dsh-continual-evolve: A Continual Self-Evolution Plugin for DeepSeek Harness
zk-andy/dsh-continual-evolve
Continual self-evolution for DeepSeek Harness: versioned, auditable, rollback-safe state refined from session trajectories with a benchmark-driven validation loop.
Let DeepSeek Harness agents turn session experience into versioned, auditable, rollback-safe persistent state for true self-evolution.
dsh plugin --profile web add github:zk-andy/dsh-continual-evolve- Category
- Skills & Agents
- Platform
- DSH-Plugin
- Author
- zk-andy
- Distribution
- Plugin
dsh-continual-evolve Key Features
dsh-continual-evolve Repository
dsh-continual-evolve Repo Summary
dsh-continual-evolve is a continual self-evolution plugin (DSH plugin) for DeepSeek Harness, maintained by ZK-Andy under the MIT license, last updated in 2026-08. It addresses the problem that agents accumulate reusable experience in every session—repeated failures, durable facts, reusable procedures—and then forget it at the next turn or session. The plugin turns this experience into versioned, auditable, rollback-safe persistent harness state. Core capabilities include: versioned entries keyed by kind (prompt / memory / skill / subagent), an evidence trail (each refinement appends an event carrying trigger / changes / evidence / outcome), deterministic rollback (inverse edits generated from applied results, no LLM re-guessing), code-enforced safety (schema validation, atomic writes, corrupt-file degrade, optimistic concurrency, immutable base system prompt), and local (session) and global (cross-session) scopes with merge semantics.
What core features does the dsh-continual-evolve DSH plugin support?
The plugin implements a full evolution loop: a pure-core engine, model tools and the /evolve command, an automatic review gate (turn-interval + compaction checkpoints, human approval for global edits), real system-prompt injection (prompt notes + delegation specs, zero token cost when empty), and a benchmark-driven validation loop (code-owned scoring, non-regressive acceptance, rubric ACL). Design provenance draws from penguin-harness (concept proof but zero code-level enforcement), prime-agent /refine (versioned entries, atomic persistence, optimistic concurrency, inverse-op rollback), and academic work (Self-Harness, AHE, HarnessOpt-Bench), ensuring that "the model proposes, the code guarantees"—all mechanical safety properties are enforced in code, never by asking the model to behave.
How do I install the dsh-continual-evolve DSH plugin?
Use the following command with the DSH plugin manager:
dsh plugin --profile web add github:zk-andy/dsh-continual-evolve
Which DeepSeek Harness components does dsh-continual-evolve depend on?
The plugin depends on official DSH component packages, including @deepseek-ai/dsh-home-paths, @deepseek-ai/dsh-llm, @deepseek-ai/dsh-tools, @deepseek-ai/dsh-agent, @deepseek-ai/dsh-commands, and @deepseek-ai/dsh-system-prompt, and is integrated as a bundle manifest. It requires Node.js ^22.19 or >=24 and passes 238 tests.
What is the current status and roadmap of dsh-continual-evolve?
The project status is "all phases complete; in long-term maintenance." Phases 1–3 shipped the full evolution loop, and subsequent enhancements include the memory layer (ranked injection, trajectory citations, archive), per-installation rubric keys, and plugin-owned file logging. The roadmap lists shipped and candidate features, detailed in the repository documentation.