The /learning-agents skill is the dispatch entry point for the plugin. It routes user input to the appropriate sub-skill based on the first token of the arguments. It also defines the public-facing command interface.
The plugin MUST provide a top-level skill named learning-agents that serves as the dispatch entry point. This skill MUST be user-invocable.
The skill MUST split $ARGUMENTS on the first whitespace. The first token is the sub-command (case-insensitive); the remainder is passed to the sub-skill.
The skill MUST accept both underscores and dashes in sub-command names as equivalent (e.g., report_issue and report-issue MUST both route to the same sub-skill).
The create <name> [template-path] sub-command MUST invoke Skill learning-agents:create-agent <name> [template-path]. The template-path argument is optional; when provided, it MUST be passed through to the sub-skill.
The learn sub-command MUST invoke Skill learning-agents:learn. Arguments after learn MUST be ignored.
The report_issue <agentId> <details> sub-command MUST:
- Search
.deepwork/tmp/agent_sessions/for a subdirectory whose name contains the providedagentId - If no match is found, inform the user
- If multiple matches exist, use the most recently modified one
- Invoke
Skill learning-agents:report-issue <resolved-path> <details>
When no arguments are provided or input does not match any known sub-command, the skill MUST display a help message listing all available sub-commands with descriptions and examples.
The skill MUST always route to the appropriate sub-skill. It MUST NOT implement sub-command logic inline.
The skill MUST pass arguments through to sub-skills exactly as provided by the user (after extracting the sub-command token).
The skill MUST support exactly three sub-commands:
create-- create a new LearningAgentlearn-- run learning cycle on pending sessionsreport_issue-- report an issue with an agent
The plugin MUST provide the following skills, each in its own directory under skills/:
learning-agents-- dispatch entry pointcreate-agent-- agent scaffoldinglearn-- learning cycle orchestrationidentify-- issue identification (non-user-invocable)investigate-issues-- issue investigation (non-user-invocable)incorporate-learnings-- learning incorporation (non-user-invocable)report-issue-- issue file creation (non-user-invocable)prompt-review-- prompt engineering review (user-invocable)
The prompt-review skill MUST be independently invocable (it is NOT routed through the /learning-agents dispatcher). It MUST be a separate, standalone skill.
The dispatch skill MUST display a list of existing LearningAgents by listing the contents of .deepwork/learning-agents/. If no agents exist, it MUST display a fallback indicator (e.g., "(none)"). This listing MUST be generated dynamically at skill load time.