How henrymao2004/misevolve works
MisEvolve is a benchmark for measuring persistent safety failures in self-improving agent skills, with SafeEvolve governance as an evaluated intervention. Its principal workflow runs scheduled episode tasks through an agent harness and an evolution method, judges behavior and authored artifacts, and records results. The README documents benchmark scheduling and persistence probes, but benchmark files are absent from the supplied tree. The graph therefore shows the documented workflow without inventing their implementation; source excerpts are partial, so some internal links are omitted.
Main entry points: methods Method harnesses · run.py Run orchestration · harness.py Episode lifecycle · _backbone.py Method LLM access
Components 30
Experiment runtime
- Run orchestration
src/misevolve/run.py - Episode lifecycle
src/misevolve/harness.py - Benchmark episodes
- Backbone config
src/misevolve/backbones.py - Method LLM access
src/misevolve/methods/_backbone.py - Episode results
Agent execution
- Agent runner contract
src/misevolve/harness.py - Container sandbox
src/misevolve/ahz_sandbox.py - Trajectory normalization
src/misevolve/adapters/_normalize.py - Claude Code adapter
src/misevolve/adapters/claudecode_runner.py - Codex adapter
src/misevolve/adapters/codex_runner.py - Hermes adapter
src/misevolve/adapters/hermes_runner.py - OpenClaw adapter
src/misevolve/adapters/openclaw_runner.py
Skill evolution
- Method harnesses
src/misevolve/methods - AutoSkill
src/misevolve/methods/autoskill.py - EvoSkill
src/misevolve/methods/evoskill.py - SkillClaw
src/misevolve/methods/skillclaw.py - SkillOpt
src/misevolve/methods/skillopt.py - SkillsVote
src/misevolve/methods/skillsvote.py - Hermes native
src/misevolve/methods/hermes_native.py - No evolution
src/misevolve/methods/no_evolution.py
Governance and judging
- Governance wrapper
src/misevolve/governance/base.py - SafeEvolve policy
src/misevolve/governance/safeevolve.py - Audit and repair
src/misevolve/governance/audit.py - Skill lineage
src/misevolve/governance/lineage.py - Governed stores
src/misevolve/governance/stores.py - Behavioral judge
src/misevolve/judges/behavioral.py - Artifact judge
src/misevolve/judges/artifact.py - Judge configuration
src/misevolve/judges/config.py
Other components
- Benchmark operator
Connections 44
- Benchmark operator → Run orchestration: starts run
README.md - Run orchestration → Episode lifecycle: runs episode
src/misevolve/run.py - Benchmark episodes → Episode lifecycle: supplies tasks
README.md - Episode lifecycle → Agent runner contract: executes tasks
src/misevolve/harness.py - Episode lifecycle → Method harnesses: evolves skills
README.md - Episode lifecycle → Behavioral judge: judges trajectories
src/misevolve/harness.py - Run orchestration → Behavioral judge: configures judge
src/misevolve/run.py - Run orchestration → Artifact judge: configures judge
src/misevolve/run.py - Episode lifecycle → Episode results: produces rows
src/misevolve/harness.py - Backbone config → Run orchestration: resolves settings
README.md - Agent runner contract → Container sandbox: runs isolated tasks
README.md - Claude Code adapter → Agent runner contract: implements
src/misevolve/adapters/claudecode_runner.py - Codex adapter → Agent runner contract: implements
src/misevolve/adapters/codex_runner.py - Hermes adapter → Agent runner contract: implements
src/misevolve/adapters/hermes_runner.py - OpenClaw adapter → Agent runner contract: implements
src/misevolve/adapters/openclaw_runner.py - Claude Code adapter → Container sandbox: dispatches task
src/misevolve/adapters/claudecode_runner.py - Codex adapter → Container sandbox: dispatches task
src/misevolve/adapters/codex_runner.py - Hermes adapter → Container sandbox: dispatches task
src/misevolve/adapters/hermes_runner.py - Claude Code adapter → Trajectory normalization: normalizes output
src/misevolve/adapters/claudecode_runner.py - Codex adapter → Trajectory normalization: normalizes output
src/misevolve/adapters/codex_runner.py - Hermes adapter → Trajectory normalization: normalizes output
src/misevolve/adapters/hermes_runner.py - OpenClaw adapter → Trajectory normalization: normalizes output
src/misevolve/adapters/openclaw_runner.py - Method harnesses → AutoSkill: registers method
src/misevolve/methods/__init__.py - Method harnesses → EvoSkill: registers method
src/misevolve/methods/__init__.py - Method harnesses → SkillClaw: registers method
src/misevolve/methods/__init__.py - Method harnesses → SkillOpt: registers method
src/misevolve/methods/__init__.py - Method harnesses → SkillsVote: registers method
src/misevolve/methods/__init__.py - Method harnesses → Hermes native: registers method
src/misevolve/methods/__init__.py - Method harnesses → No evolution: registers method
src/misevolve/methods/__init__.py - AutoSkill → Method LLM access: calls model
src/misevolve/methods/autoskill.py - EvoSkill → Method LLM access: calls model
src/misevolve/methods/evoskill.py - SkillClaw → Method LLM access: calls model
src/misevolve/methods/skillclaw.py - SkillOpt → Method LLM access: calls model
src/misevolve/methods/skillopt.py - SkillsVote → Method LLM access: calls model
src/misevolve/methods/skillsvote.py - Method LLM access → Backbone config: uses resolved model
src/misevolve/methods/_backbone.py - Governance wrapper → Method harnesses: wraps method
src/misevolve/governance/base.py - SafeEvolve policy → Governance wrapper: implements policy
src/misevolve/governance/safeevolve.py - SafeEvolve policy → Audit and repair: audits and repairs
src/misevolve/governance/safeevolve.py - SafeEvolve policy → Skill lineage: tracks lineage
src/misevolve/governance/safeevolve.py - Governance wrapper → Governed stores: uses store
src/misevolve/governance/base.py - Run orchestration → Governance wrapper: selects governance
src/misevolve/run.py - Behavioral judge → Judge configuration: loads settings
src/misevolve/judges/behavioral.py - Artifact judge → Judge configuration: loads settings
src/misevolve/judges/artifact.py - Run orchestration → Artifact judge: judges authored skills
src/misevolve/run.py