إنتقل إلى المحتوى الرئيسي

Plugin Evals: Test Your Plugin with claude plugin eval

متقدّم
What you'll learn
  • Understand what a plugin eval measures that `claude plugin validate` and a manual test cannot: does the skill *trigger*, and does it beat a bare model
  • Read the three numbers in every result — WITH, W/OUT, and Δ — and know why some graders are deliberately excluded from the score
  • Pick among the six grader types (four free, two judge-based) and write rubrics that give a stable signal
  • Mock a plugin's MCP servers so a suite is repeatable and never touches the real service
  • Wire a CI gate with pinned models, a cost ceiling, and the exit-code contract

Until now a plugin author had two tools: claude plugin validate, which checks that manifests and frontmatter parse, and their own eyes. Neither answers the question that decides whether a plugin is worth installing: when a user types a natural request, does Claude pick the skill, and is the result better than what Claude would have done anyway? claude plugin eval answers exactly that. It runs each case in a throwaway session with only your plugin loaded, runs it again with no plugin at all, grades both, and reports the difference.

This page is for people who already have a working plugin or skill. If you only need to know the concept, read the first two sections and the quiz.

The mental model: two arms and a delta​

Every case is a prompt plus one or more graders. A grader is a pass/fail check on what Claude produced. For each case Claude Code runs:

  • the with-arm: a fresh headless session with only your plugin loaded, the prompt sent, Claude allowed to work until it finishes or hits the turn/time cap, then graders applied;
  • the without-arm: the same prompt, same number of runs, no plugin.

Each arm runs the case three times by default, because a single run of a non-deterministic agent tells you almost nothing. A run's score is the fraction of graders that passed (weighted, if you set weights). The case's score is the mean across runs. You get:

ColumnMeaning
WITHMean score with the plugin loaded
W/OUTMean score with no plugin
ΔWITH − W/OUT — what the plugin actually contributed

The non-obvious part: a high WITH score alone proves nothing. If a case scores 1.0 in both arms, Claude solved it without you, and your plugin is dead weight for that prompt. The Δ is the number that justifies a plugin's existence. Anthropic's own docs name the most common first finding: a Δ near zero with a failing skill-fired grader, which means Claude never chose your skill on natural phrasing. That is a description problem in SKILL.md, not a code problem, and no amount of manual testing where you invoke the skill by name would have caught it.

Why some graders are excluded from the score​

A check like "the Skill tool was invoked" can never pass in the without-arm. If it counted, it would drag W/OUT toward zero and inflate Δ for free. So in a two-arm run Claude Code excludes from the score, in both arms, every tool_used grader whose tool is Skill, plus anything you mark arm: with-only. They still appear in the report as pass/fail indicators (the report badges them as a "plugin-fired indicator"), and the JSON marks them scored: false. Two edge rules follow:

  • If every grader in a case would be excluded, they are scored normally instead, otherwise there would be nothing left to score.
  • arm: both forces a grader to count in both arms. That is what you want for a "must not invoke the skill" check on a decoy prompt, written as tool_used with min: 0 and max: 0.

Consequence worth remembering: the same suite can report a different absolute score under --ablation none (single arm, nothing excluded) than under the default two-arm mode. Compare like with like when you chart trends.

Anatomy of a case​

A suite lives in evals/ inside the plugin (or another directory you name). Each case is a folder holding a prompt.md, a case.yaml, or both, plus a graders/ folder with one Markdown file per grader.

my-plugin/
├── .claude-plugin/plugin.json
├── skills/...
└── evals/
    ├── drafts-commit-message/
    │   ├── prompt.md            # frontmatter: run limits; body: the user's request
    │   └── graders/
    │       ├── criteria.md      # type: llm — rubric in the body
    │       └── skill-fired.md   # type: tool_used, tool: Skill
    ├── ignores-unrelated-request/
    │   └── ...
    ├── mocks/                   # optional suite-wide MCP mocks
    └── results/                 # written by each run — add to .gitignore

The prompt body is sent to Claude exactly as written. Two things bite first-time authors:

  • Each run starts in an empty working directory with a throwaway home, no user settings, no CLAUDE.md, no personal MCP servers, no other plugins, and only an allowlist of environment variables (basics like PATH, provider auth, most ANTHROPIC_*/CLAUDE_CODE_*, and anything named EVAL_*). If the task needs files, put them in the prompt, scaffold them, or ship them in the plugin.
  • @path mentions in the prompt are not expanded into attachments. Grant Read in allowed_tools if Claude must open a file.

The prompt.md frontmatter fields that matter most:

FieldDefaultNote
runs31 to 50 per arm; --runs overrides
max_turns10Up to 200. Hitting it is logged as a run error and usually lowers the score, so be generous
timeout_seconds300Up to 3600 per run
allowed_tools[]Read-only tools are granted just by listing them; anything else needs a CLI grant
modelsession default--model overrides; pin it in CI
env{}Keys must match EVAL_[A-Z0-9_]* or the run fails
pluginsnearest enclosing pluginSet ["../.."] if auto-detection misses your plugin

case.yaml carries the same fields (execution ones under execution:) plus the three that reference other files: context.scaffold_script (a Bash script that seeds the workspace, only run with --scaffold), context.history_file (a .jsonl transcript to resume, so your prompt becomes the next turn), and context.add_dirs (fixture directories Claude may read).

The six grader types​

Four are computed from the transcript and files and cost nothing. Two call a judge model and add to the bill.

TypeFree?Passes when
regexyesA JavaScript regex is found in the target (last_message by default, or trace, files, a specific file, or mock_calls). match: not_contains for absence, match: "count:N" for an exact count. Case-insensitivity goes in flags: i; inline (?i) is not supported
tool_usedyesCalls to tool whose JSON-encoded input matches input_match number between min (default 1) and max (unlimited). min: 0, max: 0 asserts a tool was never called
tool_orderyesBoth tools were called and the first before match precedes the first after match
file_existsyesA file Claude created matches the path glob (or none does, with exists: false). Files a scaffold created or Claude merely edited do not count
llmnoA judge model votes PASS on your rubric in at least two of three votes
baselinenoA judge finds the run satisfies the criteria at least as well as a reference transcript you saved as baseline_file

There are no custom-code graders. If you need to check that a build or test passed, have the prompt tell Claude to run it and write the outcome to a file, grade that file with regex, and assert the command ran with a tool_used grader whose input_match names it.

The skill-fired grader that almost every case wants looks like this (replace the skill name; the pattern also matches the namespaced plugin:skill form):

graders/skill-fired.md — did my skill actually run?

---
type: tool_used
tool: Skill
input_match: '"skill"\s*:\s*"(?:[\w-]+:)?your-skill-name"'
---

And the judge rubric next to it, written as concrete PASS/FAIL conditions rather than adjectives:

graders/criteria.md — a rubric that gives a stable verdict

---
type: llm
weight: 2
---

PASS if the reply is a single conventional-commit subject line under 72 characters,
starts with "refactor:", and mentions both the rename (getUser -> fetchUser) and the
number of call sites updated.
FAIL if the reply contains more than one candidate message, asks a clarifying
question, or omits the rename.

Habits that keep scores stable​

These are the practices the official docs recommend, and they match what anyone who has run LLM-as-judge pipelines learns the hard way:

  • Grade long output with regex over the file, not with an llm judge. Judge variance grows with the length of what it reads. Keep llm graders for short final messages.
  • One grader on the result, one on the path. A regex/llm/file_exists grader tells you the answer was right; a tool_used/tool_order grader tells you your plugin produced it. You need both to interpret Δ.
  • Suspect the judge before the plugin when Δ is negative but the skill fired. The default judge is a small fast model; it can fail a correct answer for being formatted differently from the rubric. Re-run with --judge-model sonnet and tighten the rubric so formatting does not decide.
  • Iterate on one case with one arm, then confirm at three runs: --case <name> --runs 1 --ablation none. With one arm the table shows SCORE and PASS% instead of WITH/W/OUT/Δ.

Mock the MCP servers​

If your plugin's skills call MCP tools, a run never starts the real servers unless you ask. Instead Claude Code registers a stand-in under each server's name and answers tool calls from Markdown files: evals/mocks/<server>/<tool>.md for the whole suite, or a case's own mocks/ folder to override per case. The file body is the tool result, with {{input.<field>}} substitutions and {{file:fixtures/...}} inserts. A tool with no mock file simply is not available to Claude.

evals/mocks/tracker/create_issue.md — a mock that also asserts the input

---
expect:
title: string
priority: [low, medium, high]
---

Created issue #4821: {{input.title}}

The expect: block turns a mock into an assertion: a call that violates it aborts the run with score 0 and records the server, tool, and reason. Point a grader at target: mock_calls to grade what the plugin asked the server to do. For servers whose answers depend on conversation, type: agent lets a small model play the server; clean runs save its answers under results/<timestamp>/mock-recordings/, and once you copy a recording into mocks/.replay/<server>/ later runs replay it with no model call. Commit .replay/ so CI is deterministic. _tools.json (a saved tools/list response) gives mocked tools their real descriptions and schemas instead of a permissive placeholder.

To hit real servers instead: --allow-real-servers starts the ones without mocks; --mocks off ignores mocks entirely. Either way those processes run as you, outside the sandbox, and their tools still need an explicit grant such as --allow-tools "mcp__plugin_my-plugin_github__*".

Running it​

Guided walkthrough1 of 5
  1. From the plugin root run `claude plugin eval init`. It opens an interactive session in which Claude reads the plugin, asks what a good result looks like, proposes prompts that should and should not trigger it, designs graders, pilots them once, and writes one case directory per prompt. `claude plugin eval init --bare <name>` writes a blank template instead (the only form that works without a terminal, e.g. in CI).

The full-suite command with the flags you will actually reach for:

Run the suite with a stronger judge and a cost ceiling

claude plugin eval . \
--judge-model sonnet \
--max-cost-usd 10 \
--allow-tools Read Write "Bash(npm test *)" \
-j 4

Notes on the flags: -j/--concurrency goes from 1 to 8 and only shortens wall-clock time, because all runs share your account's rate limit. --max-cost-usd is a ceiling on the list-price estimate, checked before each run starts; runs already in flight finish, so spend can overshoot by those runs, and anything left unstarted makes the command exit 2 with partial: true. Put the target (.) before --tag, --allow-tools, and --json: each of those takes a list or optional value and would swallow a target that follows it (the error "--json output path must end in .json" is that mistake).

What it costs​

Every run and every judge vote is a real model call on your account, counted against your plan's usage or your API bill. The rough arithmetic: cases × runs agent runs for the with-arm, the same again for the without-arm, plus three short judge calls per llm or baseline grader per run. The official walkthrough's single case with two graders (one llm, one tool_used) ran six agent runs in 74 seconds at a list-price estimate of about $0.41. Three levers keep a suite affordable:

  • --ablation none halves the agent runs when you are iterating on graders and do not need Δ.
  • Free graders only (regex, tool_used, tool_order, file_exists) for the every-commit suite; judge graders for the nightly one.
  • A usage-limit or rate-limit error mid-suite makes each later run end with that error and usually score 0, and the suite is not marked partial, so a throttled run can look exactly like a regression. Check the NOTES column or cases[].arms.with[].error before believing a sudden drop.

Gate CI on it​

CI job — pinned models, local report, exit code drives the build

claude plugin eval . \
--trust-plugin \
--json results.json \
--threshold 0.8 \
--model claude-sonnet-5 \
--judge-model claude-haiku-4-5 \
--no-publish \
--max-cost-usd 20

Why each flag is there:

  • --trust-plugin asserts the first-run trust decision. Without it a non-interactive job is refused with exit 1 (or hangs at the prompt if the runner allocates a TTY). Pass it only for plugins whose code and suite you would run on your own machine.
  • --model and --judge-model are pinned so a model rollout is never mistaken for a plugin regression. This matters more than it sounds: the agent under test defaults to ANTHROPIC_MODEL or Claude Code's current default, which changes with releases.
  • --threshold 0.8 because the default is 1.0, which makes the command exit 1 whenever any case is less than perfect. A 1.0 bar on a three-run mean of a non-deterministic agent is a flaky gate.
  • --json results.json writes the versioned result document (schemaVersion: 1, camelCase, new fields added without renames) and silences progress output.

The exit-code contract:

ExitMeaning
0Every case at or above the threshold and every case file loaded
1A case below threshold, a case file failed to load, no cases found, a run could not start, untrusted directory without --trust-plugin, or an invalid option
2Partial run: cost ceiling hit, or credential rejected at or before the first run. results.json still written with partial: true
130 / 143Interrupted / terminated (e.g. CI timeout). Partial results written

Report-writing or publishing problems never change the exit code. When charting trends, drop documents with partial: true and runs flagged skippedPaidGraders, since their scores are not comparable.

Fields a gating script reads from the JSON: aggregates.overallScore, aggregates.casesPassed / casesTotal, aggregates.meanDelta, and per case cases[].aggregates.score and .delta (omitted when arms are not comparable). A run with a non-null cases[].arms.with[].error (e.g. timed out after 300s) is still graded on what it produced, so an error does not imply score 0; an aborted run (a mock's expect: fired) scores 0 with error still null.

What a run can and cannot reach​

Read this before evaluating a plugin you did not write. Pointing claude plugin eval at a directory is the same trust decision as claude --plugin-dir: the plugin's skills and hooks load and run on your machine as you. The isolation limits what the agent under test can do; it is not a boundary against the plugin's own code, and a passing suite says nothing about whether the plugin is safe.

  • Isolated agent. Each run gets a throwaway home, working directory, and Claude Code config, runs as a claude -p child, cannot read the eval directory (so it never sees the prompts or graders of the case or its siblings), and has the Artifact tool switched off.
  • Managed settings still apply. Restrictions an administrator deployed to the machine apply inside a run, so results on a managed laptop can differ from an unmanaged CI runner.
  • Opt-in escape hatches. A case's scaffold_script runs as you, outside the sandbox, and only with --scaffold. Real MCP servers only with --allow-real-servers or --mocks off. A case's allowed_tools or a skill's own allowed-tools frontmatter can never widen these.
  • Network. Shell commands you grant follow the sandbox's network rules; a WebFetch(domain:…) grant reaches that domain directly; the plugin's hooks and any real MCP server you start can reach any host.

For a third-party plugin that ships hooks or needs its real servers, treat the scores as advisory unless you ran the suite in a container or on a CI runner. See Reviewing Third-Party Code for the broader checklist.

Quick reference: the gotchas​

  • Default target for regex is last_message. When you target trace, it is JSON per line, so quotes appear as \" and newlines are escaped.
  • files is a list of paths Claude created, not their contents. To grade contents, use { source: file, path: <path> } as target/focus. An llm judge sees PNG/JPEG/GIF/WebP as images and refuses other binaries (.pptx, PDF).
  • An llm judge reading trace sees only the first 12 and last 12 messages.
  • file_exists sees only files created during the run. A scaffolded or merely edited file is invisible to it; grade its contents or use tool_used on Edit.
  • Unknown keys in prompt.md frontmatter are an error, and env keys must be EVAL_*.
  • Evaluating an installed plugin by name (name@marketplace) writes results under ./evals/results/ in your current directory and skips the trust prompt.
  • The plugin-eval case format is separate from the evals/evals.json file the skill-creator plugin uses for skill evals.
اضغط Enter أو مفتاح المسافة لقلب البطاقة. استخدم مفتاحي السهمين الأيسر والأيمن للتنقل بين البطاقات.تم إظهار المصطلح.
1 / 7

Check yourself

0/6
  1. A case scores 1.00 WITH and 1.00 W/OUT. What does that tell you?
  2. Your CI job runs `claude plugin eval . --json results.json` with no other flags and exits 1 although every case scored 0.9. Why?
  3. Which grader types cost nothing extra to run?
  4. Δ is negative but the skill-fired grader passed on every with-arm run. What does the official guidance say to check first?
  5. A run's mocked MCP tool receives a call whose input violates the mock's `expect:` block. What happens?
  6. Which of these does the agent under test have access to during a run?

Sources & further reading​

Next​

  • Plugins & Marketplaces — package what you just tested, and publish it once the suite is green
  • Skills — the description field is what a failing skill-fired grader is telling you to fix
  • Reviewing Third-Party Code — before you run someone else's plugin (and its hooks) through an eval on your machine