claude plugin eval sandbox never loads a plugin's own agents/*.md for the Agent tool

Summary

claude plugin eval never registers a plugin's own custom agents (agents/*.md) for the Agent tool inside its eval sandbox, under any target form. An eval case that dispatches subagent_type: "<plugin>:<agent-name>" fails with Agent type '<plugin>:<agent-name>' not found, regardless of whether the target is a filesystem path, an installed plugin name, or <plugin>@<marketplace>.

Expected behavior

The eval "Available tools" reference documents Agent as a grantable tool type for eval cases. A plugin that ships a skill dispatching one of its own agents, plus an eval suite covering that skill, would reasonably expect the sandbox to resolve the plugin's own agent — the way an interactive session does.

Actual behavior

The Agent tool inside the eval sandbox only ever lists the built-in agent types: claude, Explore, general-purpose, Plan, statusline-setup. A plugin's own agents/*.md are never listed or resolvable, tested under all of:

• claude plugin eval <path-to-plugin>
• claude plugin eval <installed-plugin-name>
• claude plugin eval <plugin-name>@<marketplace> (via a scratch marketplace + claude plugin install)

Agent(subagent_type: "<plugin>:<agent-name>") (and the bare agent name) returns Agent type '<plugin>:<agent-name>' not found in every case.

Docs cross-check

The Plugin Evals guide states:

claude plugin eval loads the target plugin's skills and hooks and runs its eval suite on your machine, as you.

Agents are never mentioned as loaded — only skills and hooks — so this may be intended scope rather than a bug. It isn't called out explicitly, though, and it silently breaks any eval case written to exercise agent-dispatching skill behavior (no error until the eval actually runs and the dispatch fails).

Suggested fix (either resolves this)

1. Load the plugin's own agents/*.md into the eval sandbox's Agent tool, matching interactive-session behavior, or
2. Document explicitly in the Plugin Evals guide that a plugin's own custom agents are out of scope for eval sandboxes, so authors know up front to inline agent instructions into the eval prompt instead.

Related but distinct issues

• #83168 (open) — intermittent plugin-agent loading failure in interactive sessions; different symptom (intermittent, not eval-specific) and likely different root cause.
• #59881, #21841, #18982, #13605, #21560 (closed) — earlier plugin-agent loading bugs, not eval-specific.

Environment

• Claude Code CLI 2.1.280
• macOS (darwin)