Plugin agent loading fails silently with no error feedback
Summary
Custom marketplace plugins with agents cannot be loaded under any circumstances. After exhaustive troubleshooting — matching the exact format of official plugins, validating every frontmatter field, reinstalling through the CLI, clearing caches, restarting — /agents returns nothing. No error message, no warning, no diagnostic output whatsoever. The system fails silently and completely.
This is not a user error. This is a broken feature that shipped without basic validation, error handling, or any way for users to self-diagnose.
Environment
- Platform: Linux (RHEL 9)
- Claude Code CLI (latest)
- Custom marketplace via GitHub:
whatch-healtheconnections/agentic-orch
What was attempted (in order)
- Created a custom marketplace (
config-hub) pointing to a GitHub repo - Created a plugin (
my-agents) with 11 agent.mdfiles and 18 command.mdfiles in the correct directory structure - Installed via
/plugin install my-agents@config-hub /agentsshowed nothing — no agents, no errors- Reviewed and fixed all agent frontmatter to use single-line
descriptionfields matching the exact format of official plugins (feature-dev,code-simplifier) - Fixed an empty 0-byte agent file (
stale-branch-worker.md) with valid frontmatter and body - Verified
plugin.jsonmatches the documented schema —name,version,description,author,agents,commandsfields - Committed and pushed all fixes to the source repo
- Uninstalled the plugin (cleared
installed_plugins.json, deleted cache) - Reinstalled via
/plugin install my-agents@config-hub --scope local - Restarted Claude Code
/agentsstill shows nothing. Still no error message.
Every frontmatter block was verified to match this exact pattern from official plugins:
---
name: agent-name
description: Single line plain text description
model: sonnet
color: green
---
The plugin.json was verified to match the documented format:
{
"name": "my-agents",
"version": "0.1.0",
"description": "Agents, commands, and skills",
"author": { "name": "whatch-healtheconnections" },
"agents": "./agents/",
"commands": "./commands/"
}
All files are present on disk in the correct locations. The marketplace is registered. The plugin shows as installed. Nothing works.
What is broken
- Silent failure with zero diagnostics. When plugin agent loading fails, there is no error, no warning, no log entry. The user has no way to determine what went wrong. This is the most critical issue — even if the root cause is a subtle format problem, the system gives the user absolutely nothing to work with.
- No documentation on failure modes. The plugin docs describe the happy path. There is no troubleshooting guide, no list of common pitfalls, no explanation of what causes agent discovery to fail.
- No validation on install.
/plugin installsucceeds and gives no indication that the agents it just installed will not actually load. If the plugin has structural issues, they should be caught and reported at install time — not discovered later when/agentssilently returns empty.
/reload-pluginsexistence is undiscoverable. This command exists but is not surfaced anywhere in the/pluginhelp or general/help. Users restart Claude Code repeatedly instead of using it, because they don't know it exists.
- The AI inside Claude Code cannot diagnose plugin issues. Claude itself has no visibility into the plugin loading system, no access to plugin logs, and no documentation on the internals. When the user asks Claude for help, Claude is reduced to guessing — which is exactly what happened here across multiple failed attempts.
Expected behavior
/plugin installshould validate the plugin structure and report errors immediately- If individual agent files fail to parse, the error should name the file and the problem (e.g., "stale-branch-worker.md: missing required frontmatter field 'name'")
- Valid agents should still load even if one file is malformed — do not fail the entire plugin
/agentsshould indicate when plugins are installed but zero agents loaded ("1 plugin installed, 0 agents loaded — run/plugin validatefor details")- A
/plugin validateor/plugin doctorcommand should exist for diagnosing issues - Plugin loading errors should be written to a log file that users and Claude can inspect
Impact
This feature is currently unusable for custom marketplaces. A user who invests time creating agents, setting up a marketplace repo, and installing plugins will hit a wall with no feedback and no path forward. The only recourse is filing a bug report and waiting — which is not acceptable for a feature that is publicly documented and available.
This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗