eval-grader — Claude Code subagent: what it does, how to install
eval-grader — a grader for skill-eval assertions that returns PASS, FAIL or PARTIAL on each, with one line of evidence. Handy when a reviewer's write-up needs a verdict from something that did not produce it.
Elastic-2.0 ★ 155 updated 2026-10-29
/plugin marketplace add serpro69/claude-toolbox
/plugin install kk@claude-toolboxWhat eval-grader does
Hand it a reviewer's write-up plus a list of assertions and it answers with a verdict on each one: PASS, FAIL or PARTIAL, and a single line pointing at the text that settled it. The agent had no hand in producing what it grades, and it never opens the live fixtures, so the grade cannot quietly drift toward the source of truth.
Component mode is the default. Only the output text counts there, and execution traces do not upgrade a grade. Workflow mode has to be named by the controller: the supplied rubric and manifest load first, sealed evidence after that. An unrecognised mode is an input error, and the agent asks for a corrected one rather than guessing.
Read is the only tool, kept to two jobs: fetching a reviewer-output file the harness left on disk, and re-reading its own agent file or playbook. Paths such as test-files/, klaude-plugin/profiles/** and SKILL.md stay shut, because opening them brings rubric leakage back in. A claim the output text cannot support turns into PARTIAL.
What it helps with
- Grade a skill eval's assertions one by one
- Get a verdict from an agent that did not write the output
- Review a reviewer's text without touching the fixtures
- Work through a sealed evidence package in workflow mode
- Spot which behaviour a reviewer's output fails to show
Example prompts for the agent
- Grade the reviewer output against these assertions and give me PASS, FAIL or PARTIAL for each, with one line of evidence.
- Grading mode: workflow — load the rubric and manifest first, then open only the evidence files listed there.
- Check the checklist-loading assertion: if there's no quoted trigger, mark it PARTIAL.
How to install eval-grader
1Plugin marketplace
/plugin marketplace add serpro69/claude-toolbox/plugin install kk@claude-toolbox2Manually: put the files in place
.claude/agents/eval-grader.mdFor all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention
Manually: put the files in place
.codex/agents/eval-grader.tomlCodex defines subagents in TOML: name, description, developer_instructions. Move the Markdown text into developer_instructions by hand. For all projects use ~/.codex/agents/
Manually: put the files in place
.cursor/agents/eval-grader.mdCursor also reads .claude/agents/
Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean
Questions about eval-grader
How to install eval-grader in Claude Code?
Use the marketplace: /plugin marketplace add serpro69/claude-toolbox, then /plugin install kk@claude-toolbox. The agent file ends up at .claude/agents/eval-grader.md.
What does eval-grader do?
It checks a reviewer's output against the supplied assertions and returns PASS, FAIL or PARTIAL for each, with one line quoting the text or naming what is missing.
How is workflow mode different from component mode?
Component judges the output text alone and traces cannot strengthen the grade. Workflow is turned on by the controller, which makes the rubric and manifest load before any sealed evidence is opened.