Workflow Packager: Full Lifecycle CRUD Execution Engine

Problem

Agent skill inventories often drift into duplicated, stale, or ambiguous assets: similar names with inconsistent descriptions, multiple skills covering the same workflow, and old skills that are rarely used but still referenced. Manual review is slow and error-prone, while bulk deletion can break active workflows.

How It Works

Workflow Packager treats conversation evidence, skill manifests, and file-system state as input, then narrows maintenance into four auditable actions:
- Create: detect explicit repetition and latent patterns, propose a short candidate list first, and create only a minimal viable asset;
- Delete: mark zombie skills that are older than 30 days, have 0 references, and are not pinned; run a dry run before deleting confirmed unused assets;
- Merge: identify overlapping pairs using keyword clustering, trigger-word collision, and functional intent comparison, then merge into a stable target skill after cross-validation;
- Patch: make local updates to trigger words, related_skills, version numbers, or command URLs, and verify the patch.
Each run applies quantity limits, pinned protection, pre-deletion records, and a trash-based tool-fallback path to reduce irreversible mistakes.

Boundaries

It fits Agent environments that already expose skills, memories, sessions, or similar file structures. If evidence is weak, data is sensitive, or actions would touch production side effects, the skill marks the item as Skip or Need More Evidence.