertion (recommended for reusable custom assertions)

#!/usr/bin/env bun
import { defineAssertion } from '@agentv/sdk';

export default defineAssertion(({ output, trace }) => {
  const finalOutput = output ?? '';
  const pass = finalOutput.length > 0 && (trace?.eventCount ?? 0) <= 10;
  return {
    pass,
    score: pass ? 1 : 0,
    reason: 'Checks content exists and is efficient',
  };
});

Assertions support both pass: boolean and score: number (0-1). If only pass is given, score is 1 (pass) or 0 (fail).

Use defineAssertion() when you want a reusable assertion type discovered from .agentv/assertions/ and referenced by filename as type: <name>. This follows Promptfoo's normal eval terminology: custom logic is an assertion, with Promptfoo using fixed assertion types such as javascript, python, ruby, and webhook. AgentV extends that model by allowing arbitrary discovered assertion type names.

defineScriptGrader (full control)

#!/usr/bin/env bun
import { defineScriptGrader } from '@agentv/sdk';

export default defineScriptGrader(({ output, trace }) => {
  const finalOutput = output ?? '';
  const hasOutput = finalOutput.length > 0;
  const efficient = (trace?.eventCount ?? 0) <= 5;
  return {
    pass: hasOutput && efficient,
    score: hasOutput && efficient ? 1.0 : 0.5,
    reason: 'Checks content exists and tool usage is bounded',
    checks: [
      { text: 'Output is not empty', pass: hasOutput, reason: hasOutput ? 'Output text is present' : 'Output is empty' },
      { text: 'Efficient tool usage', pass: efficient, reason: efficient ? 'Trace event count is within limit' : 'Trace event count is too high' },
    ],
  };
});

Use defineScriptGrader() when the custom component is a command-backed grader with explicit score control, check arrays, workspace commands, or LLM calls through a grader provider. defineScriptGrader() scripts are referenced in YAML with type: script and command: [bun, run, grader.ts]. Plain Vitest workspace verifier files can use command: [agentv, eval, graders/check.test.ts].

Convention-Based Discovery

Place assertion files in .agentv/assertions/ — they auto-register by filename:

.agentv/assertions/min-words.ts  →  type: min-words
.agentv/assertions/sentiment.ts   →  type: sentiment

No command: needed in YAML — just use type: <filename>.

Programmatic API

Use evaluate() from @agentv/core to run evals as a library when you need application-level control. Keep YAML as the default portable surface; use specFile to point at existing evals and inline tests only when the definition belongs in code.

import { evaluate } from '@agentv/core';

const { results, summary } = await evaluate({
  tests: [
    {
      id: 'greeting',
      input: 'Say hello',
      expectedOutput: 'Hello there!',
      assert: [{ type: 'contains', value: 'hello' }],
    },
  ],
  target: { provider: 'mock_agent' },
});
console.log(`${summary.passed}/${summary.total} passed`);

Programmatic API notes:

• Inline programmatic tests use assert, not assertions.
• Use camelCase in TypeScript (expectedOutput, beforeAll, budgetUsd).
• In YAML, use evaluate_options.max_concurrency for authored eval concurrency; reserve workers for project/target runtime config.
• When bridging from Python, generate canonical YAML/JSONL or call the CLI; there is no separate first-party Python authoring SDK.

Supports inline tests or file-based usage via specFile.

defineConfig

Type-safe project configuration in agentv.config.ts:

import { defineConfig } from '@agentv/core';

export default defineConfig({
  execution: { workers: 5, maxRetries: 2 },
  output: { dir: './results' },
  limits: { maxCostUsd: 10.0 },
});

Auto-discovered from project root. Validated with Zod.

Scaffold Commands

agentv create assertion <name>  # → .agentv/assertions/<name>.ts
agentv create eval <name>       # → evals/<name>.eval.yaml + .cases.jsonl

Skill Improvement Workflow

For a complete guide to iterating on skills using evaluations — writing scenarios, running baselines, comparing results, and improving — see the Skill Improvement Workflow guide.

Observability Export

AgentV exports observability data via OpenTelemetry:

• agentv eval <file.yaml> --otel-file traces/eval.otlp.json writes a post-run OTLP JSON file that external systems such as Opik can ingest.
• agentv eval <file.yaml> --export-otel --otel-backend <name> streams live traces when a built-in or local resolver exists.

Do not invent a separate Opik-specific eval surface. Keep the eval definition in YAML and route observability through OTLP export.

Schemas

• Eval file: references/eval.schema.json
• Config: references/config.schema.json

Accessing reference files

To load a specific reference without pulling the entire skill into context:

agentv skills get agentv-eval-writer --ref eval.schema.json

Or resolve the skill directory and read files directly:

cat $(agentv skills path agentv-eval-writer)/references/eval.schema.json

Use --full to retrieve every file in the skill at once.