Agent Tests

Producer skill for agent tests: a coding agent executes a Markdown case file against a target environment and returns an auditable report. Verification is the executing agent's judgment plus the evidence it collects (browser, API, DB, logs, cloud, telemetry), rather than fixed deterministic assertions.

This skill authors and runs agent tests. Each case produces an auditable report — a report path, a final PASS/FAIL/BLOCKED/ABORTED status, and the collected evidence artifacts — for whatever invoked it to consume.

When To Use Agent Tests

Use agent tests when verification requires flexible, tool-driven judgment across UI, API, database, logs, files, network, or live-environment state, and a deterministic test would be premature, brittle, too expensive, or too narrow.

Especially useful for:

• UI changes needing visual, interactive, browser-console, network, or trace

verification.

• API and DB workflows where confidence requires live requests plus persisted

state inspection.

• Full-stack flows crossing frontend, backend, storage, jobs, external mocks,

and cleanup.

• Exploratory regression checks before converting a stable path into a standard

unit, contract, integration, YAML E2E, or other deterministic test.

Agent tests are not a replacement for deterministic tests. Prefer converting high-value stable agent flows into a standard deterministic test (unit, contract, integration, YAML E2E, or other) once the path stabilizes.

Project Layout

Agent tests live under tests/agent/ in the target repo:

tests/agent/agent-test-suites.json   suite manifest (project-owned)
tests/agent/<feature>/<case>.md       case files, one logical journey each
tests/agent/run-agent-verification.ts runner/orchestrator
agent-test-reports/                    generated reports; do not hand-edit

The repo owns its real agent-test-suites.json, case files, fixture setup, auth/session bootstrap, CI wiring, engine secrets and MCP config, staging/production mutation policies, and cleanup ownership.

Use The Local Convention First

Before authoring, check for tests/agent/agent-test-template.md. If it exists, treat it as the authoritative local convention. Do not rediscover or replace it through broad search unless the user explicitly asks to update the convention.

If no local harness exists and an agent test is the cheapest sufficient proof, scaffold it on demand by copying the bundled starters from this skill:

• <skill-dir>/assets/agent-test-template.md → tests/agent/agent-test-template.md
• <skill-dir>/assets/agent-test-suites.example.json → tests/agent/agent-test-suites.example.json
• <skill-dir>/assets/run-agent-verification.ts → tests/agent/run-agent-verification.ts

Copy these only when a project actually adopts agent tests; do not pre-seed repos that have none. Then create a real manifest from the example:

cp tests/agent/agent-test-suites.example.json tests/agent/agent-test-suites.json

Authoring A Case

Write each case from <skill-dir>/assets/agent-test-template.md (or the local template). A good case makes the executing agent unlikely to guess:

• State the requirement, risk, or behavior under test and link its sources.
• Fill real project context: URLs per environment, accounts/orgs, fixture and

mutation policy, DB/log/cloud access, production synthetic-data policy, and cleanup ownership.

• Separate environment preflight (prove the target can execute the case)

from product verification (run the checks).

• List concrete checks: browser pages/states, API routes and status codes, DB

tables/rows, audit events, log filters, telemetry queries, timing.

• Make each assertion unconditional for state the test itself creates. A

check gated on an optional affordance ("if the UI offers X, verify Y") lets the agent skip Y and still PASS whenever X isn't found — silently dropping the coverage the case exists to provide. For fixtures you control (a freshly created token, org, etc.), require the affordance and assert on it directly; reserve conditional wording for genuinely environment-dependent surfaces.

• Name the exact evidence each behavior requires.
• Keep cases portable; bind environment specifics under the testing-environments

section, not in the steps.

• Never store raw secrets. Record variable names, roles, and access patterns

only.

Running

Add a package script in the target repo, adjusted for its package manager:

{ "scripts": { "agent:verify": "tsx tests/agent/run-agent-verification.ts" } }

Run a suite or a single case:

pnpm agent:verify --target local --suite smoke --project-name "<project name>"
pnpm agent:verify --target local --case tests/agent/<feature>/<case>.md

references/runner.md documents the runner defaults, flags, matching environment variables, and the report status contract in full.

Report Status Contract

Every case report must end with exactly one line:

Status: PASS
Status: FAIL
Status: BLOCKED
Status: ABORTED

• PASS / FAIL: environment preflight completed and product verification ran.
• BLOCKED: environment/setup could not execute the case (missing DB/log access,

unreachable URL, failed login/session bootstrap, missing fixtures or approval). Do not report FAIL for an environment blocker.

• ABORTED: orchestration interruption only. Release gates ignore ABORTED and

rerun the case; never end with PASS/FAIL/BLOCKED for an interruption.

Required cases fail the runner unless they return PASS.

When a case drives a browser, collect auditable evidence — HTML report, screenshot set, video, trace, or project-standard equivalent. Text-only browser claims are not sufficient.

Output

A run leaves an auditable report at its report path, ending with exactly one Status: line, plus any collected evidence artifacts (HTML report, screenshots, video, trace, logs). That report and its evidence are the deliverable — whatever invoked this skill consumes them.

When Not To Use

• When a deterministic unit, contract, integration, or YAML E2E test is the

cheaper, more durable proof — use the standard deterministic format instead.

• When no live environment, access, or fixtures are available — record the

blocker rather than authoring an unrunnable case.