dsh-skill-curator

Every time a conversation ends, the judgments, preferences, and problem-solving paths accumulated along the way tend to vanish with the session. dsh-skill-curator turns that loss into an automated loop: every N real conversation turns (default 3), it spins up a background subagent that reads a session digest (tail verbatim plus head compression) and creates or updates SKILL.md files under the user skills directory. The whole process is zero-intrusion and requires no DSH source modifications.

Core capabilities include digest injection to minimize token overhead, a six-tool whitelist (list, read, create, patch, write support files, adopt) that grants the subagent only the minimum necessary permissions, an ownership guard that refuses to modify any skill not stamped with the author marker or listed in adoptSkills, and a two-layer fault-tolerance strategy (fallback to the session model plus exponential-backoff retries) so that a network blip never loses a curation run. It supports custom review endpoints and models, and review history is persisted to a local JSON file that survives plugin removal and reinstallation.

It is well suited to teams and individuals who rely on DSH for iterative workflows: anyone who wants to distill patterns from each conversation into reusable skills without manually maintaining SKILL.md files. The manual /skill-refine command allows an on-demand review of the current session at any time, complementing the automatic polling cycle.