phoenix-pxi-playwright

name	phoenix-pxi-playwright
description	Write, extend, and debug PXI Playwright E2E tests for Phoenix. Use when adding PXI agent frontend specs, authoring LLM-as-judge rubrics, asserting PXI tool use, persisting PXI test runs as Phoenix experiments, or debugging PXI E2E failures.
metadata	{"internal":true}

Phoenix PXI Playwright Tests

Use this skill when authoring or maintaining Playwright specs for PXI, Phoenix's built-in AI assistant. The concrete harness lives in app/tests/pxi/; this skill is the authoring guide for using and extending that harness.

Start Here

• Read the existing example spec first: app/tests/pxi/docs-smoke.spec.ts.
• Reuse the shared fixture and driver from app/tests/pxi/fixtures.ts.
• Reuse shared constants from app/tests/pxi/constants.ts and shared types from app/tests/pxi/types.ts.
• Put pure parsing/API helpers in app/tests/pxi/utils.ts rather than in specs or fixture classes.
• Reuse the generic AI SDK judge from app/tests/pxi/judge.ts.
• Reuse experiment persistence from app/tests/pxi/experimentPersistence.ts.
• Add one entry to PXI_EXPERIMENT_EXAMPLES in app/tests/pxi/experimentPersistence.ts for every new PXI spec scenario. All specs share the same dataset, and the update upload treats that registry as the complete desired set of examples.
• Do not create a bespoke PXI driver, duplicate experiment client, or duplicate PXI tool schemas in a spec.

Current Harness

The current harness provides these abstractions:

• test and expect from ./fixtures: PXI-aware Playwright fixture exports.
• constants.ts: default assistant and judge model/project constants.
• types.ts: shared PXI harness types such as PxiTurn.
• utils.ts: pure utilities for API response validation and span/tool parsing.
• pxi.open(): opens PXI for the test session.
• pxi.acknowledgeConsent(): accepts PXI consent for the test session.
• pxi.askAndWait(prompt): sends a user prompt and waits for the assistant turn. It does not require a backend TOOL span; add explicit tool assertions in the spec.
• pxi.expectNoAgentError(): asserts the visible PXI session did not surface an agent error.
• pxi.expectBackendToolSpanCalled(turn): asserts the PXI turn produced at least one persisted backend TOOL span and merges those backend tool names into turn.calledTools. Use this for server/MCP-backed tools such as docs tools, not for purely client-executed external tools.
• pxi.expectDocsToolCalled(turn): asserts the PXI turn used runtime docs tooling via Phoenix-observed tool spans.
• pxi.getMetadata(): collects PXI metadata for persistence.
• judge({ model, system, prompt, assistantText, rubric }): evaluates an assistant answer with AI SDK generateText and structured Output.object.
• evaluatePxiOutcome({ assertions, judgeInput }): runs deterministic assertions and LLM judging while preserving failed post-turn outcomes for experiment persistence.
• assertPxiOutcome(outcome): fails the Playwright test after persistence, preferring the original deterministic assertion failure when one exists.
• persistPxiExperiment({ request, record }): stores the PXI interaction, judge result, and metadata as a Phoenix experiment.

Authoring Workflow

1. Add the scenario prompt and expected output to PXI_EXPERIMENT_EXAMPLES so the shared PXI E2E dataset gets one example per test scenario.
2. Put the scenario prompt, PXI user instructions, and judge rubric in the spec file so the test is readable top-to-bottom. Import the scenario from PXI_EXPERIMENT_EXAMPLES rather than duplicating prompt strings.
3. Drive PXI through the real UI with pxi.open, pxi.acknowledgeConsent, and pxi.askAndWait.
4. Add deterministic assertions before judge assertions, such as expected text, no agent error, or expected tool use.
5. Put all post-turn deterministic assertions inside evaluatePxiOutcome, including pxi.expectBackendToolSpanCalled(turn), so failures after PXI returns an answer still get persisted.
6. After pxi.askAndWait returns a turn, persist both passing and failing outcomes. Do not let deterministic assertion failures skip experiment persistence.
7. Use evaluatePxiOutcome instead of writing per-spec try/catch blocks. It runs judge even when deterministic Playwright assertions fail, then combines the judge explanation with a sanitized, truncated Playwright assertion message in the persisted failed evaluation.
8. Run the targeted spec with isolated ports before reporting success.

Spec Pattern

import {
  persistPxiExperiment,
  PXI_EXPERIMENT_EXAMPLES,
} from "./experimentPersistence";
import { expect, test } from "./fixtures";
import { getRequiredJudgeApiKeyEnv } from "./judge";
import { assertPxiOutcome, evaluatePxiOutcome } from "./outcome";

const EXPERIMENT_EXAMPLE = PXI_EXPERIMENT_EXAMPLES.someScenario;
const USER_PROMPT = EXPERIMENT_EXAMPLE.prompt;
const JUDGE_RUBRIC = [
  "The answer satisfies the user request.",
  "The answer is grounded in the expected Phoenix context.",
  "The answer does not invent unsupported facts.",
];
const JUDGE_API_KEY_ENV = getRequiredJudgeApiKeyEnv();

test.describe("PXI scenario", () => {
  test("handles the scenario", async ({
    browserName,
    page,
    pxi,
    request,
  }, testInfo) => {
    test.skip(
      browserName !== "chromium",
      "PXI real-LLM smoke runs once in chromium."
    );
    test.skip(
      process.env.PXI_E2E !== "true",
      "Set PXI_E2E=true to run PXI E2E tests."
    );
    test.skip(
      !process.env.OPENAI_API_KEY,
      "OPENAI_API_KEY is required for the PXI assistant."
    );
    test.skip(
      !process.env[JUDGE_API_KEY_ENV],
      `${JUDGE_API_KEY_ENV} is required for the PXI E2E judge.`
    );

    await pxi.open();
    await pxi.acknowledgeConsent();

    const turn = await pxi.askAndWait(USER_PROMPT);
    const outcome = await evaluatePxiOutcome({
      assertions: async () => {
        await pxi.expectNoAgentError();
        // For server/MCP-backed tools only. Client-executed external tools can
        // be asserted through visible tool UI or final app state instead.
        await pxi.expectBackendToolSpanCalled(turn);
        expect(turn.assistantText).toContain("deterministic expected text");
      },
      judgeInput: {
        system: "You are judging a Phoenix PXI E2E answer.",
        prompt: USER_PROMPT,
        assistantText: turn.assistantText,
        rubric: JUDGE_RUBRIC,
      },
    });

    await persistPxiExperiment({
      request,
      record: {
        example: EXPERIMENT_EXAMPLE,
        assistantText: turn.assistantText,
        calledTools: turn.calledTools,
        url: page.url(),
        durationMs: turn.durationMs,
        judgeResult: outcome.judgeResult,
        playwrightProject: testInfo.project.name,
        ...pxi.getMetadata(),
      },
    });

    assertPxiOutcome(outcome);
  });
});

Judge Models

• Pass judge models as provider/model strings.
• Default: openai/gpt-4.1 via PXI_E2E_JUDGE_MODEL fallback.
• Supported providers are currently openai and anthropic.
• openai/... requires OPENAI_API_KEY.
• anthropic/... requires ANTHROPIC_API_KEY.
• The judge must use AI SDK 6 structured output idioms: generateText with Output.object({ schema }). Do not use deprecated generateObject.

Running Specs

Use isolated ports to avoid collisions with a local Phoenix dev server:

pnpm run test:e2e:pxi

Real PXI specs require external model credentials. Keep tests skipped by default unless PXI_E2E=true is set.

When PXI_E2E=true, the Playwright config runs with one worker because PXI specs share a Phoenix test server/database and may mutate agent-visible setup such as sandbox configs.

Experiment Persistence

• Persistence defaults to http://localhost:6006 so real E2E runs can be inspected in the developer's local Phoenix instance.
• Override with PXI_E2E_EXPERIMENT_BASE_URL.
• Use PXI_E2E_EXPERIMENT_BEARER_TOKEN only when persisting to an authenticated Phoenix target.
• All PXI specs share the PXI E2E Agent Tests dataset. persistPxiExperiment uploads every entry in PXI_EXPERIMENT_EXAMPLES using action: "update", so adding or removing an entry updates the dataset examples declaratively.
• Every new PXI spec must add one PXI_EXPERIMENT_EXAMPLES entry with a stable id, prompt, expected output, experiment name prefix, and experiment description, then pass that entry as record.example.
• Persist the complete test record: prompt, assistant text, called tools, duration, judge result, Playwright project, URL, and PXI metadata.
• Persist failed post-turn outcomes too. If PXI returned an answer, the experiment run should exist even when deterministic assertions or judge checks fail.
• Let evaluatePxiOutcome strip ANSI escape sequences and truncate Playwright assertion messages before they are added to judgeResult.explanation; raw Playwright error messages are terminal-formatted and too noisy for Phoenix experiment tables.

Tool Assertions

• Use pxi.expectBackendToolSpanCalled(turn) when the scenario requires a server-observed backend TOOL span. Keep the call inside evaluatePxiOutcome.
• Use pxi.expectDocsToolCalled(turn) after pxi.expectBackendToolSpanCalled(turn) for docs scenarios because it checks the backend tool names merged into turn.calledTools.
• Do not require backend TOOL spans for client-executed external tools such as UI control tools. Assert their visible tool chips, tool result text, or the resulting app state instead.

Extending The Harness

If fixtures.ts, judge.ts, or experimentPersistence.ts do not satisfy a testing use case, use internal_docs/pxi_playwright_e2e_harness.md as the design guide for extending the harness, utilities, and this skill together. That internal doc describes the intended architecture for richer PXI drivers, seed fixtures, LLM modes, outcome assertions, experiment persistence, and debugging artifacts.

Prefer small harness extensions that make future specs simpler. Do not add one-off logic to an individual spec if it should be reusable across PXI E2E scenarios.

Keep This Skill Fresh

When a session creates new PXI Playwright capabilities, discovers a reliable assertion pattern, changes judge behavior, adds seed fixtures, or learns a debugging technique, update this skill in the same branch. Future developers and agents should not have to rediscover branch-specific harness behavior.