Quality eval orchestrator

Quality-Eval Orchestrator

The single entry skill for proving software (and the agents that write it) actually works. It places the task on the gate-stage × subject map and delegates to one of 12 specialist spokes. The cross-cutting model every spoke shares — the four quality gates (write → commit → CI → ship), what "evidence" means at each, and how pass@k separates a flaky pass from a real one — lives in quality-eval-core; read it before deciding what counts as "done" or wiring a measurement loop.

Cluster map (spoke → role)

• eval-harness — formal eval-driven development: pass/fail criteria + pass@k reliability for Claude Code task completion.
• agent-eval — head-to-head comparison of coding agents (Claude Code, Aider, Codex…) by pass rate, cost, time, consistency.
• tdd-workflow — red→green→refactor discipline with 80%+ coverage (unit + integration + E2E) on new features, bugfixes, refactors.
• verification-loop — the commit-time gate: a structured "is this actually finished?" check before a PR.
• e2e-testing — Playwright patterns: Page Object Model, config, CI/CD, artifacts, flaky-test strategies (web apps).
• windows-desktop-e2e — native desktop E2E via pywinauto + Windows UIA (WPF, WinForms, Win32/MFC, Qt).
• ai-regression-testing — regression patterns for AI-written code: sandbox/mock-mode API tests, blind-spot catching where one model both writes and reviews.
• browser-qa — automated visual + interaction verification of a deployed/preview UI.
• error-handling — robust error design across TS/Python/Go: typed errors, boundaries, retries, circuit breakers, user-facing messages.
• benchmark — performance baselines + regression detection before/after a PR; stack-alternative comparison.
• plankton-code-quality — write-time gate: auto-format, lint, and Claude-powered fixes on every file edit via hooks.
• production-audit — release-time gate: local-evidence production-readiness audit ("what breaks in prod?") with no third-party data sharing.

Folded spokes (also routable)

These spokes were folded into the cluster from the wider skill library. They overlap the gates above but each adds a distinct capability — route to them when the request matches:

• evals — Anthropic-style agent-workflow evaluation: three grader types (code/model/human), transcript capture, pass@k/pass^k, capability-vs-regression suites with saturation tracking. Use when grading multi-turn agent trajectories (not just final outputs) or building a reusable eval suite; eval-harness stays the lighter pass/fail-criteria entry point.
• optimize — autonomous hill-climbing optimization loop (/optimize): metric mode (a shell command emitting a number — latency, bundle size, page speed) or eval mode (LLM-as-judge for skills/prompts/agents). Use when the goal is to iteratively improve a target, not just measure it once; complements benchmark (which establishes the baseline this loop climbs from).
• autoresearch — Karpathy-style keep-or-discard experiment loop: change one variable, measure, keep or revert, in bounded batches. Use for overnight/iterative research runs on code (train.py-style) or on business hypotheses (positioning, CTA, pricing) where each run is logged with an explicit decision.
• markdown-rendering-regression — rendered-HTML regression gate for Markdown/MDX builds (Astro docs/wiki): catches raw-Markdown leaks, frontmatter leakage, placeholder copy, local filesystem paths, internal build labels, and banned emoji icons that a green build still ships. Use after a static build produces dist/; pairs with browser-qa for visual checks.

Picked-up spokes

Vetted from the antigravity-awesome-skills library (MIT). Each fills a gap the gates above didn't cover — route to them when the request matches:

• accesslint-scan — live-page accessibility audit: runs axe over a running page/preview URL, locates every WCAG violation by selector (+ file:line when source-mapped), and returns a prioritized fix worklist without editing. Use for a full a11y audit of a deployed/local UI; the runtime counterpart to the design-time accessibility/frontend-a11y guidance.
• accesslint-diff — accessibility regression gate: diffs a live page's violations against a baseline (uncommitted-changes stash mode by default, or --branch <name>), reporting only new violations introduced vs fixed vs pre-existing. Use to prove a UI change ships no new WCAG violations before a PR; pairs with accesslint-scan.
• fix-review — remediation-verification gate: confirms that fix commits actually resolve audit/security findings at the root cause, introduce no new bugs or vulnerabilities, and carry test coverage for the fixed scenario. Use when reviewing commits that claim to close findings; complements verification-loop (general "is it done?") and security-review with a findings-to-fix focus.
• simplify-code — safe-simplification review: reviews a scoped git diff across four parallel roles (reuse, quality, efficiency, clarity) and optionally applies only high-confidence, behavior-preserving fixes. Use when asked to simplify/clean up/refactor changed code without changing behavior; a quality-only pass (it does not hunt for bugs — that's verification-loop/fix-review).
• clean-code — Clean Code reference: Robert C. Martin's principles (intention-revealing names, small single-purpose functions, comment discipline, error handling, code-smell heuristics) as a review/refactor checklist. Use for maintainability-focused PR review or legacy refactors; the language-agnostic readability rubric simplify-code and plankton-code-quality operationalize.

Routing rules by intent

Define what "done" means / measure reliability

• "How do I know the agent succeeded?" / pass@k → eval-harness (model in quality-eval-core)
• "Which coding agent/model is best for this repo?" → agent-eval
• Grade multi-turn agent workflows/transcripts, build a reusable eval suite (graders, pass^k, saturation) → evals

Write code correctly (test-first)

• New feature / bugfix / refactor → tdd-workflow
• Designing error types, retries, circuit breakers → error-handling

Verify a change is finished

• Final "is it really done?" before a PR → verification-loop
• Catch regressions in AI-modified backend/API → ai-regression-testing
• Verify fix commits actually close audit/security findings without new bugs → fix-review

Improve clarity / maintainability of a change

• Simplify / clean up / refactor a diff with safe, behavior-preserving fixes → simplify-code
• Apply Clean Code naming/function/smell heuristics in review or refactor → clean-code

Accessibility (runtime)

• Full WCAG audit of a live/preview page with a fix worklist → accesslint-scan
• Prove a UI change adds no new accessibility violations (diff vs baseline/branch) → accesslint-diff

Test behavior end-to-end

• Web app flows → e2e-testing
• Native Windows desktop app → windows-desktop-e2e
• Visual / interaction check on a deployed UI → browser-qa
• Rendered Markdown/MDX (Astro docs/wiki) leaking raw syntax, placeholders, or local paths in dist/ → markdown-rendering-regression

Performance

• Baselines, before/after PR, "it feels slow" → benchmark

Optimize / iterate toward a target

• Autonomously hill-climb a metric (latency, bundle size, page speed) or an LLM-judged quality target (skill/prompt/agent) → optimize
• Karpathy-style keep-or-discard experiment loop, overnight/bounded batches (code or business hypotheses) → autoresearch

Guard quality continuously / before shipping

• Format + lint on every edit → plankton-code-quality
• "Is it ready to ship / what breaks in prod?" → production-audit

Sibling clusters

• Language/stack-native tests — for a specific language or stack, defer to that cluster's *-testing spoke: python-testing (python-backend), rust-testing (rust), react-testing (frontend-web), golang-testing / cpp-testing / perl-testing (systems-languages), kotlin-testing (jvm), swift-protocol-di-testing (native-ios), testing-tauri-apps (tauri). This cluster owns the language-agnostic discipline (TDD, the eval harness, the four gates) — not per-stack test syntax or framework idioms.

Standard flow

1. Locate the task: which gate stage (write-time → test → verify → benchmark → ship) and which subject (code, a web/desktop UI, or an agent).
2. If it touches what counts as done, pass@k, or a measurement loop, pull the model from quality-eval-core first — gates and evidence are shared, not per-spoke.
3. Delegate to the spoke(s). Multi-step asks fan out in gate order (e.g. "ship this feature safely" → tdd-workflow → e2e-testing → verification-loop → production-audit).
4. Return: chosen spoke(s), the gate stage, what evidence will prove the outcome, and the next action.

Guardrails

See quality-eval-core. In short: evidence over assertion — never report a pass without the command output that proves it; a single green run is not a pass (use pass@k for anything non-deterministic or agent-driven). Don't let the thing being tested also silently grade itself (the AI write-and-review blind spot), and don't let a gate be satisfied by loosening its own config — flag config tampering. Pick the cheapest gate that catches the failure class: a write-time lint shouldn't be doing a production audit's job, and vice versa.

Loading spokes on demand

To keep CLI startup context lean, this cluster's spokes are not separately registered as skills — only this orchestrator and its *-core are enumerated. When you route to a spoke named above, load it on demand by reading its file:

~/.agents/skill-clusters/skills/<spoke-name>/SKILL.md (or skills/<spoke-name>/SKILL.md inside the skill-clusters repo).