Plugin commands should be loaded as keyword-triggered skills

[jpshackelford]

Problem

When loading plugins via Plugin.load(), markdown files in the commands/ directory are loaded into plugin.commands but are not merged into the agent context as skills. Only files in the skills/ directory become part of the agent's context.

This means Claude Code format plugins that use commands/ (which is the standard location for slash commands) don't work correctly with OpenHands - the agent never receives the command instructions.

Current Behavior

In plugin.py:

• _load_skills() loads from skills/ directory
• _load_commands() loads from commands/ directory

In conversation_service.py (_merge_plugin_into_request):

• Only plugin.skills is merged into the agent context
• plugin.commands is ignored

Example

Given a plugin structure:

city-weather/
├── .claude-plugin/
│   └── plugin.json
└── commands/
    └── now.md          # Weather command instructions

The now.md file is loaded but never added to the agent's skills, so the agent doesn't know how to handle weather requests using the plugin's instructions.

Proposed Solution

Treat markdown files in the commands/ directory as keyword-triggered skills using the Claude Code namespacing format:

1. When loading commands from commands/, convert them to Skill objects
2. Set the skill's trigger to a KeywordTrigger using the format /<plugin-name>:<command-name>
3. Merge these command-based skills into the agent context alongside regular skills

Example Trigger Format

For a plugin named city-weather with a command now.md:

• Trigger keyword: /city-weather:now
• When user types /city-weather:now Tokyo, the skill activates with $ARGUMENTS = Tokyo

Implementation Notes

1. In _load_commands() or a new helper, create Skill objects from CommandDefinition
2. Set trigger = KeywordTrigger(keywords=["/<plugin-name>:<command-name>"])
3. The skill content should include the command's markdown with $ARGUMENTS placeholder
4. In _merge_plugin_into_request(), merge both plugin.skills and the converted commands, but don't duplicate loading of the items in the command directory.

Context

• Related to Plugin Directory feature (#12088 in OpenHands/OpenHands)
• Discovered while testing plugin loading via PR #12338 (app server) and PR feat(agent-server): Support plugin loading when starting conversations #1651 (SDK)
• Claude Code plugins use commands/ as the standard location for slash commands

References

• Claude Code Plugin Marketplaces
• Test plugin: https://github.com/jpshackelford/openhands-sample-plugins

[jpshackelford]

For the sake of keeping things moving for our end-to-end demo, when we create a PR for this, branch from the code here #1651 so we have have agent server image with all three required PRs from software-agent-sdk.

[enyst]

cc: @jpshackelford

This issue sounds about right. Please let me note this:

"Set the skill's trigger to a KeywordTrigger using the format /:"

This is a detail I meant to mention in the other discussion, about an hypothetical autorun.inf, Agentskills format includes these commands and they include the plugin name. That, it seems to me, means the initial user message will need, once it works correctly, presumably only the command name, not an extra something for plugin name since it's already included in it.