Framework Quality Audit

Purpose

EITR's core product bet (see docs/architecture/README.md Section 1, "Introduction & Goals") is that the generated framework is the deterministic first layer of a larger SDET ecosystem, not a disposable scaffold. That means generation quality has to be checked periodically, per stack combination, with a report that is comparable across runs — same categories, same scoring rigor, same layout — so drift and regressions show up as a diff in scores over time, not as a differently-shaped essay each time.

This skill is intentionally a process + report-shape definition, not a scorecard. It carries no memorized scores or verdicts from any prior run — each invocation is a fresh, unbiased read of whatever the generator currently produces for the requested stack. Never seed a new audit with a previous run's numbers.

When to use

Invoke this skill whenever the user asks for a quality/maturity audit, review, or scorecard of an EITR-generated framework — for the TS+Playwright combo or any other supported language + automation-tool pairing. If the user names a stack, audit that stack. If they don't, ask which stack (or default to whichever sample project already exists in the repo, if any).

The 19 fixed categories (exact names, exact order, exact grouping)

Never rename, reorder, merge, split, or add/remove a category between runs — that is what makes scores comparable across audits. If a category is genuinely inapplicable to a given stack (e.g. a language with no equivalent of a given concern), still include its row and mark it N/A with a one-line reason rather than omitting it.

Group A — Framework code (10 categories)

1. CPOM-архитектура — base classes/primitives/widgets for the stack's Page-Object-Model equivalent: no assertions in base components, Now()-suffix (or language-idiomatic equivalent) convention for state getters, auto-retry semantics, collection/list handling, type-safety discipline (generics/casts isolated and documented).
2. Тесты и фикстуры — dependency-injection/fixture pattern, auth bootstrap quality, 3-tier locator priority usage, dynamic TDM usage in examples, dual-layer (UI+API) example coverage, test.step()-equivalent step demarcation in shipped examples (not just described in docs).
3. Конфиг тест-раннера (playwright.config.ts / pytest.ini / pom.xml / etc.) — projects/ environments setup, retries, trace/video/screenshot or language-equivalent diagnostics, reporters, cross-browser or cross-target coverage, sharding support.
4. CI/CD — real vs. placeholder pipelines across every CI provider the generator claims to support; whether gates (lint, type-check) are actually wired in as enforced steps; caching; sharding/matrix parallelism; whether every provider is genuine pipeline-as-code or some are instructions-only.
5. TDM-хелперы — dynamic test-data generators, collision resistance, zero-dependency discipline, teardown/cleanup isolation, bulk/seed data loading if claimed.
6. Конфиг и окружение — env var documentation completeness (does .env.example list every var the generated code actually reads), secrets handling, multi-environment support, whether generated team-shared config actually reaches version control (.gitignore correctness).
7. Линт / формат / типы — type-checker strictness, real static-analysis gates vs. decorative ones, real linter (ESLint/Ruff/Checkstyle/etc.) presence vs. a narrow hand-rolled regex script, coverage of the rules the framework's own AI-rules docs claim to enforce.
8. Документация — README/CONVENTIONS accuracy against actual generated file inventory (stale inventories count against this), concreteness of instructions, presence of a worked example that actually exists in the generated tree (not just described in prose).
9. Zero lock-in — grep the entire generated tree for the tool's own brand name in any form, including in live protocol/handshake surfaces (e.g. an MCP serverInfo.name), not just in filenames or comments. Any hit outside a legitimate self-referential config filename is a finding.
10. Мёртвый код / долг шаблонов — TODO/FIXME/HACK markers, unused/aspirational commented-out code referencing nonexistent classes, redundant/overlapping primitives, any generator-template code that looks unfinished or duplicated.

Group B — AI ecosystem (8 categories + 1 explicit gap check)

11. Паритет ассистентов (N шт.) — for every AI assistant EITR claims to support (Claude Code, Cursor, Devin Desktop, Codex, Copilot, Antigravity, Aider — update the count if the supported list changes), confirm real non-placeholder generated config exists, note any assistant that's thinner than the rest.
12. Качество агентов — read at least 1-2 generated agent definitions fully: concrete/measurable instructions vs. vague adjectives, tool scoping matched to role (not one identical tool list for every agent regardless of what it does), whether the agent is genuinely self-contained in a fresh project with zero knowledge of EITR's own internals.
13. Качество скиллов — same assessment for generated skills; check for accidental duplicate skill registrations (same capability under two names).
14. Слэш-команды — presence and native-format correctness of invocable commands across assistants; whether a user has any single place that maps one capability to its exact invocation syntax in each assistant they might use.
15. MCP-мост — read the actual embedded MCP server source (not just its config), and for every exposed tool compare the tool's description against what its implementation actually does — a tool whose description promises more than the code delivers is a finding, not just "it works."
16. TMS-интеграции — for every claimed TMS provider (Jira Xray, Azure DevOps, TestRail, Zephyr, etc.), read the actual adapter code and confirm it calls that provider's real, provider-specific API — a shared/generic fallback path masquerading as multiple providers (e.g. one generic Jira call standing in for both Xray and Zephyr) is a critical finding. Separately verify whether result-publish-back (not just fetch) is actually implemented, not just advertised.
17. Rules-файлы для конечного пользователя (the generated project's own AGENTS.md/CLAUDE.md/ equivalent, NOT EITR's own root rule files) — read fully; assess whether the rules are genuinely production-grade and specific (concrete good/bad examples, measurable constraints) versus generic boilerplate; re-confirm zero lock-in here too.
18. Свежесть AI-синтаксиса — cross-check frontmatter schemas, MCP protocol version, and per-assistant config formats against currently-known-correct conventions; flag anything that looks pinned to a stale spec revision or an unverified format guess. Note explicitly whether the repo shows any evidence of a recent live-verification pass (per the project's own AI Ecosystem Freshness rule) for the items you're flagging as possibly stale.
19. Агент requirements→тест-кейсы — search the generator source for any agent/skill/MCP tool that generates test cases from requirements text, TMS ticket content, live application exploration, or a design system (e.g. Storybook). Distinguish this clearly from a requirements-quality-validator (which checks existing requirements, doesn't generate tests from them) or a DOM-crawl/page-object synthesizer (which maps existing UI, doesn't derive test cases from requirements). If genuinely absent, score 0/10 — отсутствует and say so plainly; do not round up because the surrounding ecosystem is otherwise strong.

Process

1. Determine the target stack. Confirm language + automation tool with the user if not already named. Note the current count/list of supported AI assistants and CI providers from the generator source (packages/cli/src/questionnaire/schema.ts, packages/engine/src/plan/templates/) so categories 11 and 4 use accurate, current lists rather than assumptions from a prior audit.
2. Get real generated output for that stack — never audit from memory or from templates alone.
   • First check whether a real generated sample for that exact stack already exists checked into the repo (e.g. PlaywrightTests/ for TS+Playwright). If found, verify it isn't stale: compare its .eitr/manifest.json engineVersion against packages/engine/src/version.ts.
   • If no sample exists for the requested stack, generate one fresh via the CLI into the scratchpad/sandbox directory (never into the repo root — see CLAUDE.md Section 4 cleanliness rule) and clean it up after the audit unless the user asks to keep it.
   • Either way, cross-check the sample against the current template source in packages/engine/src/plan/templates/<lang>/ to catch template-vs-output drift.
3. Run two parallel, independent read-only audits (spawn as background Agent calls so they don't block each other or bloat your context) — never do this investigation serially, and never let one agent see or bias the other's findings:
   • Agent 1 — framework code (categories 1-10). Best-fit agent type: framework-auditor if the stack is one it already knows how to navigate, otherwise general-purpose. Give it: the real sample path, the relevant template source path for the stack, and the exact category list from Group A above with the check criteria verbatim, plus an instruction to propose stack-specific killer-feature ideas (not generic ones).
   • Agent 2 — AI ecosystem (categories 11-19). Agent type general-purpose. Give it: the sample's AI-assistant directories, the MCP bridge and TMS adapter source paths, and the exact category list from Group B above with the check criteria verbatim, plus the augmentation-not-replacement design principle (agents prepare decisions, humans approve consequential/irreversible ones — no autonomous merge/auto-approve designs) as the lens for killer-feature ideas.
   • Instruct both agents explicitly: do not modify any files (read-only audit); cite concrete file:line evidence for every claim; do not pad or round up scores for categories the rest of the codebase is strong in — score each category independently on its own evidence (anti-sycophancy: this is a ruthless, highly critical audit, not a pitch); report ALL gaps found, not just the top few; and return findings as structured markdown, one section per category, each with a /10 score, evidence bullets with file references, and category-specific gaps.
4. Synthesize, don't just concatenate. Once both agents report back:
   • Compute the overall average across the 18 numerically-scored categories (exclude category 19 from the average if it scored 0 — отсутствует, and call that out explicitly as a separate headline fact rather than folding a hard zero into an average that would understate everything else).
   • Merge both agents' flat gap lists into one deduplicated, ordered list — keep every item, most severe first, mark which are minor (secondary color) vs. major/critical.
   • Merge both agents' killer-feature ideas into one list, tagged by area.
   • Sort the merged gaps into a two-column priority split: quick fixes (hours) vs. strategic investments (days-to-weeks) — a defect a user could fix same-day (typo, missing cache flag, wrong RNG call) doesn't belong next to a missing agent architecture.
5. Build the report using the exact deterministic HTML/CSS structure below. Load the artifact-design skill first (required by the Artifact tool before writing any artifact file) — but do not redesign from scratch: reuse the token system, layout, and section order below verbatim so every audit renders identically regardless of stack or date. The only things that change between runs are the stack name, date, scores, evidence bullets, gap list, killer-feature list, and priority items — never the visual system or the category taxonomy.
6. Write the response in Russian (per this repo's own CLAUDE.md Section 6 language rule), concise, with the artifact link and a short critical summary — not a restatement of the whole report in chat.

Deterministic report template (HTML skeleton — fill placeholders, keep everything else identical)

The full HTML skeleton (CSS tokens, section order, and {{PLACEHOLDER}}/<!-- REPEAT --> markers) lives in references/html-skeleton.md — read it in full before building any report artifact. Reuse it verbatim; only the stack name, date, scores, evidence bullets, gap list, killer-feature list, and priority items change between runs, never the visual system or the category taxonomy.

Scoring discipline (anti-sycompancy guard)

• A category score is earned by evidence in that category alone. Do not let a strong overall impression of the codebase pull a weakly-evidenced category up, and do not let one bad category drag down a category that has its own solid evidence.
• Every score below 10 needs a stated, concrete reason in the card's evidence bullets — "not quite 10" without a named defect is not acceptable per this repo's own Uncompromising Engineering Rigor mandate.
• Category 19 (requirements→test-case generation agent) is a binary existence check dressed as a 1-10 category: if genuinely absent, it is 0/10, not 3/10 for "partial credit" from an adjacent agent that does something else. Do not average it into a softer number.
• Treat "the docs say it enforces X" and "the code actually enforces X" as two different claims — always verify the second before crediting a category for it.

After publishing

Report back to the user in Russian: the artifact link, the overall score with one ruthless-but-fair sentence of context, and the 2-4 single most consequential findings (not a re-listing of all 19 categories — the artifact is the detailed record, the chat message is the executive takeaway).