菜单
Use when debugging e2e-test/e2e-map/e2e-walkthrough failures, maintaining these skills, or evaluating execution results. Triggers on "fix e2e skill", "e2e skill broke", "e2e gap analysis".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when reviewing a GitHub PR and posting inline review comments. Triggered by PR number, PR URL, 'review this PR', 'review current branch PR'.
Use when compiling E2E flow YAML to standalone bash test scripts. Triggers on "/e2e-compile", "compile flow", "recompile flow".
Use when generating, verifying, or smoke-testing E2E flow YAML from plans, specs, PRs, or mapped pages. Supports browser and CLI-only (Execute external) flows. Triggers on "e2e flow", "generate flow", "verify flow", "smoke test", "cli flow".
Use when asked to 'gemini review', 'gemini challenge', 'ask gemini', or want an independent cross-model second opinion / adversarial review / consultation from a different AI vendor (Google's Antigravity CLI, `agy`) on a diff, plan, or codebase question.
Use when a GitHub PR has review comments or automated feedback that needs triage, validation, and structured response before pushing fixes.
Use when reviewing a teammate's plan, Linear project/issue, design doc, or Slack-thread proposal AS the EM before build starts — applies Kent's EM lenses, verifies the single load-bearing assumption against real code, de-dups against existing reviews, runs a mandatory cross-model pressure-test, and delivers a calibrated, venue-ready message. Triggers: "review this plan/project as EM", "EM review", "plan review", or a Linear project/issue URL paired with "review".
| name | e2e-skill-ops |
| description | Use when debugging e2e-test/e2e-map/e2e-walkthrough failures, maintaining these skills, or evaluating execution results. Triggers on "fix e2e skill", "e2e skill broke", "e2e gap analysis". |
Meta-skill for the e2e-map / e2e-test / e2e-walkthrough pipeline. Enforces known-issue lookup, cross-skill impact scanning, post-fix verification, and persistent learning.
When NOT to use:
/e2e-skill-ops --debug # Diagnose e2e execution failure
/e2e-skill-ops --maintain # Sync changes across skills after edit
/e2e-skill-ops --add-feature # Add capability to one or more e2e skills
/e2e-skill-ops --evaluate # Gap analysis: execution results vs skill expectations
No args: show available modes with status, then ask which to run:
E2E Skill Operations
────────────────────
Modes:
/e2e-skill-ops --debug — diagnose e2e execution failure
/e2e-skill-ops --maintain — sync changes across skills after edit
/e2e-skill-ops --add-feature — add capability to one or more e2e skills
/e2e-skill-ops --evaluate — gap analysis: execution results vs expectations
Recent findings: {count from skill-quality-findings.md, or "none yet"}
Which mode? (or describe what happened — I'll pick the right one)
Unknown mode (e.g., --fix): show the same menu.
Multiple modes (e.g., --debug --maintain): run in the order given, sequentially.
Mode selection from description: When the user describes a problem instead of picking a mode, select using this priority:
--debug--maintain--evaluate--add-featureEvery mode enforces these. Skipping any one is a violation.
| # | Rule | Why |
|---|---|---|
| 1 | Search before diagnose — query journal + .claude/e2e/reports/skill-quality-findings.md + MEMORY.md + ${CLAUDE_PLUGIN_ROOT}/references/learned-patterns.md + .claude/e2e-lessons.md for prior findings. If skill-quality-findings.md doesn't exist, create it with a # E2E Skill Quality Findings heading. | Prevents re-discovering known issues |
| 2 | 3-skill impact scan — check all files in the Impact Matrix below | Prevents silent drift between skills |
| 3 | Verify after fix — run the failing scenario (or subagent pressure test) after changes | Prevents "fix and forget" |
| 4 | Write back findings (D1+D2) — append to skill-quality-findings.md + journal. D1: general patterns → auto-append to ${CLAUDE_PLUGIN_ROOT}/references/learned-patterns.md. D2 (--evaluate mode): project-specific patterns → gated write to .claude/e2e-lessons.md or project CLAUDE.md (see ${CLAUDE_PLUGIN_ROOT}/references/knowledge-capture.md). | Prevents session amnesia + accumulates knowledge |
| 5 | Propose, don't ship plugin file changes — present changes to any file under e2e-pipeline/skills/, e2e-pipeline/agents/, or e2e-pipeline/references/ for human approval before writing. Only project-local files (.claude/e2e/) can be applied directly. "It's just reference.md, not SKILL.md" is not an exemption — reference files are read by agents and affect execution. | Skills, agents, and references are shared infrastructure |
When ANY e2e skill changes, scan every row:
| File | Check for |
|---|---|
e2e-pipeline plugin: skills/e2e-test/SKILL.md | Auth flow, expect grammar, action patterns, common mistakes |
e2e-pipeline plugin: skills/e2e-map/SKILL.md | Mapping format, selector conventions, common mistakes |
e2e-pipeline plugin: skills/e2e-walkthrough/SKILL.md | Auth flow, self-repair, common mistakes |
e2e-pipeline plugin: skills/e2e-flow/SKILL.md | Codebase scan, agent dispatch, smoke mode, common mistakes |
e2e-pipeline plugin: skills/*/reference.md | Execution details, output procedures, templates |
e2e-pipeline plugin: agents/ | e2e-test-runner.md, e2e-mapper.md, e2e-trace-analyzer.md, e2e-flow-writer.md, e2e-flow-verifier.md, e2e-media-processor.md — startup flow, auth, recording, correction logic, media generation |
e2e-pipeline plugin: references/ | commands.md, common-patterns.md |
<project>/.claude/e2e/mappings/*.yaml | Selector conventions match skill guidance |
<project>/.claude/e2e/reports/skill-quality-findings.md | Past findings for this exact pattern |
Missing file rule: If a matrix row target doesn't exist at the expected path, that is a finding — record it in skill-quality-findings.md as a structural gap. Do NOT silently skip missing rows. Verify the file was intentionally removed or relocated before dismissing.
Quick skip rule: Purely cosmetic changes (typo, formatting) — scan only "common mistakes" in the other two skills. All other changes require full scan.
Each mode has detailed checklists in reference.md.
| Mode | Purpose | Key steps |
|---|---|---|
--debug | Diagnose e2e failure | Search known → classify symptom → fix → verify → write back |
--maintain | Sync after skill edit | Describe change → impact scan → propose all → verify → record |
--add-feature | Extend pipeline | Feasibility PoC → impact scan → design → implement → RED test → evaluate |
--evaluate | Gap analysis post-run | Prerequisite: scan .claude/e2e/reports/ for report files. If no reports exist, STOP — tell user to run /e2e-test first, do NOT produce an empty findings report. Then: collect results → classify failures → identify skill gaps → prioritize → report. For comprehensive skill/agent re-verification, cross-reference VERIFICATION-SOP.md checklists. |
You are NOT done until all conditions are met. Proposing a fix is not finishing.
| Mode | NOT DONE until... |
|---|---|
--debug | (1) Fix applied to file, (2) re-ran failing step — passes, (3) finding written, (4) D1 patterns appended if applicable |
--maintain | (1) All changes applied (after approval), (2) ran a relevant flow — no regression, (3) findings updated, (4) D1 patterns appended if applicable |
--add-feature | (1) Feature implemented, (2) subagent RED test executed — behavior matches, (3) feature documented, (4) D1 patterns appended if applicable |
--evaluate | (1) Findings report written, (2) fix priorities presented, (3) recurring vs new compared, (4) D1+D2 knowledge capture |
Evidence format — don't say "verified", show:
Verification: `/e2e-test catalog-browse` -> 7/7 PASS (previously 5/7 FAIL on step 3, 6)
If verification fails, loop back to --debug. Write-back is not optional — at minimum a journal entry.
| Thought | Reality |
|---|---|
| "I know what this error is, skip the search" | That's what all 3 baseline agents did. Search first. |
| "Only one skill needs updating" | OTP change touched 3 skills + 1 reference. Always scan. |
| "Fix works, we're done" | No verification = no confidence. Run the failing step. |
| "I'll remember this for next time" | You won't. Write it to findings.md. |
| "This is just a mapping fix, not a skill issue" | Mapping patterns come from skill guidance. Check if the skill taught the wrong pattern. |
| "I proposed the fix, human can verify later" | Proposing is step 4 of 7. You own verification. |
| "Context is running low, skip write-back" | Journal entry is ~500 tokens. Always enough. No excuse. |
| "It's just reference.md, not SKILL.md" | Reference files are read by agents at startup. A bad reference = bad agent behavior. Same approval gate. |
| "No test results exist, so evaluate is done" | No results = prerequisite not met. Tell user to run tests first. An empty report is not a findings report. |
| "This file doesn't exist, skip this matrix row" | Missing file IS a finding. Record it. Don't silently skip. |
| "User described a problem, I know which mode" | Unless it clearly matches the selection table, show the menu and ask. Guessing wrong wastes the entire workflow. |
| Need | Command |
|---|---|
| Search past findings | Read .claude/e2e/reports/skill-quality-findings.md |
| Search journal | mcp__private-journal__search_journal query: "e2e skill" |
| List all e2e skills | ls ~/.claude/plugins/*/e2e-pipeline/*/skills/*/SKILL.md ~/.claude/plugins/local/e2e-pipeline/skills/*/SKILL.md 2>/dev/null |
| List project mappings | ls .claude/e2e/mappings/*.yaml |
| List project flows | ls .claude/e2e/flows/*.yaml |
| Verify a single flow | /e2e-test <flow-name> |
| Verify full suite | /e2e-test --suite smoke |
| View trace interactively | npx playwright show-trace <trace.zip> |