Getting it into your agent
It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.
git clone --depth 1 https://github.com/luanpdd/kit-mcpnpx agentmods add agents/luanpdd/kit-mcp/llm-eval-harness-writerWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/agents/luanpdd/kit-mcp/llm-eval-harness-writer)<a href="https://agentmods.dev/agents/luanpdd/kit-mcp/llm-eval-harness-writer"><img src="https://agentmods.dev/badge/agents/luanpdd/kit-mcp/llm-eval-harness-writer/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/agents/luanpdd/kit-mcp/llm-eval-harness-writer"><img src="https://agentmods.dev/badge/agents/luanpdd/kit-mcp/llm-eval-harness-writer.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00056 | $0.06361 |
| Opus 5 | $0.00028 | $0.03180 |
| Sonnet 5 | $0.00011 | $0.01272 |
| Haiku 4.5 | $0.00006 | $0.00636 |
Grade A, and why
llm-eval-harness-writer scanned grade A with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 8d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Runs shell commandslowCapability
Expected in a hook, worth knowing in a rule or an instructions file.
import { execSync } from "node:child_process" How it starts
The opening of the file, as written. The whole thing — 405 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Você é o escritor de harness de eval de LLM. Recebe um target_prompt (ou tool/agent que chama um LLM) e produz uma suite de avaliação contra rubrica: um golden dataset de casos rotulados, uma rubrica de critérios pontuáveis, um runner LLM-as-judge determinístico (temperature=0 + seed fixo), um agregador de score e um gate de CI que falha quando a qualidade regride entre versões de prompt. Você gera os arquivos (arquétipo writer); não executa a avaliação em produção.
Você consulta — não duplica:
ai-prompt-characterization— determinismo (temperature=0, seed), sanitização de PII, as 5 intents canônicasllm-as-dependency— adapterLLMProvider+FakeLLMProviderpara isolar custo/latência do judge nos testes
Compat: Full em todos os IDEs (filesystem-only; o runner roda no CI do projeto). Veja COMPATIBILITY.md.
Por que existe
Equipes mexem em prompt de produção "no olho" e descobrem a regressão pelo ticket do cliente. Snapshot tests congelam o output exato (qualquer mudança quebra, mesmo melhoria); characterization preserva comportamento atual como oracle; mutation testing acha pontos cegos do teste. Nenhum deles mede se a saída é boa. Este agent materializa a peça que falta: uma rubrica explícita de qualidade + um judge que pontua cada caso contra ela + um baseline versionado + um gate que reprova o PR quando o score médio cai.
Distinção dos vizinhos (não invadir):
legacy-characterizer→ snapshot/oracle do comportamento atual (igualdade, não qualidade).ai-mutation-tester→ pontos cegos do teste (mutants comportamentais que sobrevivem).- skill
llm-as-dependency→ adapter para tornar o LLM testável (estrutura, não medição). - Este agent → mede qualidade vs critério com judge + rubrica + gate de regressão.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 8d ago First seen · 405 lines · 56 tokens per session scan A 1764acce3eb5
llm-eval-harness-writer is an agent published in the GitHub repository luanpdd/kit-mcp (1 stars, last pushed yesterday), licensed MIT. It adds 56 tokens to every session and 6,361 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 1 finding (runs shell commands). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other agents, from other repositories
unfolding-po
PO (Product Owner) role in the Unfolding Specs process. Decomposes Features into smaller Features, creates Acceptance Tests, and identifies implicit business assumptions as Domain Model Decisions (DMDs).
evals
LLM evaluation — eval harness design, benchmark suites, automated regression, human eval orchestration.
llm2bedrock-prompt-evaluator
Run each golden prompt against the target Bedrock model via the pinned uv harness, score with LLM-as-judge, and report a pass rate. Handles throttling with backoff; returns a structured eval object, or a partial/blocked control state.
gsd-eval-planner
Designs a structured evaluation strategy for an AI phase. Identifies critical failure modes, selects eval dimensions with rubrics, recommends tooling, and specifies the reference dataset. Writes the Evaluation Strategy, Guardrails, and Production Monitoring sections of AI-SPEC.md. Spawned by /gsd:ai-integration-phase…
eval-specialist
LLM evaluation specialist. Designs frameworks to evaluate prompt quality, RAG retrieval performance, and overall app quality. Builds benchmarks, A/B tests, and regression tests.
FAI AI Search Portal Reviewer
AI Search Portal reviewer — index schema audit, search relevance testing, answer citation accuracy, facet UX review, and performance benchmarking.