dsh-skillflux

As a skill library grows, every request inherits a heavier catalog that consumes context and makes selection less predictable. SkillFlux inserts a dynamic runtime layer between the Agent and its full skill pool: for each turn it selects and mounts no more than maxActiveSkills (three by default) highly relevant skills, then releases the mounts when the turn ends, so the model always faces a lean, stable directory.

Routing is deterministic and layered. User-defined ordered rules run first, followed by English and Chinese lexical scoring against a configurable threshold, with an optional local embedding fallback that fills only unfilled catalog slots. When the local pool yields no good match, SkillFlux searches skills.sh and the public GitHub SKILL.md corpus in real time, then ranks candidates with a multi-signal quality score capped at 100, weighing task relevance (55), marketplace installs (15), GitHub stars (15), forks (5), repository activity within a configurable window (10), owner trust and license metadata (15), and cross-source corroboration (8). Every candidate is pinned to an immutable commit SHA and carries an explainable evidence level plus context-specific warnings.

Downloaded skills are governed by an independent cache layer that prunes entries idle for 90 days by default, evicts lowest-value entries when pool limits are exceeded, and protects active mounts and concurrent loads throughout. It is built for DSH developers whose skill pool is growing and who want the model-facing catalog to stay lean and selection to stay predictable, while retaining the ability to discover, vet, and mount new skills from the public ecosystem on demand.