Keeping the Library Coherent Over Time
AutoHarness can make individual skills bounded and still end up with a confusing library. Episode reflection sees one work window at a time; a collection can accumulate several narrow lessons that are individually valid but collectively hard to discover. This chapter looks at the longer-term maintenance layer: the curator that reshapes the library, the metrics that make its behavior visible, and the security and host constraints that determine how much confidence the design deserves in practice.
Episode learning has a shape problem
A reflector is asked to compare a new lesson with the existing index, but its immediate evidence is still one episode. That is enough to notice that a newly corrected testing rule resembles one skill already in play. It is not always enough to see that five older skills, created in different sessions, are really fragments of one class of work. A system can therefore avoid exact duplicates while still accumulating a forest of near-neighbors.
The curator exists for that different scale of judgment. It reads the index of agent-created skills across the global and project layers and asks whether a maintainer would keep several narrow entries or write one class-level umbrella with labeled subsections. The goal is not to delete information for the sake of a smaller count. Narrow, session-shaped details can move into references, templates, or scripts under the umbrella when they remain useful. Discoverability improves when the library represents the class of work rather than the chronology of every correction.
This is a meaningful division of labor. The reflector is close to the episode and can preserve a newly discovered lesson before it is forgotten. The curator is allowed to think about the library as a whole and can recognize a shape problem that no single episode reveals. Treating them as interchangeable would either make every reflection expensive and library-wide, or leave consolidation to a later operator who may not know why the siblings exist.
Package integrity makes the merge more demanding than copying text. An absorbed skill may contain evidence, references, a template, or a script that its body relies on. The curator must either re-home the needed support file and repair the pointers or leave the skill standalone. Provenance evidence materialized by the promoter is not ordinary content to flatten or discard. A class-level umbrella is only better if it remains a complete, navigable package.
A merge is a state transition, not silent deletion
Whole-library consolidation is the operation most likely to need recovery. A mistaken patch to one skill can be inspected and reversed as a file change. A merge can reshape many relationships at once: one skill is broadened, siblings are archived, and their future recall routes change. AutoHarness therefore takes snapshots of the skill trees before a curator run. The snapshot is not a semantic guarantee, but it gives an operator a recoverable baseline if the resulting library shape is wrong.
The shared promoter remains the write boundary. The curator can propose an umbrella, patch an existing skill, carry support files, and delete absorbed siblings, but it does not directly edit the live tree. A deletion that claims absorption must name a live, self-produced umbrella.