Crate skilllite_evolution 

SkillLite Evolution: self-evolving prompts, skills, and memory.

EVO-1: Feedback collection + evaluation system + structured memory. EVO-2: Prompt externalization + seed data mechanism. EVO-3: Evolution engine core + evolution prompt design. EVO-5: Polish + transparency (audit, degradation, CLI, time trends).

Interacts with the agent through the EvolutionLlm trait for LLM completion.

Modules

external_learner

EVO-6: External knowledge learning.

feedback

Evolution feedback collection and evaluation system (EVO-1).

memory_learner

进化 Memory：从执行反馈中沉淀事实与经历，供检索与类比。

prompt_learner

Prompt learner: extract rules and examples from execution feedback (EVO-3).

seed

Seed data management for the self-evolving engine (EVO-2 + EVO-6).

skill_synth

Skill synthesis: auto-generate, refine, and retire skills (EVO-4).

Structs

EvolutionMessage

Minimal message format for evolution LLM calls (no tool calling).

EvolutionScope

EvolutionThresholds

进化触发阈值，均由环境变量配置，未设置时使用下列默认值。

Enums

EvolutionMode

Which dimensions of evolution are enabled.

EvolutionProfile

进化触发场景：不设或 default 时与原有默认行为完全一致。

EvolutionRunResult

Result of attempting to run evolution. Distinguishes “skipped (busy)” from “no scope” from “ran (with or without changes)”.

SkillAction

Action type for skill evolution.

Traits

EvolutionLlm

LLM completion interface for evolution.

Functions

append_changelog

atomic_write

原子写入：先写 .tmp 再 rename

check_auto_rollback

create_snapshot

finish_evolution

format_evolution_changes

gatekeeper_l1_path

L1 path gatekeeper. When skills_root is Some, also allows target under skills_root/_evolved (project-level skill evolution).

gatekeeper_l1_template_integrity

gatekeeper_l2_size

gatekeeper_l3_content

log_evolution_event

mark_decisions_evolved

on_shutdown

query_changes_by_txn

restore_snapshot

run_evolution

Run a full evolution cycle.

should_evolve

should_evolve_with_mode

strip_think_blocks

Strip reasoning/thinking blocks emitted by various models. Handles <think>, <thinking>, <reasoning> tags (DeepSeek, QwQ, open-source variants). Returns the content after the last closing tag, or the original string if none found. Should be called at the LLM layer so all downstream consumers get clean output.

try_start_evolution