Problem
Alluka persists approved skills and workflows per agent, then includes every Markdown file in both directories in every system prompt. As an agent evolves, unrelated capabilities inflate context for every run and make it difficult to determine which instructions were active for a given task.
There is no library-level lifecycle for reusable capabilities: enabled state, version/source metadata, assignment to specific agents, or explicit removal from an agent's prompt.
Feature
Introduce a local reusable-skills library and explicit per-agent skill assignment.
Each skill has a stable identity, name, description, bounded instruction content, version/source metadata, enabled state, and local lifecycle. An agent receives only skills that are both enabled and assigned to it. The active skill set for a run is inspectable without disclosing unrelated skill contents.
Expected behavior
- Creating, updating, enabling, disabling, deleting, and assigning skills have visible state-specific outcomes.
- Disabled, deleted, unknown, or unassigned skills do not contribute instructions to an agent run.
- Deleting a skill removes it from all agent assignments.
- Existing approved agent skills/workflows are preserved or clearly surfaced for migration.
- Skill content stays local and is not included in general agent catalogs or settings payloads unless explicitly requested by an authorized local user.
- The prompt identifies the selected capabilities clearly enough to support run auditing.
Why this matters
Selective skills preserve the value of Alluka's agent evolution model while preventing prompt bloat and accidental cross-domain instructions. They make reusable knowledge intentional rather than an ever-growing implicit system prompt.
Reference
Problem
Alluka persists approved skills and workflows per agent, then includes every Markdown file in both directories in every system prompt. As an agent evolves, unrelated capabilities inflate context for every run and make it difficult to determine which instructions were active for a given task.
There is no library-level lifecycle for reusable capabilities: enabled state, version/source metadata, assignment to specific agents, or explicit removal from an agent's prompt.
Feature
Introduce a local reusable-skills library and explicit per-agent skill assignment.
Each skill has a stable identity, name, description, bounded instruction content, version/source metadata, enabled state, and local lifecycle. An agent receives only skills that are both enabled and assigned to it. The active skill set for a run is inspectable without disclosing unrelated skill contents.
Expected behavior
Why this matters
Selective skills preserve the value of Alluka's agent evolution model while preventing prompt bloat and accidental cross-domain instructions. They make reusable knowledge intentional rather than an ever-growing implicit system prompt.
Reference