e2e-author

E2E Author: planner -> generator orchestration

Turn a critical user flow into a reliable Playwright spec by driving the planner and generator roles. This skill orchestrates; the roles do the exploration and code generation.

Two runtime families, three execution paths, same gates: on Claude Code the roles are the named agents that e2e-setup generated via init-agents --loop=claude (Path A); on Codex those agent files are not registerable as named subagents, so each role runs as a generic subagent carrying the bundled contract from references/role-contracts.md (Path B), or in-agent sequentially when no delegation is available (Path C).

Precondition check

• Read the E2E SSOT doc (e2e/AGENTS.md or .claude/e2e-guidelines.md) for the CUF list and conventions. If absent, run e2e-setup Step 5 or capture the CUFs inline.
• The environment seed (seed.spec.ts, wherever e2e-setup put it) must exist (both paths need it). For app exploration, the planner/generator need at least one of: playwright-cli (re-run the probe in e2e-setup Step 1: a global playwright-cli, the project's npx playwright cli, or npx @playwright/cli@latest) or the .mcp.json playwright-test entry (fallback when playwright-cli is unavailable). The two are independent, so having only one is enough to proceed. Route the user to dev:e2e-setup first only when seed.spec.ts is missing, or when both playwright-cli and the MCP entry are unavailable.
• Path A only: confirm the named agents exist (.claude/agents/playwright-test-planner.md, playwright-test-generator.md). If they are missing and you are on Claude Code, run e2e-setup. Do not hand-roll them. Do not demand these Claude agent files under Codex; there the bundled contracts stand in (see Step 0).

Workflow

1. Resolve the plugin root + pick the execution path: run once. The resolver reaches the bundled role contracts on Path B/C; the path decision governs Steps 2-3. Tell the user which path you took in one sentence.

   # Claude exports CLAUDE_PLUGIN_ROOT; Codex does not. Each branch verifies
   # the target (CHK) exists before committing, so a stale env falls through.
   CHK="references/role-contracts.md"
   PLUGIN_ROOT=""
   [ -n "${CLAUDE_PLUGIN_ROOT:-}" ] && [ -e "$CLAUDE_PLUGIN_ROOT/$CHK" ] && PLUGIN_ROOT="$CLAUDE_PLUGIN_ROOT"
   [ -z "$PLUGIN_ROOT" ] && [ -e "plugins/dev/$CHK" ] && PLUGIN_ROOT="plugins/dev"
   if [ -z "$PLUGIN_ROOT" ]; then
     cache_root="${CODEX_PLUGIN_CACHE:-$HOME/.codex/plugins/cache}"
     while IFS= read -r d; do
       [ -e "$d/$CHK" ] && { PLUGIN_ROOT="$d"; break; }
     done < <(ls -1d "$cache_root"/*/dev/*/ 2>/dev/null | awk -F/ '{print $(NF-1)"\t"$(0)}' | sort -t. -k1,1rn -k2,2rn -k3,3rn | cut -f2- | sed 's#/$##')
   fi
   { [ -n "$PLUGIN_ROOT" ] && [ -e "$PLUGIN_ROOT/$CHK" ]; } || { echo "e2e-author: role contracts not resolved (need $CHK)" >&2; exit 1; }
   echo "PLUGIN_ROOT=$PLUGIN_ROOT"
   

   • Path A: the named agents playwright-test-planner / -generator are registered (Claude Code). Dispatch them by name (Steps 2-3, unchanged).
   • Path B: named agents are not registerable but a generic subagent tool is available (Agent). Dispatch one generic subagent per role, carrying the matching contract from ${PLUGIN_ROOT}/references/role-contracts.md inline.
   • Path C: no delegation available. Run each role yourself, in order, following the same contract with your own tools (Bash for npx playwright, the playwright-test MCP server for browser drive).