bmad-tea-evaluate — скилл для Claude Code: оценка агента
Скилл bmad-tea-evaluate для Claude Code проводит поведенческую оценку агента и помогает закрыть найденные пробелы. Запускается по просьбе оценить агента или подтвердить, что он ведёт себя правильно.
MIT 866 установок ★ 118 обновлён 2026-11-13
npx skills add bmad-code-org/bmad-method-test-architecture-enterprise --skill bmad-tea-evaluate -a claude-codeЧто делает bmad-tea-evaluate
Скилл проверяет агента по поведенческим требованиям и формирует оценку на этапах разработки и на отложенной выборке. Он выявляет пробелы, помогает их исправить и завершает работу планом для CI, который будет применять эту проверку.
Работа проходит через последовательные этапы: осмотр цели, сбор требований и проектирование набора проб. Для оценки нужны настроенный TEA в проекте и uv; решение о результате скилл получает через CLI eval-quality, а не вычисляет сам. Перед третьим этапом требуется подтверждение требований со стороны пользователя.
При продолжении работы скилл читает сохранённые записи оценки и связанные файлы проекта. Если конфигурации TEA нет, он просит сначала выполнить bmad setup tea и останавливается.
Какие задачи решает
- проверить поведение агента по заданным требованиям
- найти пробелы в поведенческой оценке и подготовить исправления
- составить план CI для проверки агента
Примеры запросов агенту
- Оцени этого агента и покажи, где его поведение не совпадает с требованиями.
- Хочу доказательство, что агент выполняет эти сценарии правильно.
Как установить bmad-tea-evaluate
1Через Skills CLI (Vercel)
npx skills add bmad-code-org/bmad-method-test-architecture-enterprise --skill bmad-tea-evaluate -a claude-code2Вручную: положить файлы
~/.claude/skills/bmad-tea-evaluate/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/bmad-tea-evaluate/
1Через Skills CLI (Vercel)
npx skills add bmad-code-org/bmad-method-test-architecture-enterprise --skill bmad-tea-evaluate -a codex2Вручную: положить файлы
.agents/skills/bmad-tea-evaluate/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add bmad-code-org/bmad-method-test-architecture-enterprise --skill bmad-tea-evaluate -a cursor2Вручную: положить файлы
.cursor/skills/bmad-tea-evaluate/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install bmad-code-org/bmad-method-test-architecture-enterprise skills/bmad-tea-evaluate2Вручную: положить файлы
.github/skills/bmad-tea-evaluate/SKILL.md1Через Skills CLI (Vercel)
npx skills add bmad-code-org/bmad-method-test-architecture-enterprise --skill bmad-tea-evaluate -a gemini-cli2Командой платформы
gemini skills install https://github.com/bmad-code-org/bmad-method-test-architecture-enterprise --path skills/bmad-tea-evaluateВручную: положить файлы
.devin/skills/bmad-tea-evaluate/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Через Skills CLI (Vercel)
npx skills add bmad-code-org/bmad-method-test-architecture-enterprise --skill bmad-tea-evaluate -a opencodeClaude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о bmad-tea-evaluate
Как установить bmad-tea-evaluate в Claude Code?
Выполни npx skills add bmad-code-org/bmad-method-test-architecture-enterprise --skill bmad-tea-evaluate -a claude-code. Файл скилла будет доступен по пути ~/.claude/skills/bmad-tea-evaluate/SKILL.md.
Что делает bmad-tea-evaluate?
Скилл проводит оценку поведения агента по требованиям, фиксирует пробелы и помогает подготовить исправления. В конце он формирует план CI для этой проверки.
Что нужно для работы bmad-tea-evaluate?
В проекте должны быть настроены TEA и каталог _bmad, а также нужен uv. Оценку и вердикт выполняет CLI eval-quality.
Исходный текст
SKILL.md · MIT
---
name: bmad-tea-evaluate
description: 'Run scored behavioral evaluations and repair gaps. Use when the user says "evaluate this agent" or "I want proof this behaves correctly"'
---
# Evaluate
## Overview
**Goal:** Take an adopter from a described target through a scored development and held-out evaluation with named gaps and repairs, and finish it with the CI plan that enforces it.
**Role:** You are the Master Test Architect.
You will continue to operate with your given name, identity, and communication_style, merged with the details of this role description.
## Conventions
- Bare paths (e.g. `references/inspection.md`) resolve from the skill root.
- `{skill-root}` resolves to this skill's installed directory (where `customize.toml` lives).
- `{project-root}` is the nearest ancestor of the working directory that contains an `_bmad` directory.
- `{skill-name}` resolves to the skill directory's basename.
- Resolve sibling files such as `references/...` and `assets/...` from `{skill-root}`.
## Execution Authority
Resolve intent before activation, customization resolution or any hook. Ordinary evaluation keeps its current workflow unchanged. Native clean/mutated trials, run comparison and baseline acceptance retain their existing authority and meaning.
An explicit native compare/accept request keeps that command's authority, including a request that mentions consistency while comparing saved runs. Saved-result consistency is an experiment route only when the user asks to analyze variability in existing results. Resolve contextual words such as “these results” from the conversation before choosing a route.
- Saved-result consistency or explicitly requested offline experiment analysis: load `references/configuration-experiments.md` and take its offline route directly. Suppress prepend, append and completion hooks. Do not run configuration resolver scripts, model calls, judge calls or fresh trials. Return after reporting the saved evidence.
- Fresh repeatability or a one-factor configuration experiment: load `references/configuration-experiments.md` directly. Suppress prepend, append and completion hooks throughout this route. No hook can grant execution or spending authority.
- “Is it repeatable?” without a saved/fresh choice: clarify saved records versus fresh runs before executing anything.
- Ordinary evaluation, ordinary offline record evaluation or native compare/accept (including comparing saved runs): proceed with On Activation below. Do not load the experiment guide, analyze experiments or add trials after the normal result.
## On Activation
Before resolving customizations or running hooks, inspect `{project-root}/_bmad/custom/` for `bmad-testarch-evaluate.toml` and `bmad-testarch-evaluate.user.toml`. If either exists, stop and tell the user to run `npx --package=bmad-method-test-architecture-enterprise@4 tea-migrate-customizations --project-root {project-root}`. Resolve any reported manual merge, then start a fresh chat. Do not continue with default settings while legacy overrides remain.
### Step 1: Resolve the Workflow Block
Run: `uv run {project-root}/_bmad/scripts/resolve_customization.py --skill {skill-root} --project-root {project-root} --key workflow`
**If the script fails**, resolve the `workflow` block yourself by reading these three files in base → team → user order and applying the same structural merge rules as the resolver:
1. `{skill-root}/customize.toml` — defaults
2. `{project-root}/_bmad/custom/{skill-name}.toml` — team overrides
3. `{project-root}/_bmad/custom/{skill-name}.user.toml` — personal overrides
Any missing file is skipped. Scalars override, tables deep-merge, arrays of tables keyed by `code` or `id` replace matching entries and append new entries, and all other arrays append.
### Step 2: Execute Prepend Steps
Execute each entry in `{workflow.activation_steps_prepend}` in order before proceeding.
### Step 3: Load Persistent Facts
Treat every entry in `{workflow.persistent_facts}` as foundational context you carry for the rest of the workflow run. Entries prefixed `file:` are paths or globs resolved from `{project-root}` — expand them and load every matching file in lexical path order as facts. All other entries are facts verbatim.
### Step 4: Load Config
Run `uv run {project-root}/_bmad/scripts/resolve_config.py --project-root {project-root} --key core --key modules.tea`. If the script fails, merge `{project-root}/_bmad/config.toml`, `{project-root}/_bmad/custom/config.toml` and `{project-root}/_bmad/custom/config.user.toml` yourself, in that order, with the merge rules above. If `_bmad/config.toml` is missing or has no `modules.tea` table, tell the user TEA is not set up, ask them to run `bmad setup tea` first, and stop.
Each `core` and `modules.tea` key is available as `{<key>}`, and as `config.<key>` in later steps. Replace `{project-root}` inside a value with the project root. Setup answers are strings: read `"true"` and `"false"` as booleans.
`{tea_evaluations_folder}` is `{workflow.evaluations_folder}`, relative to `{project-root}`.
### Step 5: Greet the User
Greet `{user_name}`, speaking in `{communication_language}`.
### Step 6: Execute Append Steps
Execute each entry in `{workflow.activation_steps_append}` in order.
Activation is complete. Begin the workflow below.
## Workflow
Evaluate is one continuous loop over twelve stages, run inline rather than as separate step files: each stage's craft lives in its own `references/` guide, loaded when that stage is reached. Work under `{tea_evaluations_folder}` unless the adopter names another location. On resume, read the existing inspection record, `evaluation.json`, any `requirements.md`, any `authoring-log.md`, any `<evaluation-folder>/ci/evaluation-ci-plan.json`, and any `{test_artifacts}/evaluate/<evaluationId>/gap-report.md` before choosing a stage. Before Stage 3, require the adopter's confirmation in the statement and compare `requirements.digest` with eval-quality's `digestBytes` over the committed file; otherwise return to intake. Continue from the first incomplete stage, or the adopter's requested stage when its prerequisites hold. `assets/README.md` says how to fill each template under `assets/`.
Read `references/authoring-routing.md` before delegating any authoring work. Use its bounded worker tasks and authoring ledger throughout the loop when the host supports subagents.
Never improvise the stage's craft yourself, and never compute a verdict, score, or pass/fail decision outside `eval-quality`'s own CLI (AD-6): a verdict this skill computed itself would not be one `eval-quality` sealed.
### Stage 1: Inspection
Inspect the target and identify its kind. Load `references/inspection.md`.
### Stage 2: Intake
Capture behavioral requirements and constraints the adopter confirms. Load `references/intake.md`.
### Stage 3: Corpus
Design the probe corpus. Load `references/corpus.md`.
### Stage 4: Contract
Author the Behavioral Evaluation Contract. Load `references/contract.md`. Stamp its `sourceSpecDigest` with eval-quality's `digestBytes` over the confirmed `requirements.md` bytes, matching `evaluation.json` `requirements.digest`.
### Stage 5: Oracles
Design oracles and rubrics. Load `references/oracles.md`.
### Stage 6: Adapters
Scaffold the execution-target registry and adapters. Load `references/adapters.md`.
When the contract, oracles and registry are filled, copy `assets/evaluator-conditions.template.json` to `policy/evaluator-conditions.json` and fill the target's `modelSnapshot` and `systemPromptDigest`. A `tea-skill-runner` registry needs a real model snapshot. For a deterministic evaluator with a rubric, declare `evaluation.json.judge` and fill `judge.modelSnapshot`; remove the unused template `judge` block for other cases. Ensure `policy/scoring-policy.json` has been copied from the installed template and filled if any probe takes the `controlled-mutation`, `historical`, or `gameability` route. Ask the adopter for its `policyId` and thresholds, then set `evaluation.json.trials` at or above `minimumTrialCount`.
Before the first `tea-evaluate check`, load `references/mutation.md` and author every `mutations/<mutationId>.mutation.json` named by a Stage 3 `controlled-mutation` probe. Use an adopter-owned target, a single exact replacement, and the probe's observable defect signature. Refresh `corpus-index.json` after writing those files. For an adopter installation, run `npm exec --prefix {tea_evaluations_folder} -- tea-evaluate digest --evaluation <evaluation-folder>`; for TeA's own package, run `node cli/evaluate.js digest --evaluation <evaluation-folder>` from the repository root. Stage 6 preflight qualifies the nominated probes; Stage 8 inspects their manifestation and rollback evidence and expands the mutation set.
For an adopter installation, run `npm exec --prefix {tea_evaluations_folder} -- tea-evaluate check --evaluation <evaluation-folder>`, then `npm exec --prefix {tea_evaluations_folder} -- eval-quality compile --in <evaluation-folder>/contract.json --out <evaluation-folder>/compiled-contract.json`, then `npm exec --prefix {tea_evaluations_folder} -- eval-quality seal --in <evaluation-folder>/contract.json --out <evaluation-folder>/sealed-brief.json`. For TeA's own package, run `node cli/evaluate.js check --evaluation <evaluation-folder>`, then `./node_modules/.bin/eval-quality compile --in <evaluation-folder>/contract.json --out <evaluation-folder>/compiled-contract.json`, then `./node_modules/.bin/eval-quality seal --in <evaluation-folder>/contract.json --out <evaluation-folder>/sealed-brief.json` from the repository root. The local CLI and binary use that root's one eval-quality installation. Stop at the first nonzero exit. Report its command, exit code and stderr before changing the artifact and rerunning that stage.
After seal succeeds, copy `assets/evaluation-folder.gitignore` to `<evaluation-folder>/.gitignore` before the first preflight creates `runs/`. Before a live preflight, follow the Git isolation procedure in `references/run.md`; Git worktrees share refs even when their working files are separate. For an adopter installation, run `npm exec --prefix {tea_evaluations_folder} -- tea-evaluate preflight --evaluation <evaluation-folder> --partition development`; for TeA's own package, run `node cli/evaluate.js preflight --evaluation <evaluation-folder> --partition development` from the repository root. This verifies the selected registry can launch and satisfy its clean control. The flag is required only when `evaluation.json` declares a `partitionPlan`, and a preflight with no `--partition` over one exits 64. Held-out preflight (`--partition held-out`) runs only after the development review. Stop and report the command, exit code and stderr on failure; complete adapter wiring only after preflight succeeds.
### Stage 7: Evaluator
Choose or build the evaluation layer. Load `references/evaluator.md`.
### Stage 8: Mutation
Apply and roll back controlled mutations. Load `references/mutation.md`.
### Stage 9: Harness
Scaffold the runner and scoring policy. Load `references/harness.md`.
### Stage 10: Run
Run the clean and mutated arms and score them. Load `references/run.md`.
### Stage 11: Gaps
Interpret a weak result and close the gap. Load `references/gaps.md`.
### Stage 12: CI
Inspect the adopter's repository, place each check in a tier, write `<evaluation-folder>/ci/evaluation-ci-plan.json` and hand it to `bmad-tea-framework` with CI setup scope. Load `references/ci.md`.
## On Complete
Run: `uv run {project-root}/_bmad/scripts/resolve_customization.py --skill {skill-root} --project-root {project-root} --key workflow.on_complete`
If the resolver succeeds and returns a non-empty `workflow.on_complete`, execute that value as the final terminal instruction before exiting.
If the resolver fails, returns no output, or resolves an empty value, skip the hook and exit normally.
© bmad-code-org · MIT · bmad-code-org/bmad-method-test-architecture-enterprise