sk-improve-agent

To safely improve agent behavior without mutating canonical targets, use an evaluator-first loop that derives a dynamic profile, scores improvements across 5 dimensions, and records append-only artifacts for promotion decisions. This ensures changes remain evidence-backed and reversible.

Dynamic profiling for agent evaluation derives scoring rules directly from the agent's own frontmatter and policy. It enables on-the-fly evaluation by generating a target-specific profile, ensuring the 5-dimension scoring accurately reflects the unique characteristics of any agent in the repository.

You score agent performance using a 5-dimension framework covering structural integrity, rule coherence, integration consistency, output quality, and system fitness. This quantifies improvements before promotion, ensuring that only evidence-backed changes are promoted to the canonical target.

Yes, you can evaluate any agent defined in the .opencode/agent directory. The system dynamically profiles the target agent path and applies deterministic benchmarks to generate a packet-local candidate with associated evidence, regardless of the specific agent configuration.

Yes, the workflow supports guarded promotion and rollback using append-only evidence and memory artifacts. This ensures auditable decisions and safe recovery, keeping runtime mirrors aligned while preserving evaluation truth if a promoted change needs to be reversed.

You should avoid using an evaluator-first loop if your workflow requires immediate canonical edits without traceable benchmarks or append-only evidence. The system is designed for bounded, auditable improvements, so direct mutation workflows bypass its core safety boundaries.