*"hj8tdZddlmZddlmZmZmZmZddlm Z ddl m Z ddl m Z ddlmZddlmZmZgd Zy ) u evolve_server — **population-based** asynchronous skill evolution engine for SkillClaw. Periodically (or on-demand) fetches session interaction data from shared storage, builds a dual-layer representation for each session (lossless programmatic trajectory + LLM trajectory-aware analysis), aggregates sessions by skill, and lets the LLM decide whether to evolve, optimise, or skip each skill — with full trajectory context. Pipeline:: Shared Storage Sessions → Dual-layer preprocessing: A. Programmatic trajectory (step-by-step path, zero info loss) B. LLM trajectory-aware analysis (causal chains, skill effectiveness) C. Metadata extraction (skills_referenced, avg_prm, tool_errors) → Aggregate sessions by skill → Per-skill evolution (LLM sees trajectory + analysis; decides: improve / optimize / skip) → No-skill session handling (LLM sees trajectory + analysis; decides: create / skip) → Shared Storage Skills Usage:: python -m evolve_server # periodic (default 10 min) python -m evolve_server --once # single pass python -m evolve_server --port 8787 # with HTTP trigger )EvolveServerConfig)FAILURE_LABELS NO_SKILL_KEYDecisionAction FailureType)AsyncLLMClient)SkillIDRegistry)AgentEvolveServer) EvolveServer) LocalBucket MockBucket) r r rrr r r rrrrN)__doc__ core.configrcore.constantsrrrrcore.llm_clientrcore.skill_registryr engines.agentr engines.workflowr storage.mock_bucketr r __all__1/root/.hermes/SkillClaw/evolve_server/__init__.pyrs+:,UU+0,*8 r