Playwright Test

mp-playwright-test

Run reliable browser verification over a defined scope. This skill owns scope → surfaces and orchestration. Read ${CLAUDE_SKILL_DIR}/../shared/PLAYWRIGHT_TESTING.md now — the reliability rules (sanity-gate, assert-don't-eyeball, programmatic auth, never networkidle) live there and are followed verbatim. $ARGUMENTS

Rules

• Raw Playwright only — the project's installed playwright dep, run as a Node script. Browser MCP work belongs to the exploratory mp-chrome-devtools-tester agent, not this reliability path.
• Verify only — assert and screenshot, leaving source untouched.
• This skill encodes policy; the runner command, dev-server port, auth endpoint, and seed users come from the project's AGENTS.md / memory, not from here.

Step 1: Resolve scope

Parse $ARGUMENTS:

• uncommitted (or empty) → working-tree mode: git diff --name-only HEAD (+ untracked).
• pr → PR mode: gh pr view --json number,headRefName then gh pr diff --name-only for the current branch's open PR.
• anything else → verbal mode: treat the text as a description of the app area to test.

Step 2: Map scope → surfaces

• working-tree / PR mode — from the changed file list, keep UI-affecting files (routes, components, styles, layouts). Resolve each to the route(s) that render it. Drop pure backend/logic changes: they have no visual surface and are out of scope for this skill.
• verbal mode — translate the described area into the concrete route(s) and interactions to exercise (run a quick Explore if the routes are not obvious).

If mapping yields zero UI surfaces, stop and report that there is nothing to visually verify.

Step 3: Discover project specifics

Per ${CLAUDE_SKILL_DIR}/../shared/PLAYWRIGHT_TESTING.md § Discover project specifics, read the project's AGENTS.md / CLAUDE.md / memory for: the Playwright runner/helper (e.g. scripts/shot.mjs), the dev-server start command and exact port, the sign-in API + seed users, and where credentials live (.local/, .env.local). If no runner script exists, the verifier writes a minimal one from the shared skeleton.

Step 4: Verify in a sub-agent

Spawn a read-only claude sub-agent with model: "sonnet" to run the verification. Give it: the surface list (with the route + what changed for each), the discovered runner/port/auth details, and the instruction to Read ${CLAUDE_SKILL_DIR}/../shared/PLAYWRIGHT_TESTING.md and follow it exactly — stale-worktree sanity-gate FIRST, then programmatic auth, explicit waits (never networkidle), one measured assertion per surface, a screenshot per surface under test-results/. It verifies every surface even if one fails, and returns the PASS/FAIL table — it does not fix anything.

The sanity-gate is load-bearing: if the running dev server does not reflect the code under test, the sub-agent must kill the stale server and start one bound to this checkout before verifying (see mp-continue for the port-zombie kill/restart pattern).

Step 5: Report

Relay the sub-agent's per-surface table: surface, PASS/FAIL/BLOCKED, measured value vs expected, and the screenshot path. Call out any surface where the sanity-gate had to restart the server. Failures are reported, not fixed — hand them back to the caller (or to /mp:execute) to resolve.