Autonomous evolution loop + cross-run memory for DeepSeek Harness (fork)

[ziyam-khan]

We built an AVO-style autonomous improvement loop on top of DeepSeek Harness as ecosystem plugins, running on our fork.

What it does

• lineage_evolve — one fresh structured-output agent per candidate produces the next variation against your objective; a monotonic commit gate folds results into a durable lineage that survives restarts
• strategy-memory mount — goal completions and blockers distill into typed cross-run lessons (hypothesis / dead-end / preference / decision) recalled into future prompt assemblies
• supervisor guard — mandatory strategy-change directives when tool calls repeat past advisory levels or error storms persist
• action-efficiency budgets — snapshot scenarios pin max model-facing actions per completed objective so harness changes that degrade efficiency fail keylessly

All six packages ship with composition tests through real services, bilingual READMEs, invariants, and config-catalog registration.

Fork: https://github.com/ziyam-khan/deepseek-harness
Docs: docs/user/develop/practice/evolution-plugins.md on the fork/main branch

Design inspired by AVO (arXiv:2603.24517) with distillation lessons from mem0, Hindsight, Honcho, and TencentDB Agent Memory.