Update to Fievel

{
  "description": "Create plain-language Fievel quality checks, check saved responses, or prepare human review without claiming unexecuted tests ran.",
  "included_files": [],
  "name": "evaluation-designer",
  "skill_md_contents": "---\nname: evaluation-designer\ndescription: Create plain-language Fievel quality checks, check saved responses, or prepare human review without claiming unexecuted tests ran.\n---\n\n# Evaluation Designer\n\n## Using this skill from the Fievel plugin\n\nThis plugin carries Fievel's workflows and Python tools. The user's project is\nseparate from the installed plugin. Find the plugin root two directories above\nthis SKILL.md; its `runtime/` directory is the bundled resource root, called R below.\nResolve R from this skill's actual installed path, never a development checkout.\n\n- Prefer Fievel's MCP tools when they are available (names start with `fievel_`).\n  Local tools such as `fievel_init`, `fievel_validate`, `fievel_check` and\n  `fievel_view` work only inside the user's workspace and never overwrite files;\n  call `fievel_setup_status` first. Remote tools such as `fievel_validate_package`\n  and `fievel_check_responses` receive package content, store nothing, and return\n  results to save as new files; tell the user that content is sent to the Fievel\n  server for processing. Without tools, use the CLI as described below.\n- Read `R/AGENTS.md` for Fievel's conversation and workflow rules. References in\n  this skill and in AGENTS.md to AGENTS.md, docs/, knowledge/, templates/, examples/,\n  schemas/, scripts/ and .agents/skills/ are relative to R, not the user's working directory.\n- Keep the current working directory in the user's chosen workspace. Run the CLI\n  by its absolute path: `python \"R/scripts/fievel.py\" <command> <project> ...`.\n  Replace R with the resolved path. Resolve project and output paths in the user's\n  workspace. Run one simple command per step: the Python interpreter\n  with the CLI path and its arguments, without chained commands, shell variables\n  or PowerShell scripts. Read and list files with the host's file tools. Simple\n  commands keep the user's permission prompts clear and predictable. Never create\n  projects, install dependencies, or edit inside R or the plugin cache. Never copy Fievel AGENTS.md over a user's repository instructions.\n- Check Python 3.10+ and `python -c \"import jsonschema\"`. If missing, explain the\n  prerequisite. When setup is authorized, create a new `.fievel-venv` in the user's\n  workspace and install `R/requirements.txt` with its Python. Reuse a working\n  environment; do not overwrite an existing environment. Use that interpreter for\n  every CLI command. Do not install globally or claim the plugin bundles Python.\n- Follow existing workspace instructions. Explain that project files remain local,\n  but material read into the conversation is handled by the user's coding agent.\n- The repository-maintenance and adapter-sync instructions in R/AGENTS.md are for\n  Fievel development, not for modifying the installed plugin during a user session.\n- Other Fievel skills in this plugin: getting-started, content-system-builder,\n  source-discovery, ai-behavior-designer, evaluation-designer and content-audit.\n  Switch to the one that fits when the user's request changes.\n\nSave projects under the user's chosen folder (default `my-content-system/` in\ntheir workspace). The user approves content; validation does not. Do not run\nmodel evaluations unless the user authorized them.\n\n## Workflow\n\nRead AGENTS.md and stages 5–6 of docs/WORKFLOW.md. Required inputs: actual design\nrequirements, realistic user situations and a project folder. Read\nknowledge/evaluations/checking-quality.md on demand. Translate “good” into inspectable\nquestions; link each check's requirement_id to a standards or behavior rule.\nCover ordinary, unknown, conflicting-source and action-request situations.\n\nUse contains/not_contains only for case-insensitive literal checks. Use human for\nmeaning, invented facts, tone, accessibility and action authorization. A phrase\nabsence cannot prove the absence of unauthorized action. Store definitions in\nevaluation.json. Drafts can be saved before coverage is complete; before execution,\nrun `python scripts/fievel.py validate <folder> --require-coverage` and address\nuncovered requirements. With no saved responses, explain that checks are defined,\nnot run.\n\nTo inspect a real saved execution report alongside the package, generate\n`python scripts/fievel.py view <folder> --report <execution.json> --out <new.html>`.\nThe viewer reproduces literal outcomes before presenting totals, and marks stale,\nunsupported or altered reports. It does not validate completed human judgments\nor establish the origin of saved responses. Do not equate reproduced with approved.\n\nWith actual saved text mapped by scenario ID, run:\n`python scripts/fievel.py check <folder> --responses <json> --out <new-report.json>`.\nPrepare a separate review using `review` with the same arguments. Human questions\nstay not_run in execution and pending in the review sheet. The reviewer records\nidentity, date, pass/fail and rationale; never fabricate completed human judgments.\nFor a self-contained evidence review, read docs/HUMAN-REVIEW.md. Use\n`review <folder> --responses <json> --evidence-config <config> --out <new.json>`.\nThe configuration supplies draft/reviewed rubrics, explicit source excerpts and\nresponse origins; never invent approved criteria or fetch source locations implicitly.\n`validate-review <report> --require-complete` checks completed records; add\n`--folder <folder>` to reject staleness. Missing responses remain not_run; pending,\nborderline and insufficient_evidence cannot count as complete. A complete record\ncan contain failures and is not content approval or authenticated human authorship.\nLegacy sheets must be regenerated in the opt-in format rather than relabeled.\n\nAI judgments, if requested conversationally, must be labeled AI suggestions and\nstored separately; no judge backend is implemented.\n\nOutput: definitions, actual report if executed, pending review, suggested improvements\nin a separate Markdown note. Compare hashes before reusing results after edits.\nRecovery: no responses or unknown IDs -> repair the mapping, never substitute fixtures\nas evidence about a live agent. Example: “Don’t invent decline reasons” -> human check.\n"
}